test(reset-password): fix runResetPasswordCli deadlock vs the #6387 non-TTY CLI (CI shard-4 wedge)

#6387 (register reset-password subcommand + non-TTY stdin path) made bin/reset-password.mjs
read a piped (non-TTY) stdin all at once — first line is the password — and stop printing
the interactive 'Enter new password' / 'Confirm new password' prompts. The unchanged test
helper only wrote the password after SEEING those prompts and only closed stdin after the
confirm prompt, so against the non-TTY CLI it never fed input nor closed stdin: the child
blocked forever on stdin-end and the whole unit shard wedged at the 25min CI timeout
(cancelled 3x). Feed the password + child.stdin.end() immediately, matching the non-TTY
contract. Now deterministic: 3/3 pass locally in ~7s (no skip, real CLI run).

CLAUDE.md

@@ -72,7 +72,7 @@ Client → /v1/chat/completions (Next.js route)
 
 API routes follow a consistent pattern: `Route → CORS preflight → Zod body validation → Optional auth (extractApiKey/isValidApiKey) → API key policy enforcement → Handler delegation (open-sse)`. No global Next.js middleware — interception is route-specific.
 
-**Combo routing** (`open-sse/services/combo.ts`): 17 strategies (priority, weighted, fill-first, round-robin, P2C, random, least-used, cost-optimized, reset-aware, reset-window, headroom, strict-random, auto, lkgp, context-optimized, context-relay, fusion). Each target calls `handleSingleModel()` which wraps `handleChatCore()` with per-target error handling and circuit breaker checks. The `fusion` strategy is the exception: it fans out to a panel of models in parallel, then a judge model synthesizes one final answer (`open-sse/services/fusion.ts`). See `docs/routing/AUTO-COMBO.md` for the 12-factor Auto-Combo scoring + the full strategy table and `docs/architecture/RESILIENCE_GUIDE.md` for the 3 resilience layers.
+**Combo routing** (`open-sse/services/combo.ts`): 18 strategies (priority, weighted, fill-first, round-robin, P2C, random, least-used, cost-optimized, reset-aware, reset-window, headroom, strict-random, auto, lkgp, context-optimized, context-relay, fusion, pipeline). Each target calls `handleSingleModel()` which wraps `handleChatCore()` with per-target error handling and circuit breaker checks. The `fusion` strategy is the exception: it fans out to a panel of models in parallel, then a judge model synthesizes one final answer (`open-sse/services/fusion.ts`). See `docs/routing/AUTO-COMBO.md` for the 12-factor Auto-Combo scoring + the full strategy table and `docs/architecture/RESILIENCE_GUIDE.md` for the 3 resilience layers.
 
 ---
 
@@ -221,7 +221,7 @@ connection continue serving other models.
 - **2 spaces**, semicolons, double quotes, 100 char width, es5 trailing commas (enforced by lint-staged via Prettier)
 - **Imports**: external → internal (`@/`, `@omniroute/open-sse`) → relative
 - **Naming**: files=camelCase/kebab, components=PascalCase, constants=UPPER_SNAKE
-- **ESLint**: `no-eval`, `no-implied-eval`, `no-new-func` = error everywhere; `no-explicit-any` = warn in `open-sse/` and `tests/`
+- **ESLint**: `no-eval`, `no-implied-eval`, `no-new-func` = error everywhere; `no-explicit-any` = **error** in `open-sse/` and `tests/` (since #6218 — pre-existing violations are frozen in `config/quality/eslint-suppressions.json`, new ones must be fixed; `npm run lint` applies the suppressions and is what CI runs)
 - **TypeScript**: `strict: false`, target ES2022, module esnext, resolution bundler. Prefer explicit types.
 
 ### Database
@@ -332,7 +332,7 @@ For any non-trivial change, read the matching deep-dive first:
 | Repo navigation                               | `docs/architecture/REPOSITORY_MAP.md`                   |
 | Architecture                                  | `docs/architecture/ARCHITECTURE.md`                     |
 | Engineering reference                         | `docs/architecture/CODEBASE_DOCUMENTATION.md`           |
-| Auto-Combo (12-factor scoring, 17 strategies) | `docs/routing/AUTO-COMBO.md`                            |
+| Auto-Combo (12-factor scoring, 18 strategies) | `docs/routing/AUTO-COMBO.md`                            |
 | Resilience (3 mechanisms)                     | `docs/architecture/RESILIENCE_GUIDE.md`                 |
 | Reasoning replay                              | `docs/routing/REASONING_REPLAY.md`                      |
 | Skills framework                              | `docs/frameworks/SKILLS.md`                             |
@@ -543,7 +543,7 @@ the stale-enforcement added in Fase 6A.3.
 21. **Release-freeze — the FROZEN release branch belongs to the release captain; development does NOT stop (parallel-cycle model, 2026-07-04).** `/generate-release` opens a marker issue labeled `release-freeze` at the start of reconciliation (Phase 0a), **immediately cuts the next cycle's branch `release/vX+1` from the frozen tip (Phase 0a.0b — bump + living release PR + re-home of open PRs)**, and closes the freeze once the release PR squash-merges to `main`. Before merging **any** PR, every campaign workflow (`/review-issues`, `/review-prs`, `/implement-features`, `/green-prs`, `/port-upstream-*`) **MUST** check `gh issue list --repo diegosouzapw/OmniRoute --label release-freeze --state open` — if a freeze is active: **NEVER merge into the frozen `release/vX.Y.Z` named in the freeze title**; instead resolve the ACTIVE development branch (the **highest** `release/v*` by semver — normally `release/vX+1`, announced in a freeze-issue comment) and **retarget the PR there** (`gh pr edit <N> --base release/vX+1`, then VERIFY with `gh pr view <N> --json baseRefName` — the edit fails silently) and merge normally. **HOLD only when the highest release/v\* branch IS the frozen one** (the short window before 0a.0b completes, or a pre-parallel-cycle release) — in that case leave the PR ready and open, tell the operator, and resume when the next branch appears or the freeze lifts. The just-shipped fixes reach `release/vX+1` via the Phase 5 sync-back (`scripts/release/sync-next-cycle.mjs`); do not try to sync mid-release. This is a **coordination signal, not a permission lock**: the release captain and the campaign sessions share the `diegosouzapw` identity, so a GitHub branch-protection lock cannot distinguish them — only this honored marker prevents the mid-release commit races that forced full CHANGELOG re-reconciliation in v3.8.40/v3.8.41 (a parallel campaign advanced `release/vX.Y.Z` by 34 commits mid-run). The release captain's own reconciliation/cycle-open pushes are exempt — they _are_ the release. Fixes that must land during a freeze (a homologation finding) follow the post-merge read-only rule: land on `main` first via `fix/release-vX.Y.Z-*`. **⛔ ONLY `/generate-release` may raise a release-freeze, and ONLY at its Phase 0a (start of generating a new version) — lifted at Phase 12c after the squash-merge to `main`.** No campaign, session, or agent may open a `release-freeze` marker at any other time — a freeze is **never** a mid-development coordination tool. If a session ever believes a freeze is genuinely, unavoidably necessary outside the `/generate-release` flow, it **MUST first ask the operator (`diegosouzapw`) in chat, explicitly alert "estou criando um freeze" and get an explicit yes** — never open, extend, or re-open a `release-freeze` autonomously. Conversely, do **not** close/lift an active `/generate-release` freeze to unblock campaign merges: it protects the captain's single clean CI run and auto-lifts at Phase 12c — closing it early re-triggers the exact commit race it prevents. Verify a freeze is legitimate before acting on it: an open `release-freeze` whose title/body references an **OPEN** release PR (`gh pr view <N> --json state`) is the authorized captain freeze — hold, don't touch.
 22. **Cross-session safety — this repo is worked by MANY parallel sessions/agents at once; never step on another's in-flight work.** Two absolute bans, both recurring incidents (this rule exists because they keep happening):
     - **(a) Never `git stash` / `git stash pop` — ANYWHERE in this repo, including inside an isolated worktree, and including inside any subagent you dispatch.** `git stash` operates on the **shared repository object store**, not the per-worktree working tree — so a stash pushed or popped in one session can silently clobber or resurrect another parallel session's uncommitted changes. This is not hypothetical: 2026-07-02 a `#5923` quotaCache change leaked into the unrelated `#2296` worktree via a global `stash pop`, and the same class reincided through a **subagent**. To compare working changes against a base ref **without** stashing, use `git show <ref>:<path>` or `git diff <ref> -- <path>`; to confirm a typecheck/lint error is pre-existing on the base, inspect the base ref directly (`git show origin/release/vX.Y.Z:<path>`) — never stash your tree away to "get it clean". **Put this ban verbatim in the prompt of every subagent that touches git** (agents don't inherit this file's context — the recurrence was a subagent).
-    - **(b) Never merge, push, rebase, or force-push a PR / branch / worktree that another session is actively working.** An open PR whose head is a live fix worktree in `.claude/worktrees/` you did **not** create (e.g. `fix-5852`/`fix-5923` carrying fresh commits, even when they share your `diegosouzapw` identity), or any branch another session owns, is **off-limits — HOLD**, and let the owning session merge it. **Before** merging or pushing to any PR you did not create *this* session, run `git worktree list` to check for a matching in-flight worktree and re-check `gh pr view <N> --json state,headRefOid`. Only the owning session merges its own in-flight PR; mid-flight merges race the owner and re-trigger the exact commit/CHANGELOG races Rule #19 and Rule #21 guard against. (Reinforces Rule #19.)
+    - **(b) Never merge, push, rebase, or force-push a PR / branch / worktree that another session is actively working.** An open PR whose head is a live fix worktree in `.claude/worktrees/` you did **not** create (e.g. `fix-5852`/`fix-5923` carrying fresh commits, even when they share your `diegosouzapw` identity), or any branch another session owns, is **off-limits — HOLD**, and let the owning session merge it. **Before** merging or pushing to any PR you did not create _this_ session, run `git worktree list` to check for a matching in-flight worktree and re-check `gh pr view <N> --json state,headRefOid`. Only the owning session merges its own in-flight PR; mid-flight merges race the owner and re-trigger the exact commit/CHANGELOG races Rule #19 and Rule #21 guard against. (Reinforces Rule #19.)
 
 ---
 

Dockerfile

@@ -70,6 +70,11 @@ RUN --mount=type=cache,id=npm-cache,target=/root/.npm \
 # See docs/ops/QUALITY_GATE_PLAYBOOK.md Parte 6.
 ENV OMNIROUTE_USE_TURBOPACK=1
 
+# Docker containers cannot run the MITM/Agent-Bridge stack (no host DNS/cert
+# access), so keep @/mitm/manager on the graceful stub (#3390). This flag is
+# Docker-only: npm/Electron/VPS builds must bundle the REAL manager (#6344).
+ENV OMNIROUTE_MITM_STUB=1
+
 # Raise the V8 heap ceiling for the build. The webpack production optimization
 # pass needs more than V8's default ceiling (~2 GB) for a codebase this size; a
 # memory-constrained Docker build otherwise dies with "FATAL ERROR: ... JavaScript

bin/cli/commands/launch-codex.mjs

@@ -19,6 +19,16 @@ const STRIPPED_CODEX_ENV_KEYS = [
 /** Placeholder so codex's `env_key` is always satisfied when the backend is open. */
 const NO_AUTH_SENTINEL = "omniroute-no-auth";
 
+// On Windows the `codex` binary is an npm `.cmd` shim that `spawn` cannot resolve
+// without a shell (bare "codex" → ENOENT). Mirror the qodercli Windows fix (#6263):
+// spawn `codex.cmd` through a shell on win32, and the bare binary elsewhere.
+export function resolveCodexSpawn(platform) {
+  if (platform === "win32") {
+    return { command: "codex.cmd", shell: true };
+  }
+  return { command: "codex", shell: undefined };
+}
+
 function stripTrailingSlash(value) {
   let s = String(value);
   let end = s.length;
@@ -126,10 +136,10 @@ export async function runLaunchCodexCommand(opts = {}, codexArgs = []) {
 
   if (!(await healthCheck(baseUrl))) {
     console.error(
-      (t("launch.notRunning") || "OmniRoute is not reachable at {port}. Start it with 'omniroute serve'.").replace(
-        "{port}",
-        baseUrl
-      )
+      (
+        t("launch.notRunning") ||
+        "OmniRoute is not reachable at {port}. Start it with 'omniroute serve'."
+      ).replace("{port}", baseUrl)
     );
     return 1;
   }
@@ -142,7 +152,12 @@ export async function runLaunchCodexCommand(opts = {}, codexArgs = []) {
   const env = buildCodexEnv(process.env, authToken);
 
   return await new Promise((resolve) => {
-    const child = spawn("codex", extraArgs, { env, stdio: "inherit" });
+    const { command: codexLaunch, shell: shellValue } = resolveCodexSpawn(process.platform);
+    const child = spawn(codexLaunch, extraArgs, {
+      env,
+      stdio: "inherit",
+      shell: shellValue,
+    });
     child.on("error", (err) => {
       if (err?.code === "ENOENT") {
         console.error(
@@ -165,10 +180,16 @@ export function registerLaunchCodex(program) {
       t("launchCodex.description") || "Launch Codex CLI pointed at OmniRoute (local or remote VPS)"
     )
     .option("--port <port>", "Local OmniRoute port (ignored when --remote is set)", "20128")
-    .option("--remote <url>", "Remote OmniRoute base URL, e.g. http://192.168.0.15:20128 (overrides --port + context)")
+    .option(
+      "--remote <url>",
+      "Remote OmniRoute base URL, e.g. http://192.168.0.15:20128 (overrides --port + context)"
+    )
     .option("--profile <name>", "Codex profile to activate (passed as --profile <name>)")
     .option("-p, --p <name>", "Alias for --profile")
-    .option("--api-key <key>", "OmniRoute API key (overrides OMNIROUTE_API_KEY env var for this invocation)")
+    .option(
+      "--api-key <key>",
+      "OmniRoute API key (overrides OMNIROUTE_API_KEY env var for this invocation)"
+    )
     .allowUnknownOption(true)
     .allowExcessArguments(true)
     .argument("[codexArgs...]", "arguments passed through to the codex binary")

bin/omniroute.mjs

@@ -6,6 +6,7 @@
  * Special bypasses (handled before Commander):
  *   --mcp                     Start MCP server over stdio
  *   reset-encrypted-columns   Recovery tool for broken encrypted credentials
+ *   reset-password            Reset the admin/management password
  *
  * All other commands are routed through Commander (bin/cli/program.mjs).
  */
@@ -210,6 +211,15 @@ if (process.argv.includes("reset-encrypted-columns")) {
   process.exit(exitCode ?? 0);
 }
 
+if (process.argv.includes("reset-password")) {
+  // bin/reset-password.mjs self-executes its `main()` on import and calls
+  // process.exit() on completion/error. Await a never-resolving promise so
+  // control never falls through to Commander (which would then reject
+  // `reset-password` as an unknown command). See #6261.
+  await import(pathToFileURL(join(ROOT, "bin", "reset-password.mjs")).href);
+  await new Promise(() => {});
+}
+
 try {
   const { createProgram } = await import(
     pathToFileURL(join(ROOT, "bin", "cli", "program.mjs")).href

bin/reset-password.mjs

@@ -5,10 +5,15 @@
  *
  * Usage:
  *   node bin/reset-password.mjs
- *   npx omniroute reset-password
+ *   omniroute reset-password
+ *
+ * Non-interactive / scripted usage (piped stdin, e.g. CI or Docker):
+ *   printf 'NewPass123\nNewPass123\n' | omniroute reset-password
+ *   printf 'NewPass123' | omniroute reset-password --password-stdin
  *
  * Resets the admin password for OmniRoute.
- * Prompts for a new password and updates the database directly.
+ * Prompts for a new password (interactive TTY) or reads it from stdin
+ * (non-TTY) and updates the database directly.
  *
  * @module bin/reset-password
  */
@@ -21,19 +26,61 @@ import { readManagementPasswordState, resetManagementPassword } from "./cli/sqli
 const DATA_DIR = resolveDataDir();
 const DB_PATH = resolveStoragePath(DATA_DIR);
 
-const rl = createInterface({
-  input: process.stdin,
-  output: process.stdout,
-});
+const MIN_PASSWORD_LENGTH = 8;
 
-function ask(question) {
-  return new Promise((resolve) => rl.question(question, resolve));
+/** Read the entire stdin stream as a UTF-8 string (used for non-TTY input). */
+function readAllStdin() {
+  return new Promise((resolve) => {
+    let data = "";
+    process.stdin.setEncoding("utf8");
+    process.stdin.on("data", (chunk) => {
+      data += chunk;
+    });
+    process.stdin.on("end", () => resolve(data));
+    process.stdin.on("error", () => resolve(data));
+    // Resuming is implied by attaching a 'data' listener, but be explicit so a
+    // paused stream (some spawn setups) still flows to EOF.
+    process.stdin.resume();
+  });
 }
 
-function exitWithError(message) {
-  console.error(message);
-  rl.close();
-  process.exit(1);
+/**
+ * Obtain the new password (and its confirmation).
+ *
+ * - `--password-stdin`: the ENTIRE stdin is the password, no confirmation.
+ * - non-TTY stdin (piped): read all of stdin once; first line is the password,
+ *   second line — when present — is the confirmation, else the first line is
+ *   reused (a single-line pipe means "no separate confirmation").
+ * - interactive TTY: two sequential prompts (unchanged behavior).
+ *
+ * The non-TTY path exists because two sequential `rl.question` promises never
+ * settle under a piped EOF — the second read blocks forever, so the reset was
+ * silently never applied (#6258).
+ */
+async function collectPassword() {
+  if (process.argv.includes("--password-stdin")) {
+    const raw = await readAllStdin();
+    const password = raw.replace(/[\r\n]+$/, "");
+    return { password, confirm: password };
+  }
+
+  if (!process.stdin.isTTY) {
+    const raw = await readAllStdin();
+    const lines = raw.split(/\r?\n/);
+    const password = lines[0] ?? "";
+    const confirm = lines[1] ? lines[1] : password;
+    return { password, confirm };
+  }
+
+  const rl = createInterface({ input: process.stdin, output: process.stdout });
+  try {
+    const ask = (question) => new Promise((resolve) => rl.question(question, resolve));
+    const password = await ask("Enter new password (min 8 chars): ");
+    const confirm = await ask("Confirm new password: ");
+    return { password, confirm };
+  } finally {
+    rl.close();
+  }
 }
 
 console.log("\n🔑 OmniRoute — Password Reset\n");
@@ -54,27 +101,34 @@ async function main() {
     console.log("ℹ️  No password is currently set.");
   }
 
-  const password = await ask("Enter new password (min 8 chars): ");
+  const { password, confirm } = await collectPassword();
 
-  if (!password || password.length < 8) {
-    exitWithError("\n❌ Password must be at least 8 characters.\n");
+  if (!password || password.length < MIN_PASSWORD_LENGTH) {
+    console.error(`\n❌ Password must be at least ${MIN_PASSWORD_LENGTH} characters.\n`);
+    process.exit(1);
   }
 
-  const confirm = await ask("Confirm new password: ");
-
   if (password !== confirm) {
-    exitWithError("\n❌ Passwords do not match.\n");
+    console.error("\n❌ Passwords do not match.\n");
+    process.exit(1);
   }
 
   await resetManagementPassword(password, DB_PATH);
-  rl.close();
 
   console.log("\n✅ Password reset successfully!");
   console.log("   Restart OmniRoute for changes to take effect.\n");
 }
 
-main().catch((err) => {
-  console.error(`\n❌ Error: ${err.message}\n`);
-  rl.close();
-  process.exit(1);
-});
+main()
+  .then(() => {
+    // Explicit exit(0) so a caller that imports this module (bin/omniroute.mjs
+    // routes `omniroute reset-password` here) terminates cleanly instead of
+    // hanging / exiting with code 13 on an unsettled wrapper await. On POSIX,
+    // console.log to a pipe is synchronous, so the success line is already
+    // flushed by the time we exit.
+    process.exit(0);
+  })
+  .catch((err) => {
+    console.error(`\n❌ Error: ${err.message}\n`);
+    process.exit(1);
+  });

config/quality/complexity-baseline.json

@@ -1,6 +1,7 @@
 {
   "_comment": "Catraca de complexidade (check-complexity.mjs, ESLint core rules complexity>=15 e max-lines-per-function>80 sobre src+open-sse+electron+bin via eslint.complexity.config.mjs). Conta total de violacoes; so pode cair. --update ratcheta.",
-  "count": 2035,
+  "count": 2050,
+  "_rebaseline_2026_07_07_v3846_release_close": "2035->2050 (+15). v3.8.46 release close (generate-release Phase 0 pre-flight): drift herdado do merge burst do ciclo (39 commits do dia + campanha /review-*). Trust-but-verify: os fixes de base-red do captain (agentSkills path.resolve #6366, catalogo cache #6408, tipagem de teste no-explicit-any, MitmProxyTab suppression) sao complexity-net-zero — check:complexity mede 2050 identico com e sem os fixes (a catraca NAO roda no fast-path PR->release, entao o ramo acumulou sem rebaselinar). Tighten via --update next cycle.",
   "_rebaseline_2026_07_04_v3844_release_close": "2026->2028 (+2). v3.8.44 release close (generate-release Phase 0/1): drift residual do fim do ciclo medido no tip pos-#6155 (merge burst final: #6155 cooling-panel + #6104 Kenari + #6139/#6128 provider-limits). Trust-but-verify: os 2 fixes de codigo do release-captain (model.ts alias boundary, auggie.ts stdin error handlers) adicionam 0 violacoes NOVAS — eslint.complexity direto nos 2 arquivos flagra apenas funcoes que ja estouravam o limite antes (runStreaming/start ja >80 linhas; resolveModelByProviderInference/getModelInfoCore pre-existentes de #5918), e resolveProviderAlias segue abaixo de 15. Logo o +2 e drift herdado do burst. Tighten via --update next cycle.",
   "_rebaseline_2026_07_03_v3844_ipfilter_release_green": "2015->2026 (+11). v3.8.44 cycle drift measured on release tip 32e4c906e during the #6131/#5975 release-green rebaseline. Inherited from the merge burst (Quality Ratchet does not run on PR->release fast-gates). route-edge-coverage +7 is my #5975 test comment; the rest is parallel-session drift. Tighten via --update next cycle.",
   "_rebaseline_2026_07_03_6007_sidecar_manifest": "2007->2015. PR #6007 (re-cut) adds sidecar/provider-manifest header wiring and the public manifest URL helper; the feature files themselves introduce 0 new complexity violations (check-complexity flags none in providerPluginManifestUrl.ts). The +8 vs the 2007 release baseline is inherited release/v3.8.44 drift absorbed at merge (check:complexity measures 2015 on the current release tip). Tighten via --update next cycle / at /generate-release Phase 0.",

config/quality/eslint-suppressions.json

@@ -242,6 +242,11 @@
       "count": 1
     }
   },
+  "src/app/(dashboard)/dashboard/settings/components/MitmProxyTab.tsx": {
+    "@next/next/no-html-link-for-pages": {
+      "count": 1
+    }
+  },
   "src/app/(dashboard)/dashboard/settings/components/ModelLockoutCard.tsx": {
     "react-hooks/exhaustive-deps": {
       "count": 1

config/quality/file-size-baseline.json

@@ -1,4 +1,5 @@
 {
+  "_rebaseline_2026_07_07_v3846_release_close": "Release v3.8.46 Phase 0 (generate-release) — drift de ciclo absorvido no fechamento (fast-gates PR->release nao rodam check:file-size). PROD god-files crescidos por merges do ciclo (nao meus; DECOMPOR idealmente, debt #3501): proxies.ts 1060->1173, chat.ts 1681->1751, ApiManagerPageClient.tsx 3058->3120, ProxyRegistryManager.tsx 1125->1437 (feature de proxy). TEST frozen: models-catalog-route.test.ts 1600->1605 (+5 do fix#2 do captain, #6408 catalogo cache), vscode-token-routes.test.ts 1212->1285 (cycle drift + os asserts effort_tiers/supportsThinking do #6241 alinhados no release-PR-CI base-red), que adiciona o import + 2 chamadas do hook __resetCatalogBuilderRunsForTest existente no setup (harness, sem asserts). Shrink estrutural rastreado no roadmap #3501.",
   "_rebaseline_2026_07_04_v3844_release_close": "Release v3.8.44 Phase 0 (generate-release): drift de ciclo absorvido no fechamento, medido no tip 415d159c8 (fast-gates PR->release nao rodam check:file-size). oauth/[provider]/[action]/route.ts 924->960 (#6054 zed keychain-import 400 gracioso; PR #6158 aberto extrai o guard e restaura o freeze — quando mergear, o frozen so encolhe), providerLimits.ts 982->998 (#6139 TOCTOU quota recovery + #6128), chat.ts 1647->1662 (#6057 per-request Auto-Combo X-OmniRoute-Mode/Budget + #6097), auth.ts 2405->2426 (#6139 + #6090 quota preflight lockouts + #5943 codex session affinity). Crescimento irreducivel em chokepoints existentes, coberto por testes por-PR; shrink estrutural rastreado no roadmap #3501.",
   "_rebaseline_2026_07_03_v3844_ipfilter_release_green": "testFrozen bumps: models-catalog-route 1507->1600, perplexity-web 959->999, route-edge-coverage 1234->1241 (last is my #5975 comment +7). v3.8.44 cycle drift measured on release tip 32e4c906e during the #6131/#5975 release-green rebaseline. Inherited from the merge burst (Quality Ratchet does not run on PR->release fast-gates). route-edge-coverage +7 is my #5975 test comment; the rest is parallel-session drift. Tighten via --update next cycle.",
   "_rebaseline_2026_07_03_v3844_residual_release_green": "Residual file-size drift on tip 716041223: providerLimits.ts 955->982 + accountFallback.ts 1790->1864 (production god-files grown by parallel-session merges e.g. #6128; ideally DECOMPOSE not rebaseline, tracked as debt) + sse-auth.test.ts 1553->1600. None mine.",
@@ -188,7 +189,7 @@
     "open-sse/utils/stream.ts": 2792,
     "src/app/(dashboard)/dashboard/HomePageClient.tsx": 1385,
     "src/app/(dashboard)/dashboard/analytics/ComboHealthTab.tsx": 1028,
-    "src/app/(dashboard)/dashboard/api-manager/ApiManagerPageClient.tsx": 3058,
+    "src/app/(dashboard)/dashboard/api-manager/ApiManagerPageClient.tsx": 3120,
     "src/app/(dashboard)/dashboard/cache/media/MediaPageClient.tsx": 1105,
     "src/app/(dashboard)/dashboard/cache/page.tsx": 845,
     "src/app/(dashboard)/dashboard/cli-code/components/CodexToolCard.tsx": 900,
@@ -201,8 +202,8 @@
     "src/app/(dashboard)/dashboard/playground/components/tabs/ApiTab.tsx": 847,
     "src/app/(dashboard)/dashboard/providers/[id]/ProviderDetailPageClient.tsx": 784,
     "src/app/(dashboard)/dashboard/providers/[id]/components/ConnectionRow.tsx": 942,
-    "src/app/(dashboard)/dashboard/providers/[id]/components/modals/AddApiKeyModal.tsx": 915,
-    "src/app/(dashboard)/dashboard/providers/[id]/components/modals/EditConnectionModal.tsx": 1231,
+    "src/app/(dashboard)/dashboard/providers/[id]/components/modals/AddApiKeyModal.tsx": 959,
+    "src/app/(dashboard)/dashboard/providers/[id]/components/modals/EditConnectionModal.tsx": 1278,
     "src/app/(dashboard)/dashboard/providers/[id]/hooks/useProviderConnections.ts": 954,
     "src/app/(dashboard)/dashboard/providers/[id]/hooks/useProviderModels.ts": 155,
     "src/app/(dashboard)/dashboard/providers/[id]/hooks/useProviderSettings.ts": 264,
@@ -215,7 +216,7 @@
     "src/app/(dashboard)/dashboard/settings/components/CompressionSettingsTab.tsx": 974,
     "src/app/(dashboard)/dashboard/settings/components/MemorySkillsTab.tsx": 898,
     "src/app/(dashboard)/dashboard/settings/components/PricingTab.tsx": 1012,
-    "src/app/(dashboard)/dashboard/settings/components/ProxyRegistryManager.tsx": 1125,
+    "src/app/(dashboard)/dashboard/settings/components/ProxyRegistryManager.tsx": 1437,
     "src/app/(dashboard)/dashboard/settings/components/ResilienceTab.tsx": 1183,
     "src/app/(dashboard)/dashboard/settings/components/RoutingTab.tsx": 1629,
     "src/app/(dashboard)/dashboard/settings/components/SystemStorageTab.tsx": 1924,
@@ -233,7 +234,7 @@
     "src/lib/db/migrationRunner.ts": 1125,
     "src/lib/db/models.ts": 1259,
     "src/lib/db/providers.ts": 1107,
-    "src/lib/db/proxies.ts": 1060,
+    "src/lib/db/proxies.ts": 1173,
     "src/lib/db/settings.ts": 1155,
     "src/lib/db/usageAnalytics.ts": 925,
     "src/lib/evals/evalRunner.ts": 961,
@@ -246,7 +247,7 @@
     "src/lib/usage/providerLimits.ts": 998,
     "src/lib/usage/usageHistory.ts": 988,
     "_rebaseline_2026_06_27_5193_5203_antigravity_oauthmodal": "Antigravity remote-login own growth: OAuthModal.tsx 960->969 (gate units). #5193 (+~4: remote paste instruction shown for all remote incl. Google + its rationale comment) and #5203 (+~5: handleManualSubmit credential-blob branch + button guard; submit logic extracted to oauthBlobSubmit.ts to minimize). Frozen set to the SUM so either merge order passes. Cohesive at the existing manual-submit chokepoint.",
-    "src/shared/components/OAuthModal.tsx": 989,
+    "src/shared/components/OAuthModal.tsx": 993,
     "src/shared/components/RequestLoggerV2.tsx": 1629,
     "src/shared/components/analytics/charts.tsx": 1558,
     "src/shared/constants/cliTools.ts": 875,
@@ -256,7 +257,7 @@
     "src/shared/services/cliRuntime.ts": 1100,
     "src/shared/validation/schemas.ts": 2523,
     "_rebaseline_2026_06_28_5275_correlation_id_extract": "Extraction of the safe CorrelationId subset of #5275 (hartmark) — request correlation id stored in call_logs (migration 109) and returned via the X-Correlation-Id response header, WITHOUT the combo/resilience or build/lazy-loading changes (those stay in #5275). Own growth: callLogs.ts 975->985 (correlation_id column on CallLogSummaryRow + read/map), usageHistory.ts 983->988 (correlationId metadata normalize), chat.ts 1575->1632 (withCorrelationId response wiring + combo-failure log carrying correlationId), chatHelpers.ts new 811 (withCorrelationId helper + reqId threading; was 791<cap pre-feature). Cohesive request/logging chokepoint wiring; structural shrink of chat.ts tracked in #3501.",
-    "src/sse/handlers/chat.ts": 1681,
+    "src/sse/handlers/chat.ts": 1751,
     "src/sse/handlers/chatHelpers.ts": 860,
     "src/sse/services/auth.ts": 2447,
     "open-sse/executors/default.ts": 877,
@@ -294,7 +295,7 @@
     "tests/unit/grok-web.test.ts": 2437,
     "tests/unit/image-generation-handler.test.ts": 2019,
     "tests/unit/model-sync-route.test.ts": 1016,
-    "tests/unit/models-catalog-route.test.ts": 1600,
+    "tests/unit/models-catalog-route.test.ts": 1605,
     "_rebaseline_pr4561_qwen_oauth_url": "Reconcile #4561 (port decolua/9router#683) already-merged growth: oauth-providers-config.test.ts 855->867 (+12, qwen.ai URL regression-pin test). Fast-gate PR->release does not run check:file-size, so this surfaced post-merge.",
     "_rebaseline_basered_codebuddy_cn": "Base-red fix (#4664 CodeBuddy CN): oauth-providers-config.test.ts 867->870 (+3) to align the EXPECTED provider list/config with the codebuddy-cn provider that #4664 added to the registry without updating this test (it asserts 'exactly once').",
     "_rebaseline_2026_06_27_5193_antigravity_test": "#5193 own test growth: oauth-providers-config.test.ts 870->873 (+3: antigravity projectId assertion + 50ms tick for the now fire-and-forget onboarding, matching the no-PKCE/no-openid flow).",
@@ -317,7 +318,7 @@
     "tests/unit/translator-openai-to-kiro.test.ts": 1234,
     "tests/unit/translator-resp-gemini-to-openai.test.ts": 1234,
     "tests/unit/usage-service-hardening.test.ts": 1633,
-    "tests/unit/vscode-token-routes.test.ts": 1212,
+    "tests/unit/vscode-token-routes.test.ts": 1285,
     "tests/unit/combo-config.test.ts": 881,
     "_rebaseline_2026_07_02_5928_base_red": "web-cookie-providers-new.test.ts 845->850: #5928 (test(security) Kimi Web URL host parse, CodeQL #689) grew the file +5 lines and merged into release/v3.8.44 WITHOUT rebaselining, leaving a fast-gates base-red that blocked every subsequent PR->release. Test growth is legitimate (a security regression test); maintainer absorbs the drift here. Frozen at 850.",
     "tests/unit/web-cookie-providers-new.test.ts": 890,
@@ -376,5 +377,8 @@
   "_rebaseline_2026_07_05_6154_copilot_catalog_helpers": "PR #6154 own growth: src/app/(dashboard)/dashboard/providers/[id]/providerPageHelpers.ts 1021->1034 (+13 = GitHub Copilot catalog refresh — model-section helper wiring for the refreshed passthrough/compatible model lists). Cohesive UI-helper growth alongside the registry/modelSpecs catalog refresh; not extractable. Covered by the PR's provider-registry-github-copilot-* unit tests. Fast-path PR->release skips check:file-size, so this bump lands with the PR (contributor backryun).",
   "_rebaseline_2026_07_05_6213_kiro_thinking_filesize": "PR #6213 own growth (kiro adaptive-thinking -> reasoning_content, +384): open-sse/translator/request/openai-to-kiro.ts 853->890 (+37 = additionalModelRequestFields builder for adaptive thinking: output_config.effort + thinking:{type:adaptive} + max_tokens, only when the request asked for thinking) and tests/unit/translator-openai-to-kiro.test.ts 1093->1234 (+141 = adaptive-thinking request/frame regression cases). The fast-path PR->release does NOT gate check:file-size on the merge, so this cohesive feature growth accumulated on the release tip (see the 2026-07-02 #5798 note for the same pattern). Superseded by the release captain's rebaseline-at-release.",
   "_rebaseline_2026_07_05_6235_doubao_dola": "PR #6235 own growth: tests/unit/web-cookie-providers-new.test.ts 850->890 (+40 = doubao-web -> Dola global provider switch regression cases: new host/cookie-domain/token-source assertions for www.dola.com). Cohesive test growth alongside the provider switch; contributor backryun. Fast-path PR->release skips check:file-size, so this bump lands with the PR.",
-  "_rebaseline_2026_07_06_v3845_release_close": "Release v3.8.45 cycle-close rebaseline (captain, sess ce897453): 13 files grown by the cycle's merged fix/feature PRs (#6216 streaming fixes + request-logger UI grew RequestLoggerV2/chat/chatHelpers/auth/stream/response-sanitizer.test; #6251/#6253 dashboard UX grew combos page/modals/wizard/ComboDefaultsTab/ProxyRegistryManager/providerPageHelpers). Growth is legitimate merged-feature code, absorbed at release per Phase 0 drift policy; all remain frozen (cannot grow further)."
+  "_rebaseline_2026_07_06_v3845_release_close": "Release v3.8.45 cycle-close rebaseline (captain, sess ce897453): 13 files grown by the cycle's merged fix/feature PRs (#6216 streaming fixes + request-logger UI grew RequestLoggerV2/chat/chatHelpers/auth/stream/response-sanitizer.test; #6251/#6253 dashboard UX grew combos page/modals/wizard/ComboDefaultsTab/ProxyRegistryManager/providerPageHelpers). Growth is legitimate merged-feature code, absorbed at release per Phase 0 drift policy; all remain frozen (cannot grow further).",
+  "_rebaseline_2026_07_06_6118_zed_oauthmodal": "PR #6118 own growth: OAuthModal.tsx 989->993 (+4 = Zed hosted native-app sign-in modal branch). Cohesive UI growth for the zed-hosted OAuth provider; not extractable. The prior 6118 comment set the note but left the frozen value at 989.",
+  "_rebaseline_2026_07_06_6351_glm_team_quota": "PR #6351 own growth (GLM team-plan quota fields threaded through the connection modals; new GlmTeamQuotaFields.tsx extracted): AddApiKeyModal.tsx ->951 (+9), EditConnectionModal.tsx ->1277 (+18). Absorbs the pre-existing session base-red on these frozen modals; release captain rebaseline-at-release supersedes.",
+  "_rebaseline_2026_07_06_6499_unique_default_name": "PR #6499 own growth: AddApiKeyModal.tsx 952->959 (+7 = a unique default connection name so a second API key for the same provider does not reuse 'main' and trigger the backend name-based upsert that silently overwrote the first connection). The pure name derivation was extracted to computeConnectionDefaultName.ts (unit-tested) to keep the growth minimal; the contributor's original full-form-reset rewrite was trimmed to a spread reset to avoid dropping the GLM team-quota fields #6351 added and to hold the frozen god-file growth down. Release captain rebaseline-at-release supersedes."
 }

config/quality/quality-baseline.json

@@ -120,7 +120,8 @@
       "_rebaseline_2026_06_26_v3837_release": "343->345. v3.8.37 cycle drift surfaced by the release-green pre-flight (the Quality Ratchet does NOT run on PR->release fast-gates, so warnings/complexity accrued unmeasured across this cycle's 76 commits — provider adds DGrid/Pioneer/xAI, headroom proxy lifecycle #4649, ~50 SSE/translator fixes, Engine Combos #5062). Trust-but-verify: this release-finalize working tree touches ONLY CHANGELOG.md, docs/i18n/*/CHANGELOG.md mirrors, and these baselines — 0 production-code change, so all drift is inherited cycle drift (`any` warn-allowed in open-sse/ + tests/). Tighten via --require-tighten next cycle."
     },
     "cognitiveComplexity": {
-      "value": 877,
+      "value": 882,
+      "_rebaseline_2026_07_07_v3846_release_close": "877->882 (+5). v3.8.46 release close (generate-release Phase 0 pre-flight): drift herdado do merge burst do ciclo. Trust-but-verify: os fixes de base-red do captain (agentSkills path.resolve #6366, catalogo cache #6408, tipagem de teste, MitmProxyTab suppression) sao cognitive-net-zero — check:cognitive-complexity mede 882 identico com e sem os fixes (a catraca NAO roda no fast-path PR->release). Tighten via --update next cycle.",
       "_rebaseline_2026_07_03_v3844_ipfilter_release_green": "861->867 (+6). v3.8.44 cycle drift measured on release tip 32e4c906e during the #6131/#5975 release-green rebaseline. Inherited from the merge burst (Quality Ratchet does not run on PR->release fast-gates). route-edge-coverage +7 is my #5975 test comment; the rest is parallel-session drift. Tighten via --update next cycle.",
       "_rebaseline_2026_07_03_v3844_review_prs_fix_batch": "860->861 (+1). Inherited v3.8.44 cycle drift surfaced by the release-green pre-flight during the /review-prs fix-batch round; check:cognitive-complexity measures 861 on the release tip 72ee80649. Negligible +1 from the round's / parallel-session merge burst (cognitive-complexity does NOT run on PR->release fast-gates). Structural shrink tracked in #3501. Tighten via --update next cycle.",
       "_rebaseline_2026_07_02_v3844_post_5939": "859->860 (+1). Same inherited post-3a3d618fe release drift as the complexity note of this date; this PR touches one CLI line (nullish default, no new function over the threshold) and one test file (not scanned). Tighten via --update next cycle.",
@@ -167,9 +168,10 @@
       "dedicatedGate": true
     },
     "bundleSize": {
-      "value": 5601,
+      "value": 6534,
       "direction": "down",
-      "dedicatedGate": true
+      "dedicatedGate": true,
+      "_rebaseline_2026_07_07_v3846_release_close": "5601->6534 (+933). v3.8.46 release close: gzip of the 4 bin/*.mjs entrypoints (size-limit + @size-limit/file) grew from this cycle's feature/fix merges pulled transitively into the CLI entrypoints (new providers, combo pipeline strategy #6396, effort/thinking standardization #6241, catalog cache-invalidation #6408). Measured 6534 locally via `check:bundle-size --ratchet` (deterministic gzip, matches CI). Legitimate cycle growth; shrink is separate debt."
     },
     "openapiBreaking": {
       "value": 0,

config/quality/test-masking-allowlist.json

@@ -16,6 +16,8 @@
   "tests/unit/combo-context-window-filter.test.ts": "v3.8.44 #6088: fix(combo) prefer known context capacity over unknown — a semântica do filtro mudou (unknown-context só sobrevive como fallback), o arquivo foi reescrito: 18 asserts pequenos (length/didFallback/ok) viraram 4 assert.deepEqual na LISTA ORDENADA completa de sobreviventes por caso (mais forte por caso). Asserts migrados ao novo contrato, não enfraquecidos. Verificado legítimo. Prune após v3.8.44 mergear para main.",
   "tests/unit/qoder-executor.test.ts": "v3.8.44 #5816: feat(qoder) drive PAT auth via qodercli — o executor migrou de chamadas HTTP diretas (api.qoder.com/api1.qoder.sh, headers Cosy-*) para o contrato stdio qodercli://; os asserts que pinavam URLs/headers da superfície aposentada foram substituídos por asserts do novo contrato (73→65). Asserts migrados à nova superfície, não enfraquecidos. Verificado legítimo. Prune após v3.8.44 mergear para main.",
   "tests/unit/qoder-jobtoken-exchange-4683.test.ts": "v3.8.44 #5816: feat(qoder) drive PAT auth via qodercli — o exchange de job token deixou de fazer o POST personal_token direto (agora via qodercli); 2 asserts da superfície HTTP aposentada removidos, demais migrados (27→25). Verificado legítimo, não mascaramento. Prune após v3.8.44 mergear para main.",
+  "tests/integration/v1-contracts-behavior.test.ts": "v3.8.46 #6303: fix(api) filter specialty model catalogs — os 10 asserts inline de SHAPE das listas embedding/image (object=='model', type=='embedding'/'image', typeof id/owned_by) foram removidos porque #6303 unificou os catálogos de especialidade e moveu a cobertura de shape para tests/unit/models-catalog-route.test.ts (que asserta embedding.type/gpt-image-2.type=='image'/owned_by de forma mais abrangente, +audio/rerank/video/music) e o comportamento de credential-hiding para tests/unit/specialty-model-catalog-routes.test.ts; net 44→42. Cobertura migrada, não enfraquecida. Verificado legítimo. Prune após v3.8.46 mergear para main.",
+  "tests/unit/check-docs-symbols.test.ts": "v3.8.46 release-PR CI: o gate check:docs-symbols (stale-enforcement) flagou as 2 ultimas entradas de KNOWN_STALE_DOC_REFS (/api/chat, /api/settings/tunnels) como CORRIGIDAS e exigiu remove-las, esvaziando o allowlist. Os 2 asserts removidos eram `assert.ok(size > 0)` (guarda non-empty) que ficaram FALSOS — um allowlist vazio e o estado valido/ideal (todo ref stale foi corrigido). A invariante significativa (/api/-shape de cada entrada presente) foi mantida; net 28->26. Nao e enfraquecimento — a assuncao non-empty tornou-se obsoleta. Prune apos v3.8.46 mergear para main.",
   "_deletedWithReplacement": {
     "_comment": "Deleções de arquivo de teste com SUBSTITUTO verificado (o gate exige que o replacement exista no HEAD e seja arquivo de teste). Uso restrito ao caso 'reescrito em outro path sem rename detectável pelo -M do git'. Cada entrada precisa de reason com PR ref e passa por revisão humana no release PR. Prune após o release mergear para main.",
     "open-sse/services/combo/__tests__/targetExhaustion.test.ts": {

docs/architecture/ARCHITECTURE.md

@@ -17,13 +17,13 @@ It provides a single OpenAI-compatible endpoint (`/v1/*`) and routes traffic acr
 
 Core capabilities:
 
-- OpenAI-compatible API surface for CLI/tools (237 providers, 73 executors)
+- OpenAI-compatible API surface for CLI/tools (237 providers, 75 executors)
 - Request/response translation across provider formats
 - Model combo fallback (multi-model sequence)
 - Structured combo steps (`provider + model + connection`) with runtime ordering by `compositeTiers`
 - Account-level fallback (multi-account per provider)
 - Quota preflight and quota-aware P2C account selection in the main chat path
-- OAuth + API-key provider connection management (17 OAuth provider modules)
+- OAuth + API-key provider connection management (19 OAuth provider modules)
 - Embedding generation via `/v1/embeddings` (6 providers, 9 models)
 - Image generation via `/v1/images/generations` (10+ providers, 20+ models)
 - Audio transcription via `/v1/audio/transcriptions` (7 providers)
@@ -66,7 +66,7 @@ Core capabilities:
 - Prompt injection guard middleware
 - Prompt compression pipeline with Caveman, RTK, stacked pipelines, compression combos, language packs, and analytics
 - ACP (Agent Communication Protocol) registry
-- Modular OAuth providers (16 individual modules under `src/lib/oauth/providers/`)
+- Modular OAuth providers (19 individual modules under `src/lib/oauth/providers/`)
 - Uninstall/full-uninstall scripts
 - OAuth environment repair action
 - WebSocket bridge for OpenAI-compatible WS clients (`/v1/ws`)

docs/architecture/CODEBASE_DOCUMENTATION.md

@@ -452,7 +452,7 @@ open-sse/
 ├── types.d.ts
 ├── config/                 Provider registries, header profiles, identity, …
 ├── handlers/               Request handlers (chat, embeddings, audio, image, …)
-├── executors/              45 provider-specific HTTP executors
+├── executors/              75 provider-specific HTTP executors
 ├── translator/             Format conversion (OpenAI ↔ Claude ↔ Gemini ↔ Cursor ↔ Kiro)
 ├── transformer/            Responses API ↔ Chat Completions stream transformer
 ├── services/               80+ service modules (combos, fallback, quotas, identity, …)
@@ -482,7 +482,7 @@ open-sse/
 
 ### 4.2 `open-sse/executors/`
 
-73 provider executors, each extending `BaseExecutor` (`base.ts`):
+75 provider executors, each extending `BaseExecutor` (`base.ts`):
 
 `antigravity`, `azure-openai`, `blackbox-web`, `chatgpt-web`, `cliproxyapi`,
 `cloudflare-ai`, `codex`, `commandCode`, `cursor`, `default`, `devin-cli`,

docs/architecture/RESILIENCE_GUIDE.md

@@ -201,7 +201,7 @@ Bounded by `comboCooldownWait` (`enabled`, `maxWaitMs` 5s, `maxAttempts` 2,
 
 ## Other Resilience Features
 
-- **17 routing strategies** (priority, weighted, round-robin, context-relay, fill-first, p2c, random, least-used, cost-optimized, reset-aware, reset-window, headroom, strict-random, auto, lkgp, context-optimized, fusion) — see [AUTO-COMBO.md](../routing/AUTO-COMBO.md).
+- **18 routing strategies** (priority, weighted, round-robin, context-relay, fill-first, p2c, random, least-used, cost-optimized, reset-aware, reset-window, headroom, strict-random, auto, lkgp, context-optimized, fusion, pipeline) — see [AUTO-COMBO.md](../routing/AUTO-COMBO.md).
 - **Reset-aware routing** (v3.8.0) — prioritizes connections by quota reset time.
 - **Background mode degradation** — Responses API `background: true` degraded to sync with warning.
 - **Dynamic tool limit detection** — backs off providers when tool count limits hit.

docs/frameworks/SEARCH_TOOLS_STUDIO.md

@@ -57,7 +57,7 @@ New tab for extracting content from a URL via `POST /v1/web/fetch` (created in p
 - Submit → fetch → render `ScrapeResult.tsx`.
 - `ScrapeResult` renders markdown preview + raw toggle.
 - Cap: if response body > **256 KB**, UI shows `(truncated, view raw)` and opens raw in a Monaco modal (D21).
-- Metadata panel: provider (firecrawl/jina-reader/tavily-search), latency, cost, response size, links count.
+- Metadata panel: provider (firecrawl/jina-reader/tavily-search/tinyfish), latency, cost, response size, links count.
 - Uses `useScrapeFetch.ts` hook.
 
 ### Compare Tab
@@ -108,7 +108,7 @@ Runs the same query/URL across up to **4 providers in parallel** (D22):
 | Field                          | Source                                                                                     |
 | ------------------------------ | ------------------------------------------------------------------------------------------ |
 | `id`, `name`                   | `searchRegistry.ts`                                                                        |
-| `kind`                         | `"search"` (12 providers) or `"fetch"` (firecrawl, jina-reader, tavily-search)             |
+| `kind`                         | `"search"` (12 providers) or `"fetch"` (firecrawl, jina-reader, tavily-search, tinyfish)   |
 | `costPerQuery`                 | Registry data                                                                              |
 | `freeMonthlyQuota`             | Registry data                                                                              |
 | `searchTypes` / `fetchFormats` | Registry data                                                                              |
@@ -136,7 +136,7 @@ Only one backend change was needed for this feature:
 
 `src/app/api/search/providers/route.ts` was extended to:
 
-- Include all 3 fetch providers (`firecrawl`, `jina-reader`, `tavily-search`) in the array.
+- Include all 4 fetch providers (`firecrawl`, `jina-reader`, `tavily-search`, `tinyfish`) in the array.
 - Add `kind: "search" | "fetch"` to every item.
 - Add `status: "configured" | "missing" | "rate_limited"` derived from live credential state.
 - Maintain backward compatibility — existing fields (`id`, `name`, etc.) unchanged.

docs/i18n/ar/CHANGELOG.md

@@ -6,6 +6,213 @@
 
 ## [3.8.31] — 2026-06-20
 
+## [3.8.46] — TBD
+
+### ✨ New Features
+
+- **feat(sse):** **hide paid-only models from `auto/*` routing** when `hidePaidModels` is on ([#6512](https://github.com/diegosouzapw/OmniRoute/issues/6512)) — follow-up to #6328/#6495. PR #6495 hid paid-only models from the `GET /v1/models` listing, but `auto/*` combos (`auto/best-coding`, `auto/glm`, …) could still pick a paid-only backend into their candidate pool → a 402/403 at request time. `createVirtualAutoCombo` now filters the candidate pool through the new pure `open-sse/services/autoCombo/paidModelFilter.ts` (`filterPaidOnlyCandidates`), applying the same free-model predicate #6495 uses in `catalog.ts` (`providerHasFreeModels(provider) && isFreeModel(provider, {id})`) whenever `settings.hidePaidModels === true`. Applied before the category/tier/family narrowing, so it covers every `auto/*` combo; an all-paid pool degrades to the existing graceful empty-pool path. **Opt-in — default OFF leaves the pool unchanged** (identity). Regression guard: `tests/unit/autoCombo/paid-model-filter-6512.test.ts` (4, incl. the default-off identity guard).
+- **feat(sse):** **provider-family auto combos** — `auto/glm`, `auto/minimax`, `auto/mimo`, `auto/zai`, `auto/gemma`, `auto/llama`, `auto/gemini` ([#6453](https://github.com/diegosouzapw/OmniRoute/issues/6453)) — new routable ids that materialize an on-demand virtual combo spanning whatever installed backends currently expose that model family, degrading gracefully as backends rotate. A new pure `open-sse/services/autoCombo/modelFamily.ts` (`detectModelFamily`) classifies by model-id prefix for six families; `zai` is instead resolved by provider id (z.ai's hosted API serves the same `glm-*` model ids as every other GLM backend, so `auto/zai` means "route to my z.ai backend specifically" vs `auto/glm`'s "any connected GLM backend"). Reuses the existing `createVirtualAutoCombo` on-demand materialization path (no DB writes) and the `/v1/models` catalog advertising loop. Regression guard: `tests/unit/autoCombo/provider-family-combos.test.ts` (11).
+- **feat(proxy):** native **proxy-pool round-robin / egress IP rotation** ([#6365](https://github.com/diegosouzapw/OmniRoute/issues/6365)) — a scope (global / provider / account) can now hold **multiple** proxies as a pool with a rotation strategy, so outbound requests cycle their egress IP instead of pinning one proxy per scope. Migration `117_proxy_pool_rotation.sql` lifts the `UNIQUE(scope, scope_id)` constraint (rebuild via the canonical rename/copy/drop; existing single assignments become 1-element pools) and adds a `proxy_scope_rotation` companion table holding the per-scope strategy + a persisted monotonic round-robin cursor. Strategies: `round-robin` (default, monotonic cursor — never `Math.random`), `random`, and `sticky-per-N-min`. Resolution (`resolveProxyForScopeFromRegistry` / `resolveProxyForConnectionFromRegistry`) now fetches the alive, position-ordered candidate set (unchanged `PROXY_ALIVE_PREDICATE`) and applies the strategy; an empty / all-dead pool still returns `null` — the #6246 fail-closed guard is **untouched** (never falls through to direct egress). Backend + DB only; dashboard pool-builder UI is a follow-up. Regression guard: `tests/unit/proxy-pool-rotation-6365.test.ts` (8, incl. fail-closed + backward-compat).
+- **feat(providers):** end-to-end **tool/function calling on the native Gemini `/v1beta` endpoint** ([#6222](https://github.com/diegosouzapw/OmniRoute/issues/6222)) — both directions of the Gemini↔OpenAI conversion now preserve tool data (previously silently dropped). Request side: `convertGeminiToInternal` (extracted to its own testable module) maps `tools[].functionDeclarations` → OpenAI `tools`, prior `functionCall` parts → assistant `tool_calls`, and `functionResponse` parts → `tool`-role messages. Response side: `convertOpenAIResponseToGemini` emits `parts[].functionCall {name,args}` from `message.tool_calls`, and the streaming `openAIChunkToGeminiChunk` accumulates fragmented `tool_calls` deltas by index into complete `functionCall` parts. The non-Gemini client paths (Claude, OpenAI-Responses) already preserved tool calls — this closes the gap specific to the native Gemini surface. Regression guard: `tests/unit/v1beta-gemini-tool-calling-6222.test.ts` (6, incl. a streaming SSE round-trip).
+- **feat(providers):** copilot-m365-web **enterprise / work tier** support ([#6334](https://github.com/diegosouzapw/OmniRoute/issues/6334)) — mirrors the EDU-tier pattern (#6210): `M365ConnectionParams` gains an `agent` field, a new opt-in `M365_ENTERPRISE_OVERRIDES` preset (`agent=work`, `scenario=officeweb`, `licenseType=Premium`) applies via `providerSpecificData.tier="enterprise"` (alias `"work"`), and `agent` is also overridable directly via `providerSpecificData.agent`. `buildWsUrl` was hardcoding `agent="web"` (the one enterprise-distinguishing param with no override path), so a Premium work account handshook then returned an empty stream. The individual and EDU paths are untouched. Kilo's dup flag vs #6210 (EDU tier) was a false positive — different tier. Regression guard: `tests/unit/copilot-m365-enterprise-6334.test.ts` (7). End-to-end confirmation on a real Premium work account is a live-VPS validation follow-up (Hard Rule #18). (thanks @Forcerecon)
+- **feat(api):** standardized, provider-agnostic **`effort` + `thinking` request params** ([#6241](https://github.com/diegosouzapw/OmniRoute/issues/6241)) — a thin standardization layer over the existing mature per-provider reasoning plumbing (no provider mapper touched). `providerChatCompletionSchema` gains a canonical `effort` (reusing the shared `none/low/medium/high/xhigh` vocabulary — the UI tiers `extra`/`max` collapse onto `xhigh`) and a boolean `thinking`. A pure `normalizeReasoningRequest` (wired once in `src/sse/handlers/chat.ts`, before any reasoning field is read) folds them onto the fields the translators already consume (`reasoning_effort` / `reasoning.effort` / `thinking`), so they fan out to Anthropic / Gemini / xAI / Responses — an explicit client `reasoning_effort` / object-shaped `thinking` always wins (backward-compatible). `/models` additively exposes `supportsThinking` + `effort_tiers` so the frontend can render the toggles (UI component is a follow-up). Regression guard: `tests/unit/effort-thinking-standardization-6241.test.ts` (12). (thanks @Iammilansoni, @shabeer)
+- **feat(combo):** new **`pipeline` (sequential) combo strategy** ([#6297](https://github.com/diegosouzapw/OmniRoute/issues/6297)) — the 18th routing strategy runs targets **in order**, threading each step's output into the next step's input, with an optional per-step `prompt` (system instruction); only the final step's response is returned. Distinct from `fusion` (parallel fan-out + judge). Implemented as a self-contained `open-sse/services/pipeline.ts` (sibling to `fusion.ts`), dispatched from `combo.ts`; the step list reuses `combo.models` order and reads an optional `prompt` off each target (backward-compatible — ignored by every other strategy). Intermediate steps run non-streaming with tools stripped (complete prose to thread forward); the final step keeps the client's `stream` flag + tools. A failing/empty/unparseable intermediate step fails the whole pipeline explicitly via a sanitized error (never silently swallowed). Kilo's dup flag vs #563 was a false positive (that's model→chain selection; this is a sequential chain). Regression guard: `tests/unit/combo-pipeline-strategy.test.ts` (5). (thanks @ofekbetzalel)
+- **feat(ci):** `check:test-masking` now flags **inline-reimplemented prod conditions** ([#6348](https://github.com/diegosouzapw/OmniRoute/issues/6348)) — a new **report-only** subcheck (v2, 6A.10 family) catches the wrong-shape contract test: a test that recomputes the condition under test inline instead of importing/exercising the real function (the #6216 class, where `=== 500` → `>= 500` stayed green because the test re-implemented the branch). For each added/modified test file it warns when the file textually duplicates a ≥3-token conditional from a production file touched in the same PR **and** does not import the symbol/module owning it, via a pure, fixture-tested `findReimplementedConditions()` with an allowlist mirroring `assertReductionAllowlist`. Report-only for now (does not fail the gate) — to be promoted to blocking after a triage cycle. Regression guard: `tests/unit/check-test-masking.test.ts` (45).
+- **feat(sse):** per-connection routing override (native vs CLIProxyAPI) ([#6339](https://github.com/diegosouzapw/OmniRoute/issues/6339)) — the previously-dead `isCliproxyapiDeepModeEnabled` helper is now wired into `resolveExecutorWithProxy`: a single connection can opt itself into the CLIProxyAPI passthrough executor via `providerSpecificData.cliproxyapiMode="claude-native"`, with precedence **connection override > provider `upstream_proxy_config` mode > default**. `resolveExecutorWithProxy` now receives the resolved connection's `providerSpecificData` (threaded from `chatCore.ts`), so one connection can deep-route while the provider's other connections stay native — no DB schema change (the toggle rides in `providerSpecificData`). Also resolves the same-provider mixing ask in #6340. Regression guard: `tests/unit/chatcore-executor-proxy.test.ts` (9). (thanks @RaviTharuma)
+- **feat(dashboard):** "Add session cookie" modal now shows a prominent **"Open ‹host› →"** link to the provider's own site ([#6268](https://github.com/diegosouzapw/OmniRoute/issues/6268)) — every `-web` cookie-session provider (chatgpt-web, claude-web, gemini-web, kimi-web, lmarena, qwen-web, m365-copilot-web, …) renders a one-click external link (opening the provider's login/home page in a new tab) so operators no longer tab away to retype the URL mid-setup. The host resolves from a pure, unit-tested `resolveWebProviderHost()` (prefers `WEB_COOKIE_PROVIDERS[id].website`, falls back to the registry `baseUrl` origin); non-web providers render exactly as before. Kilo's dup flag vs #6265 (modal-too-small-on-1080p) was a false positive — distinct concern. Regression guard: `tests/unit/resolve-web-provider-host.test.ts` (5). (thanks @chirag127)
+- **feat(providers):** add **DigitalOcean** AI (serverless inference) as an OpenAI-compatible API-key provider ([#6373](https://github.com/diegosouzapw/OmniRoute/pull/6373)) — base `https://inference.do-ai.run/v1`, wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/digitalocean/`, `src/shared/constants/providers/apikey/inference-hosts.ts`). Regression guard: `tests/unit/digitalocean-provider.test.ts`. (thanks @newnol)
+- **feat(providers):** add **Huancheng Public API** (`hcnsec`) as an OpenAI-compatible regional provider ([#6410](https://github.com/diegosouzapw/OmniRoute/pull/6410)) — Xinjiang Huancheng Cybersecurity's public LLM platform (base `https://api.hcnsec.cn/v1`, free credits via daily check-ins), wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/hcnsec/`, `src/shared/constants/providers/apikey/regional.ts`). Regression guard: `tests/unit/hcnsec-provider.test.ts`. (thanks @UnrealAryan)
+- **feat(dashboard):** the web-session credential guide now shows an **"Open {host}" link** ([#6316](https://github.com/diegosouzapw/OmniRoute/pull/6316)) to the provider's sign-in site (derived from the provider `website` via `getProviderWebsiteHost`), so you can jump straight to the page where the cookie/session must be captured. Regression guard: `tests/unit/web-session-provider-link-6316.test.ts`. (thanks @jordansilly77-stack)
+- **feat(cerebras):** add the **Gemma 4 31B** model (`gemma-4-31b`) to the Cerebras registry + pricing table ([#6331](https://github.com/diegosouzapw/OmniRoute/pull/6331)). Regression guard extends `tests/unit/t28-model-catalog-updates.test.ts`. (thanks @backryun)
+- **feat(providers):** add **Yuanbao (web)** as a cookie-session provider ([#6196](https://github.com/diegosouzapw/OmniRoute/issues/6196)) — `yuanbao-web` (Tencent Yuanbao, `yuanbao.tencent.com`) with cookie-only auth (`hy_user`/`hy_token` + public agent id), SSE→OpenAI translation incl. `reasoning_content`, exposing DeepSeek V3/R1 + Hunyuan / Hunyuan-T1. Regression guard: `tests/unit/providers-yuanbao-web.test.ts`. `together-web` was **deferred** (no verifiable web-session endpoint — needs a captured request) and `huggingchat-web` **dropped** (the existing `huggingchat` already is a web-cookie provider). (thanks @chirag127)
+- **feat(providers):** route the built-in **agentrouter** through the dynamic Claude-Code wire image ([#6056](https://github.com/diegosouzapw/OmniRoute/issues/6056)) — a small static allow-set (`CC_WIRE_IMAGE_BUILTINS` in `open-sse/services/ccWireImageBuiltins.ts`), consulted by `isClaudeCodeCompatible` / `isClaudeCodeCompatibleProvider` / `applyFingerprint`, makes agentrouter adopt the CC wire-image headers + fingerprint **while guarding the CC baseUrl/auth branches** so it keeps its own registry `baseUrl` and `x-api-key` auth. Regression guard: `tests/unit/agentrouter-cc-wire-image.test.ts` (asserts the wire image is applied AND agentrouter's baseUrl/auth are preserved). Live WAF-acceptance against agentrouter.org is a VPS validation follow-up (Hard Rule #18).
+- **feat(providers):** **bulk-add API keys for Cloudflare Workers AI** ([#6174](https://github.com/diegosouzapw/OmniRoute/issues/6174)) — `cloudflare-ai` is removed from the bulk-add exclusion list and the bulk parser gains a 3-field `name|accountId|apiKey` mode; the bulk route now builds a **per-entry** `providerSpecificData` so each key carries its own `accountId` (fixing the previous shared-object reuse), and both the create + key-validation paths receive it. Regression guard: `tests/unit/bulk-api-key-parser-cloudflare.test.ts`. (thanks @muflifadla38)
+- **feat(dashboard):** routing/settings UX clarity ([#6147](https://github.com/diegosouzapw/OmniRoute/issues/6147)) — (1) weighted combos show the **effective routing share %** next to each weight when weights don't sum to 100 (`WeightTotalBar.tsx`); (2) the status widget's user-facing **"Cloud Sync" label is renamed** to "Remote Settings Sync" (`CloudSyncStatus.tsx`; internal ids/state untouched); (3) built-in providers gain an **opt-in advanced base-URL override** (`isBaseUrlOverrideEligibleProvider`, hidden behind an "Advanced" toggle, reusing the existing `providerSpecificData.baseUrl` persistence — not globally widened). Regression guard: `tests/unit/routing-settings-ux-6147.test.ts`.
+- **feat(combo):** add an option to **disable session stickiness**, per-combo or globally — round-robin / random combos can rotate to a different connection on every request instead of pinning a whole conversation to one connection by its first-message hash. Resolution precedence per-combo `config.disableSessionStickiness` → global `settings.disableSessionStickiness` → default `false` (preserves the #3825 prompt-cache/504 fix); gates **both** stickiness call sites in `open-sse/services/combo.ts`. Exposed as a global toggle (Combo Defaults) and a per-combo Inherit/on/off control. ([#6168](https://github.com/diegosouzapw/OmniRoute/issues/6168)) Regression guard: `tests/unit/combo-disable-session-stickiness.test.ts`. (thanks @RCrushMe)
+- **feat(docker):** add the `OMNIROUTE_NO_SUDO` env flag for root-less / user-namespaced deployments — the MITM cert-trust command path (`resolveSudoSpawn` in `src/mitm/systemCommands.ts`) now strips the leading `sudo` when the flag is truthy, in addition to the existing root / sudo-missing cases, so the Proxy Agent runs without `sudo` (the operator trusts the CA manually, e.g. via `NODE_EXTRA_CA_CERTS`). Argv-array `spawn` preserved — no shell interpolation (Hard Rule #13). ([#6122](https://github.com/diegosouzapw/OmniRoute/issues/6122)) Regression guard: `tests/unit/mitm-systemCommands-no-sudo.test.ts`. (thanks @powellnorma)
+- **feat(providers):** add **Requesty** as an OpenAI-compatible gateway provider (BYOK, base `https://router.requesty.ai/v1`, ~200 free requests/day) — wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/requesty/`, `src/shared/constants/providers/apikey/gateways.ts`). ([#6120](https://github.com/diegosouzapw/OmniRoute/issues/6120)) Regression guard: `tests/unit/requesty-provider.test.ts`. (thanks @chirag127)
+- **feat(dashboard):** add **configured-only / available-only filters** to the Free Provider Rankings page ([#6150](https://github.com/diegosouzapw/OmniRoute/issues/6150)) — hide providers you haven't configured, or whose connections are all rate-limited / out of quota, via server-side query params (`?configuredOnly` / `?availableOnly` on `GET /api/free-provider-rankings`) backed by a testable lib helper reusing the in-process connection state (no Redis). Both filters default off, so the default view is unchanged; this supersedes the earlier client-side "Configured Only" toggle (#6245) with an available-only dimension and unit-tested logic. Regression guard: `tests/unit/freeProviderRankings-filters.test.ts`.
+
+### 🔧 Bug Fixes
+
+- **fix(dashboard):** adding a second API key connection for the same provider no longer silently overwrites the first — the Add-API-key modal now derives a unique default connection name (`main`, then `main-2`, `main-3`, …) so the backend name-based upsert can't collide (#6499 — thanks @dilneiss).
+- **fix(compression):** the session-dedup engine now also deduplicates a large multi-line block repeated **within a single message** (intra-message dedup), not just across turns; the compression-preview API surfaces a `fallbackReason`, and the fusion panel reports how many models were rate-limited vs failed on total-panel failure (#6501 — thanks @chirag127).
+- **fix(compression):** a stacked-pipeline step naming an unregistered engine now surfaces a `validationErrors` entry instead of silently no-op'ing, so misconfigured pipelines are visible in the preview API (#6506 — thanks @chirag127).
+- **feat(usage):** add a **Codex reset-credit redemption** flow to the Provider Limits UI ([#6361](https://github.com/diegosouzapw/OmniRoute/pull/6361)) — a `useCodexResetCreditRedemption` hook + `/api/usage/codex-reset-credit` route + `codexResetCredits` lib let you redeem banked Codex reset credits from the quota card. Regression guard: `tests/unit/codex-reset-credits.test.ts`. (thanks @JxnLexn)
+- **feat(glm):** add **team-plan quota settings** for `glm-cn` connections ([#6351](https://github.com/diegosouzapw/OmniRoute/pull/6351)) — a dedicated `GlmTeamQuotaFields` form section (team quota id / limits) threaded through the Add/Edit connection modals, persisted via `providerSpecificData`, with the GLM usage service reading the team quota. Regression guards: `tests/unit/glm-team-quota.test.ts`, `provider-specific-data-schema.test.ts`. (thanks @hao3039032)
+- **feat(providers):** add **TinyFish** web-fetch/search support ([#6349](https://github.com/diegosouzapw/OmniRoute/pull/6349)) — a `tinyfish-fetch` executor + `/v1/web/fetch` route + MCP web-fetch tool, registered as a specialty-media provider with request-validation and a search-provider catalog entry. Regression guards: `tests/unit/executor-tinyfish-fetch.test.ts`, `web-fetch-handler.test.ts`, `mcp-web-fetch-tool.test.ts`, `provider-validation-tinyfish.test.ts`. (thanks @dtybnrj)
+- **fix(cli):** `omniroute launch-codex` now spawns `codex.cmd` through a shell on Windows (the npm `.cmd` shim is unresolvable by bare `spawn` → ENOENT), mirroring the qodercli Windows fix (#6263) ([#6312](https://github.com/diegosouzapw/OmniRoute/pull/6312)). Regression guard: `tests/unit/launch-codex-windows-spawn-6312.test.ts`. (thanks @swingtempo)
+- **fix(codex):** isolate the **Spark** quota from the shared Codex quota and stabilize the quota UI ordering / hydration so per-scope limits render consistently ([#6336](https://github.com/diegosouzapw/OmniRoute/pull/6336)). Regression guards: `tests/unit/codex-quota-selection-hydration.test.ts`, `provider-limits-ui.test.ts` + 3 more. (thanks @xz-dev)
+- **feat(api):** add a `hidePaidModels` setting that filters paid-only models out of the `/v1/models` catalog. Regression guard: `tests/unit/models-catalog-hide-paid.test.ts`. (thanks @chirag127)
+- **fix(api-manager):** the fallback model picker now preserves combos instead of dropping them when a primary model is unavailable ([#6443](https://github.com/diegosouzapw/OmniRoute/pull/6443)). Regression guard: `tests/unit/api-manager-page-static.test.ts`. (thanks @jmengit)
+- **fix(providers):** recoverable Antigravity / Cloud-Code (Gemini Code Assist) `403` responses ([#6452](https://github.com/diegosouzapw/OmniRoute/pull/6452)) are now classified as a retryable project-config error instead of a terminal account ban, so a fixable project/API-disabled 403 no longer forces a ~1-year cooldown / full OAuth reconnect. Regression guard: `tests/unit/errorclassifier-antigravity-403.test.ts`. (thanks @developerjillur)
+- **fix(mitm):** `sanitizeHeaders` now redacts `Set-Cookie` response headers so upstream session cookies never leak into logs / diagnostics ([#6451](https://github.com/diegosouzapw/OmniRoute/pull/6451)). Regression guard: `tests/unit/mitm-sanitize-headers.test.ts`. (thanks @developerjillur)
+- **fix(api):** `/api/compression/preview` now accepts `mode: "caveman"` and correctly handles stacked / zero-compression previews ([#6425](https://github.com/diegosouzapw/OmniRoute/issues/6425)). Regression guard: `tests/unit/api/compression-preview-caveman-and-stacked-6425.test.ts`. (thanks @chirag127)
+- **feat(providers):** add **Zed** hosted LLM aggregator as a native-app provider ([#6118](https://github.com/diegosouzapw/OmniRoute/pull/6118)) — OAuth sign-in via the Zed hosted flow, registered through the shared provider registry + executor. Regression guards: `tests/unit/zed-oauth-provider.test.ts`, `zed-import-utils.test.ts`, `zed-docker-detect.test.ts`, `mitm-handler-zed.test.ts`. VPS-validated via live operator login (Hard Rule #18).
+- **fix(oauth):** the Kiro SSO-cache auto-import now **preserves the IDC region** — cross-region Amazon Q / Kiro profiles imported from the SSO cache are no longer collapsed to the default region ([#6113](https://github.com/diegosouzapw/OmniRoute/pull/6113)). Regression guard: `tests/unit/kiro-auto-import-idc-2059.test.ts`. VPS-validated via live operator login (Hard Rule #18).
+- **fix(dashboard):** passthrough model aliases no longer collide when two namespaced model ids share a last segment (port from 9router#1850, [#6431](https://github.com/diegosouzapw/OmniRoute/pull/6431)). `enx/gpt-5.5` and `enx/codebuddy/gpt-5.5` both auto-generated the alias `gpt-5.5`, so the second model could never be added (the UI just alerted "alias already exists"). Aliases are now disambiguated deterministically — bare last segment when free, then parent-qualified (`codebuddy-gpt-5.5`), then a numeric suffix — while re-adding the exact same model id is still blocked. Regression guard: `tests/unit/passthrough-alias-1850.test.ts`. (thanks @arpicato)
+- **fix(translator):** preserve a Gemini `functionResponse` co-located with other parts (another `functionCall`, or trailing `text`) in the same content when translating **Gemini → OpenAI** ([#6376](https://github.com/diegosouzapw/OmniRoute/pull/6376)). `convertGeminiContent()` early-returned the tool message on the first `functionResponse` part, dropping any co-located parts; such contents are now pre-split (one tool message per `functionResponse`, emitted first, plus one message for the remaining parts). Regression guard: `tests/unit/gemini-to-openai-function-response.test.ts`. (thanks @warelik)
+- **fix(headroom):** detect a python interpreter managed by **mise / pyenv / asdf / conda** (port from 9router#2353, [#6382](https://github.com/diegosouzapw/OmniRoute/pull/6382)). Headroom's python probe (`src/lib/headroom/detect.ts`) searched a hardcoded `PATH`, but version managers expose their interpreters via shim dirs that only join `PATH` through interactive-shell activation — which the non-interactive server never runs, so a managed python (≥3.10) was invisible and Headroom reported it missing. The search path now prepends the well-known shim/bin dirs (`~/.local/share/mise/shims`, `~/.pyenv/shims`, `~/.asdf/shims`, `$CONDA_PREFIX/bin`, `~/.local/bin`, respecting `MISE_DATA_DIR`/`PYENV_ROOT`/`ASDF_DATA_DIR` when set), and a new `HEADROOM_PYTHON` env override lets operators point straight at their interpreter (mirroring `HEADROOM_URL`). Still shell-free (`execFileSync`). Regression guard: `tests/unit/headroom-detect.test.ts` (5). (thanks @loopyd)
+- **fix(executors):** strip the OpenAI-Codex/Claude-CLI `client_metadata` passthrough field for **NVIDIA** requests (port from 9router#1887, [#6411](https://github.com/diegosouzapw/OmniRoute/pull/6411)). NVIDIA's OpenAI-compatible wrapper rejects it with `400 Unsupported parameter`, the same class already handled for `cerebras`/`mistral`; `nvidia` (executor `default`) was missing from the strip allowlist so Codex/Claude-Code passthrough requests 400'd. Regression guard: `tests/unit/executor-default-strip-client-metadata.test.ts` (+nvidia case). (thanks @phidinhmanh)
+- **fix(translator):** strip the Claude-style `thinking` field for **NVIDIA `z-ai/glm-5.2`** (port from 9router#2023, [#6413](https://github.com/diegosouzapw/OmniRoute/pull/6413)). NVIDIA's OpenAI-compatible wrapper 400s on `thinking` (a Claude-format client routed here leaves a `thinking:{type:"adaptive"}`); the existing strip rule only dropped `reasoning`. Same class already handled for `minimax-m2.7`. Regression guard: `tests/unit/nvidia-minimax-thinking-strip.test.ts` (+glm-5.2 case). (thanks @phidinhmanh)
+- **fix(translator):** suppress the streamed `</think>` close marker for the **Antigravity IDE** client (port from 9router#1061, [#6415](https://github.com/diegosouzapw/OmniRoute/pull/6415)). On thinking-only turns Antigravity rendered a bare `</think>` as the sole visible content, tripping its loop-detection and wasting requests. Antigravity's UA (`vscode/<v> (Antigravity/<v>)`) is added to the marker-suppress allowlist (alongside OpenCode); Claude Code / Cursor still get the marker, and `x-omniroute-thinking-marker: on` force-restores it. Regression guard: `tests/unit/think-close-marker-suppress-5245.test.ts`. (thanks @abdofallah)
+- **fix(executors):** strip nested `reasoning_content` from messages for **Mistral** (port from 9router#1649, [#6417](https://github.com/diegosouzapw/OmniRoute/pull/6417)). Mistral's API returns `422 extra_forbidden` when an assistant message carries `reasoning_content` (replayed thinking from a prior turn, e.g. via the Codex `/responses` path); the generic top-level 400 field-downgrade retry never covered the nested per-message field. `DefaultExecutor` now strips it for provider `mistral` only, so DeepSeek (which requires replayed `reasoning_content`) is unaffected. Regression guard: `tests/unit/mistral-strip-reasoning-content-1649.test.ts`. (thanks @xxy9468615)
+- **fix(executors):** strip the `client_metadata` passthrough field on the **OpenCode** path (port from 9router#1442, [#6418](https://github.com/diegosouzapw/OmniRoute/pull/6418)). OpenCode upstreams (e.g. `kimi-k2.6` via opencode-go) reject it with `400 "Extra inputs are not permitted, field: 'client_metadata'"`; the DefaultExecutor strip only covered cerebras/mistral and `OpencodeExecutor` extends `BaseExecutor` directly, so nothing removed it there. Regression guard: `tests/unit/opencode-strip-client-metadata-1442.test.ts`. (thanks @yanpaing007)
+- **fix(executors):** inject the `reasoning_content` echo for the native **Moonshot Kimi** provider (port from 9router#1480, [#6419](https://github.com/diegosouzapw/OmniRoute/pull/6419)). Kimi (executor `default`) is a thinking-mode upstream that 400s with "reasoning_content must be passed back" when a prior assistant turn lacks it; the placeholder injection was only wired into the OpenCode meta-provider, so direct multi-turn Kimi conversations failed. Scoped to `kimi` (gateway-served models matching the thinking-model name pattern are unaffected). Regression guard: `tests/unit/kimi-native-reasoning-injected-1480.test.ts`. (thanks @2220258345)
+- **fix(executors):** recover from a strict gateway's `context_management: Extra inputs are not permitted` 400 (port from 9router#1468, [#6420](https://github.com/diegosouzapw/OmniRoute/pull/6420)). **Claude Code** always sends a top-level `context_management` field; strict anthropic-compatible gateways reject it. The dedicated context-editing 400-fallback only fired when OmniRoute's own `contextEditing` feature was enabled (default off), so a client-sent field passed through untouched and 400'd. `context_management` is now in the generic reactive field-strip list, so it's stripped-and-retried once regardless of the feature flag (with correct request re-signing for claude-compatible relays). Regression guard: `tests/unit/provider-field-strips.test.ts`. (thanks @ohahe52-dot)
+- **fix(network):** enable **RFC 8305 Happy Eyeballs** (`autoSelectFamily`) on the direct-egress undici dispatcher (port from 9router#1237, [#6423](https://github.com/diegosouzapw/OmniRoute/pull/6423)). When DNS returns both IPv6 (AAAA) and IPv4 (A) and the IPv6 route is broken (e.g. a NAT64 `64:ff9b::` prefix without routing), undici tried IPv6 first and hung until `ETIMEDOUT` (then a 502 + account lockout), even though `curl` reached the same host. The direct dispatcher now races both families and uses whichever connects first. Proxy paths pin family via `proxyTls` and are unaffected. Regression guard: `tests/unit/direct-dispatcher-pipelining-4580.test.ts`. (thanks @adentdk)
+- **fix(combo):** round-robin now advances the rotation pointer past the model that **actually served**, not the eagerly-scheduled one (port from 9router#948, [#6428](https://github.com/diegosouzapw/OmniRoute/pull/6428)). With `stickyLimit: 1` (true round-robin), when the scheduled model failed and a _different_ model served via fallback, the counter had already advanced +1 from the scheduled index — so the next request reused the fallback-served model, degrading round-robin into hot-spotting on whichever model was healthy. The pointer now advances to the served index + 1 (mirroring the sticky-limit>1 path). Session-stickiness (#3825) and distribution are preserved. Regression guard: `tests/unit/combo-rr-fallback-advance-948.test.ts`. (thanks @binsarjr)
+- **fix(sse):** a non-string `model` field is now rejected with a `400` before the resolver, instead of crashing downstream `.toLowerCase()`/`.split()` calls into an empty-body `500` that escapes the error sanitizer ([#6407](https://github.com/diegosouzapw/OmniRoute/issues/6407)). Regression guard: `tests/unit/chat-non-string-model-6407.test.ts`. (thanks @chirag127)
+- **fix(api):** unknown `/api/*` routes now return a JSON `404` (instead of the dashboard HTML shell) and scalar chat params (`model`/`temperature`/etc.) are validated **before** the provider lookup so malformed requests fail fast with a clear `400` ([#6424](https://github.com/diegosouzapw/OmniRoute/issues/6424), [#6412](https://github.com/diegosouzapw/OmniRoute/issues/6412)). Regression guards: `tests/unit/api/api-catchall-json-404.test.ts`, `tests/unit/chat-early-schema-validation-6412.test.ts`. (thanks @chirag127)
+- **fix(api):** `/v1/chat/completions` now rejects a non-JSON `Content-Type` with a `400` before parsing the body ([#6414](https://github.com/diegosouzapw/OmniRoute/issues/6414)). Regression guard: `tests/unit/v1-chat-completions-content-type-6414.test.ts`. (thanks @chirag127)
+- **fix(api):** the `X-OmniRoute-Compression` response header is now echoed on `/v1/chat/completions` and `/v1/completions` ([#6422](https://github.com/diegosouzapw/OmniRoute/issues/6422)). Regression guard: `tests/unit/compression-header-echo-6422.test.ts`. (thanks @chirag127)
+- **fix(api):** concurrent `GET /v1/models` requests are coalesced into a single catalog build ([#6408](https://github.com/diegosouzapw/OmniRoute/issues/6408)). Regression guard: `tests/unit/v1-models-concurrent-6408.test.ts`. (thanks @chirag127)
+- **fix(api):** `/v1/completions` now echoes the requested `body.model` in its JSON + streamed responses ([#6429](https://github.com/diegosouzapw/OmniRoute/pull/6429)). Regression guard: `tests/unit/completions-body-model-echo.test.ts`. (thanks @chirag127)
+- **fix(api):** env-var master keys now see the full `/v1/models` catalog ([#6406](https://github.com/diegosouzapw/OmniRoute/issues/6406)). Regression guard: `tests/unit/models-catalog-envkey-6406.test.ts`. (thanks @chirag127)
+- **fix(api):** non-streaming `/v1/completions` responses now echo `body.model` aligned with the `X-OmniRoute-Model` header ([#6426](https://github.com/diegosouzapw/OmniRoute/issues/6426)). Regression guard: `tests/unit/v1-completions-model-header-match-6426.test.ts`. (thanks @chirag127)
+- **fix(api):** unknown `/v1/*` routes now return a JSON `404 not_found` instead of the Next.js dashboard HTML shell ([#6405](https://github.com/diegosouzapw/OmniRoute/issues/6405)). Regression guard: `tests/unit/api/v1-catchall-json-404.test.ts`. (thanks @chirag127)
+- **fix(api):** the per-connection provider models route now degrades to the shipped catalog when a provider's `/models` endpoint answers with a redirect ([#6267](https://github.com/diegosouzapw/OmniRoute/issues/6267)) — a `qwen-web` import failed with a raw `Redirect blocked … (307)` 503. `safeOutboundFetch` throws `REDIRECT_BLOCKED` on the 307, `getSafeOutboundFetchErrorStatus` maps it to 503, and `buildDiscoveryErrorFallbackResponse` treated every 503 as a hard error — so the non-empty `getModelsByProviderId("qwen-web")` catalog was never surfaced. A models-endpoint redirect is not a fixable-config error (unlike `URL_GUARD_BLOCKED`/`INVALID_URL`, which stay hard errors), so it now falls back to the local/cached catalog before the 503 short-circuit. General fix — covers any config-driven provider that 307s. Regression guard: `tests/unit/provider-models-qwen-web-redirect-6267.test.ts`. (thanks @chirag127)
+- **fix(api):** the per-connection provider models route (MCP `list_models_catalog` + the dashboard import view) now merges USER-ADDED custom models into its response ([#6247](https://github.com/diegosouzapw/OmniRoute/issues/6247)) — custom models live in the `key_value` namespace `customModels`, which the live REST `/api/v1/models` already merges, but `src/app/api/providers/[id]/models/route.ts` never read `getCustomModels`, so custom models were dropped on both the discovery-success and local_catalog paths. They are now folded into the returned model list (deduped by id, stamped `owned_by: provider`), fixing MCP + the dashboard import view in one place. Regression guard: `tests/unit/provider-models-custom-merge-6247.test.ts`. (thanks @RCrushMe)
+
+- **fix(providers):** GitLab Duo tool-calling follow-up turns no longer fail upstream with `422 {"detail":"Validation error"}` (tokens 0/0, rejected pre-inference). The #6234 tool-result-feedback fix serialized the **entire** multi-turn conversation into GitLab's single-file `code_suggestions` (`small_file`) generation endpoint — folded history that turn-N sent as an oversized `current_file.content_above_cursor` **and** duplicated verbatim into `user_instruction`, tripping the AI-Gateway's `small_file` validation guard. The executor now **bounds** that prompt: it keeps system + latest user message + the most-recent tool round (dropping older turns), caps oversized tool results, and stops duplicating the full prompt into `user_instruction` (which now carries only the short latest user message) — while still feeding the most-recent tool result back so the agent continues (`open-sse/executors/gitlab.ts`, [#6220](https://github.com/diegosouzapw/OmniRoute/issues/6220)). The unit test covers the bounding logic; the upstream 422→200 clearing is VPS-only (Hard Rule #18). Regression guard: `tests/unit/gitlab-tool-exchange-bounded-6220.test.ts`.
+
+- **fix(i18n):** the provider-detail (`/dashboard/providers/[id]`) connection-status filter labels no longer render as `__MISSING__:All` / `__MISSING__:Active` / `__MISSING__:Error` / `__MISSING__:Banned` / `__MISSING__:CreditsExhausted` in non-English locales (notably pt-BR) ([#6290](https://github.com/diegosouzapw/OmniRoute/issues/6290)). Root cause was **not** the namespace mismatch the issue guessed — the `providers.filter*` keys resolve correctly in `en.json`; the debt lived in the locale mirrors (`src/i18n/messages/*.json`), where these five keys carried the `__MISSING__:` sync sentinel in ~15 locales and were absent entirely in ~26 others, so next-intl found the key and echoed the sentinel verbatim. All 40 non-English/-Chinese mirrors now ship real translations for the five `providers.filter*` labels. Regression guard: `tests/unit/i18n-provider-filter-keys-6290.test.ts`. (thanks @diegosouzapw)
+
+- **fix(providers):** the `copilot-m365-web` streaming executor now emits `debug`-level WebSocket diagnostics ([#6210](https://github.com/diegosouzapw/OmniRoute/issues/6210)) — the outbound WS URL (with the `access_token` **redacted** via `redactWsUrl()`), handshake success/failure, and each received SignalR frame's `type`/`target`. Previously the streaming path logged nothing, so an empty `content:null` response (the M365 Education / Starter tier symptom fixed in #6234) was undiagnosable even at `APP_LOG_LEVEL=debug`. The change is debug-level and side-effect-free — it does not alter streaming behavior or the frame parser, and the token never reaches the logs. Regression guard: `tests/unit/copilot-m365-web-logging-6210.test.ts` (thanks @qpeyba)
+
+- **fix(resilience):** a round-robin combo no longer returns `503 all upstream accounts are unavailable` when a compatibility-rejected target is actually healthy ([#6238](https://github.com/diegosouzapw/OmniRoute/issues/6238)). `filterTargetsByRequestCompatibility` drops request-incompatible targets (tool/vision/structured-output unsupported, or below the required context window) **before** any availability check runs, and its `compatible.length === 0` safety net only fired when _all_ targets were filtered — not when the kept targets later all turned out runtime-unavailable (circuit-open / cooldown / no credentials). So a combo could 503 while a compat-rejected-but-healthy provider sat unused. `handleRoundRobinCombo` now keeps the compat-rejected set and, when every compat-kept target was skipped without a single real attempt, probes those rejected targets as a **last-resort fallback tier** (via the new pure `open-sse/services/combo/comboCompatFallback.ts`) before crystallizing the 503. Regression guard: `tests/unit/combo-roundrobin-compat-fallback-6238.test.ts`. (thanks @ThongAccount)
+
+- **fix(startup):** best-effort self-heal for a corrupted Turbopack dev cache on Windows ([#6289](https://github.com/diegosouzapw/OmniRoute/issues/6289)). On Windows, `pnpm dev` can fail at startup when Turbopack `mmap`s a persistent-cache SST file and the OS refuses the mapping (`os error 1455` — "paging file too small"), which Turbopack surfaces as a misleading `Module not found: Can't resolve '@/shared/utils/machine'`. This is a **known upstream Turbopack cache-corruption bug — not our code**. The dev launcher (`scripts/dev/run-next.mjs`) now wraps `nextApp.prepare()` and, when it rejects with that signature (`isTurbopackCacheCorruption` in the new `scripts/dev/turbopackCacheHeal.mjs`), purges `.build/next/**/cache/turbopack` and retries **once** with a clear log. **Caveat — best-effort only:** the corruption often surfaces as a runtime overlay rather than a `prepare()` rejection, so this cannot always intercept it; the reliable remedy remains manually deleting the Turbopack cache dir. Regression guard: `tests/unit/turbopack-cache-heal-6289.test.ts`. (thanks @chirag127)
+
+- **fix(providers):** qodercli PAT auth no longer fails with `spawn qodercli ENOENT` on Windows ([#6263](https://github.com/diegosouzapw/OmniRoute/issues/6263)) — `spawnQoderCli` spawned the bare `qodercli` name with `shell:false` and an unenriched env, so the npm `.cmd` wrapper under `%APPDATA%\npm` (a user-PATH directory) was never resolved. It now resolves the absolute `.cmd`/`.exe` path through the existing `getCliRuntimeStatus("qoder")` resolver in `src/shared/services/cliRuntime.ts` (memoized), spawns with `shell` when the target is a `.cmd`/`.bat`, and uses the cliRuntime-enriched env (PATH + PATHEXT + APPDATA); the ENOENT error now lists the searched paths plus the `CLI_QODER_BIN` override. End-to-end spawn on a real Windows host is host-only (Hard Rule #18); the path-resolution logic is unit-tested. Regression guard: `tests/unit/qodercli-windows-resolve-6263.test.ts`. (thanks @chirag127)
+
+- **fix(sse):** the reasoning-token buffer no longer inflates **probe-sized `max_tokens`** ([#6274](https://github.com/diegosouzapw/OmniRoute/issues/6274)) — Claude Code's `/model` capability check sends `max_tokens: 1`, but for a thinking-capable model with a large output cap (e.g. `glm-5.2`) the #3587 headroom heuristic (`max(current + 1000, ceil(current * 1.5))`) rewrote it to `1001` and forwarded that upstream, wasting tokens on a request that was never a genuine reasoning budget. `resolveReasoningBufferedMaxTokens()` (`open-sse/services/reasoningTokenBuffer.ts`) now short-circuits and returns the caller's value verbatim when it is below the new `REASONING_BUFFER_MIN_TRIGGER` (256) threshold — a tiny explicit limit is a probe, not a reasoning request. Real budgets still receive the #3587 headroom unchanged, and the guard runs after the existing capability checks so unknown / non-reasoning models keep returning `null`. Regression guard: `tests/unit/reasoning-token-buffer-6274.test.ts`. (thanks @brightfiscalband)
+
+- **fix(cli):** `omniroute reset-password` now works as a real subcommand, and password resets over piped (non-TTY) stdin actually apply ([#6261](https://github.com/diegosouzapw/OmniRoute/issues/6261), [#6258](https://github.com/diegosouzapw/OmniRoute/issues/6258)). Two coupled defects: (1) **#6261** — `bin/omniroute.mjs` routed everything through Commander with only two pre-Commander bypasses (`--mcp`, `reset-encrypted-columns`), so `omniroute reset-password` was rejected as an unknown command; only the separate `omniroute-reset-password` bin worked, while the docs falsely advertised the subcommand (incl. a bogus "legacy alias still works"). A pre-Commander bypass mirroring `reset-encrypted-columns` now dynamically imports `bin/reset-password.mjs` (which self-executes) before Commander parses; the three doc lines were corrected. (2) **#6258** — `bin/reset-password.mjs` issued two sequential `rl.question` prompts; under piped stdin the second read never settled at EOF, so `main()` never reached `resetManagementPassword` and the reset was a silent no-op (both prompts printed, no success, password unchanged). The CLI now detects non-TTY stdin and reads it once (first line = password, second line = confirm if present, else reused), adds a `--password-stdin` flag (entire stdin is the password, no confirmation), and exits `0` explicitly so the success line always flushes; interactive TTY behavior is unchanged. Regression guard: `tests/unit/reset-password-cli-6261-6258.test.ts` (3). (thanks @chirag127)
+
+- **fix(db):** the mass-migration **safety abort** now tells the operator how to bypass it and stops flooding the log ([#6260](https://github.com/diegosouzapw/OmniRoute/issues/6260)) — after restoring a backup that wiped the migration tracking table, `runMigrations()` threw the abort on every downstream `ensureDbInitialized()`, re-logging the full banner 11+ times, and the message never mentioned the existing `OMNIROUTE_MAX_PENDING_MIGRATIONS` escape hatch. The abort text now appends a bypass hint (set `OMNIROUTE_MAX_PENDING_MIGRATIONS=0` in `server.env` / `DATA_DIR/.env`), and a new `MigrationSafetyAbortError` is memoized so repeated calls in the same process throw the same instance and emit a single concise line instead of the full cascade. Regression guard: `tests/unit/migration-safety-abort-6260.test.ts`. (thanks @chirag127)
+
+- **fix(auth):** importing a **distinct** Codex/ChatGPT OAuth `auth.json` is no longer falsely rejected as "already exists" when it belongs to a different user in the same workspace ([#6301](https://github.com/diegosouzapw/OmniRoute/issues/6301)). `findExistingCodexConnection` (in `src/lib/oauth/utils/codexAuthImport.ts`) deduped **only** on `providerSpecificData.workspaceId === accountId`, where `accountId` is the shared `chatgpt_account_id`/`tokens.account_id` — so two members of the same ChatGPT Team collapsed onto a single connection (409 `duplicate_account`). The id_token's `https://api.openai.com/auth` claim carries a per-user `chatgpt_user_id` alongside the workspace id (the device-flow path already persisted it as `chatgptUserId`, but the import path did not). Now `parseAndValidateCodexAuth` extracts `userId` (`chatgpt_user_id` → `user_id` → JWT `sub`) into `ParsedCodexAuth`, the create/update paths persist `chatgptUserId` in `providerSpecificData` (mirroring `codex.ts`), and dedup keys on `workspaceId` **AND** `chatgptUserId` — with a backward-compat fallback to legacy accountId-only matching when no stored connection for that workspace records a `chatgptUserId`, so genuinely-same accounts still dedup. Regression guard: `tests/unit/codex-auth-import-userid-dedup-6301.test.ts` (4). (thanks @anungma)
+
+- **fix(providers):** importing models for the **venice-web** provider no longer fails with a red "Provider venice-web does not support models listing" ([#6269](https://github.com/diegosouzapw/OmniRoute/issues/6269)). `venice-web` is a web-cookie provider with an executor but no upstream `/v1/models` endpoint and no registry `models`, so the models route fell through to the tail `400`. Mirroring the `jules`/`linkup-search`/`ollama-search` fix (#5569), it now ships a static local catalog entry in `src/lib/providers/staticModels.ts` — seeding the current Venice lineup (`venice-uncensored`, `llama-3.3-70b`, `qwen3-235b`, `qwen3-4b`, `deepseek-r1-671b`; Venice rotates its catalog, see docs.venice.ai/models/overview) — so the route returns `200` with `source:"local_catalog"`, `intentional:true`. Regression guard: `tests/unit/static-models-venice-web-6269.test.ts`. (thanks @chirag127)
+
+- **fix(api):** the specialty model catalogs (`/v1/embeddings`, `/v1/images`, `/v1/music`, `/v1/videos` model lists) are now derived from the **unified catalog filtered by a predicate** (`getSpecialtyModelsResponse`) instead of ad-hoc per-route logic, so they consistently respect active-credential visibility and stay in sync with the main catalog ([#6303](https://github.com/diegosouzapw/OmniRoute/pull/6303)). Regression guard: `tests/unit/specialty-model-catalog-routes.test.ts`. (thanks @makcimbx)
+- **fix(api):** the agent-bridge server route now resolves the MITM manager via a **dynamic `import("@/mitm/manager.runtime")`** so Turbopack does not statically pull the stub (or over-bundle the manager), and the agent-skills generator anchors its output base path with `path.join(process.cwd(), …)` so Turbopack's static analyzer stops tracing the whole project root ([#6329](https://github.com/diegosouzapw/OmniRoute/issues/6329), [#6366](https://github.com/diegosouzapw/OmniRoute/pull/6366)). Regression guard: `tests/unit/agent-bridge-server-route-dynamic-import.test.ts`. (thanks @Iammilansoni)
+- **fix(api):** internal probes (combo-test, cloud-sync verify) now pick a **management-scoped / allow-all API key** instead of naively grabbing `getApiKeys()[0]` — a restricted `self:usage` first row made the probe fail with "Model X is not allowed for this API key" even when the combo path was healthy (`pickApiKeyForInternalUse` in `src/lib/db/apiKeys.ts`). The API-manager model editor also falls back to `/api/models?all=true` when `/v1/models` is catalog-protected ([#6372](https://github.com/diegosouzapw/OmniRoute/pull/6372)). Regression guard: `tests/unit/pick-internal-api-key-6372.test.ts`. (thanks @jmengit)
+- **fix(live-ws):** the Live Dashboard WebSocket server now **rejects on bind failure** (e.g. `EADDRINUSE` when the API bridge already holds the port) instead of letting the error surface as an unhandled `error` event that crash-loops the process — the `error` listener is attached to `wss` (not `server`) and releases the EventBus subscription on a failed start ([#6324](https://github.com/diegosouzapw/OmniRoute/issues/6324)). Regression guard: `tests/unit/live-ws-eaddrinuse-6324.test.ts`. (thanks @vinayakkulkarni)
+- **fix(dashboard):** the Home provider-topology widget now trusts the live provider-metrics snapshot — it uses `topology.errorProvider` and live `activeRequests` directly instead of re-deriving state from a stale `lastErrorAt` or applying a frontend timeout filter, so the topology reflects real-time provider health ([#6322](https://github.com/diegosouzapw/OmniRoute/pull/6322)). Regression guard: `tests/unit/home-provider-topology-live-state.test.ts`. (thanks @xz-dev)
+- **fix(sse):** strip zero-width markers from streamed **tool-call arguments** — a follow-up to [#5857](https://github.com/diegosouzapw/OmniRoute/pull/5857). That PR removed injected zero-width joiners (U+200D) from streamed assistant text/reasoning but deliberately left tool-call argument JSON byte-exact. The request-side obfuscation (`open-sse/services/claudeCodeObfuscation.ts`) injects ZWJ into agent words — including the temp path inside the Bash tool description — and Claude models copy that verbatim into generated commands, which are delivered as tool-call arguments rather than assistant text. As a result the ZWJ survived and corrupted code blocks (e.g. a temp path rendered with an invisible joiner). Now `open-sse/handlers/responseSanitizer.ts` strips zero-width code points from tool-call argument strings at every emit site (OpenAI non-stream/stream chat `tool_calls` + legacy `function_call`, native Responses `function_call` items, the OpenAI→Responses conversion, and the native Responses streaming `response.function_call_arguments.delta/.done` events). Only zero-width code points are removed; JSON structure and all other bytes stay identical (no parse/restringify), so normal arguments remain byte-exact. Regression guard: 6 new cases in `tests/unit/response-sanitizer.test.ts` (suite 50/50).
+
+- **fix(nodejs):** the default app log path now resolves under `DATA_DIR` (`~/.omniroute/logs/application/app.log`) instead of `process.cwd()` ([#6197](https://github.com/diegosouzapw/OmniRoute/issues/6197)) — the globally-installed CLI runs from an arbitrary working directory, so anchoring the default to cwd made file logging silently write to (or no-op under) an unrelated directory, contradicting the documented `.env.example` default. `getAppLogFilePath()` now computes the default lazily via the pure `resolveDataDir()` resolver (honours a per-process `DATA_DIR`, no directory-creation side effect); an explicit `APP_LOG_FILE_PATH` still wins. Regression guard: `tests/unit/logenv-datadir-path-6197.test.ts` (3).
+
+- **fix(docker):** AgentBridge/`startMitm` no longer aborts in containers/headless when the Antigravity-default DNS step can't write `/etc/hosts` ([#6127](https://github.com/diegosouzapw/OmniRoute/issues/6127)), and the privileged command's stderr now reaches `app.log` instead of only a bare exit code hitting the toast ([#6198](https://github.com/diegosouzapw/OmniRoute/issues/6198)). The default DNS step (`addDNSEntry`) was called unguarded while cert install and the two sibling DNS steps were each best-effort — in the runtime Docker image (`USER node`, no `sudo`, read-only `/etc/hosts`) it threw `Command failed with code 1` out of `startMitmInternal` and killed the whole start, discarding the stderr. The three DNS steps are extracted into a best-effort `provisionDnsEntries()` where each failure is logged with the full `err` (stderr included, folded in by `systemCommands.ts`) and never aborts the start. Regression guard: `tests/unit/mitm-dns-graceful-degrade-6127.test.ts` (4).
+
+- **fix(providers):** copilot-m365-web now supports the M365 Education "Starter / OfficeWebIncludedCopilot" tier and no longer returns an empty `content:null` stream ([#6210](https://github.com/diegosouzapw/OmniRoute/issues/6210)). Two gaps: (1) `buildWsUrl()` hardcoded the individual-consumer scenario (`OfficeWebPaidConsumerCopilot`, `isEdu=false`) — the EDU tier is now opt-in via `providerSpecificData.tier="edu"`, emitting `scenario=OfficeWebIncludedCopilot`/`isEdu=true` (the individual path is unchanged); (2) the EDU/GPT-5.5 path streams deltas via `arguments[0].writeAtCursor` (incremental) instead of only `messages[].text` (accumulated snapshots), which the parser dropped — a new `accumulateBotContent()` folds both formats, with `type:2 item.result.message` as a last-resort fallback. Regression guard: `tests/unit/copilot-m365-edu-writeatcursor-6210.test.ts` (10). (thanks @qpeyba)
+
+- **fix(providers):** GitLab Duo executor now feeds tool results back into the prompt instead of looping ([#6220](https://github.com/diegosouzapw/OmniRoute/issues/6220)) — `buildPrompt()` branched only on `system`/`user` and took `userParts.at(-1)`, silently dropping the `assistant{tool_calls}` + `tool{result}` turns the client appended, so the reconstructed prompt was byte-identical to turn 1 and the model re-emitted the same `<tool>` call forever. When a tool exchange is present the full conversation is now serialized, folding each tool result back keyed by its `tool_call_id`; simple conversations keep the legacy shape. Complements the tool_call emission from [#6051](https://github.com/diegosouzapw/OmniRoute/issues/6051) (the `kilo-duplicate` label was a false positive — different, sequential defect). Regression guard: `tests/unit/gitlab-tool-result-feedback-6220.test.ts` (4).
+
+- **fix(providers):** opencode-go/opencode-zen can now synthesize the OpenCode CLI identity headers Cloudflare requires on VPS egress ([#5997](https://github.com/diegosouzapw/OmniRoute/issues/5997)) — on a datacenter VPS, `opencode.ai/zen/go/v1/chat/completions` 403s (HTML challenge) requests lacking CLI identity, while the reporter's control curl proved that `User-Agent: opencode-cli/1.0.0` + `x-opencode-client: cli` + `x-opencode-project: default` + fresh request/session UUIDs succeed. Opt-in via `OPENCODE_SYNTHESIZE_CLI_HEADERS=true` (values overridable via `OPENCODE_GO_USER_AGENT`/`OPENCODE_USER_AGENT`/`OPENCODE_CLIENT`/`OPENCODE_PROJECT`); it fills only headers the client did not already send. Kept **off by default** — the forward-only path is deliberate (fabricating a wrong value risks upstream rejection; a prior dedup regressed with `opencode/local`), so this replaces the fragile local header-injection shim without changing default behavior. Regression guard: `tests/unit/opencode-cli-headers-synthesis-5997.test.ts` (6). (thanks @aleksesipenko)
+- fix(resilience): sticky session affinity now evicts and fails over to another account when the pinned account is exhausted/unavailable (#6219)
+
+- fix(sse): Responses API passthrough now drops internal commentary-phase output before forwarding to clients (gated by RESPONSES_PASSTHROUGH_DROP_COMMENTARY, default on) (#6199)
+
+- **fix(sse):** tool-call function schemas with a root `type: null` are now coerced to `type: "object"` before dispatch ([#6359](https://github.com/diegosouzapw/OmniRoute/issues/6359)) — clients like the Codex app emit `parameters: { type: null, ... }` for some tools, which OpenAI-compatible upstreams reject with `400 Invalid schema for function '...': schema must be a JSON Schema of 'type: "object"', got 'type: null'`, failing the whole request. `toolSchemaSanitizer` already stripped the null; it now re-adds the mandatory root `"object"` type (and empty `properties`/open `additionalProperties` when absent). Combinator roots (`anyOf`/`oneOf`/`allOf`) and explicit root types are left untouched. Regression guard: 5 new cases in `tests/unit/tool-schema-sanitizer.test.mjs`.
+
+- **fix(docker):** AgentBridge no longer fails to start on npm/Electron/VPS installs with "MITM manager stub reached at runtime" ([#6344](https://github.com/diegosouzapw/OmniRoute/issues/6344)) — v3.8.45 flipped the production bundler default to Turbopack, but `next.config.mjs` aliased `@/mitm/manager` to its Docker-only degraded stub **unconditionally**. That was harmless while Docker (which sets the alias intentionally for #3390 graceful degradation) was the sole Turbopack consumer, but once every artifact built with Turbopack the stub shipped to all non-Docker users and `startMitm` threw on the first Agent-Bridge start. The alias is now opt-in via `OMNIROUTE_MITM_STUB=1` (set only by the Dockerfile) through the shared `scripts/build/mitm-stub-flag.mjs` helper; default builds bundle the real manager. Regression guard: `tests/unit/mitm-stub-alias-6344.test.mjs` (4).
+
+- **fix(proxy):** stop the v3.8.44 proxy regression that leaked the real IP and disabled healthy proxies ([#6246](https://github.com/diegosouzapw/OmniRoute/issues/6246)). Two coupled defects from the new health scheduler: (1) **IP leak** — when a proxy assigned to a connection was marked `inactive`, resolution fell through to a **direct** egress instead of blocking, exposing the operator's real IP; (2) **over-deactivation** — the sweep flipped a proxy to `inactive` on the **first** failed probe and counted our own 5s timeout / a probe-target `5xx` as the proxy's fault, so healthy paid proxies vanished from egress selection ("my proxies are not being used anymore"). Fix: the sweep decision is extracted into a pure, network-free `decideProxyHealthAction` (`src/lib/proxyHealth/decision.ts`) — by default the health check now **only counts/logs and never downgrades status** (a proxy is downgraded/removed only with `PROXY_AUTO_REMOVE=true`, after `PROXY_AUTO_REMOVE_AFTER` **consecutive** conclusive failures); probes are classified tri-state so an inconclusive result (our timeout, or a `5xx` from the probe target) never penalizes the proxy, and the probe timeout is raised 5s→15s. Separately, `safeResolveProxy` now **fails closed** via the existing policy: a connection whose assigned proxy is dead is blocked instead of leaking direct (`hasBlockingProxyAssignment`), honoring the explicit `proxy off` toggles and the `PROXY_FAIL_OPEN=true` opt-out. Existing proxies stuck `inactive` by the old behavior need a one-time manual re-activate (the operator owns proxy status). Regression guards: `tests/unit/proxy-health-decide-action-6246.test.ts`, `tests/unit/proxy-assigned-unavailable-6246.test.ts`.
+
+- **fix(proxy):** make "Test All" read-only and add bulk enable/disable ([#6246](https://github.com/diegosouzapw/OmniRoute/issues/6246)). Complements the core fail-closed / scheduler fix (#6296) with the two remaining reporter asks. (1) The **"Test All" button** (`POST /api/settings/proxies/auto-test`) used to flip a proxy to `inactive` on a failed reachability probe; since the egress selector excludes `inactive` proxies, a flaky probe (an unreachable `httpbin.org`, a proxy that blocks `HEAD`, or a slow paid proxy) silently disabled every proxy that failed — "Test All" is now **read-only by default** (only the operator sets a proxy active/inactive; opt back into the legacy test-and-set with `PROXY_HEALTH_AUTO_DEACTIVATE=true`). (2) Adds a **bulk enable/disable** proxies endpoint + toolbar action (`POST /api/settings/proxies/batch-activate`) so an operator can re-activate proxies in one click. Regression guard: `tests/unit/proxy-health-6246.test.ts`. (thanks @tenshiak)
+
+- **chatcore (tools): stop the default 128-tool cap from silently dropping opencode's `task`/MCP tools.** opencode (used as an MCP/agent host) sends a large tool list; when it exceeds the speculative `MAX_TOOLS_LIMIT` (128) default, `truncateToolList` did a blind `tools.slice(0, 128)`, dropping every tool past index 128 — including opencode's built-in `task` tool (subagent launch) and many MCP tools, so models routed through OmniRoute could no longer spawn subagents or reach part of their tools. The cap exists to avoid upstream `400`s for providers with real hard limits (e.g. grok-cli 200), so it is kept for those: detection of the opencode client (`isOpencodeClient` — any `x-opencode-*` header, or `opencode` in the user-agent) now only bypasses the **speculative 128 default**, never a known provider ceiling. Precedence is explicit — a proactive/detected provider limit always truncates (even for opencode); otherwise opencode forwards its full tool list; otherwise the unchanged 128 default applies to every other client. Refactors `getEffectiveToolLimit` into `getKnownToolLimit(provider) ?? DEFAULT_LIMIT` (byte-identical for existing callers) and fixes a cosmetic debug-log that reported the truncated count instead of the original. Regression guard: `tests/unit/tool-limit-detector.test.ts`.
+
+- **fix(mitm):** the macOS MITM-cert install check now matches the system keychain again. `security find-certificate -a -Z` prints the SHA-1 as a colon-less hex string, but the installed-check compared it against `getCertFingerprint()`'s colon-separated form, so the substring match never hit — the cert was reported as not-installed and re-prompted for the sudo install on every run. Fingerprints are now normalized (colons stripped, upper-cased) on both sides via the extracted `macCertOutputHasFingerprint` helper. Regression guard: `tests/unit/mitm-cert-mac-fingerprint.test.ts`. ([#6204](https://github.com/diegosouzapw/OmniRoute/pull/6204), closes [#6134](https://github.com/diegosouzapw/OmniRoute/issues/6134) — thanks @rianonehub)
+
+- **fix(api):** `/v1/messages/count_tokens` now counts `tool_use`, `tool_result` and `thinking` content blocks (and array-form `system` prompts) in the local-estimation path, instead of only `text`. Real agentic conversations keep ~95% of their tokens inside tool results; the previous estimate returned near-zero for them, which silently broke Claude Code's auto-compaction (context grew past the window with no compaction until the upstream API rejected the request). The real provider-side count path is unchanged. Regression guard: `tests/unit/messages-count-tokens-route.test.ts`. ([#6221](https://github.com/diegosouzapw/OmniRoute/pull/6221) — thanks @luweiCN)
+
+- **fix(antigravity):** strip a trailing assistant prefill turn for Vertex Claude models to avoid upstream 400s ([#6114](https://github.com/diegosouzapw/OmniRoute/pull/6114)). Regression guard: `tests/unit/antigravity-claude-prefill-strip.test.ts`. (thanks @anki1kr)
+
+- **fix(security):** the mutable cloud-agent routes (`/api/cloud/credentials/update`, `/api/cloud/models/alias`) now require management auth instead of being treated as public. They were classified as public API routes, so a request without management credentials could update stored cloud-agent credentials and model aliases. They are removed from the public-route set, classified as management routes in the authz pipeline, and gated by `requireManagementAuth`; cloud **read**/auth routes stay public. Regression guards: `tests/unit/cloud-write-auth.test.ts`, `tests/unit/authz/classify.test.ts`, `tests/unit/public-api-routes.test.ts`. ([#6233](https://github.com/diegosouzapw/OmniRoute/pull/6233) — thanks @vittoroliveira-dev)
+
+- **refactor(dashboard):** extract the onboarding-wizard "Open provider details" link target into a pure, unit-tested `buildProviderDetailsHref(connection)` helper. The wizard already routes by `connection.id` (the node UUID) rather than the provider category slug (#6144/#6145); this hardens that behavior behind a tested helper that guards a missing id/connection. Regression guard: `tests/unit/provider-onboarding-href.test.ts`. ([#6166](https://github.com/diegosouzapw/OmniRoute/pull/6166) — thanks @KooshaPari)
+
+- **fix(security):** the doubao synthetic device-id generator now derives its digits via an unbiased crypto-random draw (rejection sampling over `crypto.randomBytes()`) instead of a `% 10` reduction, closing a CodeQL `js/biased-cryptographic-random` finding.
+- **fix(agentSkills):** the GitHub-skills generator now resolves `outputDir` to an absolute path before writing, fixing a regression introduced by #6366 (relative-to-cwd base path) that could write generated skill files to the wrong directory.
+- **fix(security):** `/api/keys/{id}/devices` now checks the HTTP method before auth/validation, returning a `405` for non-GET/DELETE verbs instead of a misleading `401`/`500` (closes a `dast-smoke` QUERY-method finding).
+- **fix(quality):** clear the last 2 heavy quality-gate reds on the release tip (cycle pre-flight).
+- **fix(mitm):** the test suite and CI can never mutate the OS trust store — `OMNIROUTE_SKIP_SYSTEM_TRUST=1` is set globally for tests/CI so `installCert`/`uninstallCert`/`installTproxyCa` skip the privileged OS dispatch ([#6310](https://github.com/diegosouzapw/OmniRoute/pull/6310); full detail is under the [3.8.45] section below — this branch received it via the parallel-cycle sync-back).
+- **fix(api):** `POST /api/github-skills` now Zod-validates its request body; documented the new quality-gate env vars and pinned the merge-integrity GitHub Actions to a commit SHA.
+- **fix(skills):** generate the missing `omni-github-skills` registry entry and align the agent-skills catalog-count tests (follow-up to #6186).
+- **fix(quality):** clear the cycle's 11 net-new ESLint errors and make `validate-release-green` suppressions-aware.
+
+### 📝 Maintenance
+
+- **i18n(it):** add 118 missing Italian (`it`) translations (net-additive — no existing keys dropped, valid JSON), improving Italian UI coverage. ([#6212](https://github.com/diegosouzapw/OmniRoute/pull/6212) — thanks @serverless83)
+- **chore(providers):** remove deprecated **MiMo V2** model entries from the catalogs (xiaomi-mimo, opencode-go, zenmux-free, audio TTS) — the upstream V2 line is superseded by MiMo V2.5; drops `mimo-v2-tts`, `mimo-v2-pro`, `mimo-v2-omni`, `mimo-v2-flash`, `mimo-v2-flash-free` and realigns the provider-catalog tests. ([#6248](https://github.com/diegosouzapw/OmniRoute/pull/6248) — thanks @backryun)
+- **chore(release):** ~50 commits on this branch are v3.8.45 pre-flight/hardening fixes and CI-perf work that landed here via the parallel-cycle sync-back (`sync-next-cycle.mjs`, Hard Rule #21) **after** the `v3.8.45` git tag was cut, and are already fully documented under the **[3.8.45]** section below — listed here only so the per-cycle commit-coverage check (`npm run release:uncovered`) doesn't flag them as gaps. Provider/catalog/UX/backend fixes: #6041, #6078, #6108, #6135, #6148, #6149, #6154, #6158, #6161, #6162, #6163, #6164, #6165, #6170, #6177, #6178, #6181, #6186, #6187, #6191, #6193, #6194, #6195, #6200, #6205, #6208, #6209, #6211, #6213, #6223, #6224, #6225, #6226, #6227, #6228, #6229, #6230, #6235, #6291, #6292. CI/release-pipeline work: #6167, #6203, #6214, #6215, #6218, #6273, #6275, #6283, #6284, #6285, #6300, #6305.
+- **chore(release):** additional zero-ref release-cycle plumbing on this branch, kept out of `release:uncovered` on purpose (no `#N` in the commit subject to cite): opening the v3.8.46 cycle, opening/closing the v3.8.45 cycle, the finalized [3.8.45] CHANGELOG i18n sync-back to 42 mirrors, the v3.8.45 cognitive/cyclomatic and file-size drift rebaselines, ESLint stale-suppression pruning (4,273 → 4,233), and clearing test-masking/docs-all pre-flight reds for v3.8.45.
+
+### ⚡ Performance & Infrastructure
+
+- **perf(release-green):** the pre-flight validator (`scripts/quality/validate-release-green.mjs`) now runs its 4 slow suites (unit / vitest / integration / pack-artifact) **concurrently** via `Promise.all` — pre-flight wall time drops from ~the sum of the suites to ~the slowest one (~30min saved per round; Phase 0 was the nº1 bottleneck of the v3.8.45 release benchmark, 2h54 of 6h34 e2e). Guard: `tests/unit/validate-release-green.test.ts` ("runs the slow suites CONCURRENTLY"). ([#6319](https://github.com/diegosouzapw/OmniRoute/pull/6319))
+- **fix(ci):** `scripts/release/sync-next-cycle.mjs` — two defects found live in its first production run (v3.8.45 Phase 5): (1) the `git()` helper's default 1 MiB `maxBuffer` crashed with `ENOBUFS` on `git show origin/main:CHANGELOG.md` (the CHANGELOG alone is >1 MiB) — widened to 64 MiB; (2) the i18n resync only propagated the `[NEXT]` (TBD) section, leaving the just-shipped finalized section as "— TBD" in all 42 mirrors — it now also syncs `[prevVersion]` bounded by the heading below it (new exported pure helper `versionAfter`). Guards: +5 tests in `tests/unit/sync-next-cycle.test.ts` (8/8). ([#6327](https://github.com/diegosouzapw/OmniRoute/pull/6327))
+- **test(ci):** concurrency-sensitive flaky tests are **quarantined into a serial pass** (`tests/unit/serial/`, `--test-concurrency=1`, appended to every unit runner incl. sharded variants — the serial pass is sharded too so concurrent shard jobs never self-collide). Initial set: `glm-coding-plan-monthly-3580`, `quota-division-blocks`, `provider-health-autopilot`, `combo-health-autopilot` — the class behind the ~28min CI wedges/re-runs (two live 1h+ wedges cancelled during this PR's own validation). Discovery + TIA gates track the new glob; systemic root cause (async logger writing after teardown) tracked in [#6360](https://github.com/diegosouzapw/OmniRoute/issues/6360). Guard: `tests/unit/test-serial-quarantine.test.ts` (4). ([#6347](https://github.com/diegosouzapw/OmniRoute/pull/6347))
+
+### 🙌 Contributors
+
+Thanks to everyone whose work landed in v3.8.46:
+
+| Contributor                                                    | PRs / Issues           |
+| -------------------------------------------------------------- | ---------------------- |
+| [@2220258345](https://github.com/2220258345)                   | direct commit / report |
+| [@abdofallah](https://github.com/abdofallah)                   | direct commit / report |
+| [@adentdk](https://github.com/adentdk)                         | direct commit / report |
+| [@aleksesipenko](https://github.com/aleksesipenko)             | direct commit / report |
+| [@anki1kr](https://github.com/anki1kr)                         | direct commit / report |
+| [@anungma](https://github.com/anungma)                         | direct commit / report |
+| [@arpicato](https://github.com/arpicato)                       | direct commit / report |
+| [@backryun](https://github.com/backryun)                       | #6248                  |
+| [@binsarjr](https://github.com/binsarjr)                       | direct commit / report |
+| [@brightfiscalband](https://github.com/brightfiscalband)       | direct commit / report |
+| [@chirag127](https://github.com/chirag127)                     | #6501, #6506           |
+| [@developerjillur](https://github.com/developerjillur)         | direct commit / report |
+| [@dilneiss](https://github.com/dilneiss)                       | #6499                  |
+| [@dtybnrj](https://github.com/dtybnrj)                         | direct commit / report |
+| [@Forcerecon](https://github.com/Forcerecon)                   | direct commit / report |
+| [@hao3039032](https://github.com/hao3039032)                   | direct commit / report |
+| [@Iammilansoni](https://github.com/Iammilansoni)               | #6150, #6245           |
+| [@jmengit](https://github.com/jmengit)                         | direct commit / report |
+| [@jordansilly77-stack](https://github.com/jordansilly77-stack) | direct commit / report |
+| [@JxnLexn](https://github.com/JxnLexn)                         | direct commit / report |
+| [@KooshaPari](https://github.com/KooshaPari)                   | #6166                  |
+| [@loopyd](https://github.com/loopyd)                           | direct commit / report |
+| [@luweiCN](https://github.com/luweiCN)                         | #6221                  |
+| [@makcimbx](https://github.com/makcimbx)                       | direct commit / report |
+| [@muflifadla38](https://github.com/muflifadla38)               | direct commit / report |
+| [@newnol](https://github.com/newnol)                           | direct commit / report |
+| [@ofekbetzalel](https://github.com/ofekbetzalel)               | direct commit / report |
+| [@ohahe52-dot](https://github.com/ohahe52-dot)                 | direct commit / report |
+| [@phidinhmanh](https://github.com/phidinhmanh)                 | direct commit / report |
+| [@powellnorma](https://github.com/powellnorma)                 | direct commit / report |
+| [@qpeyba](https://github.com/qpeyba)                           | direct commit / report |
+| [@RaviTharuma](https://github.com/RaviTharuma)                 | direct commit / report |
+| [@RCrushMe](https://github.com/RCrushMe)                       | direct commit / report |
+| [@rianonehub](https://github.com/rianonehub)                   | #6134, #6204           |
+| [@serverless83](https://github.com/serverless83)               | #6212                  |
+| [@swingtempo](https://github.com/swingtempo)                   | direct commit / report |
+| [@tenshiak](https://github.com/tenshiak)                       | direct commit / report |
+| [@ThongAccount](https://github.com/ThongAccount)               | direct commit / report |
+| [@UnrealAryan](https://github.com/UnrealAryan)                 | direct commit / report |
+| [@vinayakkulkarni](https://github.com/vinayakkulkarni)         | direct commit / report |
+| [@vittoroliveira-dev](https://github.com/vittoroliveira-dev)   | #6233                  |
+| [@warelik](https://github.com/warelik)                         | direct commit / report |
+| [@xxy9468615](https://github.com/xxy9468615)                   | direct commit / report |
+| [@xz-dev](https://github.com/xz-dev)                           | direct commit / report |
+| [@yanpaing007](https://github.com/yanpaing007)                 | direct commit / report |
+| [@diegosouzapw](https://github.com/diegosouzapw)               | maintainer             |
+
+---
+
 ## [3.8.45] — 2026-07-06
 
 ### ✨ New Features

docs/i18n/az/CHANGELOG.md

@@ -6,6 +6,213 @@
 
 ## [3.8.31] — 2026-06-20
 
+## [3.8.46] — TBD
+
+### ✨ New Features
+
+- **feat(sse):** **hide paid-only models from `auto/*` routing** when `hidePaidModels` is on ([#6512](https://github.com/diegosouzapw/OmniRoute/issues/6512)) — follow-up to #6328/#6495. PR #6495 hid paid-only models from the `GET /v1/models` listing, but `auto/*` combos (`auto/best-coding`, `auto/glm`, …) could still pick a paid-only backend into their candidate pool → a 402/403 at request time. `createVirtualAutoCombo` now filters the candidate pool through the new pure `open-sse/services/autoCombo/paidModelFilter.ts` (`filterPaidOnlyCandidates`), applying the same free-model predicate #6495 uses in `catalog.ts` (`providerHasFreeModels(provider) && isFreeModel(provider, {id})`) whenever `settings.hidePaidModels === true`. Applied before the category/tier/family narrowing, so it covers every `auto/*` combo; an all-paid pool degrades to the existing graceful empty-pool path. **Opt-in — default OFF leaves the pool unchanged** (identity). Regression guard: `tests/unit/autoCombo/paid-model-filter-6512.test.ts` (4, incl. the default-off identity guard).
+- **feat(sse):** **provider-family auto combos** — `auto/glm`, `auto/minimax`, `auto/mimo`, `auto/zai`, `auto/gemma`, `auto/llama`, `auto/gemini` ([#6453](https://github.com/diegosouzapw/OmniRoute/issues/6453)) — new routable ids that materialize an on-demand virtual combo spanning whatever installed backends currently expose that model family, degrading gracefully as backends rotate. A new pure `open-sse/services/autoCombo/modelFamily.ts` (`detectModelFamily`) classifies by model-id prefix for six families; `zai` is instead resolved by provider id (z.ai's hosted API serves the same `glm-*` model ids as every other GLM backend, so `auto/zai` means "route to my z.ai backend specifically" vs `auto/glm`'s "any connected GLM backend"). Reuses the existing `createVirtualAutoCombo` on-demand materialization path (no DB writes) and the `/v1/models` catalog advertising loop. Regression guard: `tests/unit/autoCombo/provider-family-combos.test.ts` (11).
+- **feat(proxy):** native **proxy-pool round-robin / egress IP rotation** ([#6365](https://github.com/diegosouzapw/OmniRoute/issues/6365)) — a scope (global / provider / account) can now hold **multiple** proxies as a pool with a rotation strategy, so outbound requests cycle their egress IP instead of pinning one proxy per scope. Migration `117_proxy_pool_rotation.sql` lifts the `UNIQUE(scope, scope_id)` constraint (rebuild via the canonical rename/copy/drop; existing single assignments become 1-element pools) and adds a `proxy_scope_rotation` companion table holding the per-scope strategy + a persisted monotonic round-robin cursor. Strategies: `round-robin` (default, monotonic cursor — never `Math.random`), `random`, and `sticky-per-N-min`. Resolution (`resolveProxyForScopeFromRegistry` / `resolveProxyForConnectionFromRegistry`) now fetches the alive, position-ordered candidate set (unchanged `PROXY_ALIVE_PREDICATE`) and applies the strategy; an empty / all-dead pool still returns `null` — the #6246 fail-closed guard is **untouched** (never falls through to direct egress). Backend + DB only; dashboard pool-builder UI is a follow-up. Regression guard: `tests/unit/proxy-pool-rotation-6365.test.ts` (8, incl. fail-closed + backward-compat).
+- **feat(providers):** end-to-end **tool/function calling on the native Gemini `/v1beta` endpoint** ([#6222](https://github.com/diegosouzapw/OmniRoute/issues/6222)) — both directions of the Gemini↔OpenAI conversion now preserve tool data (previously silently dropped). Request side: `convertGeminiToInternal` (extracted to its own testable module) maps `tools[].functionDeclarations` → OpenAI `tools`, prior `functionCall` parts → assistant `tool_calls`, and `functionResponse` parts → `tool`-role messages. Response side: `convertOpenAIResponseToGemini` emits `parts[].functionCall {name,args}` from `message.tool_calls`, and the streaming `openAIChunkToGeminiChunk` accumulates fragmented `tool_calls` deltas by index into complete `functionCall` parts. The non-Gemini client paths (Claude, OpenAI-Responses) already preserved tool calls — this closes the gap specific to the native Gemini surface. Regression guard: `tests/unit/v1beta-gemini-tool-calling-6222.test.ts` (6, incl. a streaming SSE round-trip).
+- **feat(providers):** copilot-m365-web **enterprise / work tier** support ([#6334](https://github.com/diegosouzapw/OmniRoute/issues/6334)) — mirrors the EDU-tier pattern (#6210): `M365ConnectionParams` gains an `agent` field, a new opt-in `M365_ENTERPRISE_OVERRIDES` preset (`agent=work`, `scenario=officeweb`, `licenseType=Premium`) applies via `providerSpecificData.tier="enterprise"` (alias `"work"`), and `agent` is also overridable directly via `providerSpecificData.agent`. `buildWsUrl` was hardcoding `agent="web"` (the one enterprise-distinguishing param with no override path), so a Premium work account handshook then returned an empty stream. The individual and EDU paths are untouched. Kilo's dup flag vs #6210 (EDU tier) was a false positive — different tier. Regression guard: `tests/unit/copilot-m365-enterprise-6334.test.ts` (7). End-to-end confirmation on a real Premium work account is a live-VPS validation follow-up (Hard Rule #18). (thanks @Forcerecon)
+- **feat(api):** standardized, provider-agnostic **`effort` + `thinking` request params** ([#6241](https://github.com/diegosouzapw/OmniRoute/issues/6241)) — a thin standardization layer over the existing mature per-provider reasoning plumbing (no provider mapper touched). `providerChatCompletionSchema` gains a canonical `effort` (reusing the shared `none/low/medium/high/xhigh` vocabulary — the UI tiers `extra`/`max` collapse onto `xhigh`) and a boolean `thinking`. A pure `normalizeReasoningRequest` (wired once in `src/sse/handlers/chat.ts`, before any reasoning field is read) folds them onto the fields the translators already consume (`reasoning_effort` / `reasoning.effort` / `thinking`), so they fan out to Anthropic / Gemini / xAI / Responses — an explicit client `reasoning_effort` / object-shaped `thinking` always wins (backward-compatible). `/models` additively exposes `supportsThinking` + `effort_tiers` so the frontend can render the toggles (UI component is a follow-up). Regression guard: `tests/unit/effort-thinking-standardization-6241.test.ts` (12). (thanks @Iammilansoni, @shabeer)
+- **feat(combo):** new **`pipeline` (sequential) combo strategy** ([#6297](https://github.com/diegosouzapw/OmniRoute/issues/6297)) — the 18th routing strategy runs targets **in order**, threading each step's output into the next step's input, with an optional per-step `prompt` (system instruction); only the final step's response is returned. Distinct from `fusion` (parallel fan-out + judge). Implemented as a self-contained `open-sse/services/pipeline.ts` (sibling to `fusion.ts`), dispatched from `combo.ts`; the step list reuses `combo.models` order and reads an optional `prompt` off each target (backward-compatible — ignored by every other strategy). Intermediate steps run non-streaming with tools stripped (complete prose to thread forward); the final step keeps the client's `stream` flag + tools. A failing/empty/unparseable intermediate step fails the whole pipeline explicitly via a sanitized error (never silently swallowed). Kilo's dup flag vs #563 was a false positive (that's model→chain selection; this is a sequential chain). Regression guard: `tests/unit/combo-pipeline-strategy.test.ts` (5). (thanks @ofekbetzalel)
+- **feat(ci):** `check:test-masking` now flags **inline-reimplemented prod conditions** ([#6348](https://github.com/diegosouzapw/OmniRoute/issues/6348)) — a new **report-only** subcheck (v2, 6A.10 family) catches the wrong-shape contract test: a test that recomputes the condition under test inline instead of importing/exercising the real function (the #6216 class, where `=== 500` → `>= 500` stayed green because the test re-implemented the branch). For each added/modified test file it warns when the file textually duplicates a ≥3-token conditional from a production file touched in the same PR **and** does not import the symbol/module owning it, via a pure, fixture-tested `findReimplementedConditions()` with an allowlist mirroring `assertReductionAllowlist`. Report-only for now (does not fail the gate) — to be promoted to blocking after a triage cycle. Regression guard: `tests/unit/check-test-masking.test.ts` (45).
+- **feat(sse):** per-connection routing override (native vs CLIProxyAPI) ([#6339](https://github.com/diegosouzapw/OmniRoute/issues/6339)) — the previously-dead `isCliproxyapiDeepModeEnabled` helper is now wired into `resolveExecutorWithProxy`: a single connection can opt itself into the CLIProxyAPI passthrough executor via `providerSpecificData.cliproxyapiMode="claude-native"`, with precedence **connection override > provider `upstream_proxy_config` mode > default**. `resolveExecutorWithProxy` now receives the resolved connection's `providerSpecificData` (threaded from `chatCore.ts`), so one connection can deep-route while the provider's other connections stay native — no DB schema change (the toggle rides in `providerSpecificData`). Also resolves the same-provider mixing ask in #6340. Regression guard: `tests/unit/chatcore-executor-proxy.test.ts` (9). (thanks @RaviTharuma)
+- **feat(dashboard):** "Add session cookie" modal now shows a prominent **"Open ‹host› →"** link to the provider's own site ([#6268](https://github.com/diegosouzapw/OmniRoute/issues/6268)) — every `-web` cookie-session provider (chatgpt-web, claude-web, gemini-web, kimi-web, lmarena, qwen-web, m365-copilot-web, …) renders a one-click external link (opening the provider's login/home page in a new tab) so operators no longer tab away to retype the URL mid-setup. The host resolves from a pure, unit-tested `resolveWebProviderHost()` (prefers `WEB_COOKIE_PROVIDERS[id].website`, falls back to the registry `baseUrl` origin); non-web providers render exactly as before. Kilo's dup flag vs #6265 (modal-too-small-on-1080p) was a false positive — distinct concern. Regression guard: `tests/unit/resolve-web-provider-host.test.ts` (5). (thanks @chirag127)
+- **feat(providers):** add **DigitalOcean** AI (serverless inference) as an OpenAI-compatible API-key provider ([#6373](https://github.com/diegosouzapw/OmniRoute/pull/6373)) — base `https://inference.do-ai.run/v1`, wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/digitalocean/`, `src/shared/constants/providers/apikey/inference-hosts.ts`). Regression guard: `tests/unit/digitalocean-provider.test.ts`. (thanks @newnol)
+- **feat(providers):** add **Huancheng Public API** (`hcnsec`) as an OpenAI-compatible regional provider ([#6410](https://github.com/diegosouzapw/OmniRoute/pull/6410)) — Xinjiang Huancheng Cybersecurity's public LLM platform (base `https://api.hcnsec.cn/v1`, free credits via daily check-ins), wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/hcnsec/`, `src/shared/constants/providers/apikey/regional.ts`). Regression guard: `tests/unit/hcnsec-provider.test.ts`. (thanks @UnrealAryan)
+- **feat(dashboard):** the web-session credential guide now shows an **"Open {host}" link** ([#6316](https://github.com/diegosouzapw/OmniRoute/pull/6316)) to the provider's sign-in site (derived from the provider `website` via `getProviderWebsiteHost`), so you can jump straight to the page where the cookie/session must be captured. Regression guard: `tests/unit/web-session-provider-link-6316.test.ts`. (thanks @jordansilly77-stack)
+- **feat(cerebras):** add the **Gemma 4 31B** model (`gemma-4-31b`) to the Cerebras registry + pricing table ([#6331](https://github.com/diegosouzapw/OmniRoute/pull/6331)). Regression guard extends `tests/unit/t28-model-catalog-updates.test.ts`. (thanks @backryun)
+- **feat(providers):** add **Yuanbao (web)** as a cookie-session provider ([#6196](https://github.com/diegosouzapw/OmniRoute/issues/6196)) — `yuanbao-web` (Tencent Yuanbao, `yuanbao.tencent.com`) with cookie-only auth (`hy_user`/`hy_token` + public agent id), SSE→OpenAI translation incl. `reasoning_content`, exposing DeepSeek V3/R1 + Hunyuan / Hunyuan-T1. Regression guard: `tests/unit/providers-yuanbao-web.test.ts`. `together-web` was **deferred** (no verifiable web-session endpoint — needs a captured request) and `huggingchat-web` **dropped** (the existing `huggingchat` already is a web-cookie provider). (thanks @chirag127)
+- **feat(providers):** route the built-in **agentrouter** through the dynamic Claude-Code wire image ([#6056](https://github.com/diegosouzapw/OmniRoute/issues/6056)) — a small static allow-set (`CC_WIRE_IMAGE_BUILTINS` in `open-sse/services/ccWireImageBuiltins.ts`), consulted by `isClaudeCodeCompatible` / `isClaudeCodeCompatibleProvider` / `applyFingerprint`, makes agentrouter adopt the CC wire-image headers + fingerprint **while guarding the CC baseUrl/auth branches** so it keeps its own registry `baseUrl` and `x-api-key` auth. Regression guard: `tests/unit/agentrouter-cc-wire-image.test.ts` (asserts the wire image is applied AND agentrouter's baseUrl/auth are preserved). Live WAF-acceptance against agentrouter.org is a VPS validation follow-up (Hard Rule #18).
+- **feat(providers):** **bulk-add API keys for Cloudflare Workers AI** ([#6174](https://github.com/diegosouzapw/OmniRoute/issues/6174)) — `cloudflare-ai` is removed from the bulk-add exclusion list and the bulk parser gains a 3-field `name|accountId|apiKey` mode; the bulk route now builds a **per-entry** `providerSpecificData` so each key carries its own `accountId` (fixing the previous shared-object reuse), and both the create + key-validation paths receive it. Regression guard: `tests/unit/bulk-api-key-parser-cloudflare.test.ts`. (thanks @muflifadla38)
+- **feat(dashboard):** routing/settings UX clarity ([#6147](https://github.com/diegosouzapw/OmniRoute/issues/6147)) — (1) weighted combos show the **effective routing share %** next to each weight when weights don't sum to 100 (`WeightTotalBar.tsx`); (2) the status widget's user-facing **"Cloud Sync" label is renamed** to "Remote Settings Sync" (`CloudSyncStatus.tsx`; internal ids/state untouched); (3) built-in providers gain an **opt-in advanced base-URL override** (`isBaseUrlOverrideEligibleProvider`, hidden behind an "Advanced" toggle, reusing the existing `providerSpecificData.baseUrl` persistence — not globally widened). Regression guard: `tests/unit/routing-settings-ux-6147.test.ts`.
+- **feat(combo):** add an option to **disable session stickiness**, per-combo or globally — round-robin / random combos can rotate to a different connection on every request instead of pinning a whole conversation to one connection by its first-message hash. Resolution precedence per-combo `config.disableSessionStickiness` → global `settings.disableSessionStickiness` → default `false` (preserves the #3825 prompt-cache/504 fix); gates **both** stickiness call sites in `open-sse/services/combo.ts`. Exposed as a global toggle (Combo Defaults) and a per-combo Inherit/on/off control. ([#6168](https://github.com/diegosouzapw/OmniRoute/issues/6168)) Regression guard: `tests/unit/combo-disable-session-stickiness.test.ts`. (thanks @RCrushMe)
+- **feat(docker):** add the `OMNIROUTE_NO_SUDO` env flag for root-less / user-namespaced deployments — the MITM cert-trust command path (`resolveSudoSpawn` in `src/mitm/systemCommands.ts`) now strips the leading `sudo` when the flag is truthy, in addition to the existing root / sudo-missing cases, so the Proxy Agent runs without `sudo` (the operator trusts the CA manually, e.g. via `NODE_EXTRA_CA_CERTS`). Argv-array `spawn` preserved — no shell interpolation (Hard Rule #13). ([#6122](https://github.com/diegosouzapw/OmniRoute/issues/6122)) Regression guard: `tests/unit/mitm-systemCommands-no-sudo.test.ts`. (thanks @powellnorma)
+- **feat(providers):** add **Requesty** as an OpenAI-compatible gateway provider (BYOK, base `https://router.requesty.ai/v1`, ~200 free requests/day) — wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/requesty/`, `src/shared/constants/providers/apikey/gateways.ts`). ([#6120](https://github.com/diegosouzapw/OmniRoute/issues/6120)) Regression guard: `tests/unit/requesty-provider.test.ts`. (thanks @chirag127)
+- **feat(dashboard):** add **configured-only / available-only filters** to the Free Provider Rankings page ([#6150](https://github.com/diegosouzapw/OmniRoute/issues/6150)) — hide providers you haven't configured, or whose connections are all rate-limited / out of quota, via server-side query params (`?configuredOnly` / `?availableOnly` on `GET /api/free-provider-rankings`) backed by a testable lib helper reusing the in-process connection state (no Redis). Both filters default off, so the default view is unchanged; this supersedes the earlier client-side "Configured Only" toggle (#6245) with an available-only dimension and unit-tested logic. Regression guard: `tests/unit/freeProviderRankings-filters.test.ts`.
+
+### 🔧 Bug Fixes
+
+- **fix(dashboard):** adding a second API key connection for the same provider no longer silently overwrites the first — the Add-API-key modal now derives a unique default connection name (`main`, then `main-2`, `main-3`, …) so the backend name-based upsert can't collide (#6499 — thanks @dilneiss).
+- **fix(compression):** the session-dedup engine now also deduplicates a large multi-line block repeated **within a single message** (intra-message dedup), not just across turns; the compression-preview API surfaces a `fallbackReason`, and the fusion panel reports how many models were rate-limited vs failed on total-panel failure (#6501 — thanks @chirag127).
+- **fix(compression):** a stacked-pipeline step naming an unregistered engine now surfaces a `validationErrors` entry instead of silently no-op'ing, so misconfigured pipelines are visible in the preview API (#6506 — thanks @chirag127).
+- **feat(usage):** add a **Codex reset-credit redemption** flow to the Provider Limits UI ([#6361](https://github.com/diegosouzapw/OmniRoute/pull/6361)) — a `useCodexResetCreditRedemption` hook + `/api/usage/codex-reset-credit` route + `codexResetCredits` lib let you redeem banked Codex reset credits from the quota card. Regression guard: `tests/unit/codex-reset-credits.test.ts`. (thanks @JxnLexn)
+- **feat(glm):** add **team-plan quota settings** for `glm-cn` connections ([#6351](https://github.com/diegosouzapw/OmniRoute/pull/6351)) — a dedicated `GlmTeamQuotaFields` form section (team quota id / limits) threaded through the Add/Edit connection modals, persisted via `providerSpecificData`, with the GLM usage service reading the team quota. Regression guards: `tests/unit/glm-team-quota.test.ts`, `provider-specific-data-schema.test.ts`. (thanks @hao3039032)
+- **feat(providers):** add **TinyFish** web-fetch/search support ([#6349](https://github.com/diegosouzapw/OmniRoute/pull/6349)) — a `tinyfish-fetch` executor + `/v1/web/fetch` route + MCP web-fetch tool, registered as a specialty-media provider with request-validation and a search-provider catalog entry. Regression guards: `tests/unit/executor-tinyfish-fetch.test.ts`, `web-fetch-handler.test.ts`, `mcp-web-fetch-tool.test.ts`, `provider-validation-tinyfish.test.ts`. (thanks @dtybnrj)
+- **fix(cli):** `omniroute launch-codex` now spawns `codex.cmd` through a shell on Windows (the npm `.cmd` shim is unresolvable by bare `spawn` → ENOENT), mirroring the qodercli Windows fix (#6263) ([#6312](https://github.com/diegosouzapw/OmniRoute/pull/6312)). Regression guard: `tests/unit/launch-codex-windows-spawn-6312.test.ts`. (thanks @swingtempo)
+- **fix(codex):** isolate the **Spark** quota from the shared Codex quota and stabilize the quota UI ordering / hydration so per-scope limits render consistently ([#6336](https://github.com/diegosouzapw/OmniRoute/pull/6336)). Regression guards: `tests/unit/codex-quota-selection-hydration.test.ts`, `provider-limits-ui.test.ts` + 3 more. (thanks @xz-dev)
+- **feat(api):** add a `hidePaidModels` setting that filters paid-only models out of the `/v1/models` catalog. Regression guard: `tests/unit/models-catalog-hide-paid.test.ts`. (thanks @chirag127)
+- **fix(api-manager):** the fallback model picker now preserves combos instead of dropping them when a primary model is unavailable ([#6443](https://github.com/diegosouzapw/OmniRoute/pull/6443)). Regression guard: `tests/unit/api-manager-page-static.test.ts`. (thanks @jmengit)
+- **fix(providers):** recoverable Antigravity / Cloud-Code (Gemini Code Assist) `403` responses ([#6452](https://github.com/diegosouzapw/OmniRoute/pull/6452)) are now classified as a retryable project-config error instead of a terminal account ban, so a fixable project/API-disabled 403 no longer forces a ~1-year cooldown / full OAuth reconnect. Regression guard: `tests/unit/errorclassifier-antigravity-403.test.ts`. (thanks @developerjillur)
+- **fix(mitm):** `sanitizeHeaders` now redacts `Set-Cookie` response headers so upstream session cookies never leak into logs / diagnostics ([#6451](https://github.com/diegosouzapw/OmniRoute/pull/6451)). Regression guard: `tests/unit/mitm-sanitize-headers.test.ts`. (thanks @developerjillur)
+- **fix(api):** `/api/compression/preview` now accepts `mode: "caveman"` and correctly handles stacked / zero-compression previews ([#6425](https://github.com/diegosouzapw/OmniRoute/issues/6425)). Regression guard: `tests/unit/api/compression-preview-caveman-and-stacked-6425.test.ts`. (thanks @chirag127)
+- **feat(providers):** add **Zed** hosted LLM aggregator as a native-app provider ([#6118](https://github.com/diegosouzapw/OmniRoute/pull/6118)) — OAuth sign-in via the Zed hosted flow, registered through the shared provider registry + executor. Regression guards: `tests/unit/zed-oauth-provider.test.ts`, `zed-import-utils.test.ts`, `zed-docker-detect.test.ts`, `mitm-handler-zed.test.ts`. VPS-validated via live operator login (Hard Rule #18).
+- **fix(oauth):** the Kiro SSO-cache auto-import now **preserves the IDC region** — cross-region Amazon Q / Kiro profiles imported from the SSO cache are no longer collapsed to the default region ([#6113](https://github.com/diegosouzapw/OmniRoute/pull/6113)). Regression guard: `tests/unit/kiro-auto-import-idc-2059.test.ts`. VPS-validated via live operator login (Hard Rule #18).
+- **fix(dashboard):** passthrough model aliases no longer collide when two namespaced model ids share a last segment (port from 9router#1850, [#6431](https://github.com/diegosouzapw/OmniRoute/pull/6431)). `enx/gpt-5.5` and `enx/codebuddy/gpt-5.5` both auto-generated the alias `gpt-5.5`, so the second model could never be added (the UI just alerted "alias already exists"). Aliases are now disambiguated deterministically — bare last segment when free, then parent-qualified (`codebuddy-gpt-5.5`), then a numeric suffix — while re-adding the exact same model id is still blocked. Regression guard: `tests/unit/passthrough-alias-1850.test.ts`. (thanks @arpicato)
+- **fix(translator):** preserve a Gemini `functionResponse` co-located with other parts (another `functionCall`, or trailing `text`) in the same content when translating **Gemini → OpenAI** ([#6376](https://github.com/diegosouzapw/OmniRoute/pull/6376)). `convertGeminiContent()` early-returned the tool message on the first `functionResponse` part, dropping any co-located parts; such contents are now pre-split (one tool message per `functionResponse`, emitted first, plus one message for the remaining parts). Regression guard: `tests/unit/gemini-to-openai-function-response.test.ts`. (thanks @warelik)
+- **fix(headroom):** detect a python interpreter managed by **mise / pyenv / asdf / conda** (port from 9router#2353, [#6382](https://github.com/diegosouzapw/OmniRoute/pull/6382)). Headroom's python probe (`src/lib/headroom/detect.ts`) searched a hardcoded `PATH`, but version managers expose their interpreters via shim dirs that only join `PATH` through interactive-shell activation — which the non-interactive server never runs, so a managed python (≥3.10) was invisible and Headroom reported it missing. The search path now prepends the well-known shim/bin dirs (`~/.local/share/mise/shims`, `~/.pyenv/shims`, `~/.asdf/shims`, `$CONDA_PREFIX/bin`, `~/.local/bin`, respecting `MISE_DATA_DIR`/`PYENV_ROOT`/`ASDF_DATA_DIR` when set), and a new `HEADROOM_PYTHON` env override lets operators point straight at their interpreter (mirroring `HEADROOM_URL`). Still shell-free (`execFileSync`). Regression guard: `tests/unit/headroom-detect.test.ts` (5). (thanks @loopyd)
+- **fix(executors):** strip the OpenAI-Codex/Claude-CLI `client_metadata` passthrough field for **NVIDIA** requests (port from 9router#1887, [#6411](https://github.com/diegosouzapw/OmniRoute/pull/6411)). NVIDIA's OpenAI-compatible wrapper rejects it with `400 Unsupported parameter`, the same class already handled for `cerebras`/`mistral`; `nvidia` (executor `default`) was missing from the strip allowlist so Codex/Claude-Code passthrough requests 400'd. Regression guard: `tests/unit/executor-default-strip-client-metadata.test.ts` (+nvidia case). (thanks @phidinhmanh)
+- **fix(translator):** strip the Claude-style `thinking` field for **NVIDIA `z-ai/glm-5.2`** (port from 9router#2023, [#6413](https://github.com/diegosouzapw/OmniRoute/pull/6413)). NVIDIA's OpenAI-compatible wrapper 400s on `thinking` (a Claude-format client routed here leaves a `thinking:{type:"adaptive"}`); the existing strip rule only dropped `reasoning`. Same class already handled for `minimax-m2.7`. Regression guard: `tests/unit/nvidia-minimax-thinking-strip.test.ts` (+glm-5.2 case). (thanks @phidinhmanh)
+- **fix(translator):** suppress the streamed `</think>` close marker for the **Antigravity IDE** client (port from 9router#1061, [#6415](https://github.com/diegosouzapw/OmniRoute/pull/6415)). On thinking-only turns Antigravity rendered a bare `</think>` as the sole visible content, tripping its loop-detection and wasting requests. Antigravity's UA (`vscode/<v> (Antigravity/<v>)`) is added to the marker-suppress allowlist (alongside OpenCode); Claude Code / Cursor still get the marker, and `x-omniroute-thinking-marker: on` force-restores it. Regression guard: `tests/unit/think-close-marker-suppress-5245.test.ts`. (thanks @abdofallah)
+- **fix(executors):** strip nested `reasoning_content` from messages for **Mistral** (port from 9router#1649, [#6417](https://github.com/diegosouzapw/OmniRoute/pull/6417)). Mistral's API returns `422 extra_forbidden` when an assistant message carries `reasoning_content` (replayed thinking from a prior turn, e.g. via the Codex `/responses` path); the generic top-level 400 field-downgrade retry never covered the nested per-message field. `DefaultExecutor` now strips it for provider `mistral` only, so DeepSeek (which requires replayed `reasoning_content`) is unaffected. Regression guard: `tests/unit/mistral-strip-reasoning-content-1649.test.ts`. (thanks @xxy9468615)
+- **fix(executors):** strip the `client_metadata` passthrough field on the **OpenCode** path (port from 9router#1442, [#6418](https://github.com/diegosouzapw/OmniRoute/pull/6418)). OpenCode upstreams (e.g. `kimi-k2.6` via opencode-go) reject it with `400 "Extra inputs are not permitted, field: 'client_metadata'"`; the DefaultExecutor strip only covered cerebras/mistral and `OpencodeExecutor` extends `BaseExecutor` directly, so nothing removed it there. Regression guard: `tests/unit/opencode-strip-client-metadata-1442.test.ts`. (thanks @yanpaing007)
+- **fix(executors):** inject the `reasoning_content` echo for the native **Moonshot Kimi** provider (port from 9router#1480, [#6419](https://github.com/diegosouzapw/OmniRoute/pull/6419)). Kimi (executor `default`) is a thinking-mode upstream that 400s with "reasoning_content must be passed back" when a prior assistant turn lacks it; the placeholder injection was only wired into the OpenCode meta-provider, so direct multi-turn Kimi conversations failed. Scoped to `kimi` (gateway-served models matching the thinking-model name pattern are unaffected). Regression guard: `tests/unit/kimi-native-reasoning-injected-1480.test.ts`. (thanks @2220258345)
+- **fix(executors):** recover from a strict gateway's `context_management: Extra inputs are not permitted` 400 (port from 9router#1468, [#6420](https://github.com/diegosouzapw/OmniRoute/pull/6420)). **Claude Code** always sends a top-level `context_management` field; strict anthropic-compatible gateways reject it. The dedicated context-editing 400-fallback only fired when OmniRoute's own `contextEditing` feature was enabled (default off), so a client-sent field passed through untouched and 400'd. `context_management` is now in the generic reactive field-strip list, so it's stripped-and-retried once regardless of the feature flag (with correct request re-signing for claude-compatible relays). Regression guard: `tests/unit/provider-field-strips.test.ts`. (thanks @ohahe52-dot)
+- **fix(network):** enable **RFC 8305 Happy Eyeballs** (`autoSelectFamily`) on the direct-egress undici dispatcher (port from 9router#1237, [#6423](https://github.com/diegosouzapw/OmniRoute/pull/6423)). When DNS returns both IPv6 (AAAA) and IPv4 (A) and the IPv6 route is broken (e.g. a NAT64 `64:ff9b::` prefix without routing), undici tried IPv6 first and hung until `ETIMEDOUT` (then a 502 + account lockout), even though `curl` reached the same host. The direct dispatcher now races both families and uses whichever connects first. Proxy paths pin family via `proxyTls` and are unaffected. Regression guard: `tests/unit/direct-dispatcher-pipelining-4580.test.ts`. (thanks @adentdk)
+- **fix(combo):** round-robin now advances the rotation pointer past the model that **actually served**, not the eagerly-scheduled one (port from 9router#948, [#6428](https://github.com/diegosouzapw/OmniRoute/pull/6428)). With `stickyLimit: 1` (true round-robin), when the scheduled model failed and a _different_ model served via fallback, the counter had already advanced +1 from the scheduled index — so the next request reused the fallback-served model, degrading round-robin into hot-spotting on whichever model was healthy. The pointer now advances to the served index + 1 (mirroring the sticky-limit>1 path). Session-stickiness (#3825) and distribution are preserved. Regression guard: `tests/unit/combo-rr-fallback-advance-948.test.ts`. (thanks @binsarjr)
+- **fix(sse):** a non-string `model` field is now rejected with a `400` before the resolver, instead of crashing downstream `.toLowerCase()`/`.split()` calls into an empty-body `500` that escapes the error sanitizer ([#6407](https://github.com/diegosouzapw/OmniRoute/issues/6407)). Regression guard: `tests/unit/chat-non-string-model-6407.test.ts`. (thanks @chirag127)
+- **fix(api):** unknown `/api/*` routes now return a JSON `404` (instead of the dashboard HTML shell) and scalar chat params (`model`/`temperature`/etc.) are validated **before** the provider lookup so malformed requests fail fast with a clear `400` ([#6424](https://github.com/diegosouzapw/OmniRoute/issues/6424), [#6412](https://github.com/diegosouzapw/OmniRoute/issues/6412)). Regression guards: `tests/unit/api/api-catchall-json-404.test.ts`, `tests/unit/chat-early-schema-validation-6412.test.ts`. (thanks @chirag127)
+- **fix(api):** `/v1/chat/completions` now rejects a non-JSON `Content-Type` with a `400` before parsing the body ([#6414](https://github.com/diegosouzapw/OmniRoute/issues/6414)). Regression guard: `tests/unit/v1-chat-completions-content-type-6414.test.ts`. (thanks @chirag127)
+- **fix(api):** the `X-OmniRoute-Compression` response header is now echoed on `/v1/chat/completions` and `/v1/completions` ([#6422](https://github.com/diegosouzapw/OmniRoute/issues/6422)). Regression guard: `tests/unit/compression-header-echo-6422.test.ts`. (thanks @chirag127)
+- **fix(api):** concurrent `GET /v1/models` requests are coalesced into a single catalog build ([#6408](https://github.com/diegosouzapw/OmniRoute/issues/6408)). Regression guard: `tests/unit/v1-models-concurrent-6408.test.ts`. (thanks @chirag127)
+- **fix(api):** `/v1/completions` now echoes the requested `body.model` in its JSON + streamed responses ([#6429](https://github.com/diegosouzapw/OmniRoute/pull/6429)). Regression guard: `tests/unit/completions-body-model-echo.test.ts`. (thanks @chirag127)
+- **fix(api):** env-var master keys now see the full `/v1/models` catalog ([#6406](https://github.com/diegosouzapw/OmniRoute/issues/6406)). Regression guard: `tests/unit/models-catalog-envkey-6406.test.ts`. (thanks @chirag127)
+- **fix(api):** non-streaming `/v1/completions` responses now echo `body.model` aligned with the `X-OmniRoute-Model` header ([#6426](https://github.com/diegosouzapw/OmniRoute/issues/6426)). Regression guard: `tests/unit/v1-completions-model-header-match-6426.test.ts`. (thanks @chirag127)
+- **fix(api):** unknown `/v1/*` routes now return a JSON `404 not_found` instead of the Next.js dashboard HTML shell ([#6405](https://github.com/diegosouzapw/OmniRoute/issues/6405)). Regression guard: `tests/unit/api/v1-catchall-json-404.test.ts`. (thanks @chirag127)
+- **fix(api):** the per-connection provider models route now degrades to the shipped catalog when a provider's `/models` endpoint answers with a redirect ([#6267](https://github.com/diegosouzapw/OmniRoute/issues/6267)) — a `qwen-web` import failed with a raw `Redirect blocked … (307)` 503. `safeOutboundFetch` throws `REDIRECT_BLOCKED` on the 307, `getSafeOutboundFetchErrorStatus` maps it to 503, and `buildDiscoveryErrorFallbackResponse` treated every 503 as a hard error — so the non-empty `getModelsByProviderId("qwen-web")` catalog was never surfaced. A models-endpoint redirect is not a fixable-config error (unlike `URL_GUARD_BLOCKED`/`INVALID_URL`, which stay hard errors), so it now falls back to the local/cached catalog before the 503 short-circuit. General fix — covers any config-driven provider that 307s. Regression guard: `tests/unit/provider-models-qwen-web-redirect-6267.test.ts`. (thanks @chirag127)
+- **fix(api):** the per-connection provider models route (MCP `list_models_catalog` + the dashboard import view) now merges USER-ADDED custom models into its response ([#6247](https://github.com/diegosouzapw/OmniRoute/issues/6247)) — custom models live in the `key_value` namespace `customModels`, which the live REST `/api/v1/models` already merges, but `src/app/api/providers/[id]/models/route.ts` never read `getCustomModels`, so custom models were dropped on both the discovery-success and local_catalog paths. They are now folded into the returned model list (deduped by id, stamped `owned_by: provider`), fixing MCP + the dashboard import view in one place. Regression guard: `tests/unit/provider-models-custom-merge-6247.test.ts`. (thanks @RCrushMe)
+
+- **fix(providers):** GitLab Duo tool-calling follow-up turns no longer fail upstream with `422 {"detail":"Validation error"}` (tokens 0/0, rejected pre-inference). The #6234 tool-result-feedback fix serialized the **entire** multi-turn conversation into GitLab's single-file `code_suggestions` (`small_file`) generation endpoint — folded history that turn-N sent as an oversized `current_file.content_above_cursor` **and** duplicated verbatim into `user_instruction`, tripping the AI-Gateway's `small_file` validation guard. The executor now **bounds** that prompt: it keeps system + latest user message + the most-recent tool round (dropping older turns), caps oversized tool results, and stops duplicating the full prompt into `user_instruction` (which now carries only the short latest user message) — while still feeding the most-recent tool result back so the agent continues (`open-sse/executors/gitlab.ts`, [#6220](https://github.com/diegosouzapw/OmniRoute/issues/6220)). The unit test covers the bounding logic; the upstream 422→200 clearing is VPS-only (Hard Rule #18). Regression guard: `tests/unit/gitlab-tool-exchange-bounded-6220.test.ts`.
+
+- **fix(i18n):** the provider-detail (`/dashboard/providers/[id]`) connection-status filter labels no longer render as `__MISSING__:All` / `__MISSING__:Active` / `__MISSING__:Error` / `__MISSING__:Banned` / `__MISSING__:CreditsExhausted` in non-English locales (notably pt-BR) ([#6290](https://github.com/diegosouzapw/OmniRoute/issues/6290)). Root cause was **not** the namespace mismatch the issue guessed — the `providers.filter*` keys resolve correctly in `en.json`; the debt lived in the locale mirrors (`src/i18n/messages/*.json`), where these five keys carried the `__MISSING__:` sync sentinel in ~15 locales and were absent entirely in ~26 others, so next-intl found the key and echoed the sentinel verbatim. All 40 non-English/-Chinese mirrors now ship real translations for the five `providers.filter*` labels. Regression guard: `tests/unit/i18n-provider-filter-keys-6290.test.ts`. (thanks @diegosouzapw)
+
+- **fix(providers):** the `copilot-m365-web` streaming executor now emits `debug`-level WebSocket diagnostics ([#6210](https://github.com/diegosouzapw/OmniRoute/issues/6210)) — the outbound WS URL (with the `access_token` **redacted** via `redactWsUrl()`), handshake success/failure, and each received SignalR frame's `type`/`target`. Previously the streaming path logged nothing, so an empty `content:null` response (the M365 Education / Starter tier symptom fixed in #6234) was undiagnosable even at `APP_LOG_LEVEL=debug`. The change is debug-level and side-effect-free — it does not alter streaming behavior or the frame parser, and the token never reaches the logs. Regression guard: `tests/unit/copilot-m365-web-logging-6210.test.ts` (thanks @qpeyba)
+
+- **fix(resilience):** a round-robin combo no longer returns `503 all upstream accounts are unavailable` when a compatibility-rejected target is actually healthy ([#6238](https://github.com/diegosouzapw/OmniRoute/issues/6238)). `filterTargetsByRequestCompatibility` drops request-incompatible targets (tool/vision/structured-output unsupported, or below the required context window) **before** any availability check runs, and its `compatible.length === 0` safety net only fired when _all_ targets were filtered — not when the kept targets later all turned out runtime-unavailable (circuit-open / cooldown / no credentials). So a combo could 503 while a compat-rejected-but-healthy provider sat unused. `handleRoundRobinCombo` now keeps the compat-rejected set and, when every compat-kept target was skipped without a single real attempt, probes those rejected targets as a **last-resort fallback tier** (via the new pure `open-sse/services/combo/comboCompatFallback.ts`) before crystallizing the 503. Regression guard: `tests/unit/combo-roundrobin-compat-fallback-6238.test.ts`. (thanks @ThongAccount)
+
+- **fix(startup):** best-effort self-heal for a corrupted Turbopack dev cache on Windows ([#6289](https://github.com/diegosouzapw/OmniRoute/issues/6289)). On Windows, `pnpm dev` can fail at startup when Turbopack `mmap`s a persistent-cache SST file and the OS refuses the mapping (`os error 1455` — "paging file too small"), which Turbopack surfaces as a misleading `Module not found: Can't resolve '@/shared/utils/machine'`. This is a **known upstream Turbopack cache-corruption bug — not our code**. The dev launcher (`scripts/dev/run-next.mjs`) now wraps `nextApp.prepare()` and, when it rejects with that signature (`isTurbopackCacheCorruption` in the new `scripts/dev/turbopackCacheHeal.mjs`), purges `.build/next/**/cache/turbopack` and retries **once** with a clear log. **Caveat — best-effort only:** the corruption often surfaces as a runtime overlay rather than a `prepare()` rejection, so this cannot always intercept it; the reliable remedy remains manually deleting the Turbopack cache dir. Regression guard: `tests/unit/turbopack-cache-heal-6289.test.ts`. (thanks @chirag127)
+
+- **fix(providers):** qodercli PAT auth no longer fails with `spawn qodercli ENOENT` on Windows ([#6263](https://github.com/diegosouzapw/OmniRoute/issues/6263)) — `spawnQoderCli` spawned the bare `qodercli` name with `shell:false` and an unenriched env, so the npm `.cmd` wrapper under `%APPDATA%\npm` (a user-PATH directory) was never resolved. It now resolves the absolute `.cmd`/`.exe` path through the existing `getCliRuntimeStatus("qoder")` resolver in `src/shared/services/cliRuntime.ts` (memoized), spawns with `shell` when the target is a `.cmd`/`.bat`, and uses the cliRuntime-enriched env (PATH + PATHEXT + APPDATA); the ENOENT error now lists the searched paths plus the `CLI_QODER_BIN` override. End-to-end spawn on a real Windows host is host-only (Hard Rule #18); the path-resolution logic is unit-tested. Regression guard: `tests/unit/qodercli-windows-resolve-6263.test.ts`. (thanks @chirag127)
+
+- **fix(sse):** the reasoning-token buffer no longer inflates **probe-sized `max_tokens`** ([#6274](https://github.com/diegosouzapw/OmniRoute/issues/6274)) — Claude Code's `/model` capability check sends `max_tokens: 1`, but for a thinking-capable model with a large output cap (e.g. `glm-5.2`) the #3587 headroom heuristic (`max(current + 1000, ceil(current * 1.5))`) rewrote it to `1001` and forwarded that upstream, wasting tokens on a request that was never a genuine reasoning budget. `resolveReasoningBufferedMaxTokens()` (`open-sse/services/reasoningTokenBuffer.ts`) now short-circuits and returns the caller's value verbatim when it is below the new `REASONING_BUFFER_MIN_TRIGGER` (256) threshold — a tiny explicit limit is a probe, not a reasoning request. Real budgets still receive the #3587 headroom unchanged, and the guard runs after the existing capability checks so unknown / non-reasoning models keep returning `null`. Regression guard: `tests/unit/reasoning-token-buffer-6274.test.ts`. (thanks @brightfiscalband)
+
+- **fix(cli):** `omniroute reset-password` now works as a real subcommand, and password resets over piped (non-TTY) stdin actually apply ([#6261](https://github.com/diegosouzapw/OmniRoute/issues/6261), [#6258](https://github.com/diegosouzapw/OmniRoute/issues/6258)). Two coupled defects: (1) **#6261** — `bin/omniroute.mjs` routed everything through Commander with only two pre-Commander bypasses (`--mcp`, `reset-encrypted-columns`), so `omniroute reset-password` was rejected as an unknown command; only the separate `omniroute-reset-password` bin worked, while the docs falsely advertised the subcommand (incl. a bogus "legacy alias still works"). A pre-Commander bypass mirroring `reset-encrypted-columns` now dynamically imports `bin/reset-password.mjs` (which self-executes) before Commander parses; the three doc lines were corrected. (2) **#6258** — `bin/reset-password.mjs` issued two sequential `rl.question` prompts; under piped stdin the second read never settled at EOF, so `main()` never reached `resetManagementPassword` and the reset was a silent no-op (both prompts printed, no success, password unchanged). The CLI now detects non-TTY stdin and reads it once (first line = password, second line = confirm if present, else reused), adds a `--password-stdin` flag (entire stdin is the password, no confirmation), and exits `0` explicitly so the success line always flushes; interactive TTY behavior is unchanged. Regression guard: `tests/unit/reset-password-cli-6261-6258.test.ts` (3). (thanks @chirag127)
+
+- **fix(db):** the mass-migration **safety abort** now tells the operator how to bypass it and stops flooding the log ([#6260](https://github.com/diegosouzapw/OmniRoute/issues/6260)) — after restoring a backup that wiped the migration tracking table, `runMigrations()` threw the abort on every downstream `ensureDbInitialized()`, re-logging the full banner 11+ times, and the message never mentioned the existing `OMNIROUTE_MAX_PENDING_MIGRATIONS` escape hatch. The abort text now appends a bypass hint (set `OMNIROUTE_MAX_PENDING_MIGRATIONS=0` in `server.env` / `DATA_DIR/.env`), and a new `MigrationSafetyAbortError` is memoized so repeated calls in the same process throw the same instance and emit a single concise line instead of the full cascade. Regression guard: `tests/unit/migration-safety-abort-6260.test.ts`. (thanks @chirag127)
+
+- **fix(auth):** importing a **distinct** Codex/ChatGPT OAuth `auth.json` is no longer falsely rejected as "already exists" when it belongs to a different user in the same workspace ([#6301](https://github.com/diegosouzapw/OmniRoute/issues/6301)). `findExistingCodexConnection` (in `src/lib/oauth/utils/codexAuthImport.ts`) deduped **only** on `providerSpecificData.workspaceId === accountId`, where `accountId` is the shared `chatgpt_account_id`/`tokens.account_id` — so two members of the same ChatGPT Team collapsed onto a single connection (409 `duplicate_account`). The id_token's `https://api.openai.com/auth` claim carries a per-user `chatgpt_user_id` alongside the workspace id (the device-flow path already persisted it as `chatgptUserId`, but the import path did not). Now `parseAndValidateCodexAuth` extracts `userId` (`chatgpt_user_id` → `user_id` → JWT `sub`) into `ParsedCodexAuth`, the create/update paths persist `chatgptUserId` in `providerSpecificData` (mirroring `codex.ts`), and dedup keys on `workspaceId` **AND** `chatgptUserId` — with a backward-compat fallback to legacy accountId-only matching when no stored connection for that workspace records a `chatgptUserId`, so genuinely-same accounts still dedup. Regression guard: `tests/unit/codex-auth-import-userid-dedup-6301.test.ts` (4). (thanks @anungma)
+
+- **fix(providers):** importing models for the **venice-web** provider no longer fails with a red "Provider venice-web does not support models listing" ([#6269](https://github.com/diegosouzapw/OmniRoute/issues/6269)). `venice-web` is a web-cookie provider with an executor but no upstream `/v1/models` endpoint and no registry `models`, so the models route fell through to the tail `400`. Mirroring the `jules`/`linkup-search`/`ollama-search` fix (#5569), it now ships a static local catalog entry in `src/lib/providers/staticModels.ts` — seeding the current Venice lineup (`venice-uncensored`, `llama-3.3-70b`, `qwen3-235b`, `qwen3-4b`, `deepseek-r1-671b`; Venice rotates its catalog, see docs.venice.ai/models/overview) — so the route returns `200` with `source:"local_catalog"`, `intentional:true`. Regression guard: `tests/unit/static-models-venice-web-6269.test.ts`. (thanks @chirag127)
+
+- **fix(api):** the specialty model catalogs (`/v1/embeddings`, `/v1/images`, `/v1/music`, `/v1/videos` model lists) are now derived from the **unified catalog filtered by a predicate** (`getSpecialtyModelsResponse`) instead of ad-hoc per-route logic, so they consistently respect active-credential visibility and stay in sync with the main catalog ([#6303](https://github.com/diegosouzapw/OmniRoute/pull/6303)). Regression guard: `tests/unit/specialty-model-catalog-routes.test.ts`. (thanks @makcimbx)
+- **fix(api):** the agent-bridge server route now resolves the MITM manager via a **dynamic `import("@/mitm/manager.runtime")`** so Turbopack does not statically pull the stub (or over-bundle the manager), and the agent-skills generator anchors its output base path with `path.join(process.cwd(), …)` so Turbopack's static analyzer stops tracing the whole project root ([#6329](https://github.com/diegosouzapw/OmniRoute/issues/6329), [#6366](https://github.com/diegosouzapw/OmniRoute/pull/6366)). Regression guard: `tests/unit/agent-bridge-server-route-dynamic-import.test.ts`. (thanks @Iammilansoni)
+- **fix(api):** internal probes (combo-test, cloud-sync verify) now pick a **management-scoped / allow-all API key** instead of naively grabbing `getApiKeys()[0]` — a restricted `self:usage` first row made the probe fail with "Model X is not allowed for this API key" even when the combo path was healthy (`pickApiKeyForInternalUse` in `src/lib/db/apiKeys.ts`). The API-manager model editor also falls back to `/api/models?all=true` when `/v1/models` is catalog-protected ([#6372](https://github.com/diegosouzapw/OmniRoute/pull/6372)). Regression guard: `tests/unit/pick-internal-api-key-6372.test.ts`. (thanks @jmengit)
+- **fix(live-ws):** the Live Dashboard WebSocket server now **rejects on bind failure** (e.g. `EADDRINUSE` when the API bridge already holds the port) instead of letting the error surface as an unhandled `error` event that crash-loops the process — the `error` listener is attached to `wss` (not `server`) and releases the EventBus subscription on a failed start ([#6324](https://github.com/diegosouzapw/OmniRoute/issues/6324)). Regression guard: `tests/unit/live-ws-eaddrinuse-6324.test.ts`. (thanks @vinayakkulkarni)
+- **fix(dashboard):** the Home provider-topology widget now trusts the live provider-metrics snapshot — it uses `topology.errorProvider` and live `activeRequests` directly instead of re-deriving state from a stale `lastErrorAt` or applying a frontend timeout filter, so the topology reflects real-time provider health ([#6322](https://github.com/diegosouzapw/OmniRoute/pull/6322)). Regression guard: `tests/unit/home-provider-topology-live-state.test.ts`. (thanks @xz-dev)
+- **fix(sse):** strip zero-width markers from streamed **tool-call arguments** — a follow-up to [#5857](https://github.com/diegosouzapw/OmniRoute/pull/5857). That PR removed injected zero-width joiners (U+200D) from streamed assistant text/reasoning but deliberately left tool-call argument JSON byte-exact. The request-side obfuscation (`open-sse/services/claudeCodeObfuscation.ts`) injects ZWJ into agent words — including the temp path inside the Bash tool description — and Claude models copy that verbatim into generated commands, which are delivered as tool-call arguments rather than assistant text. As a result the ZWJ survived and corrupted code blocks (e.g. a temp path rendered with an invisible joiner). Now `open-sse/handlers/responseSanitizer.ts` strips zero-width code points from tool-call argument strings at every emit site (OpenAI non-stream/stream chat `tool_calls` + legacy `function_call`, native Responses `function_call` items, the OpenAI→Responses conversion, and the native Responses streaming `response.function_call_arguments.delta/.done` events). Only zero-width code points are removed; JSON structure and all other bytes stay identical (no parse/restringify), so normal arguments remain byte-exact. Regression guard: 6 new cases in `tests/unit/response-sanitizer.test.ts` (suite 50/50).
+
+- **fix(nodejs):** the default app log path now resolves under `DATA_DIR` (`~/.omniroute/logs/application/app.log`) instead of `process.cwd()` ([#6197](https://github.com/diegosouzapw/OmniRoute/issues/6197)) — the globally-installed CLI runs from an arbitrary working directory, so anchoring the default to cwd made file logging silently write to (or no-op under) an unrelated directory, contradicting the documented `.env.example` default. `getAppLogFilePath()` now computes the default lazily via the pure `resolveDataDir()` resolver (honours a per-process `DATA_DIR`, no directory-creation side effect); an explicit `APP_LOG_FILE_PATH` still wins. Regression guard: `tests/unit/logenv-datadir-path-6197.test.ts` (3).
+
+- **fix(docker):** AgentBridge/`startMitm` no longer aborts in containers/headless when the Antigravity-default DNS step can't write `/etc/hosts` ([#6127](https://github.com/diegosouzapw/OmniRoute/issues/6127)), and the privileged command's stderr now reaches `app.log` instead of only a bare exit code hitting the toast ([#6198](https://github.com/diegosouzapw/OmniRoute/issues/6198)). The default DNS step (`addDNSEntry`) was called unguarded while cert install and the two sibling DNS steps were each best-effort — in the runtime Docker image (`USER node`, no `sudo`, read-only `/etc/hosts`) it threw `Command failed with code 1` out of `startMitmInternal` and killed the whole start, discarding the stderr. The three DNS steps are extracted into a best-effort `provisionDnsEntries()` where each failure is logged with the full `err` (stderr included, folded in by `systemCommands.ts`) and never aborts the start. Regression guard: `tests/unit/mitm-dns-graceful-degrade-6127.test.ts` (4).
+
+- **fix(providers):** copilot-m365-web now supports the M365 Education "Starter / OfficeWebIncludedCopilot" tier and no longer returns an empty `content:null` stream ([#6210](https://github.com/diegosouzapw/OmniRoute/issues/6210)). Two gaps: (1) `buildWsUrl()` hardcoded the individual-consumer scenario (`OfficeWebPaidConsumerCopilot`, `isEdu=false`) — the EDU tier is now opt-in via `providerSpecificData.tier="edu"`, emitting `scenario=OfficeWebIncludedCopilot`/`isEdu=true` (the individual path is unchanged); (2) the EDU/GPT-5.5 path streams deltas via `arguments[0].writeAtCursor` (incremental) instead of only `messages[].text` (accumulated snapshots), which the parser dropped — a new `accumulateBotContent()` folds both formats, with `type:2 item.result.message` as a last-resort fallback. Regression guard: `tests/unit/copilot-m365-edu-writeatcursor-6210.test.ts` (10). (thanks @qpeyba)
+
+- **fix(providers):** GitLab Duo executor now feeds tool results back into the prompt instead of looping ([#6220](https://github.com/diegosouzapw/OmniRoute/issues/6220)) — `buildPrompt()` branched only on `system`/`user` and took `userParts.at(-1)`, silently dropping the `assistant{tool_calls}` + `tool{result}` turns the client appended, so the reconstructed prompt was byte-identical to turn 1 and the model re-emitted the same `<tool>` call forever. When a tool exchange is present the full conversation is now serialized, folding each tool result back keyed by its `tool_call_id`; simple conversations keep the legacy shape. Complements the tool_call emission from [#6051](https://github.com/diegosouzapw/OmniRoute/issues/6051) (the `kilo-duplicate` label was a false positive — different, sequential defect). Regression guard: `tests/unit/gitlab-tool-result-feedback-6220.test.ts` (4).
+
+- **fix(providers):** opencode-go/opencode-zen can now synthesize the OpenCode CLI identity headers Cloudflare requires on VPS egress ([#5997](https://github.com/diegosouzapw/OmniRoute/issues/5997)) — on a datacenter VPS, `opencode.ai/zen/go/v1/chat/completions` 403s (HTML challenge) requests lacking CLI identity, while the reporter's control curl proved that `User-Agent: opencode-cli/1.0.0` + `x-opencode-client: cli` + `x-opencode-project: default` + fresh request/session UUIDs succeed. Opt-in via `OPENCODE_SYNTHESIZE_CLI_HEADERS=true` (values overridable via `OPENCODE_GO_USER_AGENT`/`OPENCODE_USER_AGENT`/`OPENCODE_CLIENT`/`OPENCODE_PROJECT`); it fills only headers the client did not already send. Kept **off by default** — the forward-only path is deliberate (fabricating a wrong value risks upstream rejection; a prior dedup regressed with `opencode/local`), so this replaces the fragile local header-injection shim without changing default behavior. Regression guard: `tests/unit/opencode-cli-headers-synthesis-5997.test.ts` (6). (thanks @aleksesipenko)
+- fix(resilience): sticky session affinity now evicts and fails over to another account when the pinned account is exhausted/unavailable (#6219)
+
+- fix(sse): Responses API passthrough now drops internal commentary-phase output before forwarding to clients (gated by RESPONSES_PASSTHROUGH_DROP_COMMENTARY, default on) (#6199)
+
+- **fix(sse):** tool-call function schemas with a root `type: null` are now coerced to `type: "object"` before dispatch ([#6359](https://github.com/diegosouzapw/OmniRoute/issues/6359)) — clients like the Codex app emit `parameters: { type: null, ... }` for some tools, which OpenAI-compatible upstreams reject with `400 Invalid schema for function '...': schema must be a JSON Schema of 'type: "object"', got 'type: null'`, failing the whole request. `toolSchemaSanitizer` already stripped the null; it now re-adds the mandatory root `"object"` type (and empty `properties`/open `additionalProperties` when absent). Combinator roots (`anyOf`/`oneOf`/`allOf`) and explicit root types are left untouched. Regression guard: 5 new cases in `tests/unit/tool-schema-sanitizer.test.mjs`.
+
+- **fix(docker):** AgentBridge no longer fails to start on npm/Electron/VPS installs with "MITM manager stub reached at runtime" ([#6344](https://github.com/diegosouzapw/OmniRoute/issues/6344)) — v3.8.45 flipped the production bundler default to Turbopack, but `next.config.mjs` aliased `@/mitm/manager` to its Docker-only degraded stub **unconditionally**. That was harmless while Docker (which sets the alias intentionally for #3390 graceful degradation) was the sole Turbopack consumer, but once every artifact built with Turbopack the stub shipped to all non-Docker users and `startMitm` threw on the first Agent-Bridge start. The alias is now opt-in via `OMNIROUTE_MITM_STUB=1` (set only by the Dockerfile) through the shared `scripts/build/mitm-stub-flag.mjs` helper; default builds bundle the real manager. Regression guard: `tests/unit/mitm-stub-alias-6344.test.mjs` (4).
+
+- **fix(proxy):** stop the v3.8.44 proxy regression that leaked the real IP and disabled healthy proxies ([#6246](https://github.com/diegosouzapw/OmniRoute/issues/6246)). Two coupled defects from the new health scheduler: (1) **IP leak** — when a proxy assigned to a connection was marked `inactive`, resolution fell through to a **direct** egress instead of blocking, exposing the operator's real IP; (2) **over-deactivation** — the sweep flipped a proxy to `inactive` on the **first** failed probe and counted our own 5s timeout / a probe-target `5xx` as the proxy's fault, so healthy paid proxies vanished from egress selection ("my proxies are not being used anymore"). Fix: the sweep decision is extracted into a pure, network-free `decideProxyHealthAction` (`src/lib/proxyHealth/decision.ts`) — by default the health check now **only counts/logs and never downgrades status** (a proxy is downgraded/removed only with `PROXY_AUTO_REMOVE=true`, after `PROXY_AUTO_REMOVE_AFTER` **consecutive** conclusive failures); probes are classified tri-state so an inconclusive result (our timeout, or a `5xx` from the probe target) never penalizes the proxy, and the probe timeout is raised 5s→15s. Separately, `safeResolveProxy` now **fails closed** via the existing policy: a connection whose assigned proxy is dead is blocked instead of leaking direct (`hasBlockingProxyAssignment`), honoring the explicit `proxy off` toggles and the `PROXY_FAIL_OPEN=true` opt-out. Existing proxies stuck `inactive` by the old behavior need a one-time manual re-activate (the operator owns proxy status). Regression guards: `tests/unit/proxy-health-decide-action-6246.test.ts`, `tests/unit/proxy-assigned-unavailable-6246.test.ts`.
+
+- **fix(proxy):** make "Test All" read-only and add bulk enable/disable ([#6246](https://github.com/diegosouzapw/OmniRoute/issues/6246)). Complements the core fail-closed / scheduler fix (#6296) with the two remaining reporter asks. (1) The **"Test All" button** (`POST /api/settings/proxies/auto-test`) used to flip a proxy to `inactive` on a failed reachability probe; since the egress selector excludes `inactive` proxies, a flaky probe (an unreachable `httpbin.org`, a proxy that blocks `HEAD`, or a slow paid proxy) silently disabled every proxy that failed — "Test All" is now **read-only by default** (only the operator sets a proxy active/inactive; opt back into the legacy test-and-set with `PROXY_HEALTH_AUTO_DEACTIVATE=true`). (2) Adds a **bulk enable/disable** proxies endpoint + toolbar action (`POST /api/settings/proxies/batch-activate`) so an operator can re-activate proxies in one click. Regression guard: `tests/unit/proxy-health-6246.test.ts`. (thanks @tenshiak)
+
+- **chatcore (tools): stop the default 128-tool cap from silently dropping opencode's `task`/MCP tools.** opencode (used as an MCP/agent host) sends a large tool list; when it exceeds the speculative `MAX_TOOLS_LIMIT` (128) default, `truncateToolList` did a blind `tools.slice(0, 128)`, dropping every tool past index 128 — including opencode's built-in `task` tool (subagent launch) and many MCP tools, so models routed through OmniRoute could no longer spawn subagents or reach part of their tools. The cap exists to avoid upstream `400`s for providers with real hard limits (e.g. grok-cli 200), so it is kept for those: detection of the opencode client (`isOpencodeClient` — any `x-opencode-*` header, or `opencode` in the user-agent) now only bypasses the **speculative 128 default**, never a known provider ceiling. Precedence is explicit — a proactive/detected provider limit always truncates (even for opencode); otherwise opencode forwards its full tool list; otherwise the unchanged 128 default applies to every other client. Refactors `getEffectiveToolLimit` into `getKnownToolLimit(provider) ?? DEFAULT_LIMIT` (byte-identical for existing callers) and fixes a cosmetic debug-log that reported the truncated count instead of the original. Regression guard: `tests/unit/tool-limit-detector.test.ts`.
+
+- **fix(mitm):** the macOS MITM-cert install check now matches the system keychain again. `security find-certificate -a -Z` prints the SHA-1 as a colon-less hex string, but the installed-check compared it against `getCertFingerprint()`'s colon-separated form, so the substring match never hit — the cert was reported as not-installed and re-prompted for the sudo install on every run. Fingerprints are now normalized (colons stripped, upper-cased) on both sides via the extracted `macCertOutputHasFingerprint` helper. Regression guard: `tests/unit/mitm-cert-mac-fingerprint.test.ts`. ([#6204](https://github.com/diegosouzapw/OmniRoute/pull/6204), closes [#6134](https://github.com/diegosouzapw/OmniRoute/issues/6134) — thanks @rianonehub)
+
+- **fix(api):** `/v1/messages/count_tokens` now counts `tool_use`, `tool_result` and `thinking` content blocks (and array-form `system` prompts) in the local-estimation path, instead of only `text`. Real agentic conversations keep ~95% of their tokens inside tool results; the previous estimate returned near-zero for them, which silently broke Claude Code's auto-compaction (context grew past the window with no compaction until the upstream API rejected the request). The real provider-side count path is unchanged. Regression guard: `tests/unit/messages-count-tokens-route.test.ts`. ([#6221](https://github.com/diegosouzapw/OmniRoute/pull/6221) — thanks @luweiCN)
+
+- **fix(antigravity):** strip a trailing assistant prefill turn for Vertex Claude models to avoid upstream 400s ([#6114](https://github.com/diegosouzapw/OmniRoute/pull/6114)). Regression guard: `tests/unit/antigravity-claude-prefill-strip.test.ts`. (thanks @anki1kr)
+
+- **fix(security):** the mutable cloud-agent routes (`/api/cloud/credentials/update`, `/api/cloud/models/alias`) now require management auth instead of being treated as public. They were classified as public API routes, so a request without management credentials could update stored cloud-agent credentials and model aliases. They are removed from the public-route set, classified as management routes in the authz pipeline, and gated by `requireManagementAuth`; cloud **read**/auth routes stay public. Regression guards: `tests/unit/cloud-write-auth.test.ts`, `tests/unit/authz/classify.test.ts`, `tests/unit/public-api-routes.test.ts`. ([#6233](https://github.com/diegosouzapw/OmniRoute/pull/6233) — thanks @vittoroliveira-dev)
+
+- **refactor(dashboard):** extract the onboarding-wizard "Open provider details" link target into a pure, unit-tested `buildProviderDetailsHref(connection)` helper. The wizard already routes by `connection.id` (the node UUID) rather than the provider category slug (#6144/#6145); this hardens that behavior behind a tested helper that guards a missing id/connection. Regression guard: `tests/unit/provider-onboarding-href.test.ts`. ([#6166](https://github.com/diegosouzapw/OmniRoute/pull/6166) — thanks @KooshaPari)
+
+- **fix(security):** the doubao synthetic device-id generator now derives its digits via an unbiased crypto-random draw (rejection sampling over `crypto.randomBytes()`) instead of a `% 10` reduction, closing a CodeQL `js/biased-cryptographic-random` finding.
+- **fix(agentSkills):** the GitHub-skills generator now resolves `outputDir` to an absolute path before writing, fixing a regression introduced by #6366 (relative-to-cwd base path) that could write generated skill files to the wrong directory.
+- **fix(security):** `/api/keys/{id}/devices` now checks the HTTP method before auth/validation, returning a `405` for non-GET/DELETE verbs instead of a misleading `401`/`500` (closes a `dast-smoke` QUERY-method finding).
+- **fix(quality):** clear the last 2 heavy quality-gate reds on the release tip (cycle pre-flight).
+- **fix(mitm):** the test suite and CI can never mutate the OS trust store — `OMNIROUTE_SKIP_SYSTEM_TRUST=1` is set globally for tests/CI so `installCert`/`uninstallCert`/`installTproxyCa` skip the privileged OS dispatch ([#6310](https://github.com/diegosouzapw/OmniRoute/pull/6310); full detail is under the [3.8.45] section below — this branch received it via the parallel-cycle sync-back).
+- **fix(api):** `POST /api/github-skills` now Zod-validates its request body; documented the new quality-gate env vars and pinned the merge-integrity GitHub Actions to a commit SHA.
+- **fix(skills):** generate the missing `omni-github-skills` registry entry and align the agent-skills catalog-count tests (follow-up to #6186).
+- **fix(quality):** clear the cycle's 11 net-new ESLint errors and make `validate-release-green` suppressions-aware.
+
+### 📝 Maintenance
+
+- **i18n(it):** add 118 missing Italian (`it`) translations (net-additive — no existing keys dropped, valid JSON), improving Italian UI coverage. ([#6212](https://github.com/diegosouzapw/OmniRoute/pull/6212) — thanks @serverless83)
+- **chore(providers):** remove deprecated **MiMo V2** model entries from the catalogs (xiaomi-mimo, opencode-go, zenmux-free, audio TTS) — the upstream V2 line is superseded by MiMo V2.5; drops `mimo-v2-tts`, `mimo-v2-pro`, `mimo-v2-omni`, `mimo-v2-flash`, `mimo-v2-flash-free` and realigns the provider-catalog tests. ([#6248](https://github.com/diegosouzapw/OmniRoute/pull/6248) — thanks @backryun)
+- **chore(release):** ~50 commits on this branch are v3.8.45 pre-flight/hardening fixes and CI-perf work that landed here via the parallel-cycle sync-back (`sync-next-cycle.mjs`, Hard Rule #21) **after** the `v3.8.45` git tag was cut, and are already fully documented under the **[3.8.45]** section below — listed here only so the per-cycle commit-coverage check (`npm run release:uncovered`) doesn't flag them as gaps. Provider/catalog/UX/backend fixes: #6041, #6078, #6108, #6135, #6148, #6149, #6154, #6158, #6161, #6162, #6163, #6164, #6165, #6170, #6177, #6178, #6181, #6186, #6187, #6191, #6193, #6194, #6195, #6200, #6205, #6208, #6209, #6211, #6213, #6223, #6224, #6225, #6226, #6227, #6228, #6229, #6230, #6235, #6291, #6292. CI/release-pipeline work: #6167, #6203, #6214, #6215, #6218, #6273, #6275, #6283, #6284, #6285, #6300, #6305.
+- **chore(release):** additional zero-ref release-cycle plumbing on this branch, kept out of `release:uncovered` on purpose (no `#N` in the commit subject to cite): opening the v3.8.46 cycle, opening/closing the v3.8.45 cycle, the finalized [3.8.45] CHANGELOG i18n sync-back to 42 mirrors, the v3.8.45 cognitive/cyclomatic and file-size drift rebaselines, ESLint stale-suppression pruning (4,273 → 4,233), and clearing test-masking/docs-all pre-flight reds for v3.8.45.
+
+### ⚡ Performance & Infrastructure
+
+- **perf(release-green):** the pre-flight validator (`scripts/quality/validate-release-green.mjs`) now runs its 4 slow suites (unit / vitest / integration / pack-artifact) **concurrently** via `Promise.all` — pre-flight wall time drops from ~the sum of the suites to ~the slowest one (~30min saved per round; Phase 0 was the nº1 bottleneck of the v3.8.45 release benchmark, 2h54 of 6h34 e2e). Guard: `tests/unit/validate-release-green.test.ts` ("runs the slow suites CONCURRENTLY"). ([#6319](https://github.com/diegosouzapw/OmniRoute/pull/6319))
+- **fix(ci):** `scripts/release/sync-next-cycle.mjs` — two defects found live in its first production run (v3.8.45 Phase 5): (1) the `git()` helper's default 1 MiB `maxBuffer` crashed with `ENOBUFS` on `git show origin/main:CHANGELOG.md` (the CHANGELOG alone is >1 MiB) — widened to 64 MiB; (2) the i18n resync only propagated the `[NEXT]` (TBD) section, leaving the just-shipped finalized section as "— TBD" in all 42 mirrors — it now also syncs `[prevVersion]` bounded by the heading below it (new exported pure helper `versionAfter`). Guards: +5 tests in `tests/unit/sync-next-cycle.test.ts` (8/8). ([#6327](https://github.com/diegosouzapw/OmniRoute/pull/6327))
+- **test(ci):** concurrency-sensitive flaky tests are **quarantined into a serial pass** (`tests/unit/serial/`, `--test-concurrency=1`, appended to every unit runner incl. sharded variants — the serial pass is sharded too so concurrent shard jobs never self-collide). Initial set: `glm-coding-plan-monthly-3580`, `quota-division-blocks`, `provider-health-autopilot`, `combo-health-autopilot` — the class behind the ~28min CI wedges/re-runs (two live 1h+ wedges cancelled during this PR's own validation). Discovery + TIA gates track the new glob; systemic root cause (async logger writing after teardown) tracked in [#6360](https://github.com/diegosouzapw/OmniRoute/issues/6360). Guard: `tests/unit/test-serial-quarantine.test.ts` (4). ([#6347](https://github.com/diegosouzapw/OmniRoute/pull/6347))
+
+### 🙌 Contributors
+
+Thanks to everyone whose work landed in v3.8.46:
+
+| Contributor                                                    | PRs / Issues           |
+| -------------------------------------------------------------- | ---------------------- |
+| [@2220258345](https://github.com/2220258345)                   | direct commit / report |
+| [@abdofallah](https://github.com/abdofallah)                   | direct commit / report |
+| [@adentdk](https://github.com/adentdk)                         | direct commit / report |
+| [@aleksesipenko](https://github.com/aleksesipenko)             | direct commit / report |
+| [@anki1kr](https://github.com/anki1kr)                         | direct commit / report |
+| [@anungma](https://github.com/anungma)                         | direct commit / report |
+| [@arpicato](https://github.com/arpicato)                       | direct commit / report |
+| [@backryun](https://github.com/backryun)                       | #6248                  |
+| [@binsarjr](https://github.com/binsarjr)                       | direct commit / report |
+| [@brightfiscalband](https://github.com/brightfiscalband)       | direct commit / report |
+| [@chirag127](https://github.com/chirag127)                     | #6501, #6506           |
+| [@developerjillur](https://github.com/developerjillur)         | direct commit / report |
+| [@dilneiss](https://github.com/dilneiss)                       | #6499                  |
+| [@dtybnrj](https://github.com/dtybnrj)                         | direct commit / report |
+| [@Forcerecon](https://github.com/Forcerecon)                   | direct commit / report |
+| [@hao3039032](https://github.com/hao3039032)                   | direct commit / report |
+| [@Iammilansoni](https://github.com/Iammilansoni)               | #6150, #6245           |
+| [@jmengit](https://github.com/jmengit)                         | direct commit / report |
+| [@jordansilly77-stack](https://github.com/jordansilly77-stack) | direct commit / report |
+| [@JxnLexn](https://github.com/JxnLexn)                         | direct commit / report |
+| [@KooshaPari](https://github.com/KooshaPari)                   | #6166                  |
+| [@loopyd](https://github.com/loopyd)                           | direct commit / report |
+| [@luweiCN](https://github.com/luweiCN)                         | #6221                  |
+| [@makcimbx](https://github.com/makcimbx)                       | direct commit / report |
+| [@muflifadla38](https://github.com/muflifadla38)               | direct commit / report |
+| [@newnol](https://github.com/newnol)                           | direct commit / report |
+| [@ofekbetzalel](https://github.com/ofekbetzalel)               | direct commit / report |
+| [@ohahe52-dot](https://github.com/ohahe52-dot)                 | direct commit / report |
+| [@phidinhmanh](https://github.com/phidinhmanh)                 | direct commit / report |
+| [@powellnorma](https://github.com/powellnorma)                 | direct commit / report |
+| [@qpeyba](https://github.com/qpeyba)                           | direct commit / report |
+| [@RaviTharuma](https://github.com/RaviTharuma)                 | direct commit / report |
+| [@RCrushMe](https://github.com/RCrushMe)                       | direct commit / report |
+| [@rianonehub](https://github.com/rianonehub)                   | #6134, #6204           |
+| [@serverless83](https://github.com/serverless83)               | #6212                  |
+| [@swingtempo](https://github.com/swingtempo)                   | direct commit / report |
+| [@tenshiak](https://github.com/tenshiak)                       | direct commit / report |
+| [@ThongAccount](https://github.com/ThongAccount)               | direct commit / report |
+| [@UnrealAryan](https://github.com/UnrealAryan)                 | direct commit / report |
+| [@vinayakkulkarni](https://github.com/vinayakkulkarni)         | direct commit / report |
+| [@vittoroliveira-dev](https://github.com/vittoroliveira-dev)   | #6233                  |
+| [@warelik](https://github.com/warelik)                         | direct commit / report |
+| [@xxy9468615](https://github.com/xxy9468615)                   | direct commit / report |
+| [@xz-dev](https://github.com/xz-dev)                           | direct commit / report |
+| [@yanpaing007](https://github.com/yanpaing007)                 | direct commit / report |
+| [@diegosouzapw](https://github.com/diegosouzapw)               | maintainer             |
+
+---
+
 ## [3.8.45] — 2026-07-06
 
 ### ✨ New Features

docs/i18n/bg/CHANGELOG.md

@@ -6,6 +6,213 @@
 
 ## [3.8.31] — 2026-06-20
 
+## [3.8.46] — TBD
+
+### ✨ New Features
+
+- **feat(sse):** **hide paid-only models from `auto/*` routing** when `hidePaidModels` is on ([#6512](https://github.com/diegosouzapw/OmniRoute/issues/6512)) — follow-up to #6328/#6495. PR #6495 hid paid-only models from the `GET /v1/models` listing, but `auto/*` combos (`auto/best-coding`, `auto/glm`, …) could still pick a paid-only backend into their candidate pool → a 402/403 at request time. `createVirtualAutoCombo` now filters the candidate pool through the new pure `open-sse/services/autoCombo/paidModelFilter.ts` (`filterPaidOnlyCandidates`), applying the same free-model predicate #6495 uses in `catalog.ts` (`providerHasFreeModels(provider) && isFreeModel(provider, {id})`) whenever `settings.hidePaidModels === true`. Applied before the category/tier/family narrowing, so it covers every `auto/*` combo; an all-paid pool degrades to the existing graceful empty-pool path. **Opt-in — default OFF leaves the pool unchanged** (identity). Regression guard: `tests/unit/autoCombo/paid-model-filter-6512.test.ts` (4, incl. the default-off identity guard).
+- **feat(sse):** **provider-family auto combos** — `auto/glm`, `auto/minimax`, `auto/mimo`, `auto/zai`, `auto/gemma`, `auto/llama`, `auto/gemini` ([#6453](https://github.com/diegosouzapw/OmniRoute/issues/6453)) — new routable ids that materialize an on-demand virtual combo spanning whatever installed backends currently expose that model family, degrading gracefully as backends rotate. A new pure `open-sse/services/autoCombo/modelFamily.ts` (`detectModelFamily`) classifies by model-id prefix for six families; `zai` is instead resolved by provider id (z.ai's hosted API serves the same `glm-*` model ids as every other GLM backend, so `auto/zai` means "route to my z.ai backend specifically" vs `auto/glm`'s "any connected GLM backend"). Reuses the existing `createVirtualAutoCombo` on-demand materialization path (no DB writes) and the `/v1/models` catalog advertising loop. Regression guard: `tests/unit/autoCombo/provider-family-combos.test.ts` (11).
+- **feat(proxy):** native **proxy-pool round-robin / egress IP rotation** ([#6365](https://github.com/diegosouzapw/OmniRoute/issues/6365)) — a scope (global / provider / account) can now hold **multiple** proxies as a pool with a rotation strategy, so outbound requests cycle their egress IP instead of pinning one proxy per scope. Migration `117_proxy_pool_rotation.sql` lifts the `UNIQUE(scope, scope_id)` constraint (rebuild via the canonical rename/copy/drop; existing single assignments become 1-element pools) and adds a `proxy_scope_rotation` companion table holding the per-scope strategy + a persisted monotonic round-robin cursor. Strategies: `round-robin` (default, monotonic cursor — never `Math.random`), `random`, and `sticky-per-N-min`. Resolution (`resolveProxyForScopeFromRegistry` / `resolveProxyForConnectionFromRegistry`) now fetches the alive, position-ordered candidate set (unchanged `PROXY_ALIVE_PREDICATE`) and applies the strategy; an empty / all-dead pool still returns `null` — the #6246 fail-closed guard is **untouched** (never falls through to direct egress). Backend + DB only; dashboard pool-builder UI is a follow-up. Regression guard: `tests/unit/proxy-pool-rotation-6365.test.ts` (8, incl. fail-closed + backward-compat).
+- **feat(providers):** end-to-end **tool/function calling on the native Gemini `/v1beta` endpoint** ([#6222](https://github.com/diegosouzapw/OmniRoute/issues/6222)) — both directions of the Gemini↔OpenAI conversion now preserve tool data (previously silently dropped). Request side: `convertGeminiToInternal` (extracted to its own testable module) maps `tools[].functionDeclarations` → OpenAI `tools`, prior `functionCall` parts → assistant `tool_calls`, and `functionResponse` parts → `tool`-role messages. Response side: `convertOpenAIResponseToGemini` emits `parts[].functionCall {name,args}` from `message.tool_calls`, and the streaming `openAIChunkToGeminiChunk` accumulates fragmented `tool_calls` deltas by index into complete `functionCall` parts. The non-Gemini client paths (Claude, OpenAI-Responses) already preserved tool calls — this closes the gap specific to the native Gemini surface. Regression guard: `tests/unit/v1beta-gemini-tool-calling-6222.test.ts` (6, incl. a streaming SSE round-trip).
+- **feat(providers):** copilot-m365-web **enterprise / work tier** support ([#6334](https://github.com/diegosouzapw/OmniRoute/issues/6334)) — mirrors the EDU-tier pattern (#6210): `M365ConnectionParams` gains an `agent` field, a new opt-in `M365_ENTERPRISE_OVERRIDES` preset (`agent=work`, `scenario=officeweb`, `licenseType=Premium`) applies via `providerSpecificData.tier="enterprise"` (alias `"work"`), and `agent` is also overridable directly via `providerSpecificData.agent`. `buildWsUrl` was hardcoding `agent="web"` (the one enterprise-distinguishing param with no override path), so a Premium work account handshook then returned an empty stream. The individual and EDU paths are untouched. Kilo's dup flag vs #6210 (EDU tier) was a false positive — different tier. Regression guard: `tests/unit/copilot-m365-enterprise-6334.test.ts` (7). End-to-end confirmation on a real Premium work account is a live-VPS validation follow-up (Hard Rule #18). (thanks @Forcerecon)
+- **feat(api):** standardized, provider-agnostic **`effort` + `thinking` request params** ([#6241](https://github.com/diegosouzapw/OmniRoute/issues/6241)) — a thin standardization layer over the existing mature per-provider reasoning plumbing (no provider mapper touched). `providerChatCompletionSchema` gains a canonical `effort` (reusing the shared `none/low/medium/high/xhigh` vocabulary — the UI tiers `extra`/`max` collapse onto `xhigh`) and a boolean `thinking`. A pure `normalizeReasoningRequest` (wired once in `src/sse/handlers/chat.ts`, before any reasoning field is read) folds them onto the fields the translators already consume (`reasoning_effort` / `reasoning.effort` / `thinking`), so they fan out to Anthropic / Gemini / xAI / Responses — an explicit client `reasoning_effort` / object-shaped `thinking` always wins (backward-compatible). `/models` additively exposes `supportsThinking` + `effort_tiers` so the frontend can render the toggles (UI component is a follow-up). Regression guard: `tests/unit/effort-thinking-standardization-6241.test.ts` (12). (thanks @Iammilansoni, @shabeer)
+- **feat(combo):** new **`pipeline` (sequential) combo strategy** ([#6297](https://github.com/diegosouzapw/OmniRoute/issues/6297)) — the 18th routing strategy runs targets **in order**, threading each step's output into the next step's input, with an optional per-step `prompt` (system instruction); only the final step's response is returned. Distinct from `fusion` (parallel fan-out + judge). Implemented as a self-contained `open-sse/services/pipeline.ts` (sibling to `fusion.ts`), dispatched from `combo.ts`; the step list reuses `combo.models` order and reads an optional `prompt` off each target (backward-compatible — ignored by every other strategy). Intermediate steps run non-streaming with tools stripped (complete prose to thread forward); the final step keeps the client's `stream` flag + tools. A failing/empty/unparseable intermediate step fails the whole pipeline explicitly via a sanitized error (never silently swallowed). Kilo's dup flag vs #563 was a false positive (that's model→chain selection; this is a sequential chain). Regression guard: `tests/unit/combo-pipeline-strategy.test.ts` (5). (thanks @ofekbetzalel)
+- **feat(ci):** `check:test-masking` now flags **inline-reimplemented prod conditions** ([#6348](https://github.com/diegosouzapw/OmniRoute/issues/6348)) — a new **report-only** subcheck (v2, 6A.10 family) catches the wrong-shape contract test: a test that recomputes the condition under test inline instead of importing/exercising the real function (the #6216 class, where `=== 500` → `>= 500` stayed green because the test re-implemented the branch). For each added/modified test file it warns when the file textually duplicates a ≥3-token conditional from a production file touched in the same PR **and** does not import the symbol/module owning it, via a pure, fixture-tested `findReimplementedConditions()` with an allowlist mirroring `assertReductionAllowlist`. Report-only for now (does not fail the gate) — to be promoted to blocking after a triage cycle. Regression guard: `tests/unit/check-test-masking.test.ts` (45).
+- **feat(sse):** per-connection routing override (native vs CLIProxyAPI) ([#6339](https://github.com/diegosouzapw/OmniRoute/issues/6339)) — the previously-dead `isCliproxyapiDeepModeEnabled` helper is now wired into `resolveExecutorWithProxy`: a single connection can opt itself into the CLIProxyAPI passthrough executor via `providerSpecificData.cliproxyapiMode="claude-native"`, with precedence **connection override > provider `upstream_proxy_config` mode > default**. `resolveExecutorWithProxy` now receives the resolved connection's `providerSpecificData` (threaded from `chatCore.ts`), so one connection can deep-route while the provider's other connections stay native — no DB schema change (the toggle rides in `providerSpecificData`). Also resolves the same-provider mixing ask in #6340. Regression guard: `tests/unit/chatcore-executor-proxy.test.ts` (9). (thanks @RaviTharuma)
+- **feat(dashboard):** "Add session cookie" modal now shows a prominent **"Open ‹host› →"** link to the provider's own site ([#6268](https://github.com/diegosouzapw/OmniRoute/issues/6268)) — every `-web` cookie-session provider (chatgpt-web, claude-web, gemini-web, kimi-web, lmarena, qwen-web, m365-copilot-web, …) renders a one-click external link (opening the provider's login/home page in a new tab) so operators no longer tab away to retype the URL mid-setup. The host resolves from a pure, unit-tested `resolveWebProviderHost()` (prefers `WEB_COOKIE_PROVIDERS[id].website`, falls back to the registry `baseUrl` origin); non-web providers render exactly as before. Kilo's dup flag vs #6265 (modal-too-small-on-1080p) was a false positive — distinct concern. Regression guard: `tests/unit/resolve-web-provider-host.test.ts` (5). (thanks @chirag127)
+- **feat(providers):** add **DigitalOcean** AI (serverless inference) as an OpenAI-compatible API-key provider ([#6373](https://github.com/diegosouzapw/OmniRoute/pull/6373)) — base `https://inference.do-ai.run/v1`, wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/digitalocean/`, `src/shared/constants/providers/apikey/inference-hosts.ts`). Regression guard: `tests/unit/digitalocean-provider.test.ts`. (thanks @newnol)
+- **feat(providers):** add **Huancheng Public API** (`hcnsec`) as an OpenAI-compatible regional provider ([#6410](https://github.com/diegosouzapw/OmniRoute/pull/6410)) — Xinjiang Huancheng Cybersecurity's public LLM platform (base `https://api.hcnsec.cn/v1`, free credits via daily check-ins), wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/hcnsec/`, `src/shared/constants/providers/apikey/regional.ts`). Regression guard: `tests/unit/hcnsec-provider.test.ts`. (thanks @UnrealAryan)
+- **feat(dashboard):** the web-session credential guide now shows an **"Open {host}" link** ([#6316](https://github.com/diegosouzapw/OmniRoute/pull/6316)) to the provider's sign-in site (derived from the provider `website` via `getProviderWebsiteHost`), so you can jump straight to the page where the cookie/session must be captured. Regression guard: `tests/unit/web-session-provider-link-6316.test.ts`. (thanks @jordansilly77-stack)
+- **feat(cerebras):** add the **Gemma 4 31B** model (`gemma-4-31b`) to the Cerebras registry + pricing table ([#6331](https://github.com/diegosouzapw/OmniRoute/pull/6331)). Regression guard extends `tests/unit/t28-model-catalog-updates.test.ts`. (thanks @backryun)
+- **feat(providers):** add **Yuanbao (web)** as a cookie-session provider ([#6196](https://github.com/diegosouzapw/OmniRoute/issues/6196)) — `yuanbao-web` (Tencent Yuanbao, `yuanbao.tencent.com`) with cookie-only auth (`hy_user`/`hy_token` + public agent id), SSE→OpenAI translation incl. `reasoning_content`, exposing DeepSeek V3/R1 + Hunyuan / Hunyuan-T1. Regression guard: `tests/unit/providers-yuanbao-web.test.ts`. `together-web` was **deferred** (no verifiable web-session endpoint — needs a captured request) and `huggingchat-web` **dropped** (the existing `huggingchat` already is a web-cookie provider). (thanks @chirag127)
+- **feat(providers):** route the built-in **agentrouter** through the dynamic Claude-Code wire image ([#6056](https://github.com/diegosouzapw/OmniRoute/issues/6056)) — a small static allow-set (`CC_WIRE_IMAGE_BUILTINS` in `open-sse/services/ccWireImageBuiltins.ts`), consulted by `isClaudeCodeCompatible` / `isClaudeCodeCompatibleProvider` / `applyFingerprint`, makes agentrouter adopt the CC wire-image headers + fingerprint **while guarding the CC baseUrl/auth branches** so it keeps its own registry `baseUrl` and `x-api-key` auth. Regression guard: `tests/unit/agentrouter-cc-wire-image.test.ts` (asserts the wire image is applied AND agentrouter's baseUrl/auth are preserved). Live WAF-acceptance against agentrouter.org is a VPS validation follow-up (Hard Rule #18).
+- **feat(providers):** **bulk-add API keys for Cloudflare Workers AI** ([#6174](https://github.com/diegosouzapw/OmniRoute/issues/6174)) — `cloudflare-ai` is removed from the bulk-add exclusion list and the bulk parser gains a 3-field `name|accountId|apiKey` mode; the bulk route now builds a **per-entry** `providerSpecificData` so each key carries its own `accountId` (fixing the previous shared-object reuse), and both the create + key-validation paths receive it. Regression guard: `tests/unit/bulk-api-key-parser-cloudflare.test.ts`. (thanks @muflifadla38)
+- **feat(dashboard):** routing/settings UX clarity ([#6147](https://github.com/diegosouzapw/OmniRoute/issues/6147)) — (1) weighted combos show the **effective routing share %** next to each weight when weights don't sum to 100 (`WeightTotalBar.tsx`); (2) the status widget's user-facing **"Cloud Sync" label is renamed** to "Remote Settings Sync" (`CloudSyncStatus.tsx`; internal ids/state untouched); (3) built-in providers gain an **opt-in advanced base-URL override** (`isBaseUrlOverrideEligibleProvider`, hidden behind an "Advanced" toggle, reusing the existing `providerSpecificData.baseUrl` persistence — not globally widened). Regression guard: `tests/unit/routing-settings-ux-6147.test.ts`.
+- **feat(combo):** add an option to **disable session stickiness**, per-combo or globally — round-robin / random combos can rotate to a different connection on every request instead of pinning a whole conversation to one connection by its first-message hash. Resolution precedence per-combo `config.disableSessionStickiness` → global `settings.disableSessionStickiness` → default `false` (preserves the #3825 prompt-cache/504 fix); gates **both** stickiness call sites in `open-sse/services/combo.ts`. Exposed as a global toggle (Combo Defaults) and a per-combo Inherit/on/off control. ([#6168](https://github.com/diegosouzapw/OmniRoute/issues/6168)) Regression guard: `tests/unit/combo-disable-session-stickiness.test.ts`. (thanks @RCrushMe)
+- **feat(docker):** add the `OMNIROUTE_NO_SUDO` env flag for root-less / user-namespaced deployments — the MITM cert-trust command path (`resolveSudoSpawn` in `src/mitm/systemCommands.ts`) now strips the leading `sudo` when the flag is truthy, in addition to the existing root / sudo-missing cases, so the Proxy Agent runs without `sudo` (the operator trusts the CA manually, e.g. via `NODE_EXTRA_CA_CERTS`). Argv-array `spawn` preserved — no shell interpolation (Hard Rule #13). ([#6122](https://github.com/diegosouzapw/OmniRoute/issues/6122)) Regression guard: `tests/unit/mitm-systemCommands-no-sudo.test.ts`. (thanks @powellnorma)
+- **feat(providers):** add **Requesty** as an OpenAI-compatible gateway provider (BYOK, base `https://router.requesty.ai/v1`, ~200 free requests/day) — wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/requesty/`, `src/shared/constants/providers/apikey/gateways.ts`). ([#6120](https://github.com/diegosouzapw/OmniRoute/issues/6120)) Regression guard: `tests/unit/requesty-provider.test.ts`. (thanks @chirag127)
+- **feat(dashboard):** add **configured-only / available-only filters** to the Free Provider Rankings page ([#6150](https://github.com/diegosouzapw/OmniRoute/issues/6150)) — hide providers you haven't configured, or whose connections are all rate-limited / out of quota, via server-side query params (`?configuredOnly` / `?availableOnly` on `GET /api/free-provider-rankings`) backed by a testable lib helper reusing the in-process connection state (no Redis). Both filters default off, so the default view is unchanged; this supersedes the earlier client-side "Configured Only" toggle (#6245) with an available-only dimension and unit-tested logic. Regression guard: `tests/unit/freeProviderRankings-filters.test.ts`.
+
+### 🔧 Bug Fixes
+
+- **fix(dashboard):** adding a second API key connection for the same provider no longer silently overwrites the first — the Add-API-key modal now derives a unique default connection name (`main`, then `main-2`, `main-3`, …) so the backend name-based upsert can't collide (#6499 — thanks @dilneiss).
+- **fix(compression):** the session-dedup engine now also deduplicates a large multi-line block repeated **within a single message** (intra-message dedup), not just across turns; the compression-preview API surfaces a `fallbackReason`, and the fusion panel reports how many models were rate-limited vs failed on total-panel failure (#6501 — thanks @chirag127).
+- **fix(compression):** a stacked-pipeline step naming an unregistered engine now surfaces a `validationErrors` entry instead of silently no-op'ing, so misconfigured pipelines are visible in the preview API (#6506 — thanks @chirag127).
+- **feat(usage):** add a **Codex reset-credit redemption** flow to the Provider Limits UI ([#6361](https://github.com/diegosouzapw/OmniRoute/pull/6361)) — a `useCodexResetCreditRedemption` hook + `/api/usage/codex-reset-credit` route + `codexResetCredits` lib let you redeem banked Codex reset credits from the quota card. Regression guard: `tests/unit/codex-reset-credits.test.ts`. (thanks @JxnLexn)
+- **feat(glm):** add **team-plan quota settings** for `glm-cn` connections ([#6351](https://github.com/diegosouzapw/OmniRoute/pull/6351)) — a dedicated `GlmTeamQuotaFields` form section (team quota id / limits) threaded through the Add/Edit connection modals, persisted via `providerSpecificData`, with the GLM usage service reading the team quota. Regression guards: `tests/unit/glm-team-quota.test.ts`, `provider-specific-data-schema.test.ts`. (thanks @hao3039032)
+- **feat(providers):** add **TinyFish** web-fetch/search support ([#6349](https://github.com/diegosouzapw/OmniRoute/pull/6349)) — a `tinyfish-fetch` executor + `/v1/web/fetch` route + MCP web-fetch tool, registered as a specialty-media provider with request-validation and a search-provider catalog entry. Regression guards: `tests/unit/executor-tinyfish-fetch.test.ts`, `web-fetch-handler.test.ts`, `mcp-web-fetch-tool.test.ts`, `provider-validation-tinyfish.test.ts`. (thanks @dtybnrj)
+- **fix(cli):** `omniroute launch-codex` now spawns `codex.cmd` through a shell on Windows (the npm `.cmd` shim is unresolvable by bare `spawn` → ENOENT), mirroring the qodercli Windows fix (#6263) ([#6312](https://github.com/diegosouzapw/OmniRoute/pull/6312)). Regression guard: `tests/unit/launch-codex-windows-spawn-6312.test.ts`. (thanks @swingtempo)
+- **fix(codex):** isolate the **Spark** quota from the shared Codex quota and stabilize the quota UI ordering / hydration so per-scope limits render consistently ([#6336](https://github.com/diegosouzapw/OmniRoute/pull/6336)). Regression guards: `tests/unit/codex-quota-selection-hydration.test.ts`, `provider-limits-ui.test.ts` + 3 more. (thanks @xz-dev)
+- **feat(api):** add a `hidePaidModels` setting that filters paid-only models out of the `/v1/models` catalog. Regression guard: `tests/unit/models-catalog-hide-paid.test.ts`. (thanks @chirag127)
+- **fix(api-manager):** the fallback model picker now preserves combos instead of dropping them when a primary model is unavailable ([#6443](https://github.com/diegosouzapw/OmniRoute/pull/6443)). Regression guard: `tests/unit/api-manager-page-static.test.ts`. (thanks @jmengit)
+- **fix(providers):** recoverable Antigravity / Cloud-Code (Gemini Code Assist) `403` responses ([#6452](https://github.com/diegosouzapw/OmniRoute/pull/6452)) are now classified as a retryable project-config error instead of a terminal account ban, so a fixable project/API-disabled 403 no longer forces a ~1-year cooldown / full OAuth reconnect. Regression guard: `tests/unit/errorclassifier-antigravity-403.test.ts`. (thanks @developerjillur)
+- **fix(mitm):** `sanitizeHeaders` now redacts `Set-Cookie` response headers so upstream session cookies never leak into logs / diagnostics ([#6451](https://github.com/diegosouzapw/OmniRoute/pull/6451)). Regression guard: `tests/unit/mitm-sanitize-headers.test.ts`. (thanks @developerjillur)
+- **fix(api):** `/api/compression/preview` now accepts `mode: "caveman"` and correctly handles stacked / zero-compression previews ([#6425](https://github.com/diegosouzapw/OmniRoute/issues/6425)). Regression guard: `tests/unit/api/compression-preview-caveman-and-stacked-6425.test.ts`. (thanks @chirag127)
+- **feat(providers):** add **Zed** hosted LLM aggregator as a native-app provider ([#6118](https://github.com/diegosouzapw/OmniRoute/pull/6118)) — OAuth sign-in via the Zed hosted flow, registered through the shared provider registry + executor. Regression guards: `tests/unit/zed-oauth-provider.test.ts`, `zed-import-utils.test.ts`, `zed-docker-detect.test.ts`, `mitm-handler-zed.test.ts`. VPS-validated via live operator login (Hard Rule #18).
+- **fix(oauth):** the Kiro SSO-cache auto-import now **preserves the IDC region** — cross-region Amazon Q / Kiro profiles imported from the SSO cache are no longer collapsed to the default region ([#6113](https://github.com/diegosouzapw/OmniRoute/pull/6113)). Regression guard: `tests/unit/kiro-auto-import-idc-2059.test.ts`. VPS-validated via live operator login (Hard Rule #18).
+- **fix(dashboard):** passthrough model aliases no longer collide when two namespaced model ids share a last segment (port from 9router#1850, [#6431](https://github.com/diegosouzapw/OmniRoute/pull/6431)). `enx/gpt-5.5` and `enx/codebuddy/gpt-5.5` both auto-generated the alias `gpt-5.5`, so the second model could never be added (the UI just alerted "alias already exists"). Aliases are now disambiguated deterministically — bare last segment when free, then parent-qualified (`codebuddy-gpt-5.5`), then a numeric suffix — while re-adding the exact same model id is still blocked. Regression guard: `tests/unit/passthrough-alias-1850.test.ts`. (thanks @arpicato)
+- **fix(translator):** preserve a Gemini `functionResponse` co-located with other parts (another `functionCall`, or trailing `text`) in the same content when translating **Gemini → OpenAI** ([#6376](https://github.com/diegosouzapw/OmniRoute/pull/6376)). `convertGeminiContent()` early-returned the tool message on the first `functionResponse` part, dropping any co-located parts; such contents are now pre-split (one tool message per `functionResponse`, emitted first, plus one message for the remaining parts). Regression guard: `tests/unit/gemini-to-openai-function-response.test.ts`. (thanks @warelik)
+- **fix(headroom):** detect a python interpreter managed by **mise / pyenv / asdf / conda** (port from 9router#2353, [#6382](https://github.com/diegosouzapw/OmniRoute/pull/6382)). Headroom's python probe (`src/lib/headroom/detect.ts`) searched a hardcoded `PATH`, but version managers expose their interpreters via shim dirs that only join `PATH` through interactive-shell activation — which the non-interactive server never runs, so a managed python (≥3.10) was invisible and Headroom reported it missing. The search path now prepends the well-known shim/bin dirs (`~/.local/share/mise/shims`, `~/.pyenv/shims`, `~/.asdf/shims`, `$CONDA_PREFIX/bin`, `~/.local/bin`, respecting `MISE_DATA_DIR`/`PYENV_ROOT`/`ASDF_DATA_DIR` when set), and a new `HEADROOM_PYTHON` env override lets operators point straight at their interpreter (mirroring `HEADROOM_URL`). Still shell-free (`execFileSync`). Regression guard: `tests/unit/headroom-detect.test.ts` (5). (thanks @loopyd)
+- **fix(executors):** strip the OpenAI-Codex/Claude-CLI `client_metadata` passthrough field for **NVIDIA** requests (port from 9router#1887, [#6411](https://github.com/diegosouzapw/OmniRoute/pull/6411)). NVIDIA's OpenAI-compatible wrapper rejects it with `400 Unsupported parameter`, the same class already handled for `cerebras`/`mistral`; `nvidia` (executor `default`) was missing from the strip allowlist so Codex/Claude-Code passthrough requests 400'd. Regression guard: `tests/unit/executor-default-strip-client-metadata.test.ts` (+nvidia case). (thanks @phidinhmanh)
+- **fix(translator):** strip the Claude-style `thinking` field for **NVIDIA `z-ai/glm-5.2`** (port from 9router#2023, [#6413](https://github.com/diegosouzapw/OmniRoute/pull/6413)). NVIDIA's OpenAI-compatible wrapper 400s on `thinking` (a Claude-format client routed here leaves a `thinking:{type:"adaptive"}`); the existing strip rule only dropped `reasoning`. Same class already handled for `minimax-m2.7`. Regression guard: `tests/unit/nvidia-minimax-thinking-strip.test.ts` (+glm-5.2 case). (thanks @phidinhmanh)
+- **fix(translator):** suppress the streamed `</think>` close marker for the **Antigravity IDE** client (port from 9router#1061, [#6415](https://github.com/diegosouzapw/OmniRoute/pull/6415)). On thinking-only turns Antigravity rendered a bare `</think>` as the sole visible content, tripping its loop-detection and wasting requests. Antigravity's UA (`vscode/<v> (Antigravity/<v>)`) is added to the marker-suppress allowlist (alongside OpenCode); Claude Code / Cursor still get the marker, and `x-omniroute-thinking-marker: on` force-restores it. Regression guard: `tests/unit/think-close-marker-suppress-5245.test.ts`. (thanks @abdofallah)
+- **fix(executors):** strip nested `reasoning_content` from messages for **Mistral** (port from 9router#1649, [#6417](https://github.com/diegosouzapw/OmniRoute/pull/6417)). Mistral's API returns `422 extra_forbidden` when an assistant message carries `reasoning_content` (replayed thinking from a prior turn, e.g. via the Codex `/responses` path); the generic top-level 400 field-downgrade retry never covered the nested per-message field. `DefaultExecutor` now strips it for provider `mistral` only, so DeepSeek (which requires replayed `reasoning_content`) is unaffected. Regression guard: `tests/unit/mistral-strip-reasoning-content-1649.test.ts`. (thanks @xxy9468615)
+- **fix(executors):** strip the `client_metadata` passthrough field on the **OpenCode** path (port from 9router#1442, [#6418](https://github.com/diegosouzapw/OmniRoute/pull/6418)). OpenCode upstreams (e.g. `kimi-k2.6` via opencode-go) reject it with `400 "Extra inputs are not permitted, field: 'client_metadata'"`; the DefaultExecutor strip only covered cerebras/mistral and `OpencodeExecutor` extends `BaseExecutor` directly, so nothing removed it there. Regression guard: `tests/unit/opencode-strip-client-metadata-1442.test.ts`. (thanks @yanpaing007)
+- **fix(executors):** inject the `reasoning_content` echo for the native **Moonshot Kimi** provider (port from 9router#1480, [#6419](https://github.com/diegosouzapw/OmniRoute/pull/6419)). Kimi (executor `default`) is a thinking-mode upstream that 400s with "reasoning_content must be passed back" when a prior assistant turn lacks it; the placeholder injection was only wired into the OpenCode meta-provider, so direct multi-turn Kimi conversations failed. Scoped to `kimi` (gateway-served models matching the thinking-model name pattern are unaffected). Regression guard: `tests/unit/kimi-native-reasoning-injected-1480.test.ts`. (thanks @2220258345)
+- **fix(executors):** recover from a strict gateway's `context_management: Extra inputs are not permitted` 400 (port from 9router#1468, [#6420](https://github.com/diegosouzapw/OmniRoute/pull/6420)). **Claude Code** always sends a top-level `context_management` field; strict anthropic-compatible gateways reject it. The dedicated context-editing 400-fallback only fired when OmniRoute's own `contextEditing` feature was enabled (default off), so a client-sent field passed through untouched and 400'd. `context_management` is now in the generic reactive field-strip list, so it's stripped-and-retried once regardless of the feature flag (with correct request re-signing for claude-compatible relays). Regression guard: `tests/unit/provider-field-strips.test.ts`. (thanks @ohahe52-dot)
+- **fix(network):** enable **RFC 8305 Happy Eyeballs** (`autoSelectFamily`) on the direct-egress undici dispatcher (port from 9router#1237, [#6423](https://github.com/diegosouzapw/OmniRoute/pull/6423)). When DNS returns both IPv6 (AAAA) and IPv4 (A) and the IPv6 route is broken (e.g. a NAT64 `64:ff9b::` prefix without routing), undici tried IPv6 first and hung until `ETIMEDOUT` (then a 502 + account lockout), even though `curl` reached the same host. The direct dispatcher now races both families and uses whichever connects first. Proxy paths pin family via `proxyTls` and are unaffected. Regression guard: `tests/unit/direct-dispatcher-pipelining-4580.test.ts`. (thanks @adentdk)
+- **fix(combo):** round-robin now advances the rotation pointer past the model that **actually served**, not the eagerly-scheduled one (port from 9router#948, [#6428](https://github.com/diegosouzapw/OmniRoute/pull/6428)). With `stickyLimit: 1` (true round-robin), when the scheduled model failed and a _different_ model served via fallback, the counter had already advanced +1 from the scheduled index — so the next request reused the fallback-served model, degrading round-robin into hot-spotting on whichever model was healthy. The pointer now advances to the served index + 1 (mirroring the sticky-limit>1 path). Session-stickiness (#3825) and distribution are preserved. Regression guard: `tests/unit/combo-rr-fallback-advance-948.test.ts`. (thanks @binsarjr)
+- **fix(sse):** a non-string `model` field is now rejected with a `400` before the resolver, instead of crashing downstream `.toLowerCase()`/`.split()` calls into an empty-body `500` that escapes the error sanitizer ([#6407](https://github.com/diegosouzapw/OmniRoute/issues/6407)). Regression guard: `tests/unit/chat-non-string-model-6407.test.ts`. (thanks @chirag127)
+- **fix(api):** unknown `/api/*` routes now return a JSON `404` (instead of the dashboard HTML shell) and scalar chat params (`model`/`temperature`/etc.) are validated **before** the provider lookup so malformed requests fail fast with a clear `400` ([#6424](https://github.com/diegosouzapw/OmniRoute/issues/6424), [#6412](https://github.com/diegosouzapw/OmniRoute/issues/6412)). Regression guards: `tests/unit/api/api-catchall-json-404.test.ts`, `tests/unit/chat-early-schema-validation-6412.test.ts`. (thanks @chirag127)
+- **fix(api):** `/v1/chat/completions` now rejects a non-JSON `Content-Type` with a `400` before parsing the body ([#6414](https://github.com/diegosouzapw/OmniRoute/issues/6414)). Regression guard: `tests/unit/v1-chat-completions-content-type-6414.test.ts`. (thanks @chirag127)
+- **fix(api):** the `X-OmniRoute-Compression` response header is now echoed on `/v1/chat/completions` and `/v1/completions` ([#6422](https://github.com/diegosouzapw/OmniRoute/issues/6422)). Regression guard: `tests/unit/compression-header-echo-6422.test.ts`. (thanks @chirag127)
+- **fix(api):** concurrent `GET /v1/models` requests are coalesced into a single catalog build ([#6408](https://github.com/diegosouzapw/OmniRoute/issues/6408)). Regression guard: `tests/unit/v1-models-concurrent-6408.test.ts`. (thanks @chirag127)
+- **fix(api):** `/v1/completions` now echoes the requested `body.model` in its JSON + streamed responses ([#6429](https://github.com/diegosouzapw/OmniRoute/pull/6429)). Regression guard: `tests/unit/completions-body-model-echo.test.ts`. (thanks @chirag127)
+- **fix(api):** env-var master keys now see the full `/v1/models` catalog ([#6406](https://github.com/diegosouzapw/OmniRoute/issues/6406)). Regression guard: `tests/unit/models-catalog-envkey-6406.test.ts`. (thanks @chirag127)
+- **fix(api):** non-streaming `/v1/completions` responses now echo `body.model` aligned with the `X-OmniRoute-Model` header ([#6426](https://github.com/diegosouzapw/OmniRoute/issues/6426)). Regression guard: `tests/unit/v1-completions-model-header-match-6426.test.ts`. (thanks @chirag127)
+- **fix(api):** unknown `/v1/*` routes now return a JSON `404 not_found` instead of the Next.js dashboard HTML shell ([#6405](https://github.com/diegosouzapw/OmniRoute/issues/6405)). Regression guard: `tests/unit/api/v1-catchall-json-404.test.ts`. (thanks @chirag127)
+- **fix(api):** the per-connection provider models route now degrades to the shipped catalog when a provider's `/models` endpoint answers with a redirect ([#6267](https://github.com/diegosouzapw/OmniRoute/issues/6267)) — a `qwen-web` import failed with a raw `Redirect blocked … (307)` 503. `safeOutboundFetch` throws `REDIRECT_BLOCKED` on the 307, `getSafeOutboundFetchErrorStatus` maps it to 503, and `buildDiscoveryErrorFallbackResponse` treated every 503 as a hard error — so the non-empty `getModelsByProviderId("qwen-web")` catalog was never surfaced. A models-endpoint redirect is not a fixable-config error (unlike `URL_GUARD_BLOCKED`/`INVALID_URL`, which stay hard errors), so it now falls back to the local/cached catalog before the 503 short-circuit. General fix — covers any config-driven provider that 307s. Regression guard: `tests/unit/provider-models-qwen-web-redirect-6267.test.ts`. (thanks @chirag127)
+- **fix(api):** the per-connection provider models route (MCP `list_models_catalog` + the dashboard import view) now merges USER-ADDED custom models into its response ([#6247](https://github.com/diegosouzapw/OmniRoute/issues/6247)) — custom models live in the `key_value` namespace `customModels`, which the live REST `/api/v1/models` already merges, but `src/app/api/providers/[id]/models/route.ts` never read `getCustomModels`, so custom models were dropped on both the discovery-success and local_catalog paths. They are now folded into the returned model list (deduped by id, stamped `owned_by: provider`), fixing MCP + the dashboard import view in one place. Regression guard: `tests/unit/provider-models-custom-merge-6247.test.ts`. (thanks @RCrushMe)
+
+- **fix(providers):** GitLab Duo tool-calling follow-up turns no longer fail upstream with `422 {"detail":"Validation error"}` (tokens 0/0, rejected pre-inference). The #6234 tool-result-feedback fix serialized the **entire** multi-turn conversation into GitLab's single-file `code_suggestions` (`small_file`) generation endpoint — folded history that turn-N sent as an oversized `current_file.content_above_cursor` **and** duplicated verbatim into `user_instruction`, tripping the AI-Gateway's `small_file` validation guard. The executor now **bounds** that prompt: it keeps system + latest user message + the most-recent tool round (dropping older turns), caps oversized tool results, and stops duplicating the full prompt into `user_instruction` (which now carries only the short latest user message) — while still feeding the most-recent tool result back so the agent continues (`open-sse/executors/gitlab.ts`, [#6220](https://github.com/diegosouzapw/OmniRoute/issues/6220)). The unit test covers the bounding logic; the upstream 422→200 clearing is VPS-only (Hard Rule #18). Regression guard: `tests/unit/gitlab-tool-exchange-bounded-6220.test.ts`.
+
+- **fix(i18n):** the provider-detail (`/dashboard/providers/[id]`) connection-status filter labels no longer render as `__MISSING__:All` / `__MISSING__:Active` / `__MISSING__:Error` / `__MISSING__:Banned` / `__MISSING__:CreditsExhausted` in non-English locales (notably pt-BR) ([#6290](https://github.com/diegosouzapw/OmniRoute/issues/6290)). Root cause was **not** the namespace mismatch the issue guessed — the `providers.filter*` keys resolve correctly in `en.json`; the debt lived in the locale mirrors (`src/i18n/messages/*.json`), where these five keys carried the `__MISSING__:` sync sentinel in ~15 locales and were absent entirely in ~26 others, so next-intl found the key and echoed the sentinel verbatim. All 40 non-English/-Chinese mirrors now ship real translations for the five `providers.filter*` labels. Regression guard: `tests/unit/i18n-provider-filter-keys-6290.test.ts`. (thanks @diegosouzapw)
+
+- **fix(providers):** the `copilot-m365-web` streaming executor now emits `debug`-level WebSocket diagnostics ([#6210](https://github.com/diegosouzapw/OmniRoute/issues/6210)) — the outbound WS URL (with the `access_token` **redacted** via `redactWsUrl()`), handshake success/failure, and each received SignalR frame's `type`/`target`. Previously the streaming path logged nothing, so an empty `content:null` response (the M365 Education / Starter tier symptom fixed in #6234) was undiagnosable even at `APP_LOG_LEVEL=debug`. The change is debug-level and side-effect-free — it does not alter streaming behavior or the frame parser, and the token never reaches the logs. Regression guard: `tests/unit/copilot-m365-web-logging-6210.test.ts` (thanks @qpeyba)
+
+- **fix(resilience):** a round-robin combo no longer returns `503 all upstream accounts are unavailable` when a compatibility-rejected target is actually healthy ([#6238](https://github.com/diegosouzapw/OmniRoute/issues/6238)). `filterTargetsByRequestCompatibility` drops request-incompatible targets (tool/vision/structured-output unsupported, or below the required context window) **before** any availability check runs, and its `compatible.length === 0` safety net only fired when _all_ targets were filtered — not when the kept targets later all turned out runtime-unavailable (circuit-open / cooldown / no credentials). So a combo could 503 while a compat-rejected-but-healthy provider sat unused. `handleRoundRobinCombo` now keeps the compat-rejected set and, when every compat-kept target was skipped without a single real attempt, probes those rejected targets as a **last-resort fallback tier** (via the new pure `open-sse/services/combo/comboCompatFallback.ts`) before crystallizing the 503. Regression guard: `tests/unit/combo-roundrobin-compat-fallback-6238.test.ts`. (thanks @ThongAccount)
+
+- **fix(startup):** best-effort self-heal for a corrupted Turbopack dev cache on Windows ([#6289](https://github.com/diegosouzapw/OmniRoute/issues/6289)). On Windows, `pnpm dev` can fail at startup when Turbopack `mmap`s a persistent-cache SST file and the OS refuses the mapping (`os error 1455` — "paging file too small"), which Turbopack surfaces as a misleading `Module not found: Can't resolve '@/shared/utils/machine'`. This is a **known upstream Turbopack cache-corruption bug — not our code**. The dev launcher (`scripts/dev/run-next.mjs`) now wraps `nextApp.prepare()` and, when it rejects with that signature (`isTurbopackCacheCorruption` in the new `scripts/dev/turbopackCacheHeal.mjs`), purges `.build/next/**/cache/turbopack` and retries **once** with a clear log. **Caveat — best-effort only:** the corruption often surfaces as a runtime overlay rather than a `prepare()` rejection, so this cannot always intercept it; the reliable remedy remains manually deleting the Turbopack cache dir. Regression guard: `tests/unit/turbopack-cache-heal-6289.test.ts`. (thanks @chirag127)
+
+- **fix(providers):** qodercli PAT auth no longer fails with `spawn qodercli ENOENT` on Windows ([#6263](https://github.com/diegosouzapw/OmniRoute/issues/6263)) — `spawnQoderCli` spawned the bare `qodercli` name with `shell:false` and an unenriched env, so the npm `.cmd` wrapper under `%APPDATA%\npm` (a user-PATH directory) was never resolved. It now resolves the absolute `.cmd`/`.exe` path through the existing `getCliRuntimeStatus("qoder")` resolver in `src/shared/services/cliRuntime.ts` (memoized), spawns with `shell` when the target is a `.cmd`/`.bat`, and uses the cliRuntime-enriched env (PATH + PATHEXT + APPDATA); the ENOENT error now lists the searched paths plus the `CLI_QODER_BIN` override. End-to-end spawn on a real Windows host is host-only (Hard Rule #18); the path-resolution logic is unit-tested. Regression guard: `tests/unit/qodercli-windows-resolve-6263.test.ts`. (thanks @chirag127)
+
+- **fix(sse):** the reasoning-token buffer no longer inflates **probe-sized `max_tokens`** ([#6274](https://github.com/diegosouzapw/OmniRoute/issues/6274)) — Claude Code's `/model` capability check sends `max_tokens: 1`, but for a thinking-capable model with a large output cap (e.g. `glm-5.2`) the #3587 headroom heuristic (`max(current + 1000, ceil(current * 1.5))`) rewrote it to `1001` and forwarded that upstream, wasting tokens on a request that was never a genuine reasoning budget. `resolveReasoningBufferedMaxTokens()` (`open-sse/services/reasoningTokenBuffer.ts`) now short-circuits and returns the caller's value verbatim when it is below the new `REASONING_BUFFER_MIN_TRIGGER` (256) threshold — a tiny explicit limit is a probe, not a reasoning request. Real budgets still receive the #3587 headroom unchanged, and the guard runs after the existing capability checks so unknown / non-reasoning models keep returning `null`. Regression guard: `tests/unit/reasoning-token-buffer-6274.test.ts`. (thanks @brightfiscalband)
+
+- **fix(cli):** `omniroute reset-password` now works as a real subcommand, and password resets over piped (non-TTY) stdin actually apply ([#6261](https://github.com/diegosouzapw/OmniRoute/issues/6261), [#6258](https://github.com/diegosouzapw/OmniRoute/issues/6258)). Two coupled defects: (1) **#6261** — `bin/omniroute.mjs` routed everything through Commander with only two pre-Commander bypasses (`--mcp`, `reset-encrypted-columns`), so `omniroute reset-password` was rejected as an unknown command; only the separate `omniroute-reset-password` bin worked, while the docs falsely advertised the subcommand (incl. a bogus "legacy alias still works"). A pre-Commander bypass mirroring `reset-encrypted-columns` now dynamically imports `bin/reset-password.mjs` (which self-executes) before Commander parses; the three doc lines were corrected. (2) **#6258** — `bin/reset-password.mjs` issued two sequential `rl.question` prompts; under piped stdin the second read never settled at EOF, so `main()` never reached `resetManagementPassword` and the reset was a silent no-op (both prompts printed, no success, password unchanged). The CLI now detects non-TTY stdin and reads it once (first line = password, second line = confirm if present, else reused), adds a `--password-stdin` flag (entire stdin is the password, no confirmation), and exits `0` explicitly so the success line always flushes; interactive TTY behavior is unchanged. Regression guard: `tests/unit/reset-password-cli-6261-6258.test.ts` (3). (thanks @chirag127)
+
+- **fix(db):** the mass-migration **safety abort** now tells the operator how to bypass it and stops flooding the log ([#6260](https://github.com/diegosouzapw/OmniRoute/issues/6260)) — after restoring a backup that wiped the migration tracking table, `runMigrations()` threw the abort on every downstream `ensureDbInitialized()`, re-logging the full banner 11+ times, and the message never mentioned the existing `OMNIROUTE_MAX_PENDING_MIGRATIONS` escape hatch. The abort text now appends a bypass hint (set `OMNIROUTE_MAX_PENDING_MIGRATIONS=0` in `server.env` / `DATA_DIR/.env`), and a new `MigrationSafetyAbortError` is memoized so repeated calls in the same process throw the same instance and emit a single concise line instead of the full cascade. Regression guard: `tests/unit/migration-safety-abort-6260.test.ts`. (thanks @chirag127)
+
+- **fix(auth):** importing a **distinct** Codex/ChatGPT OAuth `auth.json` is no longer falsely rejected as "already exists" when it belongs to a different user in the same workspace ([#6301](https://github.com/diegosouzapw/OmniRoute/issues/6301)). `findExistingCodexConnection` (in `src/lib/oauth/utils/codexAuthImport.ts`) deduped **only** on `providerSpecificData.workspaceId === accountId`, where `accountId` is the shared `chatgpt_account_id`/`tokens.account_id` — so two members of the same ChatGPT Team collapsed onto a single connection (409 `duplicate_account`). The id_token's `https://api.openai.com/auth` claim carries a per-user `chatgpt_user_id` alongside the workspace id (the device-flow path already persisted it as `chatgptUserId`, but the import path did not). Now `parseAndValidateCodexAuth` extracts `userId` (`chatgpt_user_id` → `user_id` → JWT `sub`) into `ParsedCodexAuth`, the create/update paths persist `chatgptUserId` in `providerSpecificData` (mirroring `codex.ts`), and dedup keys on `workspaceId` **AND** `chatgptUserId` — with a backward-compat fallback to legacy accountId-only matching when no stored connection for that workspace records a `chatgptUserId`, so genuinely-same accounts still dedup. Regression guard: `tests/unit/codex-auth-import-userid-dedup-6301.test.ts` (4). (thanks @anungma)
+
+- **fix(providers):** importing models for the **venice-web** provider no longer fails with a red "Provider venice-web does not support models listing" ([#6269](https://github.com/diegosouzapw/OmniRoute/issues/6269)). `venice-web` is a web-cookie provider with an executor but no upstream `/v1/models` endpoint and no registry `models`, so the models route fell through to the tail `400`. Mirroring the `jules`/`linkup-search`/`ollama-search` fix (#5569), it now ships a static local catalog entry in `src/lib/providers/staticModels.ts` — seeding the current Venice lineup (`venice-uncensored`, `llama-3.3-70b`, `qwen3-235b`, `qwen3-4b`, `deepseek-r1-671b`; Venice rotates its catalog, see docs.venice.ai/models/overview) — so the route returns `200` with `source:"local_catalog"`, `intentional:true`. Regression guard: `tests/unit/static-models-venice-web-6269.test.ts`. (thanks @chirag127)
+
+- **fix(api):** the specialty model catalogs (`/v1/embeddings`, `/v1/images`, `/v1/music`, `/v1/videos` model lists) are now derived from the **unified catalog filtered by a predicate** (`getSpecialtyModelsResponse`) instead of ad-hoc per-route logic, so they consistently respect active-credential visibility and stay in sync with the main catalog ([#6303](https://github.com/diegosouzapw/OmniRoute/pull/6303)). Regression guard: `tests/unit/specialty-model-catalog-routes.test.ts`. (thanks @makcimbx)
+- **fix(api):** the agent-bridge server route now resolves the MITM manager via a **dynamic `import("@/mitm/manager.runtime")`** so Turbopack does not statically pull the stub (or over-bundle the manager), and the agent-skills generator anchors its output base path with `path.join(process.cwd(), …)` so Turbopack's static analyzer stops tracing the whole project root ([#6329](https://github.com/diegosouzapw/OmniRoute/issues/6329), [#6366](https://github.com/diegosouzapw/OmniRoute/pull/6366)). Regression guard: `tests/unit/agent-bridge-server-route-dynamic-import.test.ts`. (thanks @Iammilansoni)
+- **fix(api):** internal probes (combo-test, cloud-sync verify) now pick a **management-scoped / allow-all API key** instead of naively grabbing `getApiKeys()[0]` — a restricted `self:usage` first row made the probe fail with "Model X is not allowed for this API key" even when the combo path was healthy (`pickApiKeyForInternalUse` in `src/lib/db/apiKeys.ts`). The API-manager model editor also falls back to `/api/models?all=true` when `/v1/models` is catalog-protected ([#6372](https://github.com/diegosouzapw/OmniRoute/pull/6372)). Regression guard: `tests/unit/pick-internal-api-key-6372.test.ts`. (thanks @jmengit)
+- **fix(live-ws):** the Live Dashboard WebSocket server now **rejects on bind failure** (e.g. `EADDRINUSE` when the API bridge already holds the port) instead of letting the error surface as an unhandled `error` event that crash-loops the process — the `error` listener is attached to `wss` (not `server`) and releases the EventBus subscription on a failed start ([#6324](https://github.com/diegosouzapw/OmniRoute/issues/6324)). Regression guard: `tests/unit/live-ws-eaddrinuse-6324.test.ts`. (thanks @vinayakkulkarni)
+- **fix(dashboard):** the Home provider-topology widget now trusts the live provider-metrics snapshot — it uses `topology.errorProvider` and live `activeRequests` directly instead of re-deriving state from a stale `lastErrorAt` or applying a frontend timeout filter, so the topology reflects real-time provider health ([#6322](https://github.com/diegosouzapw/OmniRoute/pull/6322)). Regression guard: `tests/unit/home-provider-topology-live-state.test.ts`. (thanks @xz-dev)
+- **fix(sse):** strip zero-width markers from streamed **tool-call arguments** — a follow-up to [#5857](https://github.com/diegosouzapw/OmniRoute/pull/5857). That PR removed injected zero-width joiners (U+200D) from streamed assistant text/reasoning but deliberately left tool-call argument JSON byte-exact. The request-side obfuscation (`open-sse/services/claudeCodeObfuscation.ts`) injects ZWJ into agent words — including the temp path inside the Bash tool description — and Claude models copy that verbatim into generated commands, which are delivered as tool-call arguments rather than assistant text. As a result the ZWJ survived and corrupted code blocks (e.g. a temp path rendered with an invisible joiner). Now `open-sse/handlers/responseSanitizer.ts` strips zero-width code points from tool-call argument strings at every emit site (OpenAI non-stream/stream chat `tool_calls` + legacy `function_call`, native Responses `function_call` items, the OpenAI→Responses conversion, and the native Responses streaming `response.function_call_arguments.delta/.done` events). Only zero-width code points are removed; JSON structure and all other bytes stay identical (no parse/restringify), so normal arguments remain byte-exact. Regression guard: 6 new cases in `tests/unit/response-sanitizer.test.ts` (suite 50/50).
+
+- **fix(nodejs):** the default app log path now resolves under `DATA_DIR` (`~/.omniroute/logs/application/app.log`) instead of `process.cwd()` ([#6197](https://github.com/diegosouzapw/OmniRoute/issues/6197)) — the globally-installed CLI runs from an arbitrary working directory, so anchoring the default to cwd made file logging silently write to (or no-op under) an unrelated directory, contradicting the documented `.env.example` default. `getAppLogFilePath()` now computes the default lazily via the pure `resolveDataDir()` resolver (honours a per-process `DATA_DIR`, no directory-creation side effect); an explicit `APP_LOG_FILE_PATH` still wins. Regression guard: `tests/unit/logenv-datadir-path-6197.test.ts` (3).
+
+- **fix(docker):** AgentBridge/`startMitm` no longer aborts in containers/headless when the Antigravity-default DNS step can't write `/etc/hosts` ([#6127](https://github.com/diegosouzapw/OmniRoute/issues/6127)), and the privileged command's stderr now reaches `app.log` instead of only a bare exit code hitting the toast ([#6198](https://github.com/diegosouzapw/OmniRoute/issues/6198)). The default DNS step (`addDNSEntry`) was called unguarded while cert install and the two sibling DNS steps were each best-effort — in the runtime Docker image (`USER node`, no `sudo`, read-only `/etc/hosts`) it threw `Command failed with code 1` out of `startMitmInternal` and killed the whole start, discarding the stderr. The three DNS steps are extracted into a best-effort `provisionDnsEntries()` where each failure is logged with the full `err` (stderr included, folded in by `systemCommands.ts`) and never aborts the start. Regression guard: `tests/unit/mitm-dns-graceful-degrade-6127.test.ts` (4).
+
+- **fix(providers):** copilot-m365-web now supports the M365 Education "Starter / OfficeWebIncludedCopilot" tier and no longer returns an empty `content:null` stream ([#6210](https://github.com/diegosouzapw/OmniRoute/issues/6210)). Two gaps: (1) `buildWsUrl()` hardcoded the individual-consumer scenario (`OfficeWebPaidConsumerCopilot`, `isEdu=false`) — the EDU tier is now opt-in via `providerSpecificData.tier="edu"`, emitting `scenario=OfficeWebIncludedCopilot`/`isEdu=true` (the individual path is unchanged); (2) the EDU/GPT-5.5 path streams deltas via `arguments[0].writeAtCursor` (incremental) instead of only `messages[].text` (accumulated snapshots), which the parser dropped — a new `accumulateBotContent()` folds both formats, with `type:2 item.result.message` as a last-resort fallback. Regression guard: `tests/unit/copilot-m365-edu-writeatcursor-6210.test.ts` (10). (thanks @qpeyba)
+
+- **fix(providers):** GitLab Duo executor now feeds tool results back into the prompt instead of looping ([#6220](https://github.com/diegosouzapw/OmniRoute/issues/6220)) — `buildPrompt()` branched only on `system`/`user` and took `userParts.at(-1)`, silently dropping the `assistant{tool_calls}` + `tool{result}` turns the client appended, so the reconstructed prompt was byte-identical to turn 1 and the model re-emitted the same `<tool>` call forever. When a tool exchange is present the full conversation is now serialized, folding each tool result back keyed by its `tool_call_id`; simple conversations keep the legacy shape. Complements the tool_call emission from [#6051](https://github.com/diegosouzapw/OmniRoute/issues/6051) (the `kilo-duplicate` label was a false positive — different, sequential defect). Regression guard: `tests/unit/gitlab-tool-result-feedback-6220.test.ts` (4).
+
+- **fix(providers):** opencode-go/opencode-zen can now synthesize the OpenCode CLI identity headers Cloudflare requires on VPS egress ([#5997](https://github.com/diegosouzapw/OmniRoute/issues/5997)) — on a datacenter VPS, `opencode.ai/zen/go/v1/chat/completions` 403s (HTML challenge) requests lacking CLI identity, while the reporter's control curl proved that `User-Agent: opencode-cli/1.0.0` + `x-opencode-client: cli` + `x-opencode-project: default` + fresh request/session UUIDs succeed. Opt-in via `OPENCODE_SYNTHESIZE_CLI_HEADERS=true` (values overridable via `OPENCODE_GO_USER_AGENT`/`OPENCODE_USER_AGENT`/`OPENCODE_CLIENT`/`OPENCODE_PROJECT`); it fills only headers the client did not already send. Kept **off by default** — the forward-only path is deliberate (fabricating a wrong value risks upstream rejection; a prior dedup regressed with `opencode/local`), so this replaces the fragile local header-injection shim without changing default behavior. Regression guard: `tests/unit/opencode-cli-headers-synthesis-5997.test.ts` (6). (thanks @aleksesipenko)
+- fix(resilience): sticky session affinity now evicts and fails over to another account when the pinned account is exhausted/unavailable (#6219)
+
+- fix(sse): Responses API passthrough now drops internal commentary-phase output before forwarding to clients (gated by RESPONSES_PASSTHROUGH_DROP_COMMENTARY, default on) (#6199)
+
+- **fix(sse):** tool-call function schemas with a root `type: null` are now coerced to `type: "object"` before dispatch ([#6359](https://github.com/diegosouzapw/OmniRoute/issues/6359)) — clients like the Codex app emit `parameters: { type: null, ... }` for some tools, which OpenAI-compatible upstreams reject with `400 Invalid schema for function '...': schema must be a JSON Schema of 'type: "object"', got 'type: null'`, failing the whole request. `toolSchemaSanitizer` already stripped the null; it now re-adds the mandatory root `"object"` type (and empty `properties`/open `additionalProperties` when absent). Combinator roots (`anyOf`/`oneOf`/`allOf`) and explicit root types are left untouched. Regression guard: 5 new cases in `tests/unit/tool-schema-sanitizer.test.mjs`.
+
+- **fix(docker):** AgentBridge no longer fails to start on npm/Electron/VPS installs with "MITM manager stub reached at runtime" ([#6344](https://github.com/diegosouzapw/OmniRoute/issues/6344)) — v3.8.45 flipped the production bundler default to Turbopack, but `next.config.mjs` aliased `@/mitm/manager` to its Docker-only degraded stub **unconditionally**. That was harmless while Docker (which sets the alias intentionally for #3390 graceful degradation) was the sole Turbopack consumer, but once every artifact built with Turbopack the stub shipped to all non-Docker users and `startMitm` threw on the first Agent-Bridge start. The alias is now opt-in via `OMNIROUTE_MITM_STUB=1` (set only by the Dockerfile) through the shared `scripts/build/mitm-stub-flag.mjs` helper; default builds bundle the real manager. Regression guard: `tests/unit/mitm-stub-alias-6344.test.mjs` (4).
+
+- **fix(proxy):** stop the v3.8.44 proxy regression that leaked the real IP and disabled healthy proxies ([#6246](https://github.com/diegosouzapw/OmniRoute/issues/6246)). Two coupled defects from the new health scheduler: (1) **IP leak** — when a proxy assigned to a connection was marked `inactive`, resolution fell through to a **direct** egress instead of blocking, exposing the operator's real IP; (2) **over-deactivation** — the sweep flipped a proxy to `inactive` on the **first** failed probe and counted our own 5s timeout / a probe-target `5xx` as the proxy's fault, so healthy paid proxies vanished from egress selection ("my proxies are not being used anymore"). Fix: the sweep decision is extracted into a pure, network-free `decideProxyHealthAction` (`src/lib/proxyHealth/decision.ts`) — by default the health check now **only counts/logs and never downgrades status** (a proxy is downgraded/removed only with `PROXY_AUTO_REMOVE=true`, after `PROXY_AUTO_REMOVE_AFTER` **consecutive** conclusive failures); probes are classified tri-state so an inconclusive result (our timeout, or a `5xx` from the probe target) never penalizes the proxy, and the probe timeout is raised 5s→15s. Separately, `safeResolveProxy` now **fails closed** via the existing policy: a connection whose assigned proxy is dead is blocked instead of leaking direct (`hasBlockingProxyAssignment`), honoring the explicit `proxy off` toggles and the `PROXY_FAIL_OPEN=true` opt-out. Existing proxies stuck `inactive` by the old behavior need a one-time manual re-activate (the operator owns proxy status). Regression guards: `tests/unit/proxy-health-decide-action-6246.test.ts`, `tests/unit/proxy-assigned-unavailable-6246.test.ts`.
+
+- **fix(proxy):** make "Test All" read-only and add bulk enable/disable ([#6246](https://github.com/diegosouzapw/OmniRoute/issues/6246)). Complements the core fail-closed / scheduler fix (#6296) with the two remaining reporter asks. (1) The **"Test All" button** (`POST /api/settings/proxies/auto-test`) used to flip a proxy to `inactive` on a failed reachability probe; since the egress selector excludes `inactive` proxies, a flaky probe (an unreachable `httpbin.org`, a proxy that blocks `HEAD`, or a slow paid proxy) silently disabled every proxy that failed — "Test All" is now **read-only by default** (only the operator sets a proxy active/inactive; opt back into the legacy test-and-set with `PROXY_HEALTH_AUTO_DEACTIVATE=true`). (2) Adds a **bulk enable/disable** proxies endpoint + toolbar action (`POST /api/settings/proxies/batch-activate`) so an operator can re-activate proxies in one click. Regression guard: `tests/unit/proxy-health-6246.test.ts`. (thanks @tenshiak)
+
+- **chatcore (tools): stop the default 128-tool cap from silently dropping opencode's `task`/MCP tools.** opencode (used as an MCP/agent host) sends a large tool list; when it exceeds the speculative `MAX_TOOLS_LIMIT` (128) default, `truncateToolList` did a blind `tools.slice(0, 128)`, dropping every tool past index 128 — including opencode's built-in `task` tool (subagent launch) and many MCP tools, so models routed through OmniRoute could no longer spawn subagents or reach part of their tools. The cap exists to avoid upstream `400`s for providers with real hard limits (e.g. grok-cli 200), so it is kept for those: detection of the opencode client (`isOpencodeClient` — any `x-opencode-*` header, or `opencode` in the user-agent) now only bypasses the **speculative 128 default**, never a known provider ceiling. Precedence is explicit — a proactive/detected provider limit always truncates (even for opencode); otherwise opencode forwards its full tool list; otherwise the unchanged 128 default applies to every other client. Refactors `getEffectiveToolLimit` into `getKnownToolLimit(provider) ?? DEFAULT_LIMIT` (byte-identical for existing callers) and fixes a cosmetic debug-log that reported the truncated count instead of the original. Regression guard: `tests/unit/tool-limit-detector.test.ts`.
+
+- **fix(mitm):** the macOS MITM-cert install check now matches the system keychain again. `security find-certificate -a -Z` prints the SHA-1 as a colon-less hex string, but the installed-check compared it against `getCertFingerprint()`'s colon-separated form, so the substring match never hit — the cert was reported as not-installed and re-prompted for the sudo install on every run. Fingerprints are now normalized (colons stripped, upper-cased) on both sides via the extracted `macCertOutputHasFingerprint` helper. Regression guard: `tests/unit/mitm-cert-mac-fingerprint.test.ts`. ([#6204](https://github.com/diegosouzapw/OmniRoute/pull/6204), closes [#6134](https://github.com/diegosouzapw/OmniRoute/issues/6134) — thanks @rianonehub)
+
+- **fix(api):** `/v1/messages/count_tokens` now counts `tool_use`, `tool_result` and `thinking` content blocks (and array-form `system` prompts) in the local-estimation path, instead of only `text`. Real agentic conversations keep ~95% of their tokens inside tool results; the previous estimate returned near-zero for them, which silently broke Claude Code's auto-compaction (context grew past the window with no compaction until the upstream API rejected the request). The real provider-side count path is unchanged. Regression guard: `tests/unit/messages-count-tokens-route.test.ts`. ([#6221](https://github.com/diegosouzapw/OmniRoute/pull/6221) — thanks @luweiCN)
+
+- **fix(antigravity):** strip a trailing assistant prefill turn for Vertex Claude models to avoid upstream 400s ([#6114](https://github.com/diegosouzapw/OmniRoute/pull/6114)). Regression guard: `tests/unit/antigravity-claude-prefill-strip.test.ts`. (thanks @anki1kr)
+
+- **fix(security):** the mutable cloud-agent routes (`/api/cloud/credentials/update`, `/api/cloud/models/alias`) now require management auth instead of being treated as public. They were classified as public API routes, so a request without management credentials could update stored cloud-agent credentials and model aliases. They are removed from the public-route set, classified as management routes in the authz pipeline, and gated by `requireManagementAuth`; cloud **read**/auth routes stay public. Regression guards: `tests/unit/cloud-write-auth.test.ts`, `tests/unit/authz/classify.test.ts`, `tests/unit/public-api-routes.test.ts`. ([#6233](https://github.com/diegosouzapw/OmniRoute/pull/6233) — thanks @vittoroliveira-dev)
+
+- **refactor(dashboard):** extract the onboarding-wizard "Open provider details" link target into a pure, unit-tested `buildProviderDetailsHref(connection)` helper. The wizard already routes by `connection.id` (the node UUID) rather than the provider category slug (#6144/#6145); this hardens that behavior behind a tested helper that guards a missing id/connection. Regression guard: `tests/unit/provider-onboarding-href.test.ts`. ([#6166](https://github.com/diegosouzapw/OmniRoute/pull/6166) — thanks @KooshaPari)
+
+- **fix(security):** the doubao synthetic device-id generator now derives its digits via an unbiased crypto-random draw (rejection sampling over `crypto.randomBytes()`) instead of a `% 10` reduction, closing a CodeQL `js/biased-cryptographic-random` finding.
+- **fix(agentSkills):** the GitHub-skills generator now resolves `outputDir` to an absolute path before writing, fixing a regression introduced by #6366 (relative-to-cwd base path) that could write generated skill files to the wrong directory.
+- **fix(security):** `/api/keys/{id}/devices` now checks the HTTP method before auth/validation, returning a `405` for non-GET/DELETE verbs instead of a misleading `401`/`500` (closes a `dast-smoke` QUERY-method finding).
+- **fix(quality):** clear the last 2 heavy quality-gate reds on the release tip (cycle pre-flight).
+- **fix(mitm):** the test suite and CI can never mutate the OS trust store — `OMNIROUTE_SKIP_SYSTEM_TRUST=1` is set globally for tests/CI so `installCert`/`uninstallCert`/`installTproxyCa` skip the privileged OS dispatch ([#6310](https://github.com/diegosouzapw/OmniRoute/pull/6310); full detail is under the [3.8.45] section below — this branch received it via the parallel-cycle sync-back).
+- **fix(api):** `POST /api/github-skills` now Zod-validates its request body; documented the new quality-gate env vars and pinned the merge-integrity GitHub Actions to a commit SHA.
+- **fix(skills):** generate the missing `omni-github-skills` registry entry and align the agent-skills catalog-count tests (follow-up to #6186).
+- **fix(quality):** clear the cycle's 11 net-new ESLint errors and make `validate-release-green` suppressions-aware.
+
+### 📝 Maintenance
+
+- **i18n(it):** add 118 missing Italian (`it`) translations (net-additive — no existing keys dropped, valid JSON), improving Italian UI coverage. ([#6212](https://github.com/diegosouzapw/OmniRoute/pull/6212) — thanks @serverless83)
+- **chore(providers):** remove deprecated **MiMo V2** model entries from the catalogs (xiaomi-mimo, opencode-go, zenmux-free, audio TTS) — the upstream V2 line is superseded by MiMo V2.5; drops `mimo-v2-tts`, `mimo-v2-pro`, `mimo-v2-omni`, `mimo-v2-flash`, `mimo-v2-flash-free` and realigns the provider-catalog tests. ([#6248](https://github.com/diegosouzapw/OmniRoute/pull/6248) — thanks @backryun)
+- **chore(release):** ~50 commits on this branch are v3.8.45 pre-flight/hardening fixes and CI-perf work that landed here via the parallel-cycle sync-back (`sync-next-cycle.mjs`, Hard Rule #21) **after** the `v3.8.45` git tag was cut, and are already fully documented under the **[3.8.45]** section below — listed here only so the per-cycle commit-coverage check (`npm run release:uncovered`) doesn't flag them as gaps. Provider/catalog/UX/backend fixes: #6041, #6078, #6108, #6135, #6148, #6149, #6154, #6158, #6161, #6162, #6163, #6164, #6165, #6170, #6177, #6178, #6181, #6186, #6187, #6191, #6193, #6194, #6195, #6200, #6205, #6208, #6209, #6211, #6213, #6223, #6224, #6225, #6226, #6227, #6228, #6229, #6230, #6235, #6291, #6292. CI/release-pipeline work: #6167, #6203, #6214, #6215, #6218, #6273, #6275, #6283, #6284, #6285, #6300, #6305.
+- **chore(release):** additional zero-ref release-cycle plumbing on this branch, kept out of `release:uncovered` on purpose (no `#N` in the commit subject to cite): opening the v3.8.46 cycle, opening/closing the v3.8.45 cycle, the finalized [3.8.45] CHANGELOG i18n sync-back to 42 mirrors, the v3.8.45 cognitive/cyclomatic and file-size drift rebaselines, ESLint stale-suppression pruning (4,273 → 4,233), and clearing test-masking/docs-all pre-flight reds for v3.8.45.
+
+### ⚡ Performance & Infrastructure
+
+- **perf(release-green):** the pre-flight validator (`scripts/quality/validate-release-green.mjs`) now runs its 4 slow suites (unit / vitest / integration / pack-artifact) **concurrently** via `Promise.all` — pre-flight wall time drops from ~the sum of the suites to ~the slowest one (~30min saved per round; Phase 0 was the nº1 bottleneck of the v3.8.45 release benchmark, 2h54 of 6h34 e2e). Guard: `tests/unit/validate-release-green.test.ts` ("runs the slow suites CONCURRENTLY"). ([#6319](https://github.com/diegosouzapw/OmniRoute/pull/6319))
+- **fix(ci):** `scripts/release/sync-next-cycle.mjs` — two defects found live in its first production run (v3.8.45 Phase 5): (1) the `git()` helper's default 1 MiB `maxBuffer` crashed with `ENOBUFS` on `git show origin/main:CHANGELOG.md` (the CHANGELOG alone is >1 MiB) — widened to 64 MiB; (2) the i18n resync only propagated the `[NEXT]` (TBD) section, leaving the just-shipped finalized section as "— TBD" in all 42 mirrors — it now also syncs `[prevVersion]` bounded by the heading below it (new exported pure helper `versionAfter`). Guards: +5 tests in `tests/unit/sync-next-cycle.test.ts` (8/8). ([#6327](https://github.com/diegosouzapw/OmniRoute/pull/6327))
+- **test(ci):** concurrency-sensitive flaky tests are **quarantined into a serial pass** (`tests/unit/serial/`, `--test-concurrency=1`, appended to every unit runner incl. sharded variants — the serial pass is sharded too so concurrent shard jobs never self-collide). Initial set: `glm-coding-plan-monthly-3580`, `quota-division-blocks`, `provider-health-autopilot`, `combo-health-autopilot` — the class behind the ~28min CI wedges/re-runs (two live 1h+ wedges cancelled during this PR's own validation). Discovery + TIA gates track the new glob; systemic root cause (async logger writing after teardown) tracked in [#6360](https://github.com/diegosouzapw/OmniRoute/issues/6360). Guard: `tests/unit/test-serial-quarantine.test.ts` (4). ([#6347](https://github.com/diegosouzapw/OmniRoute/pull/6347))
+
+### 🙌 Contributors
+
+Thanks to everyone whose work landed in v3.8.46:
+
+| Contributor                                                    | PRs / Issues           |
+| -------------------------------------------------------------- | ---------------------- |
+| [@2220258345](https://github.com/2220258345)                   | direct commit / report |
+| [@abdofallah](https://github.com/abdofallah)                   | direct commit / report |
+| [@adentdk](https://github.com/adentdk)                         | direct commit / report |
+| [@aleksesipenko](https://github.com/aleksesipenko)             | direct commit / report |
+| [@anki1kr](https://github.com/anki1kr)                         | direct commit / report |
+| [@anungma](https://github.com/anungma)                         | direct commit / report |
+| [@arpicato](https://github.com/arpicato)                       | direct commit / report |
+| [@backryun](https://github.com/backryun)                       | #6248                  |
+| [@binsarjr](https://github.com/binsarjr)                       | direct commit / report |
+| [@brightfiscalband](https://github.com/brightfiscalband)       | direct commit / report |
+| [@chirag127](https://github.com/chirag127)                     | #6501, #6506           |
+| [@developerjillur](https://github.com/developerjillur)         | direct commit / report |
+| [@dilneiss](https://github.com/dilneiss)                       | #6499                  |
+| [@dtybnrj](https://github.com/dtybnrj)                         | direct commit / report |
+| [@Forcerecon](https://github.com/Forcerecon)                   | direct commit / report |
+| [@hao3039032](https://github.com/hao3039032)                   | direct commit / report |
+| [@Iammilansoni](https://github.com/Iammilansoni)               | #6150, #6245           |
+| [@jmengit](https://github.com/jmengit)                         | direct commit / report |
+| [@jordansilly77-stack](https://github.com/jordansilly77-stack) | direct commit / report |
+| [@JxnLexn](https://github.com/JxnLexn)                         | direct commit / report |
+| [@KooshaPari](https://github.com/KooshaPari)                   | #6166                  |
+| [@loopyd](https://github.com/loopyd)                           | direct commit / report |
+| [@luweiCN](https://github.com/luweiCN)                         | #6221                  |
+| [@makcimbx](https://github.com/makcimbx)                       | direct commit / report |
+| [@muflifadla38](https://github.com/muflifadla38)               | direct commit / report |
+| [@newnol](https://github.com/newnol)                           | direct commit / report |
+| [@ofekbetzalel](https://github.com/ofekbetzalel)               | direct commit / report |
+| [@ohahe52-dot](https://github.com/ohahe52-dot)                 | direct commit / report |
+| [@phidinhmanh](https://github.com/phidinhmanh)                 | direct commit / report |
+| [@powellnorma](https://github.com/powellnorma)                 | direct commit / report |
+| [@qpeyba](https://github.com/qpeyba)                           | direct commit / report |
+| [@RaviTharuma](https://github.com/RaviTharuma)                 | direct commit / report |
+| [@RCrushMe](https://github.com/RCrushMe)                       | direct commit / report |
+| [@rianonehub](https://github.com/rianonehub)                   | #6134, #6204           |
+| [@serverless83](https://github.com/serverless83)               | #6212                  |
+| [@swingtempo](https://github.com/swingtempo)                   | direct commit / report |
+| [@tenshiak](https://github.com/tenshiak)                       | direct commit / report |
+| [@ThongAccount](https://github.com/ThongAccount)               | direct commit / report |
+| [@UnrealAryan](https://github.com/UnrealAryan)                 | direct commit / report |
+| [@vinayakkulkarni](https://github.com/vinayakkulkarni)         | direct commit / report |
+| [@vittoroliveira-dev](https://github.com/vittoroliveira-dev)   | #6233                  |
+| [@warelik](https://github.com/warelik)                         | direct commit / report |
+| [@xxy9468615](https://github.com/xxy9468615)                   | direct commit / report |
+| [@xz-dev](https://github.com/xz-dev)                           | direct commit / report |
+| [@yanpaing007](https://github.com/yanpaing007)                 | direct commit / report |
+| [@diegosouzapw](https://github.com/diegosouzapw)               | maintainer             |
+
+---
+
 ## [3.8.45] — 2026-07-06
 
 ### ✨ New Features

docs/i18n/bn/CHANGELOG.md

@@ -6,6 +6,213 @@
 
 ## [3.8.31] — 2026-06-20
 
+## [3.8.46] — TBD
+
+### ✨ New Features
+
+- **feat(sse):** **hide paid-only models from `auto/*` routing** when `hidePaidModels` is on ([#6512](https://github.com/diegosouzapw/OmniRoute/issues/6512)) — follow-up to #6328/#6495. PR #6495 hid paid-only models from the `GET /v1/models` listing, but `auto/*` combos (`auto/best-coding`, `auto/glm`, …) could still pick a paid-only backend into their candidate pool → a 402/403 at request time. `createVirtualAutoCombo` now filters the candidate pool through the new pure `open-sse/services/autoCombo/paidModelFilter.ts` (`filterPaidOnlyCandidates`), applying the same free-model predicate #6495 uses in `catalog.ts` (`providerHasFreeModels(provider) && isFreeModel(provider, {id})`) whenever `settings.hidePaidModels === true`. Applied before the category/tier/family narrowing, so it covers every `auto/*` combo; an all-paid pool degrades to the existing graceful empty-pool path. **Opt-in — default OFF leaves the pool unchanged** (identity). Regression guard: `tests/unit/autoCombo/paid-model-filter-6512.test.ts` (4, incl. the default-off identity guard).
+- **feat(sse):** **provider-family auto combos** — `auto/glm`, `auto/minimax`, `auto/mimo`, `auto/zai`, `auto/gemma`, `auto/llama`, `auto/gemini` ([#6453](https://github.com/diegosouzapw/OmniRoute/issues/6453)) — new routable ids that materialize an on-demand virtual combo spanning whatever installed backends currently expose that model family, degrading gracefully as backends rotate. A new pure `open-sse/services/autoCombo/modelFamily.ts` (`detectModelFamily`) classifies by model-id prefix for six families; `zai` is instead resolved by provider id (z.ai's hosted API serves the same `glm-*` model ids as every other GLM backend, so `auto/zai` means "route to my z.ai backend specifically" vs `auto/glm`'s "any connected GLM backend"). Reuses the existing `createVirtualAutoCombo` on-demand materialization path (no DB writes) and the `/v1/models` catalog advertising loop. Regression guard: `tests/unit/autoCombo/provider-family-combos.test.ts` (11).
+- **feat(proxy):** native **proxy-pool round-robin / egress IP rotation** ([#6365](https://github.com/diegosouzapw/OmniRoute/issues/6365)) — a scope (global / provider / account) can now hold **multiple** proxies as a pool with a rotation strategy, so outbound requests cycle their egress IP instead of pinning one proxy per scope. Migration `117_proxy_pool_rotation.sql` lifts the `UNIQUE(scope, scope_id)` constraint (rebuild via the canonical rename/copy/drop; existing single assignments become 1-element pools) and adds a `proxy_scope_rotation` companion table holding the per-scope strategy + a persisted monotonic round-robin cursor. Strategies: `round-robin` (default, monotonic cursor — never `Math.random`), `random`, and `sticky-per-N-min`. Resolution (`resolveProxyForScopeFromRegistry` / `resolveProxyForConnectionFromRegistry`) now fetches the alive, position-ordered candidate set (unchanged `PROXY_ALIVE_PREDICATE`) and applies the strategy; an empty / all-dead pool still returns `null` — the #6246 fail-closed guard is **untouched** (never falls through to direct egress). Backend + DB only; dashboard pool-builder UI is a follow-up. Regression guard: `tests/unit/proxy-pool-rotation-6365.test.ts` (8, incl. fail-closed + backward-compat).
+- **feat(providers):** end-to-end **tool/function calling on the native Gemini `/v1beta` endpoint** ([#6222](https://github.com/diegosouzapw/OmniRoute/issues/6222)) — both directions of the Gemini↔OpenAI conversion now preserve tool data (previously silently dropped). Request side: `convertGeminiToInternal` (extracted to its own testable module) maps `tools[].functionDeclarations` → OpenAI `tools`, prior `functionCall` parts → assistant `tool_calls`, and `functionResponse` parts → `tool`-role messages. Response side: `convertOpenAIResponseToGemini` emits `parts[].functionCall {name,args}` from `message.tool_calls`, and the streaming `openAIChunkToGeminiChunk` accumulates fragmented `tool_calls` deltas by index into complete `functionCall` parts. The non-Gemini client paths (Claude, OpenAI-Responses) already preserved tool calls — this closes the gap specific to the native Gemini surface. Regression guard: `tests/unit/v1beta-gemini-tool-calling-6222.test.ts` (6, incl. a streaming SSE round-trip).
+- **feat(providers):** copilot-m365-web **enterprise / work tier** support ([#6334](https://github.com/diegosouzapw/OmniRoute/issues/6334)) — mirrors the EDU-tier pattern (#6210): `M365ConnectionParams` gains an `agent` field, a new opt-in `M365_ENTERPRISE_OVERRIDES` preset (`agent=work`, `scenario=officeweb`, `licenseType=Premium`) applies via `providerSpecificData.tier="enterprise"` (alias `"work"`), and `agent` is also overridable directly via `providerSpecificData.agent`. `buildWsUrl` was hardcoding `agent="web"` (the one enterprise-distinguishing param with no override path), so a Premium work account handshook then returned an empty stream. The individual and EDU paths are untouched. Kilo's dup flag vs #6210 (EDU tier) was a false positive — different tier. Regression guard: `tests/unit/copilot-m365-enterprise-6334.test.ts` (7). End-to-end confirmation on a real Premium work account is a live-VPS validation follow-up (Hard Rule #18). (thanks @Forcerecon)
+- **feat(api):** standardized, provider-agnostic **`effort` + `thinking` request params** ([#6241](https://github.com/diegosouzapw/OmniRoute/issues/6241)) — a thin standardization layer over the existing mature per-provider reasoning plumbing (no provider mapper touched). `providerChatCompletionSchema` gains a canonical `effort` (reusing the shared `none/low/medium/high/xhigh` vocabulary — the UI tiers `extra`/`max` collapse onto `xhigh`) and a boolean `thinking`. A pure `normalizeReasoningRequest` (wired once in `src/sse/handlers/chat.ts`, before any reasoning field is read) folds them onto the fields the translators already consume (`reasoning_effort` / `reasoning.effort` / `thinking`), so they fan out to Anthropic / Gemini / xAI / Responses — an explicit client `reasoning_effort` / object-shaped `thinking` always wins (backward-compatible). `/models` additively exposes `supportsThinking` + `effort_tiers` so the frontend can render the toggles (UI component is a follow-up). Regression guard: `tests/unit/effort-thinking-standardization-6241.test.ts` (12). (thanks @Iammilansoni, @shabeer)
+- **feat(combo):** new **`pipeline` (sequential) combo strategy** ([#6297](https://github.com/diegosouzapw/OmniRoute/issues/6297)) — the 18th routing strategy runs targets **in order**, threading each step's output into the next step's input, with an optional per-step `prompt` (system instruction); only the final step's response is returned. Distinct from `fusion` (parallel fan-out + judge). Implemented as a self-contained `open-sse/services/pipeline.ts` (sibling to `fusion.ts`), dispatched from `combo.ts`; the step list reuses `combo.models` order and reads an optional `prompt` off each target (backward-compatible — ignored by every other strategy). Intermediate steps run non-streaming with tools stripped (complete prose to thread forward); the final step keeps the client's `stream` flag + tools. A failing/empty/unparseable intermediate step fails the whole pipeline explicitly via a sanitized error (never silently swallowed). Kilo's dup flag vs #563 was a false positive (that's model→chain selection; this is a sequential chain). Regression guard: `tests/unit/combo-pipeline-strategy.test.ts` (5). (thanks @ofekbetzalel)
+- **feat(ci):** `check:test-masking` now flags **inline-reimplemented prod conditions** ([#6348](https://github.com/diegosouzapw/OmniRoute/issues/6348)) — a new **report-only** subcheck (v2, 6A.10 family) catches the wrong-shape contract test: a test that recomputes the condition under test inline instead of importing/exercising the real function (the #6216 class, where `=== 500` → `>= 500` stayed green because the test re-implemented the branch). For each added/modified test file it warns when the file textually duplicates a ≥3-token conditional from a production file touched in the same PR **and** does not import the symbol/module owning it, via a pure, fixture-tested `findReimplementedConditions()` with an allowlist mirroring `assertReductionAllowlist`. Report-only for now (does not fail the gate) — to be promoted to blocking after a triage cycle. Regression guard: `tests/unit/check-test-masking.test.ts` (45).
+- **feat(sse):** per-connection routing override (native vs CLIProxyAPI) ([#6339](https://github.com/diegosouzapw/OmniRoute/issues/6339)) — the previously-dead `isCliproxyapiDeepModeEnabled` helper is now wired into `resolveExecutorWithProxy`: a single connection can opt itself into the CLIProxyAPI passthrough executor via `providerSpecificData.cliproxyapiMode="claude-native"`, with precedence **connection override > provider `upstream_proxy_config` mode > default**. `resolveExecutorWithProxy` now receives the resolved connection's `providerSpecificData` (threaded from `chatCore.ts`), so one connection can deep-route while the provider's other connections stay native — no DB schema change (the toggle rides in `providerSpecificData`). Also resolves the same-provider mixing ask in #6340. Regression guard: `tests/unit/chatcore-executor-proxy.test.ts` (9). (thanks @RaviTharuma)
+- **feat(dashboard):** "Add session cookie" modal now shows a prominent **"Open ‹host› →"** link to the provider's own site ([#6268](https://github.com/diegosouzapw/OmniRoute/issues/6268)) — every `-web` cookie-session provider (chatgpt-web, claude-web, gemini-web, kimi-web, lmarena, qwen-web, m365-copilot-web, …) renders a one-click external link (opening the provider's login/home page in a new tab) so operators no longer tab away to retype the URL mid-setup. The host resolves from a pure, unit-tested `resolveWebProviderHost()` (prefers `WEB_COOKIE_PROVIDERS[id].website`, falls back to the registry `baseUrl` origin); non-web providers render exactly as before. Kilo's dup flag vs #6265 (modal-too-small-on-1080p) was a false positive — distinct concern. Regression guard: `tests/unit/resolve-web-provider-host.test.ts` (5). (thanks @chirag127)
+- **feat(providers):** add **DigitalOcean** AI (serverless inference) as an OpenAI-compatible API-key provider ([#6373](https://github.com/diegosouzapw/OmniRoute/pull/6373)) — base `https://inference.do-ai.run/v1`, wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/digitalocean/`, `src/shared/constants/providers/apikey/inference-hosts.ts`). Regression guard: `tests/unit/digitalocean-provider.test.ts`. (thanks @newnol)
+- **feat(providers):** add **Huancheng Public API** (`hcnsec`) as an OpenAI-compatible regional provider ([#6410](https://github.com/diegosouzapw/OmniRoute/pull/6410)) — Xinjiang Huancheng Cybersecurity's public LLM platform (base `https://api.hcnsec.cn/v1`, free credits via daily check-ins), wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/hcnsec/`, `src/shared/constants/providers/apikey/regional.ts`). Regression guard: `tests/unit/hcnsec-provider.test.ts`. (thanks @UnrealAryan)
+- **feat(dashboard):** the web-session credential guide now shows an **"Open {host}" link** ([#6316](https://github.com/diegosouzapw/OmniRoute/pull/6316)) to the provider's sign-in site (derived from the provider `website` via `getProviderWebsiteHost`), so you can jump straight to the page where the cookie/session must be captured. Regression guard: `tests/unit/web-session-provider-link-6316.test.ts`. (thanks @jordansilly77-stack)
+- **feat(cerebras):** add the **Gemma 4 31B** model (`gemma-4-31b`) to the Cerebras registry + pricing table ([#6331](https://github.com/diegosouzapw/OmniRoute/pull/6331)). Regression guard extends `tests/unit/t28-model-catalog-updates.test.ts`. (thanks @backryun)
+- **feat(providers):** add **Yuanbao (web)** as a cookie-session provider ([#6196](https://github.com/diegosouzapw/OmniRoute/issues/6196)) — `yuanbao-web` (Tencent Yuanbao, `yuanbao.tencent.com`) with cookie-only auth (`hy_user`/`hy_token` + public agent id), SSE→OpenAI translation incl. `reasoning_content`, exposing DeepSeek V3/R1 + Hunyuan / Hunyuan-T1. Regression guard: `tests/unit/providers-yuanbao-web.test.ts`. `together-web` was **deferred** (no verifiable web-session endpoint — needs a captured request) and `huggingchat-web` **dropped** (the existing `huggingchat` already is a web-cookie provider). (thanks @chirag127)
+- **feat(providers):** route the built-in **agentrouter** through the dynamic Claude-Code wire image ([#6056](https://github.com/diegosouzapw/OmniRoute/issues/6056)) — a small static allow-set (`CC_WIRE_IMAGE_BUILTINS` in `open-sse/services/ccWireImageBuiltins.ts`), consulted by `isClaudeCodeCompatible` / `isClaudeCodeCompatibleProvider` / `applyFingerprint`, makes agentrouter adopt the CC wire-image headers + fingerprint **while guarding the CC baseUrl/auth branches** so it keeps its own registry `baseUrl` and `x-api-key` auth. Regression guard: `tests/unit/agentrouter-cc-wire-image.test.ts` (asserts the wire image is applied AND agentrouter's baseUrl/auth are preserved). Live WAF-acceptance against agentrouter.org is a VPS validation follow-up (Hard Rule #18).
+- **feat(providers):** **bulk-add API keys for Cloudflare Workers AI** ([#6174](https://github.com/diegosouzapw/OmniRoute/issues/6174)) — `cloudflare-ai` is removed from the bulk-add exclusion list and the bulk parser gains a 3-field `name|accountId|apiKey` mode; the bulk route now builds a **per-entry** `providerSpecificData` so each key carries its own `accountId` (fixing the previous shared-object reuse), and both the create + key-validation paths receive it. Regression guard: `tests/unit/bulk-api-key-parser-cloudflare.test.ts`. (thanks @muflifadla38)
+- **feat(dashboard):** routing/settings UX clarity ([#6147](https://github.com/diegosouzapw/OmniRoute/issues/6147)) — (1) weighted combos show the **effective routing share %** next to each weight when weights don't sum to 100 (`WeightTotalBar.tsx`); (2) the status widget's user-facing **"Cloud Sync" label is renamed** to "Remote Settings Sync" (`CloudSyncStatus.tsx`; internal ids/state untouched); (3) built-in providers gain an **opt-in advanced base-URL override** (`isBaseUrlOverrideEligibleProvider`, hidden behind an "Advanced" toggle, reusing the existing `providerSpecificData.baseUrl` persistence — not globally widened). Regression guard: `tests/unit/routing-settings-ux-6147.test.ts`.
+- **feat(combo):** add an option to **disable session stickiness**, per-combo or globally — round-robin / random combos can rotate to a different connection on every request instead of pinning a whole conversation to one connection by its first-message hash. Resolution precedence per-combo `config.disableSessionStickiness` → global `settings.disableSessionStickiness` → default `false` (preserves the #3825 prompt-cache/504 fix); gates **both** stickiness call sites in `open-sse/services/combo.ts`. Exposed as a global toggle (Combo Defaults) and a per-combo Inherit/on/off control. ([#6168](https://github.com/diegosouzapw/OmniRoute/issues/6168)) Regression guard: `tests/unit/combo-disable-session-stickiness.test.ts`. (thanks @RCrushMe)
+- **feat(docker):** add the `OMNIROUTE_NO_SUDO` env flag for root-less / user-namespaced deployments — the MITM cert-trust command path (`resolveSudoSpawn` in `src/mitm/systemCommands.ts`) now strips the leading `sudo` when the flag is truthy, in addition to the existing root / sudo-missing cases, so the Proxy Agent runs without `sudo` (the operator trusts the CA manually, e.g. via `NODE_EXTRA_CA_CERTS`). Argv-array `spawn` preserved — no shell interpolation (Hard Rule #13). ([#6122](https://github.com/diegosouzapw/OmniRoute/issues/6122)) Regression guard: `tests/unit/mitm-systemCommands-no-sudo.test.ts`. (thanks @powellnorma)
+- **feat(providers):** add **Requesty** as an OpenAI-compatible gateway provider (BYOK, base `https://router.requesty.ai/v1`, ~200 free requests/day) — wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/requesty/`, `src/shared/constants/providers/apikey/gateways.ts`). ([#6120](https://github.com/diegosouzapw/OmniRoute/issues/6120)) Regression guard: `tests/unit/requesty-provider.test.ts`. (thanks @chirag127)
+- **feat(dashboard):** add **configured-only / available-only filters** to the Free Provider Rankings page ([#6150](https://github.com/diegosouzapw/OmniRoute/issues/6150)) — hide providers you haven't configured, or whose connections are all rate-limited / out of quota, via server-side query params (`?configuredOnly` / `?availableOnly` on `GET /api/free-provider-rankings`) backed by a testable lib helper reusing the in-process connection state (no Redis). Both filters default off, so the default view is unchanged; this supersedes the earlier client-side "Configured Only" toggle (#6245) with an available-only dimension and unit-tested logic. Regression guard: `tests/unit/freeProviderRankings-filters.test.ts`.
+
+### 🔧 Bug Fixes
+
+- **fix(dashboard):** adding a second API key connection for the same provider no longer silently overwrites the first — the Add-API-key modal now derives a unique default connection name (`main`, then `main-2`, `main-3`, …) so the backend name-based upsert can't collide (#6499 — thanks @dilneiss).
+- **fix(compression):** the session-dedup engine now also deduplicates a large multi-line block repeated **within a single message** (intra-message dedup), not just across turns; the compression-preview API surfaces a `fallbackReason`, and the fusion panel reports how many models were rate-limited vs failed on total-panel failure (#6501 — thanks @chirag127).
+- **fix(compression):** a stacked-pipeline step naming an unregistered engine now surfaces a `validationErrors` entry instead of silently no-op'ing, so misconfigured pipelines are visible in the preview API (#6506 — thanks @chirag127).
+- **feat(usage):** add a **Codex reset-credit redemption** flow to the Provider Limits UI ([#6361](https://github.com/diegosouzapw/OmniRoute/pull/6361)) — a `useCodexResetCreditRedemption` hook + `/api/usage/codex-reset-credit` route + `codexResetCredits` lib let you redeem banked Codex reset credits from the quota card. Regression guard: `tests/unit/codex-reset-credits.test.ts`. (thanks @JxnLexn)
+- **feat(glm):** add **team-plan quota settings** for `glm-cn` connections ([#6351](https://github.com/diegosouzapw/OmniRoute/pull/6351)) — a dedicated `GlmTeamQuotaFields` form section (team quota id / limits) threaded through the Add/Edit connection modals, persisted via `providerSpecificData`, with the GLM usage service reading the team quota. Regression guards: `tests/unit/glm-team-quota.test.ts`, `provider-specific-data-schema.test.ts`. (thanks @hao3039032)
+- **feat(providers):** add **TinyFish** web-fetch/search support ([#6349](https://github.com/diegosouzapw/OmniRoute/pull/6349)) — a `tinyfish-fetch` executor + `/v1/web/fetch` route + MCP web-fetch tool, registered as a specialty-media provider with request-validation and a search-provider catalog entry. Regression guards: `tests/unit/executor-tinyfish-fetch.test.ts`, `web-fetch-handler.test.ts`, `mcp-web-fetch-tool.test.ts`, `provider-validation-tinyfish.test.ts`. (thanks @dtybnrj)
+- **fix(cli):** `omniroute launch-codex` now spawns `codex.cmd` through a shell on Windows (the npm `.cmd` shim is unresolvable by bare `spawn` → ENOENT), mirroring the qodercli Windows fix (#6263) ([#6312](https://github.com/diegosouzapw/OmniRoute/pull/6312)). Regression guard: `tests/unit/launch-codex-windows-spawn-6312.test.ts`. (thanks @swingtempo)
+- **fix(codex):** isolate the **Spark** quota from the shared Codex quota and stabilize the quota UI ordering / hydration so per-scope limits render consistently ([#6336](https://github.com/diegosouzapw/OmniRoute/pull/6336)). Regression guards: `tests/unit/codex-quota-selection-hydration.test.ts`, `provider-limits-ui.test.ts` + 3 more. (thanks @xz-dev)
+- **feat(api):** add a `hidePaidModels` setting that filters paid-only models out of the `/v1/models` catalog. Regression guard: `tests/unit/models-catalog-hide-paid.test.ts`. (thanks @chirag127)
+- **fix(api-manager):** the fallback model picker now preserves combos instead of dropping them when a primary model is unavailable ([#6443](https://github.com/diegosouzapw/OmniRoute/pull/6443)). Regression guard: `tests/unit/api-manager-page-static.test.ts`. (thanks @jmengit)
+- **fix(providers):** recoverable Antigravity / Cloud-Code (Gemini Code Assist) `403` responses ([#6452](https://github.com/diegosouzapw/OmniRoute/pull/6452)) are now classified as a retryable project-config error instead of a terminal account ban, so a fixable project/API-disabled 403 no longer forces a ~1-year cooldown / full OAuth reconnect. Regression guard: `tests/unit/errorclassifier-antigravity-403.test.ts`. (thanks @developerjillur)
+- **fix(mitm):** `sanitizeHeaders` now redacts `Set-Cookie` response headers so upstream session cookies never leak into logs / diagnostics ([#6451](https://github.com/diegosouzapw/OmniRoute/pull/6451)). Regression guard: `tests/unit/mitm-sanitize-headers.test.ts`. (thanks @developerjillur)
+- **fix(api):** `/api/compression/preview` now accepts `mode: "caveman"` and correctly handles stacked / zero-compression previews ([#6425](https://github.com/diegosouzapw/OmniRoute/issues/6425)). Regression guard: `tests/unit/api/compression-preview-caveman-and-stacked-6425.test.ts`. (thanks @chirag127)
+- **feat(providers):** add **Zed** hosted LLM aggregator as a native-app provider ([#6118](https://github.com/diegosouzapw/OmniRoute/pull/6118)) — OAuth sign-in via the Zed hosted flow, registered through the shared provider registry + executor. Regression guards: `tests/unit/zed-oauth-provider.test.ts`, `zed-import-utils.test.ts`, `zed-docker-detect.test.ts`, `mitm-handler-zed.test.ts`. VPS-validated via live operator login (Hard Rule #18).
+- **fix(oauth):** the Kiro SSO-cache auto-import now **preserves the IDC region** — cross-region Amazon Q / Kiro profiles imported from the SSO cache are no longer collapsed to the default region ([#6113](https://github.com/diegosouzapw/OmniRoute/pull/6113)). Regression guard: `tests/unit/kiro-auto-import-idc-2059.test.ts`. VPS-validated via live operator login (Hard Rule #18).
+- **fix(dashboard):** passthrough model aliases no longer collide when two namespaced model ids share a last segment (port from 9router#1850, [#6431](https://github.com/diegosouzapw/OmniRoute/pull/6431)). `enx/gpt-5.5` and `enx/codebuddy/gpt-5.5` both auto-generated the alias `gpt-5.5`, so the second model could never be added (the UI just alerted "alias already exists"). Aliases are now disambiguated deterministically — bare last segment when free, then parent-qualified (`codebuddy-gpt-5.5`), then a numeric suffix — while re-adding the exact same model id is still blocked. Regression guard: `tests/unit/passthrough-alias-1850.test.ts`. (thanks @arpicato)
+- **fix(translator):** preserve a Gemini `functionResponse` co-located with other parts (another `functionCall`, or trailing `text`) in the same content when translating **Gemini → OpenAI** ([#6376](https://github.com/diegosouzapw/OmniRoute/pull/6376)). `convertGeminiContent()` early-returned the tool message on the first `functionResponse` part, dropping any co-located parts; such contents are now pre-split (one tool message per `functionResponse`, emitted first, plus one message for the remaining parts). Regression guard: `tests/unit/gemini-to-openai-function-response.test.ts`. (thanks @warelik)
+- **fix(headroom):** detect a python interpreter managed by **mise / pyenv / asdf / conda** (port from 9router#2353, [#6382](https://github.com/diegosouzapw/OmniRoute/pull/6382)). Headroom's python probe (`src/lib/headroom/detect.ts`) searched a hardcoded `PATH`, but version managers expose their interpreters via shim dirs that only join `PATH` through interactive-shell activation — which the non-interactive server never runs, so a managed python (≥3.10) was invisible and Headroom reported it missing. The search path now prepends the well-known shim/bin dirs (`~/.local/share/mise/shims`, `~/.pyenv/shims`, `~/.asdf/shims`, `$CONDA_PREFIX/bin`, `~/.local/bin`, respecting `MISE_DATA_DIR`/`PYENV_ROOT`/`ASDF_DATA_DIR` when set), and a new `HEADROOM_PYTHON` env override lets operators point straight at their interpreter (mirroring `HEADROOM_URL`). Still shell-free (`execFileSync`). Regression guard: `tests/unit/headroom-detect.test.ts` (5). (thanks @loopyd)
+- **fix(executors):** strip the OpenAI-Codex/Claude-CLI `client_metadata` passthrough field for **NVIDIA** requests (port from 9router#1887, [#6411](https://github.com/diegosouzapw/OmniRoute/pull/6411)). NVIDIA's OpenAI-compatible wrapper rejects it with `400 Unsupported parameter`, the same class already handled for `cerebras`/`mistral`; `nvidia` (executor `default`) was missing from the strip allowlist so Codex/Claude-Code passthrough requests 400'd. Regression guard: `tests/unit/executor-default-strip-client-metadata.test.ts` (+nvidia case). (thanks @phidinhmanh)
+- **fix(translator):** strip the Claude-style `thinking` field for **NVIDIA `z-ai/glm-5.2`** (port from 9router#2023, [#6413](https://github.com/diegosouzapw/OmniRoute/pull/6413)). NVIDIA's OpenAI-compatible wrapper 400s on `thinking` (a Claude-format client routed here leaves a `thinking:{type:"adaptive"}`); the existing strip rule only dropped `reasoning`. Same class already handled for `minimax-m2.7`. Regression guard: `tests/unit/nvidia-minimax-thinking-strip.test.ts` (+glm-5.2 case). (thanks @phidinhmanh)
+- **fix(translator):** suppress the streamed `</think>` close marker for the **Antigravity IDE** client (port from 9router#1061, [#6415](https://github.com/diegosouzapw/OmniRoute/pull/6415)). On thinking-only turns Antigravity rendered a bare `</think>` as the sole visible content, tripping its loop-detection and wasting requests. Antigravity's UA (`vscode/<v> (Antigravity/<v>)`) is added to the marker-suppress allowlist (alongside OpenCode); Claude Code / Cursor still get the marker, and `x-omniroute-thinking-marker: on` force-restores it. Regression guard: `tests/unit/think-close-marker-suppress-5245.test.ts`. (thanks @abdofallah)
+- **fix(executors):** strip nested `reasoning_content` from messages for **Mistral** (port from 9router#1649, [#6417](https://github.com/diegosouzapw/OmniRoute/pull/6417)). Mistral's API returns `422 extra_forbidden` when an assistant message carries `reasoning_content` (replayed thinking from a prior turn, e.g. via the Codex `/responses` path); the generic top-level 400 field-downgrade retry never covered the nested per-message field. `DefaultExecutor` now strips it for provider `mistral` only, so DeepSeek (which requires replayed `reasoning_content`) is unaffected. Regression guard: `tests/unit/mistral-strip-reasoning-content-1649.test.ts`. (thanks @xxy9468615)
+- **fix(executors):** strip the `client_metadata` passthrough field on the **OpenCode** path (port from 9router#1442, [#6418](https://github.com/diegosouzapw/OmniRoute/pull/6418)). OpenCode upstreams (e.g. `kimi-k2.6` via opencode-go) reject it with `400 "Extra inputs are not permitted, field: 'client_metadata'"`; the DefaultExecutor strip only covered cerebras/mistral and `OpencodeExecutor` extends `BaseExecutor` directly, so nothing removed it there. Regression guard: `tests/unit/opencode-strip-client-metadata-1442.test.ts`. (thanks @yanpaing007)
+- **fix(executors):** inject the `reasoning_content` echo for the native **Moonshot Kimi** provider (port from 9router#1480, [#6419](https://github.com/diegosouzapw/OmniRoute/pull/6419)). Kimi (executor `default`) is a thinking-mode upstream that 400s with "reasoning_content must be passed back" when a prior assistant turn lacks it; the placeholder injection was only wired into the OpenCode meta-provider, so direct multi-turn Kimi conversations failed. Scoped to `kimi` (gateway-served models matching the thinking-model name pattern are unaffected). Regression guard: `tests/unit/kimi-native-reasoning-injected-1480.test.ts`. (thanks @2220258345)
+- **fix(executors):** recover from a strict gateway's `context_management: Extra inputs are not permitted` 400 (port from 9router#1468, [#6420](https://github.com/diegosouzapw/OmniRoute/pull/6420)). **Claude Code** always sends a top-level `context_management` field; strict anthropic-compatible gateways reject it. The dedicated context-editing 400-fallback only fired when OmniRoute's own `contextEditing` feature was enabled (default off), so a client-sent field passed through untouched and 400'd. `context_management` is now in the generic reactive field-strip list, so it's stripped-and-retried once regardless of the feature flag (with correct request re-signing for claude-compatible relays). Regression guard: `tests/unit/provider-field-strips.test.ts`. (thanks @ohahe52-dot)
+- **fix(network):** enable **RFC 8305 Happy Eyeballs** (`autoSelectFamily`) on the direct-egress undici dispatcher (port from 9router#1237, [#6423](https://github.com/diegosouzapw/OmniRoute/pull/6423)). When DNS returns both IPv6 (AAAA) and IPv4 (A) and the IPv6 route is broken (e.g. a NAT64 `64:ff9b::` prefix without routing), undici tried IPv6 first and hung until `ETIMEDOUT` (then a 502 + account lockout), even though `curl` reached the same host. The direct dispatcher now races both families and uses whichever connects first. Proxy paths pin family via `proxyTls` and are unaffected. Regression guard: `tests/unit/direct-dispatcher-pipelining-4580.test.ts`. (thanks @adentdk)
+- **fix(combo):** round-robin now advances the rotation pointer past the model that **actually served**, not the eagerly-scheduled one (port from 9router#948, [#6428](https://github.com/diegosouzapw/OmniRoute/pull/6428)). With `stickyLimit: 1` (true round-robin), when the scheduled model failed and a _different_ model served via fallback, the counter had already advanced +1 from the scheduled index — so the next request reused the fallback-served model, degrading round-robin into hot-spotting on whichever model was healthy. The pointer now advances to the served index + 1 (mirroring the sticky-limit>1 path). Session-stickiness (#3825) and distribution are preserved. Regression guard: `tests/unit/combo-rr-fallback-advance-948.test.ts`. (thanks @binsarjr)
+- **fix(sse):** a non-string `model` field is now rejected with a `400` before the resolver, instead of crashing downstream `.toLowerCase()`/`.split()` calls into an empty-body `500` that escapes the error sanitizer ([#6407](https://github.com/diegosouzapw/OmniRoute/issues/6407)). Regression guard: `tests/unit/chat-non-string-model-6407.test.ts`. (thanks @chirag127)
+- **fix(api):** unknown `/api/*` routes now return a JSON `404` (instead of the dashboard HTML shell) and scalar chat params (`model`/`temperature`/etc.) are validated **before** the provider lookup so malformed requests fail fast with a clear `400` ([#6424](https://github.com/diegosouzapw/OmniRoute/issues/6424), [#6412](https://github.com/diegosouzapw/OmniRoute/issues/6412)). Regression guards: `tests/unit/api/api-catchall-json-404.test.ts`, `tests/unit/chat-early-schema-validation-6412.test.ts`. (thanks @chirag127)
+- **fix(api):** `/v1/chat/completions` now rejects a non-JSON `Content-Type` with a `400` before parsing the body ([#6414](https://github.com/diegosouzapw/OmniRoute/issues/6414)). Regression guard: `tests/unit/v1-chat-completions-content-type-6414.test.ts`. (thanks @chirag127)
+- **fix(api):** the `X-OmniRoute-Compression` response header is now echoed on `/v1/chat/completions` and `/v1/completions` ([#6422](https://github.com/diegosouzapw/OmniRoute/issues/6422)). Regression guard: `tests/unit/compression-header-echo-6422.test.ts`. (thanks @chirag127)
+- **fix(api):** concurrent `GET /v1/models` requests are coalesced into a single catalog build ([#6408](https://github.com/diegosouzapw/OmniRoute/issues/6408)). Regression guard: `tests/unit/v1-models-concurrent-6408.test.ts`. (thanks @chirag127)
+- **fix(api):** `/v1/completions` now echoes the requested `body.model` in its JSON + streamed responses ([#6429](https://github.com/diegosouzapw/OmniRoute/pull/6429)). Regression guard: `tests/unit/completions-body-model-echo.test.ts`. (thanks @chirag127)
+- **fix(api):** env-var master keys now see the full `/v1/models` catalog ([#6406](https://github.com/diegosouzapw/OmniRoute/issues/6406)). Regression guard: `tests/unit/models-catalog-envkey-6406.test.ts`. (thanks @chirag127)
+- **fix(api):** non-streaming `/v1/completions` responses now echo `body.model` aligned with the `X-OmniRoute-Model` header ([#6426](https://github.com/diegosouzapw/OmniRoute/issues/6426)). Regression guard: `tests/unit/v1-completions-model-header-match-6426.test.ts`. (thanks @chirag127)
+- **fix(api):** unknown `/v1/*` routes now return a JSON `404 not_found` instead of the Next.js dashboard HTML shell ([#6405](https://github.com/diegosouzapw/OmniRoute/issues/6405)). Regression guard: `tests/unit/api/v1-catchall-json-404.test.ts`. (thanks @chirag127)
+- **fix(api):** the per-connection provider models route now degrades to the shipped catalog when a provider's `/models` endpoint answers with a redirect ([#6267](https://github.com/diegosouzapw/OmniRoute/issues/6267)) — a `qwen-web` import failed with a raw `Redirect blocked … (307)` 503. `safeOutboundFetch` throws `REDIRECT_BLOCKED` on the 307, `getSafeOutboundFetchErrorStatus` maps it to 503, and `buildDiscoveryErrorFallbackResponse` treated every 503 as a hard error — so the non-empty `getModelsByProviderId("qwen-web")` catalog was never surfaced. A models-endpoint redirect is not a fixable-config error (unlike `URL_GUARD_BLOCKED`/`INVALID_URL`, which stay hard errors), so it now falls back to the local/cached catalog before the 503 short-circuit. General fix — covers any config-driven provider that 307s. Regression guard: `tests/unit/provider-models-qwen-web-redirect-6267.test.ts`. (thanks @chirag127)
+- **fix(api):** the per-connection provider models route (MCP `list_models_catalog` + the dashboard import view) now merges USER-ADDED custom models into its response ([#6247](https://github.com/diegosouzapw/OmniRoute/issues/6247)) — custom models live in the `key_value` namespace `customModels`, which the live REST `/api/v1/models` already merges, but `src/app/api/providers/[id]/models/route.ts` never read `getCustomModels`, so custom models were dropped on both the discovery-success and local_catalog paths. They are now folded into the returned model list (deduped by id, stamped `owned_by: provider`), fixing MCP + the dashboard import view in one place. Regression guard: `tests/unit/provider-models-custom-merge-6247.test.ts`. (thanks @RCrushMe)
+
+- **fix(providers):** GitLab Duo tool-calling follow-up turns no longer fail upstream with `422 {"detail":"Validation error"}` (tokens 0/0, rejected pre-inference). The #6234 tool-result-feedback fix serialized the **entire** multi-turn conversation into GitLab's single-file `code_suggestions` (`small_file`) generation endpoint — folded history that turn-N sent as an oversized `current_file.content_above_cursor` **and** duplicated verbatim into `user_instruction`, tripping the AI-Gateway's `small_file` validation guard. The executor now **bounds** that prompt: it keeps system + latest user message + the most-recent tool round (dropping older turns), caps oversized tool results, and stops duplicating the full prompt into `user_instruction` (which now carries only the short latest user message) — while still feeding the most-recent tool result back so the agent continues (`open-sse/executors/gitlab.ts`, [#6220](https://github.com/diegosouzapw/OmniRoute/issues/6220)). The unit test covers the bounding logic; the upstream 422→200 clearing is VPS-only (Hard Rule #18). Regression guard: `tests/unit/gitlab-tool-exchange-bounded-6220.test.ts`.
+
+- **fix(i18n):** the provider-detail (`/dashboard/providers/[id]`) connection-status filter labels no longer render as `__MISSING__:All` / `__MISSING__:Active` / `__MISSING__:Error` / `__MISSING__:Banned` / `__MISSING__:CreditsExhausted` in non-English locales (notably pt-BR) ([#6290](https://github.com/diegosouzapw/OmniRoute/issues/6290)). Root cause was **not** the namespace mismatch the issue guessed — the `providers.filter*` keys resolve correctly in `en.json`; the debt lived in the locale mirrors (`src/i18n/messages/*.json`), where these five keys carried the `__MISSING__:` sync sentinel in ~15 locales and were absent entirely in ~26 others, so next-intl found the key and echoed the sentinel verbatim. All 40 non-English/-Chinese mirrors now ship real translations for the five `providers.filter*` labels. Regression guard: `tests/unit/i18n-provider-filter-keys-6290.test.ts`. (thanks @diegosouzapw)
+
+- **fix(providers):** the `copilot-m365-web` streaming executor now emits `debug`-level WebSocket diagnostics ([#6210](https://github.com/diegosouzapw/OmniRoute/issues/6210)) — the outbound WS URL (with the `access_token` **redacted** via `redactWsUrl()`), handshake success/failure, and each received SignalR frame's `type`/`target`. Previously the streaming path logged nothing, so an empty `content:null` response (the M365 Education / Starter tier symptom fixed in #6234) was undiagnosable even at `APP_LOG_LEVEL=debug`. The change is debug-level and side-effect-free — it does not alter streaming behavior or the frame parser, and the token never reaches the logs. Regression guard: `tests/unit/copilot-m365-web-logging-6210.test.ts` (thanks @qpeyba)
+
+- **fix(resilience):** a round-robin combo no longer returns `503 all upstream accounts are unavailable` when a compatibility-rejected target is actually healthy ([#6238](https://github.com/diegosouzapw/OmniRoute/issues/6238)). `filterTargetsByRequestCompatibility` drops request-incompatible targets (tool/vision/structured-output unsupported, or below the required context window) **before** any availability check runs, and its `compatible.length === 0` safety net only fired when _all_ targets were filtered — not when the kept targets later all turned out runtime-unavailable (circuit-open / cooldown / no credentials). So a combo could 503 while a compat-rejected-but-healthy provider sat unused. `handleRoundRobinCombo` now keeps the compat-rejected set and, when every compat-kept target was skipped without a single real attempt, probes those rejected targets as a **last-resort fallback tier** (via the new pure `open-sse/services/combo/comboCompatFallback.ts`) before crystallizing the 503. Regression guard: `tests/unit/combo-roundrobin-compat-fallback-6238.test.ts`. (thanks @ThongAccount)
+
+- **fix(startup):** best-effort self-heal for a corrupted Turbopack dev cache on Windows ([#6289](https://github.com/diegosouzapw/OmniRoute/issues/6289)). On Windows, `pnpm dev` can fail at startup when Turbopack `mmap`s a persistent-cache SST file and the OS refuses the mapping (`os error 1455` — "paging file too small"), which Turbopack surfaces as a misleading `Module not found: Can't resolve '@/shared/utils/machine'`. This is a **known upstream Turbopack cache-corruption bug — not our code**. The dev launcher (`scripts/dev/run-next.mjs`) now wraps `nextApp.prepare()` and, when it rejects with that signature (`isTurbopackCacheCorruption` in the new `scripts/dev/turbopackCacheHeal.mjs`), purges `.build/next/**/cache/turbopack` and retries **once** with a clear log. **Caveat — best-effort only:** the corruption often surfaces as a runtime overlay rather than a `prepare()` rejection, so this cannot always intercept it; the reliable remedy remains manually deleting the Turbopack cache dir. Regression guard: `tests/unit/turbopack-cache-heal-6289.test.ts`. (thanks @chirag127)
+
+- **fix(providers):** qodercli PAT auth no longer fails with `spawn qodercli ENOENT` on Windows ([#6263](https://github.com/diegosouzapw/OmniRoute/issues/6263)) — `spawnQoderCli` spawned the bare `qodercli` name with `shell:false` and an unenriched env, so the npm `.cmd` wrapper under `%APPDATA%\npm` (a user-PATH directory) was never resolved. It now resolves the absolute `.cmd`/`.exe` path through the existing `getCliRuntimeStatus("qoder")` resolver in `src/shared/services/cliRuntime.ts` (memoized), spawns with `shell` when the target is a `.cmd`/`.bat`, and uses the cliRuntime-enriched env (PATH + PATHEXT + APPDATA); the ENOENT error now lists the searched paths plus the `CLI_QODER_BIN` override. End-to-end spawn on a real Windows host is host-only (Hard Rule #18); the path-resolution logic is unit-tested. Regression guard: `tests/unit/qodercli-windows-resolve-6263.test.ts`. (thanks @chirag127)
+
+- **fix(sse):** the reasoning-token buffer no longer inflates **probe-sized `max_tokens`** ([#6274](https://github.com/diegosouzapw/OmniRoute/issues/6274)) — Claude Code's `/model` capability check sends `max_tokens: 1`, but for a thinking-capable model with a large output cap (e.g. `glm-5.2`) the #3587 headroom heuristic (`max(current + 1000, ceil(current * 1.5))`) rewrote it to `1001` and forwarded that upstream, wasting tokens on a request that was never a genuine reasoning budget. `resolveReasoningBufferedMaxTokens()` (`open-sse/services/reasoningTokenBuffer.ts`) now short-circuits and returns the caller's value verbatim when it is below the new `REASONING_BUFFER_MIN_TRIGGER` (256) threshold — a tiny explicit limit is a probe, not a reasoning request. Real budgets still receive the #3587 headroom unchanged, and the guard runs after the existing capability checks so unknown / non-reasoning models keep returning `null`. Regression guard: `tests/unit/reasoning-token-buffer-6274.test.ts`. (thanks @brightfiscalband)
+
+- **fix(cli):** `omniroute reset-password` now works as a real subcommand, and password resets over piped (non-TTY) stdin actually apply ([#6261](https://github.com/diegosouzapw/OmniRoute/issues/6261), [#6258](https://github.com/diegosouzapw/OmniRoute/issues/6258)). Two coupled defects: (1) **#6261** — `bin/omniroute.mjs` routed everything through Commander with only two pre-Commander bypasses (`--mcp`, `reset-encrypted-columns`), so `omniroute reset-password` was rejected as an unknown command; only the separate `omniroute-reset-password` bin worked, while the docs falsely advertised the subcommand (incl. a bogus "legacy alias still works"). A pre-Commander bypass mirroring `reset-encrypted-columns` now dynamically imports `bin/reset-password.mjs` (which self-executes) before Commander parses; the three doc lines were corrected. (2) **#6258** — `bin/reset-password.mjs` issued two sequential `rl.question` prompts; under piped stdin the second read never settled at EOF, so `main()` never reached `resetManagementPassword` and the reset was a silent no-op (both prompts printed, no success, password unchanged). The CLI now detects non-TTY stdin and reads it once (first line = password, second line = confirm if present, else reused), adds a `--password-stdin` flag (entire stdin is the password, no confirmation), and exits `0` explicitly so the success line always flushes; interactive TTY behavior is unchanged. Regression guard: `tests/unit/reset-password-cli-6261-6258.test.ts` (3). (thanks @chirag127)
+
+- **fix(db):** the mass-migration **safety abort** now tells the operator how to bypass it and stops flooding the log ([#6260](https://github.com/diegosouzapw/OmniRoute/issues/6260)) — after restoring a backup that wiped the migration tracking table, `runMigrations()` threw the abort on every downstream `ensureDbInitialized()`, re-logging the full banner 11+ times, and the message never mentioned the existing `OMNIROUTE_MAX_PENDING_MIGRATIONS` escape hatch. The abort text now appends a bypass hint (set `OMNIROUTE_MAX_PENDING_MIGRATIONS=0` in `server.env` / `DATA_DIR/.env`), and a new `MigrationSafetyAbortError` is memoized so repeated calls in the same process throw the same instance and emit a single concise line instead of the full cascade. Regression guard: `tests/unit/migration-safety-abort-6260.test.ts`. (thanks @chirag127)
+
+- **fix(auth):** importing a **distinct** Codex/ChatGPT OAuth `auth.json` is no longer falsely rejected as "already exists" when it belongs to a different user in the same workspace ([#6301](https://github.com/diegosouzapw/OmniRoute/issues/6301)). `findExistingCodexConnection` (in `src/lib/oauth/utils/codexAuthImport.ts`) deduped **only** on `providerSpecificData.workspaceId === accountId`, where `accountId` is the shared `chatgpt_account_id`/`tokens.account_id` — so two members of the same ChatGPT Team collapsed onto a single connection (409 `duplicate_account`). The id_token's `https://api.openai.com/auth` claim carries a per-user `chatgpt_user_id` alongside the workspace id (the device-flow path already persisted it as `chatgptUserId`, but the import path did not). Now `parseAndValidateCodexAuth` extracts `userId` (`chatgpt_user_id` → `user_id` → JWT `sub`) into `ParsedCodexAuth`, the create/update paths persist `chatgptUserId` in `providerSpecificData` (mirroring `codex.ts`), and dedup keys on `workspaceId` **AND** `chatgptUserId` — with a backward-compat fallback to legacy accountId-only matching when no stored connection for that workspace records a `chatgptUserId`, so genuinely-same accounts still dedup. Regression guard: `tests/unit/codex-auth-import-userid-dedup-6301.test.ts` (4). (thanks @anungma)
+
+- **fix(providers):** importing models for the **venice-web** provider no longer fails with a red "Provider venice-web does not support models listing" ([#6269](https://github.com/diegosouzapw/OmniRoute/issues/6269)). `venice-web` is a web-cookie provider with an executor but no upstream `/v1/models` endpoint and no registry `models`, so the models route fell through to the tail `400`. Mirroring the `jules`/`linkup-search`/`ollama-search` fix (#5569), it now ships a static local catalog entry in `src/lib/providers/staticModels.ts` — seeding the current Venice lineup (`venice-uncensored`, `llama-3.3-70b`, `qwen3-235b`, `qwen3-4b`, `deepseek-r1-671b`; Venice rotates its catalog, see docs.venice.ai/models/overview) — so the route returns `200` with `source:"local_catalog"`, `intentional:true`. Regression guard: `tests/unit/static-models-venice-web-6269.test.ts`. (thanks @chirag127)
+
+- **fix(api):** the specialty model catalogs (`/v1/embeddings`, `/v1/images`, `/v1/music`, `/v1/videos` model lists) are now derived from the **unified catalog filtered by a predicate** (`getSpecialtyModelsResponse`) instead of ad-hoc per-route logic, so they consistently respect active-credential visibility and stay in sync with the main catalog ([#6303](https://github.com/diegosouzapw/OmniRoute/pull/6303)). Regression guard: `tests/unit/specialty-model-catalog-routes.test.ts`. (thanks @makcimbx)
+- **fix(api):** the agent-bridge server route now resolves the MITM manager via a **dynamic `import("@/mitm/manager.runtime")`** so Turbopack does not statically pull the stub (or over-bundle the manager), and the agent-skills generator anchors its output base path with `path.join(process.cwd(), …)` so Turbopack's static analyzer stops tracing the whole project root ([#6329](https://github.com/diegosouzapw/OmniRoute/issues/6329), [#6366](https://github.com/diegosouzapw/OmniRoute/pull/6366)). Regression guard: `tests/unit/agent-bridge-server-route-dynamic-import.test.ts`. (thanks @Iammilansoni)
+- **fix(api):** internal probes (combo-test, cloud-sync verify) now pick a **management-scoped / allow-all API key** instead of naively grabbing `getApiKeys()[0]` — a restricted `self:usage` first row made the probe fail with "Model X is not allowed for this API key" even when the combo path was healthy (`pickApiKeyForInternalUse` in `src/lib/db/apiKeys.ts`). The API-manager model editor also falls back to `/api/models?all=true` when `/v1/models` is catalog-protected ([#6372](https://github.com/diegosouzapw/OmniRoute/pull/6372)). Regression guard: `tests/unit/pick-internal-api-key-6372.test.ts`. (thanks @jmengit)
+- **fix(live-ws):** the Live Dashboard WebSocket server now **rejects on bind failure** (e.g. `EADDRINUSE` when the API bridge already holds the port) instead of letting the error surface as an unhandled `error` event that crash-loops the process — the `error` listener is attached to `wss` (not `server`) and releases the EventBus subscription on a failed start ([#6324](https://github.com/diegosouzapw/OmniRoute/issues/6324)). Regression guard: `tests/unit/live-ws-eaddrinuse-6324.test.ts`. (thanks @vinayakkulkarni)
+- **fix(dashboard):** the Home provider-topology widget now trusts the live provider-metrics snapshot — it uses `topology.errorProvider` and live `activeRequests` directly instead of re-deriving state from a stale `lastErrorAt` or applying a frontend timeout filter, so the topology reflects real-time provider health ([#6322](https://github.com/diegosouzapw/OmniRoute/pull/6322)). Regression guard: `tests/unit/home-provider-topology-live-state.test.ts`. (thanks @xz-dev)
+- **fix(sse):** strip zero-width markers from streamed **tool-call arguments** — a follow-up to [#5857](https://github.com/diegosouzapw/OmniRoute/pull/5857). That PR removed injected zero-width joiners (U+200D) from streamed assistant text/reasoning but deliberately left tool-call argument JSON byte-exact. The request-side obfuscation (`open-sse/services/claudeCodeObfuscation.ts`) injects ZWJ into agent words — including the temp path inside the Bash tool description — and Claude models copy that verbatim into generated commands, which are delivered as tool-call arguments rather than assistant text. As a result the ZWJ survived and corrupted code blocks (e.g. a temp path rendered with an invisible joiner). Now `open-sse/handlers/responseSanitizer.ts` strips zero-width code points from tool-call argument strings at every emit site (OpenAI non-stream/stream chat `tool_calls` + legacy `function_call`, native Responses `function_call` items, the OpenAI→Responses conversion, and the native Responses streaming `response.function_call_arguments.delta/.done` events). Only zero-width code points are removed; JSON structure and all other bytes stay identical (no parse/restringify), so normal arguments remain byte-exact. Regression guard: 6 new cases in `tests/unit/response-sanitizer.test.ts` (suite 50/50).
+
+- **fix(nodejs):** the default app log path now resolves under `DATA_DIR` (`~/.omniroute/logs/application/app.log`) instead of `process.cwd()` ([#6197](https://github.com/diegosouzapw/OmniRoute/issues/6197)) — the globally-installed CLI runs from an arbitrary working directory, so anchoring the default to cwd made file logging silently write to (or no-op under) an unrelated directory, contradicting the documented `.env.example` default. `getAppLogFilePath()` now computes the default lazily via the pure `resolveDataDir()` resolver (honours a per-process `DATA_DIR`, no directory-creation side effect); an explicit `APP_LOG_FILE_PATH` still wins. Regression guard: `tests/unit/logenv-datadir-path-6197.test.ts` (3).
+
+- **fix(docker):** AgentBridge/`startMitm` no longer aborts in containers/headless when the Antigravity-default DNS step can't write `/etc/hosts` ([#6127](https://github.com/diegosouzapw/OmniRoute/issues/6127)), and the privileged command's stderr now reaches `app.log` instead of only a bare exit code hitting the toast ([#6198](https://github.com/diegosouzapw/OmniRoute/issues/6198)). The default DNS step (`addDNSEntry`) was called unguarded while cert install and the two sibling DNS steps were each best-effort — in the runtime Docker image (`USER node`, no `sudo`, read-only `/etc/hosts`) it threw `Command failed with code 1` out of `startMitmInternal` and killed the whole start, discarding the stderr. The three DNS steps are extracted into a best-effort `provisionDnsEntries()` where each failure is logged with the full `err` (stderr included, folded in by `systemCommands.ts`) and never aborts the start. Regression guard: `tests/unit/mitm-dns-graceful-degrade-6127.test.ts` (4).
+
+- **fix(providers):** copilot-m365-web now supports the M365 Education "Starter / OfficeWebIncludedCopilot" tier and no longer returns an empty `content:null` stream ([#6210](https://github.com/diegosouzapw/OmniRoute/issues/6210)). Two gaps: (1) `buildWsUrl()` hardcoded the individual-consumer scenario (`OfficeWebPaidConsumerCopilot`, `isEdu=false`) — the EDU tier is now opt-in via `providerSpecificData.tier="edu"`, emitting `scenario=OfficeWebIncludedCopilot`/`isEdu=true` (the individual path is unchanged); (2) the EDU/GPT-5.5 path streams deltas via `arguments[0].writeAtCursor` (incremental) instead of only `messages[].text` (accumulated snapshots), which the parser dropped — a new `accumulateBotContent()` folds both formats, with `type:2 item.result.message` as a last-resort fallback. Regression guard: `tests/unit/copilot-m365-edu-writeatcursor-6210.test.ts` (10). (thanks @qpeyba)
+
+- **fix(providers):** GitLab Duo executor now feeds tool results back into the prompt instead of looping ([#6220](https://github.com/diegosouzapw/OmniRoute/issues/6220)) — `buildPrompt()` branched only on `system`/`user` and took `userParts.at(-1)`, silently dropping the `assistant{tool_calls}` + `tool{result}` turns the client appended, so the reconstructed prompt was byte-identical to turn 1 and the model re-emitted the same `<tool>` call forever. When a tool exchange is present the full conversation is now serialized, folding each tool result back keyed by its `tool_call_id`; simple conversations keep the legacy shape. Complements the tool_call emission from [#6051](https://github.com/diegosouzapw/OmniRoute/issues/6051) (the `kilo-duplicate` label was a false positive — different, sequential defect). Regression guard: `tests/unit/gitlab-tool-result-feedback-6220.test.ts` (4).
+
+- **fix(providers):** opencode-go/opencode-zen can now synthesize the OpenCode CLI identity headers Cloudflare requires on VPS egress ([#5997](https://github.com/diegosouzapw/OmniRoute/issues/5997)) — on a datacenter VPS, `opencode.ai/zen/go/v1/chat/completions` 403s (HTML challenge) requests lacking CLI identity, while the reporter's control curl proved that `User-Agent: opencode-cli/1.0.0` + `x-opencode-client: cli` + `x-opencode-project: default` + fresh request/session UUIDs succeed. Opt-in via `OPENCODE_SYNTHESIZE_CLI_HEADERS=true` (values overridable via `OPENCODE_GO_USER_AGENT`/`OPENCODE_USER_AGENT`/`OPENCODE_CLIENT`/`OPENCODE_PROJECT`); it fills only headers the client did not already send. Kept **off by default** — the forward-only path is deliberate (fabricating a wrong value risks upstream rejection; a prior dedup regressed with `opencode/local`), so this replaces the fragile local header-injection shim without changing default behavior. Regression guard: `tests/unit/opencode-cli-headers-synthesis-5997.test.ts` (6). (thanks @aleksesipenko)
+- fix(resilience): sticky session affinity now evicts and fails over to another account when the pinned account is exhausted/unavailable (#6219)
+
+- fix(sse): Responses API passthrough now drops internal commentary-phase output before forwarding to clients (gated by RESPONSES_PASSTHROUGH_DROP_COMMENTARY, default on) (#6199)
+
+- **fix(sse):** tool-call function schemas with a root `type: null` are now coerced to `type: "object"` before dispatch ([#6359](https://github.com/diegosouzapw/OmniRoute/issues/6359)) — clients like the Codex app emit `parameters: { type: null, ... }` for some tools, which OpenAI-compatible upstreams reject with `400 Invalid schema for function '...': schema must be a JSON Schema of 'type: "object"', got 'type: null'`, failing the whole request. `toolSchemaSanitizer` already stripped the null; it now re-adds the mandatory root `"object"` type (and empty `properties`/open `additionalProperties` when absent). Combinator roots (`anyOf`/`oneOf`/`allOf`) and explicit root types are left untouched. Regression guard: 5 new cases in `tests/unit/tool-schema-sanitizer.test.mjs`.
+
+- **fix(docker):** AgentBridge no longer fails to start on npm/Electron/VPS installs with "MITM manager stub reached at runtime" ([#6344](https://github.com/diegosouzapw/OmniRoute/issues/6344)) — v3.8.45 flipped the production bundler default to Turbopack, but `next.config.mjs` aliased `@/mitm/manager` to its Docker-only degraded stub **unconditionally**. That was harmless while Docker (which sets the alias intentionally for #3390 graceful degradation) was the sole Turbopack consumer, but once every artifact built with Turbopack the stub shipped to all non-Docker users and `startMitm` threw on the first Agent-Bridge start. The alias is now opt-in via `OMNIROUTE_MITM_STUB=1` (set only by the Dockerfile) through the shared `scripts/build/mitm-stub-flag.mjs` helper; default builds bundle the real manager. Regression guard: `tests/unit/mitm-stub-alias-6344.test.mjs` (4).
+
+- **fix(proxy):** stop the v3.8.44 proxy regression that leaked the real IP and disabled healthy proxies ([#6246](https://github.com/diegosouzapw/OmniRoute/issues/6246)). Two coupled defects from the new health scheduler: (1) **IP leak** — when a proxy assigned to a connection was marked `inactive`, resolution fell through to a **direct** egress instead of blocking, exposing the operator's real IP; (2) **over-deactivation** — the sweep flipped a proxy to `inactive` on the **first** failed probe and counted our own 5s timeout / a probe-target `5xx` as the proxy's fault, so healthy paid proxies vanished from egress selection ("my proxies are not being used anymore"). Fix: the sweep decision is extracted into a pure, network-free `decideProxyHealthAction` (`src/lib/proxyHealth/decision.ts`) — by default the health check now **only counts/logs and never downgrades status** (a proxy is downgraded/removed only with `PROXY_AUTO_REMOVE=true`, after `PROXY_AUTO_REMOVE_AFTER` **consecutive** conclusive failures); probes are classified tri-state so an inconclusive result (our timeout, or a `5xx` from the probe target) never penalizes the proxy, and the probe timeout is raised 5s→15s. Separately, `safeResolveProxy` now **fails closed** via the existing policy: a connection whose assigned proxy is dead is blocked instead of leaking direct (`hasBlockingProxyAssignment`), honoring the explicit `proxy off` toggles and the `PROXY_FAIL_OPEN=true` opt-out. Existing proxies stuck `inactive` by the old behavior need a one-time manual re-activate (the operator owns proxy status). Regression guards: `tests/unit/proxy-health-decide-action-6246.test.ts`, `tests/unit/proxy-assigned-unavailable-6246.test.ts`.
+
+- **fix(proxy):** make "Test All" read-only and add bulk enable/disable ([#6246](https://github.com/diegosouzapw/OmniRoute/issues/6246)). Complements the core fail-closed / scheduler fix (#6296) with the two remaining reporter asks. (1) The **"Test All" button** (`POST /api/settings/proxies/auto-test`) used to flip a proxy to `inactive` on a failed reachability probe; since the egress selector excludes `inactive` proxies, a flaky probe (an unreachable `httpbin.org`, a proxy that blocks `HEAD`, or a slow paid proxy) silently disabled every proxy that failed — "Test All" is now **read-only by default** (only the operator sets a proxy active/inactive; opt back into the legacy test-and-set with `PROXY_HEALTH_AUTO_DEACTIVATE=true`). (2) Adds a **bulk enable/disable** proxies endpoint + toolbar action (`POST /api/settings/proxies/batch-activate`) so an operator can re-activate proxies in one click. Regression guard: `tests/unit/proxy-health-6246.test.ts`. (thanks @tenshiak)
+
+- **chatcore (tools): stop the default 128-tool cap from silently dropping opencode's `task`/MCP tools.** opencode (used as an MCP/agent host) sends a large tool list; when it exceeds the speculative `MAX_TOOLS_LIMIT` (128) default, `truncateToolList` did a blind `tools.slice(0, 128)`, dropping every tool past index 128 — including opencode's built-in `task` tool (subagent launch) and many MCP tools, so models routed through OmniRoute could no longer spawn subagents or reach part of their tools. The cap exists to avoid upstream `400`s for providers with real hard limits (e.g. grok-cli 200), so it is kept for those: detection of the opencode client (`isOpencodeClient` — any `x-opencode-*` header, or `opencode` in the user-agent) now only bypasses the **speculative 128 default**, never a known provider ceiling. Precedence is explicit — a proactive/detected provider limit always truncates (even for opencode); otherwise opencode forwards its full tool list; otherwise the unchanged 128 default applies to every other client. Refactors `getEffectiveToolLimit` into `getKnownToolLimit(provider) ?? DEFAULT_LIMIT` (byte-identical for existing callers) and fixes a cosmetic debug-log that reported the truncated count instead of the original. Regression guard: `tests/unit/tool-limit-detector.test.ts`.
+
+- **fix(mitm):** the macOS MITM-cert install check now matches the system keychain again. `security find-certificate -a -Z` prints the SHA-1 as a colon-less hex string, but the installed-check compared it against `getCertFingerprint()`'s colon-separated form, so the substring match never hit — the cert was reported as not-installed and re-prompted for the sudo install on every run. Fingerprints are now normalized (colons stripped, upper-cased) on both sides via the extracted `macCertOutputHasFingerprint` helper. Regression guard: `tests/unit/mitm-cert-mac-fingerprint.test.ts`. ([#6204](https://github.com/diegosouzapw/OmniRoute/pull/6204), closes [#6134](https://github.com/diegosouzapw/OmniRoute/issues/6134) — thanks @rianonehub)
+
+- **fix(api):** `/v1/messages/count_tokens` now counts `tool_use`, `tool_result` and `thinking` content blocks (and array-form `system` prompts) in the local-estimation path, instead of only `text`. Real agentic conversations keep ~95% of their tokens inside tool results; the previous estimate returned near-zero for them, which silently broke Claude Code's auto-compaction (context grew past the window with no compaction until the upstream API rejected the request). The real provider-side count path is unchanged. Regression guard: `tests/unit/messages-count-tokens-route.test.ts`. ([#6221](https://github.com/diegosouzapw/OmniRoute/pull/6221) — thanks @luweiCN)
+
+- **fix(antigravity):** strip a trailing assistant prefill turn for Vertex Claude models to avoid upstream 400s ([#6114](https://github.com/diegosouzapw/OmniRoute/pull/6114)). Regression guard: `tests/unit/antigravity-claude-prefill-strip.test.ts`. (thanks @anki1kr)
+
+- **fix(security):** the mutable cloud-agent routes (`/api/cloud/credentials/update`, `/api/cloud/models/alias`) now require management auth instead of being treated as public. They were classified as public API routes, so a request without management credentials could update stored cloud-agent credentials and model aliases. They are removed from the public-route set, classified as management routes in the authz pipeline, and gated by `requireManagementAuth`; cloud **read**/auth routes stay public. Regression guards: `tests/unit/cloud-write-auth.test.ts`, `tests/unit/authz/classify.test.ts`, `tests/unit/public-api-routes.test.ts`. ([#6233](https://github.com/diegosouzapw/OmniRoute/pull/6233) — thanks @vittoroliveira-dev)
+
+- **refactor(dashboard):** extract the onboarding-wizard "Open provider details" link target into a pure, unit-tested `buildProviderDetailsHref(connection)` helper. The wizard already routes by `connection.id` (the node UUID) rather than the provider category slug (#6144/#6145); this hardens that behavior behind a tested helper that guards a missing id/connection. Regression guard: `tests/unit/provider-onboarding-href.test.ts`. ([#6166](https://github.com/diegosouzapw/OmniRoute/pull/6166) — thanks @KooshaPari)
+
+- **fix(security):** the doubao synthetic device-id generator now derives its digits via an unbiased crypto-random draw (rejection sampling over `crypto.randomBytes()`) instead of a `% 10` reduction, closing a CodeQL `js/biased-cryptographic-random` finding.
+- **fix(agentSkills):** the GitHub-skills generator now resolves `outputDir` to an absolute path before writing, fixing a regression introduced by #6366 (relative-to-cwd base path) that could write generated skill files to the wrong directory.
+- **fix(security):** `/api/keys/{id}/devices` now checks the HTTP method before auth/validation, returning a `405` for non-GET/DELETE verbs instead of a misleading `401`/`500` (closes a `dast-smoke` QUERY-method finding).
+- **fix(quality):** clear the last 2 heavy quality-gate reds on the release tip (cycle pre-flight).
+- **fix(mitm):** the test suite and CI can never mutate the OS trust store — `OMNIROUTE_SKIP_SYSTEM_TRUST=1` is set globally for tests/CI so `installCert`/`uninstallCert`/`installTproxyCa` skip the privileged OS dispatch ([#6310](https://github.com/diegosouzapw/OmniRoute/pull/6310); full detail is under the [3.8.45] section below — this branch received it via the parallel-cycle sync-back).
+- **fix(api):** `POST /api/github-skills` now Zod-validates its request body; documented the new quality-gate env vars and pinned the merge-integrity GitHub Actions to a commit SHA.
+- **fix(skills):** generate the missing `omni-github-skills` registry entry and align the agent-skills catalog-count tests (follow-up to #6186).
+- **fix(quality):** clear the cycle's 11 net-new ESLint errors and make `validate-release-green` suppressions-aware.
+
+### 📝 Maintenance
+
+- **i18n(it):** add 118 missing Italian (`it`) translations (net-additive — no existing keys dropped, valid JSON), improving Italian UI coverage. ([#6212](https://github.com/diegosouzapw/OmniRoute/pull/6212) — thanks @serverless83)
+- **chore(providers):** remove deprecated **MiMo V2** model entries from the catalogs (xiaomi-mimo, opencode-go, zenmux-free, audio TTS) — the upstream V2 line is superseded by MiMo V2.5; drops `mimo-v2-tts`, `mimo-v2-pro`, `mimo-v2-omni`, `mimo-v2-flash`, `mimo-v2-flash-free` and realigns the provider-catalog tests. ([#6248](https://github.com/diegosouzapw/OmniRoute/pull/6248) — thanks @backryun)
+- **chore(release):** ~50 commits on this branch are v3.8.45 pre-flight/hardening fixes and CI-perf work that landed here via the parallel-cycle sync-back (`sync-next-cycle.mjs`, Hard Rule #21) **after** the `v3.8.45` git tag was cut, and are already fully documented under the **[3.8.45]** section below — listed here only so the per-cycle commit-coverage check (`npm run release:uncovered`) doesn't flag them as gaps. Provider/catalog/UX/backend fixes: #6041, #6078, #6108, #6135, #6148, #6149, #6154, #6158, #6161, #6162, #6163, #6164, #6165, #6170, #6177, #6178, #6181, #6186, #6187, #6191, #6193, #6194, #6195, #6200, #6205, #6208, #6209, #6211, #6213, #6223, #6224, #6225, #6226, #6227, #6228, #6229, #6230, #6235, #6291, #6292. CI/release-pipeline work: #6167, #6203, #6214, #6215, #6218, #6273, #6275, #6283, #6284, #6285, #6300, #6305.
+- **chore(release):** additional zero-ref release-cycle plumbing on this branch, kept out of `release:uncovered` on purpose (no `#N` in the commit subject to cite): opening the v3.8.46 cycle, opening/closing the v3.8.45 cycle, the finalized [3.8.45] CHANGELOG i18n sync-back to 42 mirrors, the v3.8.45 cognitive/cyclomatic and file-size drift rebaselines, ESLint stale-suppression pruning (4,273 → 4,233), and clearing test-masking/docs-all pre-flight reds for v3.8.45.
+
+### ⚡ Performance & Infrastructure
+
+- **perf(release-green):** the pre-flight validator (`scripts/quality/validate-release-green.mjs`) now runs its 4 slow suites (unit / vitest / integration / pack-artifact) **concurrently** via `Promise.all` — pre-flight wall time drops from ~the sum of the suites to ~the slowest one (~30min saved per round; Phase 0 was the nº1 bottleneck of the v3.8.45 release benchmark, 2h54 of 6h34 e2e). Guard: `tests/unit/validate-release-green.test.ts` ("runs the slow suites CONCURRENTLY"). ([#6319](https://github.com/diegosouzapw/OmniRoute/pull/6319))
+- **fix(ci):** `scripts/release/sync-next-cycle.mjs` — two defects found live in its first production run (v3.8.45 Phase 5): (1) the `git()` helper's default 1 MiB `maxBuffer` crashed with `ENOBUFS` on `git show origin/main:CHANGELOG.md` (the CHANGELOG alone is >1 MiB) — widened to 64 MiB; (2) the i18n resync only propagated the `[NEXT]` (TBD) section, leaving the just-shipped finalized section as "— TBD" in all 42 mirrors — it now also syncs `[prevVersion]` bounded by the heading below it (new exported pure helper `versionAfter`). Guards: +5 tests in `tests/unit/sync-next-cycle.test.ts` (8/8). ([#6327](https://github.com/diegosouzapw/OmniRoute/pull/6327))
+- **test(ci):** concurrency-sensitive flaky tests are **quarantined into a serial pass** (`tests/unit/serial/`, `--test-concurrency=1`, appended to every unit runner incl. sharded variants — the serial pass is sharded too so concurrent shard jobs never self-collide). Initial set: `glm-coding-plan-monthly-3580`, `quota-division-blocks`, `provider-health-autopilot`, `combo-health-autopilot` — the class behind the ~28min CI wedges/re-runs (two live 1h+ wedges cancelled during this PR's own validation). Discovery + TIA gates track the new glob; systemic root cause (async logger writing after teardown) tracked in [#6360](https://github.com/diegosouzapw/OmniRoute/issues/6360). Guard: `tests/unit/test-serial-quarantine.test.ts` (4). ([#6347](https://github.com/diegosouzapw/OmniRoute/pull/6347))
+
+### 🙌 Contributors
+
+Thanks to everyone whose work landed in v3.8.46:
+
+| Contributor                                                    | PRs / Issues           |
+| -------------------------------------------------------------- | ---------------------- |
+| [@2220258345](https://github.com/2220258345)                   | direct commit / report |
+| [@abdofallah](https://github.com/abdofallah)                   | direct commit / report |
+| [@adentdk](https://github.com/adentdk)                         | direct commit / report |
+| [@aleksesipenko](https://github.com/aleksesipenko)             | direct commit / report |
+| [@anki1kr](https://github.com/anki1kr)                         | direct commit / report |
+| [@anungma](https://github.com/anungma)                         | direct commit / report |
+| [@arpicato](https://github.com/arpicato)                       | direct commit / report |
+| [@backryun](https://github.com/backryun)                       | #6248                  |
+| [@binsarjr](https://github.com/binsarjr)                       | direct commit / report |
+| [@brightfiscalband](https://github.com/brightfiscalband)       | direct commit / report |
+| [@chirag127](https://github.com/chirag127)                     | #6501, #6506           |
+| [@developerjillur](https://github.com/developerjillur)         | direct commit / report |
+| [@dilneiss](https://github.com/dilneiss)                       | #6499                  |
+| [@dtybnrj](https://github.com/dtybnrj)                         | direct commit / report |
+| [@Forcerecon](https://github.com/Forcerecon)                   | direct commit / report |
+| [@hao3039032](https://github.com/hao3039032)                   | direct commit / report |
+| [@Iammilansoni](https://github.com/Iammilansoni)               | #6150, #6245           |
+| [@jmengit](https://github.com/jmengit)                         | direct commit / report |
+| [@jordansilly77-stack](https://github.com/jordansilly77-stack) | direct commit / report |
+| [@JxnLexn](https://github.com/JxnLexn)                         | direct commit / report |
+| [@KooshaPari](https://github.com/KooshaPari)                   | #6166                  |
+| [@loopyd](https://github.com/loopyd)                           | direct commit / report |
+| [@luweiCN](https://github.com/luweiCN)                         | #6221                  |
+| [@makcimbx](https://github.com/makcimbx)                       | direct commit / report |
+| [@muflifadla38](https://github.com/muflifadla38)               | direct commit / report |
+| [@newnol](https://github.com/newnol)                           | direct commit / report |
+| [@ofekbetzalel](https://github.com/ofekbetzalel)               | direct commit / report |
+| [@ohahe52-dot](https://github.com/ohahe52-dot)                 | direct commit / report |
+| [@phidinhmanh](https://github.com/phidinhmanh)                 | direct commit / report |
+| [@powellnorma](https://github.com/powellnorma)                 | direct commit / report |
+| [@qpeyba](https://github.com/qpeyba)                           | direct commit / report |
+| [@RaviTharuma](https://github.com/RaviTharuma)                 | direct commit / report |
+| [@RCrushMe](https://github.com/RCrushMe)                       | direct commit / report |
+| [@rianonehub](https://github.com/rianonehub)                   | #6134, #6204           |
+| [@serverless83](https://github.com/serverless83)               | #6212                  |
+| [@swingtempo](https://github.com/swingtempo)                   | direct commit / report |
+| [@tenshiak](https://github.com/tenshiak)                       | direct commit / report |
+| [@ThongAccount](https://github.com/ThongAccount)               | direct commit / report |
+| [@UnrealAryan](https://github.com/UnrealAryan)                 | direct commit / report |
+| [@vinayakkulkarni](https://github.com/vinayakkulkarni)         | direct commit / report |
+| [@vittoroliveira-dev](https://github.com/vittoroliveira-dev)   | #6233                  |
+| [@warelik](https://github.com/warelik)                         | direct commit / report |
+| [@xxy9468615](https://github.com/xxy9468615)                   | direct commit / report |
+| [@xz-dev](https://github.com/xz-dev)                           | direct commit / report |
+| [@yanpaing007](https://github.com/yanpaing007)                 | direct commit / report |
+| [@diegosouzapw](https://github.com/diegosouzapw)               | maintainer             |
+
+---
+
 ## [3.8.45] — 2026-07-06
 
 ### ✨ New Features

docs/i18n/cs/CHANGELOG.md

@@ -6,6 +6,213 @@
 
 ## [3.8.31] — 2026-06-20
 
+## [3.8.46] — TBD
+
+### ✨ New Features
+
+- **feat(sse):** **hide paid-only models from `auto/*` routing** when `hidePaidModels` is on ([#6512](https://github.com/diegosouzapw/OmniRoute/issues/6512)) — follow-up to #6328/#6495. PR #6495 hid paid-only models from the `GET /v1/models` listing, but `auto/*` combos (`auto/best-coding`, `auto/glm`, …) could still pick a paid-only backend into their candidate pool → a 402/403 at request time. `createVirtualAutoCombo` now filters the candidate pool through the new pure `open-sse/services/autoCombo/paidModelFilter.ts` (`filterPaidOnlyCandidates`), applying the same free-model predicate #6495 uses in `catalog.ts` (`providerHasFreeModels(provider) && isFreeModel(provider, {id})`) whenever `settings.hidePaidModels === true`. Applied before the category/tier/family narrowing, so it covers every `auto/*` combo; an all-paid pool degrades to the existing graceful empty-pool path. **Opt-in — default OFF leaves the pool unchanged** (identity). Regression guard: `tests/unit/autoCombo/paid-model-filter-6512.test.ts` (4, incl. the default-off identity guard).
+- **feat(sse):** **provider-family auto combos** — `auto/glm`, `auto/minimax`, `auto/mimo`, `auto/zai`, `auto/gemma`, `auto/llama`, `auto/gemini` ([#6453](https://github.com/diegosouzapw/OmniRoute/issues/6453)) — new routable ids that materialize an on-demand virtual combo spanning whatever installed backends currently expose that model family, degrading gracefully as backends rotate. A new pure `open-sse/services/autoCombo/modelFamily.ts` (`detectModelFamily`) classifies by model-id prefix for six families; `zai` is instead resolved by provider id (z.ai's hosted API serves the same `glm-*` model ids as every other GLM backend, so `auto/zai` means "route to my z.ai backend specifically" vs `auto/glm`'s "any connected GLM backend"). Reuses the existing `createVirtualAutoCombo` on-demand materialization path (no DB writes) and the `/v1/models` catalog advertising loop. Regression guard: `tests/unit/autoCombo/provider-family-combos.test.ts` (11).
+- **feat(proxy):** native **proxy-pool round-robin / egress IP rotation** ([#6365](https://github.com/diegosouzapw/OmniRoute/issues/6365)) — a scope (global / provider / account) can now hold **multiple** proxies as a pool with a rotation strategy, so outbound requests cycle their egress IP instead of pinning one proxy per scope. Migration `117_proxy_pool_rotation.sql` lifts the `UNIQUE(scope, scope_id)` constraint (rebuild via the canonical rename/copy/drop; existing single assignments become 1-element pools) and adds a `proxy_scope_rotation` companion table holding the per-scope strategy + a persisted monotonic round-robin cursor. Strategies: `round-robin` (default, monotonic cursor — never `Math.random`), `random`, and `sticky-per-N-min`. Resolution (`resolveProxyForScopeFromRegistry` / `resolveProxyForConnectionFromRegistry`) now fetches the alive, position-ordered candidate set (unchanged `PROXY_ALIVE_PREDICATE`) and applies the strategy; an empty / all-dead pool still returns `null` — the #6246 fail-closed guard is **untouched** (never falls through to direct egress). Backend + DB only; dashboard pool-builder UI is a follow-up. Regression guard: `tests/unit/proxy-pool-rotation-6365.test.ts` (8, incl. fail-closed + backward-compat).
+- **feat(providers):** end-to-end **tool/function calling on the native Gemini `/v1beta` endpoint** ([#6222](https://github.com/diegosouzapw/OmniRoute/issues/6222)) — both directions of the Gemini↔OpenAI conversion now preserve tool data (previously silently dropped). Request side: `convertGeminiToInternal` (extracted to its own testable module) maps `tools[].functionDeclarations` → OpenAI `tools`, prior `functionCall` parts → assistant `tool_calls`, and `functionResponse` parts → `tool`-role messages. Response side: `convertOpenAIResponseToGemini` emits `parts[].functionCall {name,args}` from `message.tool_calls`, and the streaming `openAIChunkToGeminiChunk` accumulates fragmented `tool_calls` deltas by index into complete `functionCall` parts. The non-Gemini client paths (Claude, OpenAI-Responses) already preserved tool calls — this closes the gap specific to the native Gemini surface. Regression guard: `tests/unit/v1beta-gemini-tool-calling-6222.test.ts` (6, incl. a streaming SSE round-trip).
+- **feat(providers):** copilot-m365-web **enterprise / work tier** support ([#6334](https://github.com/diegosouzapw/OmniRoute/issues/6334)) — mirrors the EDU-tier pattern (#6210): `M365ConnectionParams` gains an `agent` field, a new opt-in `M365_ENTERPRISE_OVERRIDES` preset (`agent=work`, `scenario=officeweb`, `licenseType=Premium`) applies via `providerSpecificData.tier="enterprise"` (alias `"work"`), and `agent` is also overridable directly via `providerSpecificData.agent`. `buildWsUrl` was hardcoding `agent="web"` (the one enterprise-distinguishing param with no override path), so a Premium work account handshook then returned an empty stream. The individual and EDU paths are untouched. Kilo's dup flag vs #6210 (EDU tier) was a false positive — different tier. Regression guard: `tests/unit/copilot-m365-enterprise-6334.test.ts` (7). End-to-end confirmation on a real Premium work account is a live-VPS validation follow-up (Hard Rule #18). (thanks @Forcerecon)
+- **feat(api):** standardized, provider-agnostic **`effort` + `thinking` request params** ([#6241](https://github.com/diegosouzapw/OmniRoute/issues/6241)) — a thin standardization layer over the existing mature per-provider reasoning plumbing (no provider mapper touched). `providerChatCompletionSchema` gains a canonical `effort` (reusing the shared `none/low/medium/high/xhigh` vocabulary — the UI tiers `extra`/`max` collapse onto `xhigh`) and a boolean `thinking`. A pure `normalizeReasoningRequest` (wired once in `src/sse/handlers/chat.ts`, before any reasoning field is read) folds them onto the fields the translators already consume (`reasoning_effort` / `reasoning.effort` / `thinking`), so they fan out to Anthropic / Gemini / xAI / Responses — an explicit client `reasoning_effort` / object-shaped `thinking` always wins (backward-compatible). `/models` additively exposes `supportsThinking` + `effort_tiers` so the frontend can render the toggles (UI component is a follow-up). Regression guard: `tests/unit/effort-thinking-standardization-6241.test.ts` (12). (thanks @Iammilansoni, @shabeer)
+- **feat(combo):** new **`pipeline` (sequential) combo strategy** ([#6297](https://github.com/diegosouzapw/OmniRoute/issues/6297)) — the 18th routing strategy runs targets **in order**, threading each step's output into the next step's input, with an optional per-step `prompt` (system instruction); only the final step's response is returned. Distinct from `fusion` (parallel fan-out + judge). Implemented as a self-contained `open-sse/services/pipeline.ts` (sibling to `fusion.ts`), dispatched from `combo.ts`; the step list reuses `combo.models` order and reads an optional `prompt` off each target (backward-compatible — ignored by every other strategy). Intermediate steps run non-streaming with tools stripped (complete prose to thread forward); the final step keeps the client's `stream` flag + tools. A failing/empty/unparseable intermediate step fails the whole pipeline explicitly via a sanitized error (never silently swallowed). Kilo's dup flag vs #563 was a false positive (that's model→chain selection; this is a sequential chain). Regression guard: `tests/unit/combo-pipeline-strategy.test.ts` (5). (thanks @ofekbetzalel)
+- **feat(ci):** `check:test-masking` now flags **inline-reimplemented prod conditions** ([#6348](https://github.com/diegosouzapw/OmniRoute/issues/6348)) — a new **report-only** subcheck (v2, 6A.10 family) catches the wrong-shape contract test: a test that recomputes the condition under test inline instead of importing/exercising the real function (the #6216 class, where `=== 500` → `>= 500` stayed green because the test re-implemented the branch). For each added/modified test file it warns when the file textually duplicates a ≥3-token conditional from a production file touched in the same PR **and** does not import the symbol/module owning it, via a pure, fixture-tested `findReimplementedConditions()` with an allowlist mirroring `assertReductionAllowlist`. Report-only for now (does not fail the gate) — to be promoted to blocking after a triage cycle. Regression guard: `tests/unit/check-test-masking.test.ts` (45).
+- **feat(sse):** per-connection routing override (native vs CLIProxyAPI) ([#6339](https://github.com/diegosouzapw/OmniRoute/issues/6339)) — the previously-dead `isCliproxyapiDeepModeEnabled` helper is now wired into `resolveExecutorWithProxy`: a single connection can opt itself into the CLIProxyAPI passthrough executor via `providerSpecificData.cliproxyapiMode="claude-native"`, with precedence **connection override > provider `upstream_proxy_config` mode > default**. `resolveExecutorWithProxy` now receives the resolved connection's `providerSpecificData` (threaded from `chatCore.ts`), so one connection can deep-route while the provider's other connections stay native — no DB schema change (the toggle rides in `providerSpecificData`). Also resolves the same-provider mixing ask in #6340. Regression guard: `tests/unit/chatcore-executor-proxy.test.ts` (9). (thanks @RaviTharuma)
+- **feat(dashboard):** "Add session cookie" modal now shows a prominent **"Open ‹host› →"** link to the provider's own site ([#6268](https://github.com/diegosouzapw/OmniRoute/issues/6268)) — every `-web` cookie-session provider (chatgpt-web, claude-web, gemini-web, kimi-web, lmarena, qwen-web, m365-copilot-web, …) renders a one-click external link (opening the provider's login/home page in a new tab) so operators no longer tab away to retype the URL mid-setup. The host resolves from a pure, unit-tested `resolveWebProviderHost()` (prefers `WEB_COOKIE_PROVIDERS[id].website`, falls back to the registry `baseUrl` origin); non-web providers render exactly as before. Kilo's dup flag vs #6265 (modal-too-small-on-1080p) was a false positive — distinct concern. Regression guard: `tests/unit/resolve-web-provider-host.test.ts` (5). (thanks @chirag127)
+- **feat(providers):** add **DigitalOcean** AI (serverless inference) as an OpenAI-compatible API-key provider ([#6373](https://github.com/diegosouzapw/OmniRoute/pull/6373)) — base `https://inference.do-ai.run/v1`, wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/digitalocean/`, `src/shared/constants/providers/apikey/inference-hosts.ts`). Regression guard: `tests/unit/digitalocean-provider.test.ts`. (thanks @newnol)
+- **feat(providers):** add **Huancheng Public API** (`hcnsec`) as an OpenAI-compatible regional provider ([#6410](https://github.com/diegosouzapw/OmniRoute/pull/6410)) — Xinjiang Huancheng Cybersecurity's public LLM platform (base `https://api.hcnsec.cn/v1`, free credits via daily check-ins), wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/hcnsec/`, `src/shared/constants/providers/apikey/regional.ts`). Regression guard: `tests/unit/hcnsec-provider.test.ts`. (thanks @UnrealAryan)
+- **feat(dashboard):** the web-session credential guide now shows an **"Open {host}" link** ([#6316](https://github.com/diegosouzapw/OmniRoute/pull/6316)) to the provider's sign-in site (derived from the provider `website` via `getProviderWebsiteHost`), so you can jump straight to the page where the cookie/session must be captured. Regression guard: `tests/unit/web-session-provider-link-6316.test.ts`. (thanks @jordansilly77-stack)
+- **feat(cerebras):** add the **Gemma 4 31B** model (`gemma-4-31b`) to the Cerebras registry + pricing table ([#6331](https://github.com/diegosouzapw/OmniRoute/pull/6331)). Regression guard extends `tests/unit/t28-model-catalog-updates.test.ts`. (thanks @backryun)
+- **feat(providers):** add **Yuanbao (web)** as a cookie-session provider ([#6196](https://github.com/diegosouzapw/OmniRoute/issues/6196)) — `yuanbao-web` (Tencent Yuanbao, `yuanbao.tencent.com`) with cookie-only auth (`hy_user`/`hy_token` + public agent id), SSE→OpenAI translation incl. `reasoning_content`, exposing DeepSeek V3/R1 + Hunyuan / Hunyuan-T1. Regression guard: `tests/unit/providers-yuanbao-web.test.ts`. `together-web` was **deferred** (no verifiable web-session endpoint — needs a captured request) and `huggingchat-web` **dropped** (the existing `huggingchat` already is a web-cookie provider). (thanks @chirag127)
+- **feat(providers):** route the built-in **agentrouter** through the dynamic Claude-Code wire image ([#6056](https://github.com/diegosouzapw/OmniRoute/issues/6056)) — a small static allow-set (`CC_WIRE_IMAGE_BUILTINS` in `open-sse/services/ccWireImageBuiltins.ts`), consulted by `isClaudeCodeCompatible` / `isClaudeCodeCompatibleProvider` / `applyFingerprint`, makes agentrouter adopt the CC wire-image headers + fingerprint **while guarding the CC baseUrl/auth branches** so it keeps its own registry `baseUrl` and `x-api-key` auth. Regression guard: `tests/unit/agentrouter-cc-wire-image.test.ts` (asserts the wire image is applied AND agentrouter's baseUrl/auth are preserved). Live WAF-acceptance against agentrouter.org is a VPS validation follow-up (Hard Rule #18).
+- **feat(providers):** **bulk-add API keys for Cloudflare Workers AI** ([#6174](https://github.com/diegosouzapw/OmniRoute/issues/6174)) — `cloudflare-ai` is removed from the bulk-add exclusion list and the bulk parser gains a 3-field `name|accountId|apiKey` mode; the bulk route now builds a **per-entry** `providerSpecificData` so each key carries its own `accountId` (fixing the previous shared-object reuse), and both the create + key-validation paths receive it. Regression guard: `tests/unit/bulk-api-key-parser-cloudflare.test.ts`. (thanks @muflifadla38)
+- **feat(dashboard):** routing/settings UX clarity ([#6147](https://github.com/diegosouzapw/OmniRoute/issues/6147)) — (1) weighted combos show the **effective routing share %** next to each weight when weights don't sum to 100 (`WeightTotalBar.tsx`); (2) the status widget's user-facing **"Cloud Sync" label is renamed** to "Remote Settings Sync" (`CloudSyncStatus.tsx`; internal ids/state untouched); (3) built-in providers gain an **opt-in advanced base-URL override** (`isBaseUrlOverrideEligibleProvider`, hidden behind an "Advanced" toggle, reusing the existing `providerSpecificData.baseUrl` persistence — not globally widened). Regression guard: `tests/unit/routing-settings-ux-6147.test.ts`.
+- **feat(combo):** add an option to **disable session stickiness**, per-combo or globally — round-robin / random combos can rotate to a different connection on every request instead of pinning a whole conversation to one connection by its first-message hash. Resolution precedence per-combo `config.disableSessionStickiness` → global `settings.disableSessionStickiness` → default `false` (preserves the #3825 prompt-cache/504 fix); gates **both** stickiness call sites in `open-sse/services/combo.ts`. Exposed as a global toggle (Combo Defaults) and a per-combo Inherit/on/off control. ([#6168](https://github.com/diegosouzapw/OmniRoute/issues/6168)) Regression guard: `tests/unit/combo-disable-session-stickiness.test.ts`. (thanks @RCrushMe)
+- **feat(docker):** add the `OMNIROUTE_NO_SUDO` env flag for root-less / user-namespaced deployments — the MITM cert-trust command path (`resolveSudoSpawn` in `src/mitm/systemCommands.ts`) now strips the leading `sudo` when the flag is truthy, in addition to the existing root / sudo-missing cases, so the Proxy Agent runs without `sudo` (the operator trusts the CA manually, e.g. via `NODE_EXTRA_CA_CERTS`). Argv-array `spawn` preserved — no shell interpolation (Hard Rule #13). ([#6122](https://github.com/diegosouzapw/OmniRoute/issues/6122)) Regression guard: `tests/unit/mitm-systemCommands-no-sudo.test.ts`. (thanks @powellnorma)
+- **feat(providers):** add **Requesty** as an OpenAI-compatible gateway provider (BYOK, base `https://router.requesty.ai/v1`, ~200 free requests/day) — wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/requesty/`, `src/shared/constants/providers/apikey/gateways.ts`). ([#6120](https://github.com/diegosouzapw/OmniRoute/issues/6120)) Regression guard: `tests/unit/requesty-provider.test.ts`. (thanks @chirag127)
+- **feat(dashboard):** add **configured-only / available-only filters** to the Free Provider Rankings page ([#6150](https://github.com/diegosouzapw/OmniRoute/issues/6150)) — hide providers you haven't configured, or whose connections are all rate-limited / out of quota, via server-side query params (`?configuredOnly` / `?availableOnly` on `GET /api/free-provider-rankings`) backed by a testable lib helper reusing the in-process connection state (no Redis). Both filters default off, so the default view is unchanged; this supersedes the earlier client-side "Configured Only" toggle (#6245) with an available-only dimension and unit-tested logic. Regression guard: `tests/unit/freeProviderRankings-filters.test.ts`.
+
+### 🔧 Bug Fixes
+
+- **fix(dashboard):** adding a second API key connection for the same provider no longer silently overwrites the first — the Add-API-key modal now derives a unique default connection name (`main`, then `main-2`, `main-3`, …) so the backend name-based upsert can't collide (#6499 — thanks @dilneiss).
+- **fix(compression):** the session-dedup engine now also deduplicates a large multi-line block repeated **within a single message** (intra-message dedup), not just across turns; the compression-preview API surfaces a `fallbackReason`, and the fusion panel reports how many models were rate-limited vs failed on total-panel failure (#6501 — thanks @chirag127).
+- **fix(compression):** a stacked-pipeline step naming an unregistered engine now surfaces a `validationErrors` entry instead of silently no-op'ing, so misconfigured pipelines are visible in the preview API (#6506 — thanks @chirag127).
+- **feat(usage):** add a **Codex reset-credit redemption** flow to the Provider Limits UI ([#6361](https://github.com/diegosouzapw/OmniRoute/pull/6361)) — a `useCodexResetCreditRedemption` hook + `/api/usage/codex-reset-credit` route + `codexResetCredits` lib let you redeem banked Codex reset credits from the quota card. Regression guard: `tests/unit/codex-reset-credits.test.ts`. (thanks @JxnLexn)
+- **feat(glm):** add **team-plan quota settings** for `glm-cn` connections ([#6351](https://github.com/diegosouzapw/OmniRoute/pull/6351)) — a dedicated `GlmTeamQuotaFields` form section (team quota id / limits) threaded through the Add/Edit connection modals, persisted via `providerSpecificData`, with the GLM usage service reading the team quota. Regression guards: `tests/unit/glm-team-quota.test.ts`, `provider-specific-data-schema.test.ts`. (thanks @hao3039032)
+- **feat(providers):** add **TinyFish** web-fetch/search support ([#6349](https://github.com/diegosouzapw/OmniRoute/pull/6349)) — a `tinyfish-fetch` executor + `/v1/web/fetch` route + MCP web-fetch tool, registered as a specialty-media provider with request-validation and a search-provider catalog entry. Regression guards: `tests/unit/executor-tinyfish-fetch.test.ts`, `web-fetch-handler.test.ts`, `mcp-web-fetch-tool.test.ts`, `provider-validation-tinyfish.test.ts`. (thanks @dtybnrj)
+- **fix(cli):** `omniroute launch-codex` now spawns `codex.cmd` through a shell on Windows (the npm `.cmd` shim is unresolvable by bare `spawn` → ENOENT), mirroring the qodercli Windows fix (#6263) ([#6312](https://github.com/diegosouzapw/OmniRoute/pull/6312)). Regression guard: `tests/unit/launch-codex-windows-spawn-6312.test.ts`. (thanks @swingtempo)
+- **fix(codex):** isolate the **Spark** quota from the shared Codex quota and stabilize the quota UI ordering / hydration so per-scope limits render consistently ([#6336](https://github.com/diegosouzapw/OmniRoute/pull/6336)). Regression guards: `tests/unit/codex-quota-selection-hydration.test.ts`, `provider-limits-ui.test.ts` + 3 more. (thanks @xz-dev)
+- **feat(api):** add a `hidePaidModels` setting that filters paid-only models out of the `/v1/models` catalog. Regression guard: `tests/unit/models-catalog-hide-paid.test.ts`. (thanks @chirag127)
+- **fix(api-manager):** the fallback model picker now preserves combos instead of dropping them when a primary model is unavailable ([#6443](https://github.com/diegosouzapw/OmniRoute/pull/6443)). Regression guard: `tests/unit/api-manager-page-static.test.ts`. (thanks @jmengit)
+- **fix(providers):** recoverable Antigravity / Cloud-Code (Gemini Code Assist) `403` responses ([#6452](https://github.com/diegosouzapw/OmniRoute/pull/6452)) are now classified as a retryable project-config error instead of a terminal account ban, so a fixable project/API-disabled 403 no longer forces a ~1-year cooldown / full OAuth reconnect. Regression guard: `tests/unit/errorclassifier-antigravity-403.test.ts`. (thanks @developerjillur)
+- **fix(mitm):** `sanitizeHeaders` now redacts `Set-Cookie` response headers so upstream session cookies never leak into logs / diagnostics ([#6451](https://github.com/diegosouzapw/OmniRoute/pull/6451)). Regression guard: `tests/unit/mitm-sanitize-headers.test.ts`. (thanks @developerjillur)
+- **fix(api):** `/api/compression/preview` now accepts `mode: "caveman"` and correctly handles stacked / zero-compression previews ([#6425](https://github.com/diegosouzapw/OmniRoute/issues/6425)). Regression guard: `tests/unit/api/compression-preview-caveman-and-stacked-6425.test.ts`. (thanks @chirag127)
+- **feat(providers):** add **Zed** hosted LLM aggregator as a native-app provider ([#6118](https://github.com/diegosouzapw/OmniRoute/pull/6118)) — OAuth sign-in via the Zed hosted flow, registered through the shared provider registry + executor. Regression guards: `tests/unit/zed-oauth-provider.test.ts`, `zed-import-utils.test.ts`, `zed-docker-detect.test.ts`, `mitm-handler-zed.test.ts`. VPS-validated via live operator login (Hard Rule #18).
+- **fix(oauth):** the Kiro SSO-cache auto-import now **preserves the IDC region** — cross-region Amazon Q / Kiro profiles imported from the SSO cache are no longer collapsed to the default region ([#6113](https://github.com/diegosouzapw/OmniRoute/pull/6113)). Regression guard: `tests/unit/kiro-auto-import-idc-2059.test.ts`. VPS-validated via live operator login (Hard Rule #18).
+- **fix(dashboard):** passthrough model aliases no longer collide when two namespaced model ids share a last segment (port from 9router#1850, [#6431](https://github.com/diegosouzapw/OmniRoute/pull/6431)). `enx/gpt-5.5` and `enx/codebuddy/gpt-5.5` both auto-generated the alias `gpt-5.5`, so the second model could never be added (the UI just alerted "alias already exists"). Aliases are now disambiguated deterministically — bare last segment when free, then parent-qualified (`codebuddy-gpt-5.5`), then a numeric suffix — while re-adding the exact same model id is still blocked. Regression guard: `tests/unit/passthrough-alias-1850.test.ts`. (thanks @arpicato)
+- **fix(translator):** preserve a Gemini `functionResponse` co-located with other parts (another `functionCall`, or trailing `text`) in the same content when translating **Gemini → OpenAI** ([#6376](https://github.com/diegosouzapw/OmniRoute/pull/6376)). `convertGeminiContent()` early-returned the tool message on the first `functionResponse` part, dropping any co-located parts; such contents are now pre-split (one tool message per `functionResponse`, emitted first, plus one message for the remaining parts). Regression guard: `tests/unit/gemini-to-openai-function-response.test.ts`. (thanks @warelik)
+- **fix(headroom):** detect a python interpreter managed by **mise / pyenv / asdf / conda** (port from 9router#2353, [#6382](https://github.com/diegosouzapw/OmniRoute/pull/6382)). Headroom's python probe (`src/lib/headroom/detect.ts`) searched a hardcoded `PATH`, but version managers expose their interpreters via shim dirs that only join `PATH` through interactive-shell activation — which the non-interactive server never runs, so a managed python (≥3.10) was invisible and Headroom reported it missing. The search path now prepends the well-known shim/bin dirs (`~/.local/share/mise/shims`, `~/.pyenv/shims`, `~/.asdf/shims`, `$CONDA_PREFIX/bin`, `~/.local/bin`, respecting `MISE_DATA_DIR`/`PYENV_ROOT`/`ASDF_DATA_DIR` when set), and a new `HEADROOM_PYTHON` env override lets operators point straight at their interpreter (mirroring `HEADROOM_URL`). Still shell-free (`execFileSync`). Regression guard: `tests/unit/headroom-detect.test.ts` (5). (thanks @loopyd)
+- **fix(executors):** strip the OpenAI-Codex/Claude-CLI `client_metadata` passthrough field for **NVIDIA** requests (port from 9router#1887, [#6411](https://github.com/diegosouzapw/OmniRoute/pull/6411)). NVIDIA's OpenAI-compatible wrapper rejects it with `400 Unsupported parameter`, the same class already handled for `cerebras`/`mistral`; `nvidia` (executor `default`) was missing from the strip allowlist so Codex/Claude-Code passthrough requests 400'd. Regression guard: `tests/unit/executor-default-strip-client-metadata.test.ts` (+nvidia case). (thanks @phidinhmanh)
+- **fix(translator):** strip the Claude-style `thinking` field for **NVIDIA `z-ai/glm-5.2`** (port from 9router#2023, [#6413](https://github.com/diegosouzapw/OmniRoute/pull/6413)). NVIDIA's OpenAI-compatible wrapper 400s on `thinking` (a Claude-format client routed here leaves a `thinking:{type:"adaptive"}`); the existing strip rule only dropped `reasoning`. Same class already handled for `minimax-m2.7`. Regression guard: `tests/unit/nvidia-minimax-thinking-strip.test.ts` (+glm-5.2 case). (thanks @phidinhmanh)
+- **fix(translator):** suppress the streamed `</think>` close marker for the **Antigravity IDE** client (port from 9router#1061, [#6415](https://github.com/diegosouzapw/OmniRoute/pull/6415)). On thinking-only turns Antigravity rendered a bare `</think>` as the sole visible content, tripping its loop-detection and wasting requests. Antigravity's UA (`vscode/<v> (Antigravity/<v>)`) is added to the marker-suppress allowlist (alongside OpenCode); Claude Code / Cursor still get the marker, and `x-omniroute-thinking-marker: on` force-restores it. Regression guard: `tests/unit/think-close-marker-suppress-5245.test.ts`. (thanks @abdofallah)
+- **fix(executors):** strip nested `reasoning_content` from messages for **Mistral** (port from 9router#1649, [#6417](https://github.com/diegosouzapw/OmniRoute/pull/6417)). Mistral's API returns `422 extra_forbidden` when an assistant message carries `reasoning_content` (replayed thinking from a prior turn, e.g. via the Codex `/responses` path); the generic top-level 400 field-downgrade retry never covered the nested per-message field. `DefaultExecutor` now strips it for provider `mistral` only, so DeepSeek (which requires replayed `reasoning_content`) is unaffected. Regression guard: `tests/unit/mistral-strip-reasoning-content-1649.test.ts`. (thanks @xxy9468615)
+- **fix(executors):** strip the `client_metadata` passthrough field on the **OpenCode** path (port from 9router#1442, [#6418](https://github.com/diegosouzapw/OmniRoute/pull/6418)). OpenCode upstreams (e.g. `kimi-k2.6` via opencode-go) reject it with `400 "Extra inputs are not permitted, field: 'client_metadata'"`; the DefaultExecutor strip only covered cerebras/mistral and `OpencodeExecutor` extends `BaseExecutor` directly, so nothing removed it there. Regression guard: `tests/unit/opencode-strip-client-metadata-1442.test.ts`. (thanks @yanpaing007)
+- **fix(executors):** inject the `reasoning_content` echo for the native **Moonshot Kimi** provider (port from 9router#1480, [#6419](https://github.com/diegosouzapw/OmniRoute/pull/6419)). Kimi (executor `default`) is a thinking-mode upstream that 400s with "reasoning_content must be passed back" when a prior assistant turn lacks it; the placeholder injection was only wired into the OpenCode meta-provider, so direct multi-turn Kimi conversations failed. Scoped to `kimi` (gateway-served models matching the thinking-model name pattern are unaffected). Regression guard: `tests/unit/kimi-native-reasoning-injected-1480.test.ts`. (thanks @2220258345)
+- **fix(executors):** recover from a strict gateway's `context_management: Extra inputs are not permitted` 400 (port from 9router#1468, [#6420](https://github.com/diegosouzapw/OmniRoute/pull/6420)). **Claude Code** always sends a top-level `context_management` field; strict anthropic-compatible gateways reject it. The dedicated context-editing 400-fallback only fired when OmniRoute's own `contextEditing` feature was enabled (default off), so a client-sent field passed through untouched and 400'd. `context_management` is now in the generic reactive field-strip list, so it's stripped-and-retried once regardless of the feature flag (with correct request re-signing for claude-compatible relays). Regression guard: `tests/unit/provider-field-strips.test.ts`. (thanks @ohahe52-dot)
+- **fix(network):** enable **RFC 8305 Happy Eyeballs** (`autoSelectFamily`) on the direct-egress undici dispatcher (port from 9router#1237, [#6423](https://github.com/diegosouzapw/OmniRoute/pull/6423)). When DNS returns both IPv6 (AAAA) and IPv4 (A) and the IPv6 route is broken (e.g. a NAT64 `64:ff9b::` prefix without routing), undici tried IPv6 first and hung until `ETIMEDOUT` (then a 502 + account lockout), even though `curl` reached the same host. The direct dispatcher now races both families and uses whichever connects first. Proxy paths pin family via `proxyTls` and are unaffected. Regression guard: `tests/unit/direct-dispatcher-pipelining-4580.test.ts`. (thanks @adentdk)
+- **fix(combo):** round-robin now advances the rotation pointer past the model that **actually served**, not the eagerly-scheduled one (port from 9router#948, [#6428](https://github.com/diegosouzapw/OmniRoute/pull/6428)). With `stickyLimit: 1` (true round-robin), when the scheduled model failed and a _different_ model served via fallback, the counter had already advanced +1 from the scheduled index — so the next request reused the fallback-served model, degrading round-robin into hot-spotting on whichever model was healthy. The pointer now advances to the served index + 1 (mirroring the sticky-limit>1 path). Session-stickiness (#3825) and distribution are preserved. Regression guard: `tests/unit/combo-rr-fallback-advance-948.test.ts`. (thanks @binsarjr)
+- **fix(sse):** a non-string `model` field is now rejected with a `400` before the resolver, instead of crashing downstream `.toLowerCase()`/`.split()` calls into an empty-body `500` that escapes the error sanitizer ([#6407](https://github.com/diegosouzapw/OmniRoute/issues/6407)). Regression guard: `tests/unit/chat-non-string-model-6407.test.ts`. (thanks @chirag127)
+- **fix(api):** unknown `/api/*` routes now return a JSON `404` (instead of the dashboard HTML shell) and scalar chat params (`model`/`temperature`/etc.) are validated **before** the provider lookup so malformed requests fail fast with a clear `400` ([#6424](https://github.com/diegosouzapw/OmniRoute/issues/6424), [#6412](https://github.com/diegosouzapw/OmniRoute/issues/6412)). Regression guards: `tests/unit/api/api-catchall-json-404.test.ts`, `tests/unit/chat-early-schema-validation-6412.test.ts`. (thanks @chirag127)
+- **fix(api):** `/v1/chat/completions` now rejects a non-JSON `Content-Type` with a `400` before parsing the body ([#6414](https://github.com/diegosouzapw/OmniRoute/issues/6414)). Regression guard: `tests/unit/v1-chat-completions-content-type-6414.test.ts`. (thanks @chirag127)
+- **fix(api):** the `X-OmniRoute-Compression` response header is now echoed on `/v1/chat/completions` and `/v1/completions` ([#6422](https://github.com/diegosouzapw/OmniRoute/issues/6422)). Regression guard: `tests/unit/compression-header-echo-6422.test.ts`. (thanks @chirag127)
+- **fix(api):** concurrent `GET /v1/models` requests are coalesced into a single catalog build ([#6408](https://github.com/diegosouzapw/OmniRoute/issues/6408)). Regression guard: `tests/unit/v1-models-concurrent-6408.test.ts`. (thanks @chirag127)
+- **fix(api):** `/v1/completions` now echoes the requested `body.model` in its JSON + streamed responses ([#6429](https://github.com/diegosouzapw/OmniRoute/pull/6429)). Regression guard: `tests/unit/completions-body-model-echo.test.ts`. (thanks @chirag127)
+- **fix(api):** env-var master keys now see the full `/v1/models` catalog ([#6406](https://github.com/diegosouzapw/OmniRoute/issues/6406)). Regression guard: `tests/unit/models-catalog-envkey-6406.test.ts`. (thanks @chirag127)
+- **fix(api):** non-streaming `/v1/completions` responses now echo `body.model` aligned with the `X-OmniRoute-Model` header ([#6426](https://github.com/diegosouzapw/OmniRoute/issues/6426)). Regression guard: `tests/unit/v1-completions-model-header-match-6426.test.ts`. (thanks @chirag127)
+- **fix(api):** unknown `/v1/*` routes now return a JSON `404 not_found` instead of the Next.js dashboard HTML shell ([#6405](https://github.com/diegosouzapw/OmniRoute/issues/6405)). Regression guard: `tests/unit/api/v1-catchall-json-404.test.ts`. (thanks @chirag127)
+- **fix(api):** the per-connection provider models route now degrades to the shipped catalog when a provider's `/models` endpoint answers with a redirect ([#6267](https://github.com/diegosouzapw/OmniRoute/issues/6267)) — a `qwen-web` import failed with a raw `Redirect blocked … (307)` 503. `safeOutboundFetch` throws `REDIRECT_BLOCKED` on the 307, `getSafeOutboundFetchErrorStatus` maps it to 503, and `buildDiscoveryErrorFallbackResponse` treated every 503 as a hard error — so the non-empty `getModelsByProviderId("qwen-web")` catalog was never surfaced. A models-endpoint redirect is not a fixable-config error (unlike `URL_GUARD_BLOCKED`/`INVALID_URL`, which stay hard errors), so it now falls back to the local/cached catalog before the 503 short-circuit. General fix — covers any config-driven provider that 307s. Regression guard: `tests/unit/provider-models-qwen-web-redirect-6267.test.ts`. (thanks @chirag127)
+- **fix(api):** the per-connection provider models route (MCP `list_models_catalog` + the dashboard import view) now merges USER-ADDED custom models into its response ([#6247](https://github.com/diegosouzapw/OmniRoute/issues/6247)) — custom models live in the `key_value` namespace `customModels`, which the live REST `/api/v1/models` already merges, but `src/app/api/providers/[id]/models/route.ts` never read `getCustomModels`, so custom models were dropped on both the discovery-success and local_catalog paths. They are now folded into the returned model list (deduped by id, stamped `owned_by: provider`), fixing MCP + the dashboard import view in one place. Regression guard: `tests/unit/provider-models-custom-merge-6247.test.ts`. (thanks @RCrushMe)
+
+- **fix(providers):** GitLab Duo tool-calling follow-up turns no longer fail upstream with `422 {"detail":"Validation error"}` (tokens 0/0, rejected pre-inference). The #6234 tool-result-feedback fix serialized the **entire** multi-turn conversation into GitLab's single-file `code_suggestions` (`small_file`) generation endpoint — folded history that turn-N sent as an oversized `current_file.content_above_cursor` **and** duplicated verbatim into `user_instruction`, tripping the AI-Gateway's `small_file` validation guard. The executor now **bounds** that prompt: it keeps system + latest user message + the most-recent tool round (dropping older turns), caps oversized tool results, and stops duplicating the full prompt into `user_instruction` (which now carries only the short latest user message) — while still feeding the most-recent tool result back so the agent continues (`open-sse/executors/gitlab.ts`, [#6220](https://github.com/diegosouzapw/OmniRoute/issues/6220)). The unit test covers the bounding logic; the upstream 422→200 clearing is VPS-only (Hard Rule #18). Regression guard: `tests/unit/gitlab-tool-exchange-bounded-6220.test.ts`.
+
+- **fix(i18n):** the provider-detail (`/dashboard/providers/[id]`) connection-status filter labels no longer render as `__MISSING__:All` / `__MISSING__:Active` / `__MISSING__:Error` / `__MISSING__:Banned` / `__MISSING__:CreditsExhausted` in non-English locales (notably pt-BR) ([#6290](https://github.com/diegosouzapw/OmniRoute/issues/6290)). Root cause was **not** the namespace mismatch the issue guessed — the `providers.filter*` keys resolve correctly in `en.json`; the debt lived in the locale mirrors (`src/i18n/messages/*.json`), where these five keys carried the `__MISSING__:` sync sentinel in ~15 locales and were absent entirely in ~26 others, so next-intl found the key and echoed the sentinel verbatim. All 40 non-English/-Chinese mirrors now ship real translations for the five `providers.filter*` labels. Regression guard: `tests/unit/i18n-provider-filter-keys-6290.test.ts`. (thanks @diegosouzapw)
+
+- **fix(providers):** the `copilot-m365-web` streaming executor now emits `debug`-level WebSocket diagnostics ([#6210](https://github.com/diegosouzapw/OmniRoute/issues/6210)) — the outbound WS URL (with the `access_token` **redacted** via `redactWsUrl()`), handshake success/failure, and each received SignalR frame's `type`/`target`. Previously the streaming path logged nothing, so an empty `content:null` response (the M365 Education / Starter tier symptom fixed in #6234) was undiagnosable even at `APP_LOG_LEVEL=debug`. The change is debug-level and side-effect-free — it does not alter streaming behavior or the frame parser, and the token never reaches the logs. Regression guard: `tests/unit/copilot-m365-web-logging-6210.test.ts` (thanks @qpeyba)
+
+- **fix(resilience):** a round-robin combo no longer returns `503 all upstream accounts are unavailable` when a compatibility-rejected target is actually healthy ([#6238](https://github.com/diegosouzapw/OmniRoute/issues/6238)). `filterTargetsByRequestCompatibility` drops request-incompatible targets (tool/vision/structured-output unsupported, or below the required context window) **before** any availability check runs, and its `compatible.length === 0` safety net only fired when _all_ targets were filtered — not when the kept targets later all turned out runtime-unavailable (circuit-open / cooldown / no credentials). So a combo could 503 while a compat-rejected-but-healthy provider sat unused. `handleRoundRobinCombo` now keeps the compat-rejected set and, when every compat-kept target was skipped without a single real attempt, probes those rejected targets as a **last-resort fallback tier** (via the new pure `open-sse/services/combo/comboCompatFallback.ts`) before crystallizing the 503. Regression guard: `tests/unit/combo-roundrobin-compat-fallback-6238.test.ts`. (thanks @ThongAccount)
+
+- **fix(startup):** best-effort self-heal for a corrupted Turbopack dev cache on Windows ([#6289](https://github.com/diegosouzapw/OmniRoute/issues/6289)). On Windows, `pnpm dev` can fail at startup when Turbopack `mmap`s a persistent-cache SST file and the OS refuses the mapping (`os error 1455` — "paging file too small"), which Turbopack surfaces as a misleading `Module not found: Can't resolve '@/shared/utils/machine'`. This is a **known upstream Turbopack cache-corruption bug — not our code**. The dev launcher (`scripts/dev/run-next.mjs`) now wraps `nextApp.prepare()` and, when it rejects with that signature (`isTurbopackCacheCorruption` in the new `scripts/dev/turbopackCacheHeal.mjs`), purges `.build/next/**/cache/turbopack` and retries **once** with a clear log. **Caveat — best-effort only:** the corruption often surfaces as a runtime overlay rather than a `prepare()` rejection, so this cannot always intercept it; the reliable remedy remains manually deleting the Turbopack cache dir. Regression guard: `tests/unit/turbopack-cache-heal-6289.test.ts`. (thanks @chirag127)
+
+- **fix(providers):** qodercli PAT auth no longer fails with `spawn qodercli ENOENT` on Windows ([#6263](https://github.com/diegosouzapw/OmniRoute/issues/6263)) — `spawnQoderCli` spawned the bare `qodercli` name with `shell:false` and an unenriched env, so the npm `.cmd` wrapper under `%APPDATA%\npm` (a user-PATH directory) was never resolved. It now resolves the absolute `.cmd`/`.exe` path through the existing `getCliRuntimeStatus("qoder")` resolver in `src/shared/services/cliRuntime.ts` (memoized), spawns with `shell` when the target is a `.cmd`/`.bat`, and uses the cliRuntime-enriched env (PATH + PATHEXT + APPDATA); the ENOENT error now lists the searched paths plus the `CLI_QODER_BIN` override. End-to-end spawn on a real Windows host is host-only (Hard Rule #18); the path-resolution logic is unit-tested. Regression guard: `tests/unit/qodercli-windows-resolve-6263.test.ts`. (thanks @chirag127)
+
+- **fix(sse):** the reasoning-token buffer no longer inflates **probe-sized `max_tokens`** ([#6274](https://github.com/diegosouzapw/OmniRoute/issues/6274)) — Claude Code's `/model` capability check sends `max_tokens: 1`, but for a thinking-capable model with a large output cap (e.g. `glm-5.2`) the #3587 headroom heuristic (`max(current + 1000, ceil(current * 1.5))`) rewrote it to `1001` and forwarded that upstream, wasting tokens on a request that was never a genuine reasoning budget. `resolveReasoningBufferedMaxTokens()` (`open-sse/services/reasoningTokenBuffer.ts`) now short-circuits and returns the caller's value verbatim when it is below the new `REASONING_BUFFER_MIN_TRIGGER` (256) threshold — a tiny explicit limit is a probe, not a reasoning request. Real budgets still receive the #3587 headroom unchanged, and the guard runs after the existing capability checks so unknown / non-reasoning models keep returning `null`. Regression guard: `tests/unit/reasoning-token-buffer-6274.test.ts`. (thanks @brightfiscalband)
+
+- **fix(cli):** `omniroute reset-password` now works as a real subcommand, and password resets over piped (non-TTY) stdin actually apply ([#6261](https://github.com/diegosouzapw/OmniRoute/issues/6261), [#6258](https://github.com/diegosouzapw/OmniRoute/issues/6258)). Two coupled defects: (1) **#6261** — `bin/omniroute.mjs` routed everything through Commander with only two pre-Commander bypasses (`--mcp`, `reset-encrypted-columns`), so `omniroute reset-password` was rejected as an unknown command; only the separate `omniroute-reset-password` bin worked, while the docs falsely advertised the subcommand (incl. a bogus "legacy alias still works"). A pre-Commander bypass mirroring `reset-encrypted-columns` now dynamically imports `bin/reset-password.mjs` (which self-executes) before Commander parses; the three doc lines were corrected. (2) **#6258** — `bin/reset-password.mjs` issued two sequential `rl.question` prompts; under piped stdin the second read never settled at EOF, so `main()` never reached `resetManagementPassword` and the reset was a silent no-op (both prompts printed, no success, password unchanged). The CLI now detects non-TTY stdin and reads it once (first line = password, second line = confirm if present, else reused), adds a `--password-stdin` flag (entire stdin is the password, no confirmation), and exits `0` explicitly so the success line always flushes; interactive TTY behavior is unchanged. Regression guard: `tests/unit/reset-password-cli-6261-6258.test.ts` (3). (thanks @chirag127)
+
+- **fix(db):** the mass-migration **safety abort** now tells the operator how to bypass it and stops flooding the log ([#6260](https://github.com/diegosouzapw/OmniRoute/issues/6260)) — after restoring a backup that wiped the migration tracking table, `runMigrations()` threw the abort on every downstream `ensureDbInitialized()`, re-logging the full banner 11+ times, and the message never mentioned the existing `OMNIROUTE_MAX_PENDING_MIGRATIONS` escape hatch. The abort text now appends a bypass hint (set `OMNIROUTE_MAX_PENDING_MIGRATIONS=0` in `server.env` / `DATA_DIR/.env`), and a new `MigrationSafetyAbortError` is memoized so repeated calls in the same process throw the same instance and emit a single concise line instead of the full cascade. Regression guard: `tests/unit/migration-safety-abort-6260.test.ts`. (thanks @chirag127)
+
+- **fix(auth):** importing a **distinct** Codex/ChatGPT OAuth `auth.json` is no longer falsely rejected as "already exists" when it belongs to a different user in the same workspace ([#6301](https://github.com/diegosouzapw/OmniRoute/issues/6301)). `findExistingCodexConnection` (in `src/lib/oauth/utils/codexAuthImport.ts`) deduped **only** on `providerSpecificData.workspaceId === accountId`, where `accountId` is the shared `chatgpt_account_id`/`tokens.account_id` — so two members of the same ChatGPT Team collapsed onto a single connection (409 `duplicate_account`). The id_token's `https://api.openai.com/auth` claim carries a per-user `chatgpt_user_id` alongside the workspace id (the device-flow path already persisted it as `chatgptUserId`, but the import path did not). Now `parseAndValidateCodexAuth` extracts `userId` (`chatgpt_user_id` → `user_id` → JWT `sub`) into `ParsedCodexAuth`, the create/update paths persist `chatgptUserId` in `providerSpecificData` (mirroring `codex.ts`), and dedup keys on `workspaceId` **AND** `chatgptUserId` — with a backward-compat fallback to legacy accountId-only matching when no stored connection for that workspace records a `chatgptUserId`, so genuinely-same accounts still dedup. Regression guard: `tests/unit/codex-auth-import-userid-dedup-6301.test.ts` (4). (thanks @anungma)
+
+- **fix(providers):** importing models for the **venice-web** provider no longer fails with a red "Provider venice-web does not support models listing" ([#6269](https://github.com/diegosouzapw/OmniRoute/issues/6269)). `venice-web` is a web-cookie provider with an executor but no upstream `/v1/models` endpoint and no registry `models`, so the models route fell through to the tail `400`. Mirroring the `jules`/`linkup-search`/`ollama-search` fix (#5569), it now ships a static local catalog entry in `src/lib/providers/staticModels.ts` — seeding the current Venice lineup (`venice-uncensored`, `llama-3.3-70b`, `qwen3-235b`, `qwen3-4b`, `deepseek-r1-671b`; Venice rotates its catalog, see docs.venice.ai/models/overview) — so the route returns `200` with `source:"local_catalog"`, `intentional:true`. Regression guard: `tests/unit/static-models-venice-web-6269.test.ts`. (thanks @chirag127)
+
+- **fix(api):** the specialty model catalogs (`/v1/embeddings`, `/v1/images`, `/v1/music`, `/v1/videos` model lists) are now derived from the **unified catalog filtered by a predicate** (`getSpecialtyModelsResponse`) instead of ad-hoc per-route logic, so they consistently respect active-credential visibility and stay in sync with the main catalog ([#6303](https://github.com/diegosouzapw/OmniRoute/pull/6303)). Regression guard: `tests/unit/specialty-model-catalog-routes.test.ts`. (thanks @makcimbx)
+- **fix(api):** the agent-bridge server route now resolves the MITM manager via a **dynamic `import("@/mitm/manager.runtime")`** so Turbopack does not statically pull the stub (or over-bundle the manager), and the agent-skills generator anchors its output base path with `path.join(process.cwd(), …)` so Turbopack's static analyzer stops tracing the whole project root ([#6329](https://github.com/diegosouzapw/OmniRoute/issues/6329), [#6366](https://github.com/diegosouzapw/OmniRoute/pull/6366)). Regression guard: `tests/unit/agent-bridge-server-route-dynamic-import.test.ts`. (thanks @Iammilansoni)
+- **fix(api):** internal probes (combo-test, cloud-sync verify) now pick a **management-scoped / allow-all API key** instead of naively grabbing `getApiKeys()[0]` — a restricted `self:usage` first row made the probe fail with "Model X is not allowed for this API key" even when the combo path was healthy (`pickApiKeyForInternalUse` in `src/lib/db/apiKeys.ts`). The API-manager model editor also falls back to `/api/models?all=true` when `/v1/models` is catalog-protected ([#6372](https://github.com/diegosouzapw/OmniRoute/pull/6372)). Regression guard: `tests/unit/pick-internal-api-key-6372.test.ts`. (thanks @jmengit)
+- **fix(live-ws):** the Live Dashboard WebSocket server now **rejects on bind failure** (e.g. `EADDRINUSE` when the API bridge already holds the port) instead of letting the error surface as an unhandled `error` event that crash-loops the process — the `error` listener is attached to `wss` (not `server`) and releases the EventBus subscription on a failed start ([#6324](https://github.com/diegosouzapw/OmniRoute/issues/6324)). Regression guard: `tests/unit/live-ws-eaddrinuse-6324.test.ts`. (thanks @vinayakkulkarni)
+- **fix(dashboard):** the Home provider-topology widget now trusts the live provider-metrics snapshot — it uses `topology.errorProvider` and live `activeRequests` directly instead of re-deriving state from a stale `lastErrorAt` or applying a frontend timeout filter, so the topology reflects real-time provider health ([#6322](https://github.com/diegosouzapw/OmniRoute/pull/6322)). Regression guard: `tests/unit/home-provider-topology-live-state.test.ts`. (thanks @xz-dev)
+- **fix(sse):** strip zero-width markers from streamed **tool-call arguments** — a follow-up to [#5857](https://github.com/diegosouzapw/OmniRoute/pull/5857). That PR removed injected zero-width joiners (U+200D) from streamed assistant text/reasoning but deliberately left tool-call argument JSON byte-exact. The request-side obfuscation (`open-sse/services/claudeCodeObfuscation.ts`) injects ZWJ into agent words — including the temp path inside the Bash tool description — and Claude models copy that verbatim into generated commands, which are delivered as tool-call arguments rather than assistant text. As a result the ZWJ survived and corrupted code blocks (e.g. a temp path rendered with an invisible joiner). Now `open-sse/handlers/responseSanitizer.ts` strips zero-width code points from tool-call argument strings at every emit site (OpenAI non-stream/stream chat `tool_calls` + legacy `function_call`, native Responses `function_call` items, the OpenAI→Responses conversion, and the native Responses streaming `response.function_call_arguments.delta/.done` events). Only zero-width code points are removed; JSON structure and all other bytes stay identical (no parse/restringify), so normal arguments remain byte-exact. Regression guard: 6 new cases in `tests/unit/response-sanitizer.test.ts` (suite 50/50).
+
+- **fix(nodejs):** the default app log path now resolves under `DATA_DIR` (`~/.omniroute/logs/application/app.log`) instead of `process.cwd()` ([#6197](https://github.com/diegosouzapw/OmniRoute/issues/6197)) — the globally-installed CLI runs from an arbitrary working directory, so anchoring the default to cwd made file logging silently write to (or no-op under) an unrelated directory, contradicting the documented `.env.example` default. `getAppLogFilePath()` now computes the default lazily via the pure `resolveDataDir()` resolver (honours a per-process `DATA_DIR`, no directory-creation side effect); an explicit `APP_LOG_FILE_PATH` still wins. Regression guard: `tests/unit/logenv-datadir-path-6197.test.ts` (3).
+
+- **fix(docker):** AgentBridge/`startMitm` no longer aborts in containers/headless when the Antigravity-default DNS step can't write `/etc/hosts` ([#6127](https://github.com/diegosouzapw/OmniRoute/issues/6127)), and the privileged command's stderr now reaches `app.log` instead of only a bare exit code hitting the toast ([#6198](https://github.com/diegosouzapw/OmniRoute/issues/6198)). The default DNS step (`addDNSEntry`) was called unguarded while cert install and the two sibling DNS steps were each best-effort — in the runtime Docker image (`USER node`, no `sudo`, read-only `/etc/hosts`) it threw `Command failed with code 1` out of `startMitmInternal` and killed the whole start, discarding the stderr. The three DNS steps are extracted into a best-effort `provisionDnsEntries()` where each failure is logged with the full `err` (stderr included, folded in by `systemCommands.ts`) and never aborts the start. Regression guard: `tests/unit/mitm-dns-graceful-degrade-6127.test.ts` (4).
+
+- **fix(providers):** copilot-m365-web now supports the M365 Education "Starter / OfficeWebIncludedCopilot" tier and no longer returns an empty `content:null` stream ([#6210](https://github.com/diegosouzapw/OmniRoute/issues/6210)). Two gaps: (1) `buildWsUrl()` hardcoded the individual-consumer scenario (`OfficeWebPaidConsumerCopilot`, `isEdu=false`) — the EDU tier is now opt-in via `providerSpecificData.tier="edu"`, emitting `scenario=OfficeWebIncludedCopilot`/`isEdu=true` (the individual path is unchanged); (2) the EDU/GPT-5.5 path streams deltas via `arguments[0].writeAtCursor` (incremental) instead of only `messages[].text` (accumulated snapshots), which the parser dropped — a new `accumulateBotContent()` folds both formats, with `type:2 item.result.message` as a last-resort fallback. Regression guard: `tests/unit/copilot-m365-edu-writeatcursor-6210.test.ts` (10). (thanks @qpeyba)
+
+- **fix(providers):** GitLab Duo executor now feeds tool results back into the prompt instead of looping ([#6220](https://github.com/diegosouzapw/OmniRoute/issues/6220)) — `buildPrompt()` branched only on `system`/`user` and took `userParts.at(-1)`, silently dropping the `assistant{tool_calls}` + `tool{result}` turns the client appended, so the reconstructed prompt was byte-identical to turn 1 and the model re-emitted the same `<tool>` call forever. When a tool exchange is present the full conversation is now serialized, folding each tool result back keyed by its `tool_call_id`; simple conversations keep the legacy shape. Complements the tool_call emission from [#6051](https://github.com/diegosouzapw/OmniRoute/issues/6051) (the `kilo-duplicate` label was a false positive — different, sequential defect). Regression guard: `tests/unit/gitlab-tool-result-feedback-6220.test.ts` (4).
+
+- **fix(providers):** opencode-go/opencode-zen can now synthesize the OpenCode CLI identity headers Cloudflare requires on VPS egress ([#5997](https://github.com/diegosouzapw/OmniRoute/issues/5997)) — on a datacenter VPS, `opencode.ai/zen/go/v1/chat/completions` 403s (HTML challenge) requests lacking CLI identity, while the reporter's control curl proved that `User-Agent: opencode-cli/1.0.0` + `x-opencode-client: cli` + `x-opencode-project: default` + fresh request/session UUIDs succeed. Opt-in via `OPENCODE_SYNTHESIZE_CLI_HEADERS=true` (values overridable via `OPENCODE_GO_USER_AGENT`/`OPENCODE_USER_AGENT`/`OPENCODE_CLIENT`/`OPENCODE_PROJECT`); it fills only headers the client did not already send. Kept **off by default** — the forward-only path is deliberate (fabricating a wrong value risks upstream rejection; a prior dedup regressed with `opencode/local`), so this replaces the fragile local header-injection shim without changing default behavior. Regression guard: `tests/unit/opencode-cli-headers-synthesis-5997.test.ts` (6). (thanks @aleksesipenko)
+- fix(resilience): sticky session affinity now evicts and fails over to another account when the pinned account is exhausted/unavailable (#6219)
+
+- fix(sse): Responses API passthrough now drops internal commentary-phase output before forwarding to clients (gated by RESPONSES_PASSTHROUGH_DROP_COMMENTARY, default on) (#6199)
+
+- **fix(sse):** tool-call function schemas with a root `type: null` are now coerced to `type: "object"` before dispatch ([#6359](https://github.com/diegosouzapw/OmniRoute/issues/6359)) — clients like the Codex app emit `parameters: { type: null, ... }` for some tools, which OpenAI-compatible upstreams reject with `400 Invalid schema for function '...': schema must be a JSON Schema of 'type: "object"', got 'type: null'`, failing the whole request. `toolSchemaSanitizer` already stripped the null; it now re-adds the mandatory root `"object"` type (and empty `properties`/open `additionalProperties` when absent). Combinator roots (`anyOf`/`oneOf`/`allOf`) and explicit root types are left untouched. Regression guard: 5 new cases in `tests/unit/tool-schema-sanitizer.test.mjs`.
+
+- **fix(docker):** AgentBridge no longer fails to start on npm/Electron/VPS installs with "MITM manager stub reached at runtime" ([#6344](https://github.com/diegosouzapw/OmniRoute/issues/6344)) — v3.8.45 flipped the production bundler default to Turbopack, but `next.config.mjs` aliased `@/mitm/manager` to its Docker-only degraded stub **unconditionally**. That was harmless while Docker (which sets the alias intentionally for #3390 graceful degradation) was the sole Turbopack consumer, but once every artifact built with Turbopack the stub shipped to all non-Docker users and `startMitm` threw on the first Agent-Bridge start. The alias is now opt-in via `OMNIROUTE_MITM_STUB=1` (set only by the Dockerfile) through the shared `scripts/build/mitm-stub-flag.mjs` helper; default builds bundle the real manager. Regression guard: `tests/unit/mitm-stub-alias-6344.test.mjs` (4).
+
+- **fix(proxy):** stop the v3.8.44 proxy regression that leaked the real IP and disabled healthy proxies ([#6246](https://github.com/diegosouzapw/OmniRoute/issues/6246)). Two coupled defects from the new health scheduler: (1) **IP leak** — when a proxy assigned to a connection was marked `inactive`, resolution fell through to a **direct** egress instead of blocking, exposing the operator's real IP; (2) **over-deactivation** — the sweep flipped a proxy to `inactive` on the **first** failed probe and counted our own 5s timeout / a probe-target `5xx` as the proxy's fault, so healthy paid proxies vanished from egress selection ("my proxies are not being used anymore"). Fix: the sweep decision is extracted into a pure, network-free `decideProxyHealthAction` (`src/lib/proxyHealth/decision.ts`) — by default the health check now **only counts/logs and never downgrades status** (a proxy is downgraded/removed only with `PROXY_AUTO_REMOVE=true`, after `PROXY_AUTO_REMOVE_AFTER` **consecutive** conclusive failures); probes are classified tri-state so an inconclusive result (our timeout, or a `5xx` from the probe target) never penalizes the proxy, and the probe timeout is raised 5s→15s. Separately, `safeResolveProxy` now **fails closed** via the existing policy: a connection whose assigned proxy is dead is blocked instead of leaking direct (`hasBlockingProxyAssignment`), honoring the explicit `proxy off` toggles and the `PROXY_FAIL_OPEN=true` opt-out. Existing proxies stuck `inactive` by the old behavior need a one-time manual re-activate (the operator owns proxy status). Regression guards: `tests/unit/proxy-health-decide-action-6246.test.ts`, `tests/unit/proxy-assigned-unavailable-6246.test.ts`.
+
+- **fix(proxy):** make "Test All" read-only and add bulk enable/disable ([#6246](https://github.com/diegosouzapw/OmniRoute/issues/6246)). Complements the core fail-closed / scheduler fix (#6296) with the two remaining reporter asks. (1) The **"Test All" button** (`POST /api/settings/proxies/auto-test`) used to flip a proxy to `inactive` on a failed reachability probe; since the egress selector excludes `inactive` proxies, a flaky probe (an unreachable `httpbin.org`, a proxy that blocks `HEAD`, or a slow paid proxy) silently disabled every proxy that failed — "Test All" is now **read-only by default** (only the operator sets a proxy active/inactive; opt back into the legacy test-and-set with `PROXY_HEALTH_AUTO_DEACTIVATE=true`). (2) Adds a **bulk enable/disable** proxies endpoint + toolbar action (`POST /api/settings/proxies/batch-activate`) so an operator can re-activate proxies in one click. Regression guard: `tests/unit/proxy-health-6246.test.ts`. (thanks @tenshiak)
+
+- **chatcore (tools): stop the default 128-tool cap from silently dropping opencode's `task`/MCP tools.** opencode (used as an MCP/agent host) sends a large tool list; when it exceeds the speculative `MAX_TOOLS_LIMIT` (128) default, `truncateToolList` did a blind `tools.slice(0, 128)`, dropping every tool past index 128 — including opencode's built-in `task` tool (subagent launch) and many MCP tools, so models routed through OmniRoute could no longer spawn subagents or reach part of their tools. The cap exists to avoid upstream `400`s for providers with real hard limits (e.g. grok-cli 200), so it is kept for those: detection of the opencode client (`isOpencodeClient` — any `x-opencode-*` header, or `opencode` in the user-agent) now only bypasses the **speculative 128 default**, never a known provider ceiling. Precedence is explicit — a proactive/detected provider limit always truncates (even for opencode); otherwise opencode forwards its full tool list; otherwise the unchanged 128 default applies to every other client. Refactors `getEffectiveToolLimit` into `getKnownToolLimit(provider) ?? DEFAULT_LIMIT` (byte-identical for existing callers) and fixes a cosmetic debug-log that reported the truncated count instead of the original. Regression guard: `tests/unit/tool-limit-detector.test.ts`.
+
+- **fix(mitm):** the macOS MITM-cert install check now matches the system keychain again. `security find-certificate -a -Z` prints the SHA-1 as a colon-less hex string, but the installed-check compared it against `getCertFingerprint()`'s colon-separated form, so the substring match never hit — the cert was reported as not-installed and re-prompted for the sudo install on every run. Fingerprints are now normalized (colons stripped, upper-cased) on both sides via the extracted `macCertOutputHasFingerprint` helper. Regression guard: `tests/unit/mitm-cert-mac-fingerprint.test.ts`. ([#6204](https://github.com/diegosouzapw/OmniRoute/pull/6204), closes [#6134](https://github.com/diegosouzapw/OmniRoute/issues/6134) — thanks @rianonehub)
+
+- **fix(api):** `/v1/messages/count_tokens` now counts `tool_use`, `tool_result` and `thinking` content blocks (and array-form `system` prompts) in the local-estimation path, instead of only `text`. Real agentic conversations keep ~95% of their tokens inside tool results; the previous estimate returned near-zero for them, which silently broke Claude Code's auto-compaction (context grew past the window with no compaction until the upstream API rejected the request). The real provider-side count path is unchanged. Regression guard: `tests/unit/messages-count-tokens-route.test.ts`. ([#6221](https://github.com/diegosouzapw/OmniRoute/pull/6221) — thanks @luweiCN)
+
+- **fix(antigravity):** strip a trailing assistant prefill turn for Vertex Claude models to avoid upstream 400s ([#6114](https://github.com/diegosouzapw/OmniRoute/pull/6114)). Regression guard: `tests/unit/antigravity-claude-prefill-strip.test.ts`. (thanks @anki1kr)
+
+- **fix(security):** the mutable cloud-agent routes (`/api/cloud/credentials/update`, `/api/cloud/models/alias`) now require management auth instead of being treated as public. They were classified as public API routes, so a request without management credentials could update stored cloud-agent credentials and model aliases. They are removed from the public-route set, classified as management routes in the authz pipeline, and gated by `requireManagementAuth`; cloud **read**/auth routes stay public. Regression guards: `tests/unit/cloud-write-auth.test.ts`, `tests/unit/authz/classify.test.ts`, `tests/unit/public-api-routes.test.ts`. ([#6233](https://github.com/diegosouzapw/OmniRoute/pull/6233) — thanks @vittoroliveira-dev)
+
+- **refactor(dashboard):** extract the onboarding-wizard "Open provider details" link target into a pure, unit-tested `buildProviderDetailsHref(connection)` helper. The wizard already routes by `connection.id` (the node UUID) rather than the provider category slug (#6144/#6145); this hardens that behavior behind a tested helper that guards a missing id/connection. Regression guard: `tests/unit/provider-onboarding-href.test.ts`. ([#6166](https://github.com/diegosouzapw/OmniRoute/pull/6166) — thanks @KooshaPari)
+
+- **fix(security):** the doubao synthetic device-id generator now derives its digits via an unbiased crypto-random draw (rejection sampling over `crypto.randomBytes()`) instead of a `% 10` reduction, closing a CodeQL `js/biased-cryptographic-random` finding.
+- **fix(agentSkills):** the GitHub-skills generator now resolves `outputDir` to an absolute path before writing, fixing a regression introduced by #6366 (relative-to-cwd base path) that could write generated skill files to the wrong directory.
+- **fix(security):** `/api/keys/{id}/devices` now checks the HTTP method before auth/validation, returning a `405` for non-GET/DELETE verbs instead of a misleading `401`/`500` (closes a `dast-smoke` QUERY-method finding).
+- **fix(quality):** clear the last 2 heavy quality-gate reds on the release tip (cycle pre-flight).
+- **fix(mitm):** the test suite and CI can never mutate the OS trust store — `OMNIROUTE_SKIP_SYSTEM_TRUST=1` is set globally for tests/CI so `installCert`/`uninstallCert`/`installTproxyCa` skip the privileged OS dispatch ([#6310](https://github.com/diegosouzapw/OmniRoute/pull/6310); full detail is under the [3.8.45] section below — this branch received it via the parallel-cycle sync-back).
+- **fix(api):** `POST /api/github-skills` now Zod-validates its request body; documented the new quality-gate env vars and pinned the merge-integrity GitHub Actions to a commit SHA.
+- **fix(skills):** generate the missing `omni-github-skills` registry entry and align the agent-skills catalog-count tests (follow-up to #6186).
+- **fix(quality):** clear the cycle's 11 net-new ESLint errors and make `validate-release-green` suppressions-aware.
+
+### 📝 Maintenance
+
+- **i18n(it):** add 118 missing Italian (`it`) translations (net-additive — no existing keys dropped, valid JSON), improving Italian UI coverage. ([#6212](https://github.com/diegosouzapw/OmniRoute/pull/6212) — thanks @serverless83)
+- **chore(providers):** remove deprecated **MiMo V2** model entries from the catalogs (xiaomi-mimo, opencode-go, zenmux-free, audio TTS) — the upstream V2 line is superseded by MiMo V2.5; drops `mimo-v2-tts`, `mimo-v2-pro`, `mimo-v2-omni`, `mimo-v2-flash`, `mimo-v2-flash-free` and realigns the provider-catalog tests. ([#6248](https://github.com/diegosouzapw/OmniRoute/pull/6248) — thanks @backryun)
+- **chore(release):** ~50 commits on this branch are v3.8.45 pre-flight/hardening fixes and CI-perf work that landed here via the parallel-cycle sync-back (`sync-next-cycle.mjs`, Hard Rule #21) **after** the `v3.8.45` git tag was cut, and are already fully documented under the **[3.8.45]** section below — listed here only so the per-cycle commit-coverage check (`npm run release:uncovered`) doesn't flag them as gaps. Provider/catalog/UX/backend fixes: #6041, #6078, #6108, #6135, #6148, #6149, #6154, #6158, #6161, #6162, #6163, #6164, #6165, #6170, #6177, #6178, #6181, #6186, #6187, #6191, #6193, #6194, #6195, #6200, #6205, #6208, #6209, #6211, #6213, #6223, #6224, #6225, #6226, #6227, #6228, #6229, #6230, #6235, #6291, #6292. CI/release-pipeline work: #6167, #6203, #6214, #6215, #6218, #6273, #6275, #6283, #6284, #6285, #6300, #6305.
+- **chore(release):** additional zero-ref release-cycle plumbing on this branch, kept out of `release:uncovered` on purpose (no `#N` in the commit subject to cite): opening the v3.8.46 cycle, opening/closing the v3.8.45 cycle, the finalized [3.8.45] CHANGELOG i18n sync-back to 42 mirrors, the v3.8.45 cognitive/cyclomatic and file-size drift rebaselines, ESLint stale-suppression pruning (4,273 → 4,233), and clearing test-masking/docs-all pre-flight reds for v3.8.45.
+
+### ⚡ Performance & Infrastructure
+
+- **perf(release-green):** the pre-flight validator (`scripts/quality/validate-release-green.mjs`) now runs its 4 slow suites (unit / vitest / integration / pack-artifact) **concurrently** via `Promise.all` — pre-flight wall time drops from ~the sum of the suites to ~the slowest one (~30min saved per round; Phase 0 was the nº1 bottleneck of the v3.8.45 release benchmark, 2h54 of 6h34 e2e). Guard: `tests/unit/validate-release-green.test.ts` ("runs the slow suites CONCURRENTLY"). ([#6319](https://github.com/diegosouzapw/OmniRoute/pull/6319))
+- **fix(ci):** `scripts/release/sync-next-cycle.mjs` — two defects found live in its first production run (v3.8.45 Phase 5): (1) the `git()` helper's default 1 MiB `maxBuffer` crashed with `ENOBUFS` on `git show origin/main:CHANGELOG.md` (the CHANGELOG alone is >1 MiB) — widened to 64 MiB; (2) the i18n resync only propagated the `[NEXT]` (TBD) section, leaving the just-shipped finalized section as "— TBD" in all 42 mirrors — it now also syncs `[prevVersion]` bounded by the heading below it (new exported pure helper `versionAfter`). Guards: +5 tests in `tests/unit/sync-next-cycle.test.ts` (8/8). ([#6327](https://github.com/diegosouzapw/OmniRoute/pull/6327))
+- **test(ci):** concurrency-sensitive flaky tests are **quarantined into a serial pass** (`tests/unit/serial/`, `--test-concurrency=1`, appended to every unit runner incl. sharded variants — the serial pass is sharded too so concurrent shard jobs never self-collide). Initial set: `glm-coding-plan-monthly-3580`, `quota-division-blocks`, `provider-health-autopilot`, `combo-health-autopilot` — the class behind the ~28min CI wedges/re-runs (two live 1h+ wedges cancelled during this PR's own validation). Discovery + TIA gates track the new glob; systemic root cause (async logger writing after teardown) tracked in [#6360](https://github.com/diegosouzapw/OmniRoute/issues/6360). Guard: `tests/unit/test-serial-quarantine.test.ts` (4). ([#6347](https://github.com/diegosouzapw/OmniRoute/pull/6347))
+
+### 🙌 Contributors
+
+Thanks to everyone whose work landed in v3.8.46:
+
+| Contributor                                                    | PRs / Issues           |
+| -------------------------------------------------------------- | ---------------------- |
+| [@2220258345](https://github.com/2220258345)                   | direct commit / report |
+| [@abdofallah](https://github.com/abdofallah)                   | direct commit / report |
+| [@adentdk](https://github.com/adentdk)                         | direct commit / report |
+| [@aleksesipenko](https://github.com/aleksesipenko)             | direct commit / report |
+| [@anki1kr](https://github.com/anki1kr)                         | direct commit / report |
+| [@anungma](https://github.com/anungma)                         | direct commit / report |
+| [@arpicato](https://github.com/arpicato)                       | direct commit / report |
+| [@backryun](https://github.com/backryun)                       | #6248                  |
+| [@binsarjr](https://github.com/binsarjr)                       | direct commit / report |
+| [@brightfiscalband](https://github.com/brightfiscalband)       | direct commit / report |
+| [@chirag127](https://github.com/chirag127)                     | #6501, #6506           |
+| [@developerjillur](https://github.com/developerjillur)         | direct commit / report |
+| [@dilneiss](https://github.com/dilneiss)                       | #6499                  |
+| [@dtybnrj](https://github.com/dtybnrj)                         | direct commit / report |
+| [@Forcerecon](https://github.com/Forcerecon)                   | direct commit / report |
+| [@hao3039032](https://github.com/hao3039032)                   | direct commit / report |
+| [@Iammilansoni](https://github.com/Iammilansoni)               | #6150, #6245           |
+| [@jmengit](https://github.com/jmengit)                         | direct commit / report |
+| [@jordansilly77-stack](https://github.com/jordansilly77-stack) | direct commit / report |
+| [@JxnLexn](https://github.com/JxnLexn)                         | direct commit / report |
+| [@KooshaPari](https://github.com/KooshaPari)                   | #6166                  |
+| [@loopyd](https://github.com/loopyd)                           | direct commit / report |
+| [@luweiCN](https://github.com/luweiCN)                         | #6221                  |
+| [@makcimbx](https://github.com/makcimbx)                       | direct commit / report |
+| [@muflifadla38](https://github.com/muflifadla38)               | direct commit / report |
+| [@newnol](https://github.com/newnol)                           | direct commit / report |
+| [@ofekbetzalel](https://github.com/ofekbetzalel)               | direct commit / report |
+| [@ohahe52-dot](https://github.com/ohahe52-dot)                 | direct commit / report |
+| [@phidinhmanh](https://github.com/phidinhmanh)                 | direct commit / report |
+| [@powellnorma](https://github.com/powellnorma)                 | direct commit / report |
+| [@qpeyba](https://github.com/qpeyba)                           | direct commit / report |
+| [@RaviTharuma](https://github.com/RaviTharuma)                 | direct commit / report |
+| [@RCrushMe](https://github.com/RCrushMe)                       | direct commit / report |
+| [@rianonehub](https://github.com/rianonehub)                   | #6134, #6204           |
+| [@serverless83](https://github.com/serverless83)               | #6212                  |
+| [@swingtempo](https://github.com/swingtempo)                   | direct commit / report |
+| [@tenshiak](https://github.com/tenshiak)                       | direct commit / report |
+| [@ThongAccount](https://github.com/ThongAccount)               | direct commit / report |
+| [@UnrealAryan](https://github.com/UnrealAryan)                 | direct commit / report |
+| [@vinayakkulkarni](https://github.com/vinayakkulkarni)         | direct commit / report |
+| [@vittoroliveira-dev](https://github.com/vittoroliveira-dev)   | #6233                  |
+| [@warelik](https://github.com/warelik)                         | direct commit / report |
+| [@xxy9468615](https://github.com/xxy9468615)                   | direct commit / report |
+| [@xz-dev](https://github.com/xz-dev)                           | direct commit / report |
+| [@yanpaing007](https://github.com/yanpaing007)                 | direct commit / report |
+| [@diegosouzapw](https://github.com/diegosouzapw)               | maintainer             |
+
+---
+
 ## [3.8.45] — 2026-07-06
 
 ### ✨ New Features

docs/i18n/da/CHANGELOG.md

@@ -6,6 +6,213 @@
 
 ## [3.8.31] — 2026-06-20
 
+## [3.8.46] — TBD
+
+### ✨ New Features
+
+- **feat(sse):** **hide paid-only models from `auto/*` routing** when `hidePaidModels` is on ([#6512](https://github.com/diegosouzapw/OmniRoute/issues/6512)) — follow-up to #6328/#6495. PR #6495 hid paid-only models from the `GET /v1/models` listing, but `auto/*` combos (`auto/best-coding`, `auto/glm`, …) could still pick a paid-only backend into their candidate pool → a 402/403 at request time. `createVirtualAutoCombo` now filters the candidate pool through the new pure `open-sse/services/autoCombo/paidModelFilter.ts` (`filterPaidOnlyCandidates`), applying the same free-model predicate #6495 uses in `catalog.ts` (`providerHasFreeModels(provider) && isFreeModel(provider, {id})`) whenever `settings.hidePaidModels === true`. Applied before the category/tier/family narrowing, so it covers every `auto/*` combo; an all-paid pool degrades to the existing graceful empty-pool path. **Opt-in — default OFF leaves the pool unchanged** (identity). Regression guard: `tests/unit/autoCombo/paid-model-filter-6512.test.ts` (4, incl. the default-off identity guard).
+- **feat(sse):** **provider-family auto combos** — `auto/glm`, `auto/minimax`, `auto/mimo`, `auto/zai`, `auto/gemma`, `auto/llama`, `auto/gemini` ([#6453](https://github.com/diegosouzapw/OmniRoute/issues/6453)) — new routable ids that materialize an on-demand virtual combo spanning whatever installed backends currently expose that model family, degrading gracefully as backends rotate. A new pure `open-sse/services/autoCombo/modelFamily.ts` (`detectModelFamily`) classifies by model-id prefix for six families; `zai` is instead resolved by provider id (z.ai's hosted API serves the same `glm-*` model ids as every other GLM backend, so `auto/zai` means "route to my z.ai backend specifically" vs `auto/glm`'s "any connected GLM backend"). Reuses the existing `createVirtualAutoCombo` on-demand materialization path (no DB writes) and the `/v1/models` catalog advertising loop. Regression guard: `tests/unit/autoCombo/provider-family-combos.test.ts` (11).
+- **feat(proxy):** native **proxy-pool round-robin / egress IP rotation** ([#6365](https://github.com/diegosouzapw/OmniRoute/issues/6365)) — a scope (global / provider / account) can now hold **multiple** proxies as a pool with a rotation strategy, so outbound requests cycle their egress IP instead of pinning one proxy per scope. Migration `117_proxy_pool_rotation.sql` lifts the `UNIQUE(scope, scope_id)` constraint (rebuild via the canonical rename/copy/drop; existing single assignments become 1-element pools) and adds a `proxy_scope_rotation` companion table holding the per-scope strategy + a persisted monotonic round-robin cursor. Strategies: `round-robin` (default, monotonic cursor — never `Math.random`), `random`, and `sticky-per-N-min`. Resolution (`resolveProxyForScopeFromRegistry` / `resolveProxyForConnectionFromRegistry`) now fetches the alive, position-ordered candidate set (unchanged `PROXY_ALIVE_PREDICATE`) and applies the strategy; an empty / all-dead pool still returns `null` — the #6246 fail-closed guard is **untouched** (never falls through to direct egress). Backend + DB only; dashboard pool-builder UI is a follow-up. Regression guard: `tests/unit/proxy-pool-rotation-6365.test.ts` (8, incl. fail-closed + backward-compat).
+- **feat(providers):** end-to-end **tool/function calling on the native Gemini `/v1beta` endpoint** ([#6222](https://github.com/diegosouzapw/OmniRoute/issues/6222)) — both directions of the Gemini↔OpenAI conversion now preserve tool data (previously silently dropped). Request side: `convertGeminiToInternal` (extracted to its own testable module) maps `tools[].functionDeclarations` → OpenAI `tools`, prior `functionCall` parts → assistant `tool_calls`, and `functionResponse` parts → `tool`-role messages. Response side: `convertOpenAIResponseToGemini` emits `parts[].functionCall {name,args}` from `message.tool_calls`, and the streaming `openAIChunkToGeminiChunk` accumulates fragmented `tool_calls` deltas by index into complete `functionCall` parts. The non-Gemini client paths (Claude, OpenAI-Responses) already preserved tool calls — this closes the gap specific to the native Gemini surface. Regression guard: `tests/unit/v1beta-gemini-tool-calling-6222.test.ts` (6, incl. a streaming SSE round-trip).
+- **feat(providers):** copilot-m365-web **enterprise / work tier** support ([#6334](https://github.com/diegosouzapw/OmniRoute/issues/6334)) — mirrors the EDU-tier pattern (#6210): `M365ConnectionParams` gains an `agent` field, a new opt-in `M365_ENTERPRISE_OVERRIDES` preset (`agent=work`, `scenario=officeweb`, `licenseType=Premium`) applies via `providerSpecificData.tier="enterprise"` (alias `"work"`), and `agent` is also overridable directly via `providerSpecificData.agent`. `buildWsUrl` was hardcoding `agent="web"` (the one enterprise-distinguishing param with no override path), so a Premium work account handshook then returned an empty stream. The individual and EDU paths are untouched. Kilo's dup flag vs #6210 (EDU tier) was a false positive — different tier. Regression guard: `tests/unit/copilot-m365-enterprise-6334.test.ts` (7). End-to-end confirmation on a real Premium work account is a live-VPS validation follow-up (Hard Rule #18). (thanks @Forcerecon)
+- **feat(api):** standardized, provider-agnostic **`effort` + `thinking` request params** ([#6241](https://github.com/diegosouzapw/OmniRoute/issues/6241)) — a thin standardization layer over the existing mature per-provider reasoning plumbing (no provider mapper touched). `providerChatCompletionSchema` gains a canonical `effort` (reusing the shared `none/low/medium/high/xhigh` vocabulary — the UI tiers `extra`/`max` collapse onto `xhigh`) and a boolean `thinking`. A pure `normalizeReasoningRequest` (wired once in `src/sse/handlers/chat.ts`, before any reasoning field is read) folds them onto the fields the translators already consume (`reasoning_effort` / `reasoning.effort` / `thinking`), so they fan out to Anthropic / Gemini / xAI / Responses — an explicit client `reasoning_effort` / object-shaped `thinking` always wins (backward-compatible). `/models` additively exposes `supportsThinking` + `effort_tiers` so the frontend can render the toggles (UI component is a follow-up). Regression guard: `tests/unit/effort-thinking-standardization-6241.test.ts` (12). (thanks @Iammilansoni, @shabeer)
+- **feat(combo):** new **`pipeline` (sequential) combo strategy** ([#6297](https://github.com/diegosouzapw/OmniRoute/issues/6297)) — the 18th routing strategy runs targets **in order**, threading each step's output into the next step's input, with an optional per-step `prompt` (system instruction); only the final step's response is returned. Distinct from `fusion` (parallel fan-out + judge). Implemented as a self-contained `open-sse/services/pipeline.ts` (sibling to `fusion.ts`), dispatched from `combo.ts`; the step list reuses `combo.models` order and reads an optional `prompt` off each target (backward-compatible — ignored by every other strategy). Intermediate steps run non-streaming with tools stripped (complete prose to thread forward); the final step keeps the client's `stream` flag + tools. A failing/empty/unparseable intermediate step fails the whole pipeline explicitly via a sanitized error (never silently swallowed). Kilo's dup flag vs #563 was a false positive (that's model→chain selection; this is a sequential chain). Regression guard: `tests/unit/combo-pipeline-strategy.test.ts` (5). (thanks @ofekbetzalel)
+- **feat(ci):** `check:test-masking` now flags **inline-reimplemented prod conditions** ([#6348](https://github.com/diegosouzapw/OmniRoute/issues/6348)) — a new **report-only** subcheck (v2, 6A.10 family) catches the wrong-shape contract test: a test that recomputes the condition under test inline instead of importing/exercising the real function (the #6216 class, where `=== 500` → `>= 500` stayed green because the test re-implemented the branch). For each added/modified test file it warns when the file textually duplicates a ≥3-token conditional from a production file touched in the same PR **and** does not import the symbol/module owning it, via a pure, fixture-tested `findReimplementedConditions()` with an allowlist mirroring `assertReductionAllowlist`. Report-only for now (does not fail the gate) — to be promoted to blocking after a triage cycle. Regression guard: `tests/unit/check-test-masking.test.ts` (45).
+- **feat(sse):** per-connection routing override (native vs CLIProxyAPI) ([#6339](https://github.com/diegosouzapw/OmniRoute/issues/6339)) — the previously-dead `isCliproxyapiDeepModeEnabled` helper is now wired into `resolveExecutorWithProxy`: a single connection can opt itself into the CLIProxyAPI passthrough executor via `providerSpecificData.cliproxyapiMode="claude-native"`, with precedence **connection override > provider `upstream_proxy_config` mode > default**. `resolveExecutorWithProxy` now receives the resolved connection's `providerSpecificData` (threaded from `chatCore.ts`), so one connection can deep-route while the provider's other connections stay native — no DB schema change (the toggle rides in `providerSpecificData`). Also resolves the same-provider mixing ask in #6340. Regression guard: `tests/unit/chatcore-executor-proxy.test.ts` (9). (thanks @RaviTharuma)
+- **feat(dashboard):** "Add session cookie" modal now shows a prominent **"Open ‹host› →"** link to the provider's own site ([#6268](https://github.com/diegosouzapw/OmniRoute/issues/6268)) — every `-web` cookie-session provider (chatgpt-web, claude-web, gemini-web, kimi-web, lmarena, qwen-web, m365-copilot-web, …) renders a one-click external link (opening the provider's login/home page in a new tab) so operators no longer tab away to retype the URL mid-setup. The host resolves from a pure, unit-tested `resolveWebProviderHost()` (prefers `WEB_COOKIE_PROVIDERS[id].website`, falls back to the registry `baseUrl` origin); non-web providers render exactly as before. Kilo's dup flag vs #6265 (modal-too-small-on-1080p) was a false positive — distinct concern. Regression guard: `tests/unit/resolve-web-provider-host.test.ts` (5). (thanks @chirag127)
+- **feat(providers):** add **DigitalOcean** AI (serverless inference) as an OpenAI-compatible API-key provider ([#6373](https://github.com/diegosouzapw/OmniRoute/pull/6373)) — base `https://inference.do-ai.run/v1`, wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/digitalocean/`, `src/shared/constants/providers/apikey/inference-hosts.ts`). Regression guard: `tests/unit/digitalocean-provider.test.ts`. (thanks @newnol)
+- **feat(providers):** add **Huancheng Public API** (`hcnsec`) as an OpenAI-compatible regional provider ([#6410](https://github.com/diegosouzapw/OmniRoute/pull/6410)) — Xinjiang Huancheng Cybersecurity's public LLM platform (base `https://api.hcnsec.cn/v1`, free credits via daily check-ins), wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/hcnsec/`, `src/shared/constants/providers/apikey/regional.ts`). Regression guard: `tests/unit/hcnsec-provider.test.ts`. (thanks @UnrealAryan)
+- **feat(dashboard):** the web-session credential guide now shows an **"Open {host}" link** ([#6316](https://github.com/diegosouzapw/OmniRoute/pull/6316)) to the provider's sign-in site (derived from the provider `website` via `getProviderWebsiteHost`), so you can jump straight to the page where the cookie/session must be captured. Regression guard: `tests/unit/web-session-provider-link-6316.test.ts`. (thanks @jordansilly77-stack)
+- **feat(cerebras):** add the **Gemma 4 31B** model (`gemma-4-31b`) to the Cerebras registry + pricing table ([#6331](https://github.com/diegosouzapw/OmniRoute/pull/6331)). Regression guard extends `tests/unit/t28-model-catalog-updates.test.ts`. (thanks @backryun)
+- **feat(providers):** add **Yuanbao (web)** as a cookie-session provider ([#6196](https://github.com/diegosouzapw/OmniRoute/issues/6196)) — `yuanbao-web` (Tencent Yuanbao, `yuanbao.tencent.com`) with cookie-only auth (`hy_user`/`hy_token` + public agent id), SSE→OpenAI translation incl. `reasoning_content`, exposing DeepSeek V3/R1 + Hunyuan / Hunyuan-T1. Regression guard: `tests/unit/providers-yuanbao-web.test.ts`. `together-web` was **deferred** (no verifiable web-session endpoint — needs a captured request) and `huggingchat-web` **dropped** (the existing `huggingchat` already is a web-cookie provider). (thanks @chirag127)
+- **feat(providers):** route the built-in **agentrouter** through the dynamic Claude-Code wire image ([#6056](https://github.com/diegosouzapw/OmniRoute/issues/6056)) — a small static allow-set (`CC_WIRE_IMAGE_BUILTINS` in `open-sse/services/ccWireImageBuiltins.ts`), consulted by `isClaudeCodeCompatible` / `isClaudeCodeCompatibleProvider` / `applyFingerprint`, makes agentrouter adopt the CC wire-image headers + fingerprint **while guarding the CC baseUrl/auth branches** so it keeps its own registry `baseUrl` and `x-api-key` auth. Regression guard: `tests/unit/agentrouter-cc-wire-image.test.ts` (asserts the wire image is applied AND agentrouter's baseUrl/auth are preserved). Live WAF-acceptance against agentrouter.org is a VPS validation follow-up (Hard Rule #18).
+- **feat(providers):** **bulk-add API keys for Cloudflare Workers AI** ([#6174](https://github.com/diegosouzapw/OmniRoute/issues/6174)) — `cloudflare-ai` is removed from the bulk-add exclusion list and the bulk parser gains a 3-field `name|accountId|apiKey` mode; the bulk route now builds a **per-entry** `providerSpecificData` so each key carries its own `accountId` (fixing the previous shared-object reuse), and both the create + key-validation paths receive it. Regression guard: `tests/unit/bulk-api-key-parser-cloudflare.test.ts`. (thanks @muflifadla38)
+- **feat(dashboard):** routing/settings UX clarity ([#6147](https://github.com/diegosouzapw/OmniRoute/issues/6147)) — (1) weighted combos show the **effective routing share %** next to each weight when weights don't sum to 100 (`WeightTotalBar.tsx`); (2) the status widget's user-facing **"Cloud Sync" label is renamed** to "Remote Settings Sync" (`CloudSyncStatus.tsx`; internal ids/state untouched); (3) built-in providers gain an **opt-in advanced base-URL override** (`isBaseUrlOverrideEligibleProvider`, hidden behind an "Advanced" toggle, reusing the existing `providerSpecificData.baseUrl` persistence — not globally widened). Regression guard: `tests/unit/routing-settings-ux-6147.test.ts`.
+- **feat(combo):** add an option to **disable session stickiness**, per-combo or globally — round-robin / random combos can rotate to a different connection on every request instead of pinning a whole conversation to one connection by its first-message hash. Resolution precedence per-combo `config.disableSessionStickiness` → global `settings.disableSessionStickiness` → default `false` (preserves the #3825 prompt-cache/504 fix); gates **both** stickiness call sites in `open-sse/services/combo.ts`. Exposed as a global toggle (Combo Defaults) and a per-combo Inherit/on/off control. ([#6168](https://github.com/diegosouzapw/OmniRoute/issues/6168)) Regression guard: `tests/unit/combo-disable-session-stickiness.test.ts`. (thanks @RCrushMe)
+- **feat(docker):** add the `OMNIROUTE_NO_SUDO` env flag for root-less / user-namespaced deployments — the MITM cert-trust command path (`resolveSudoSpawn` in `src/mitm/systemCommands.ts`) now strips the leading `sudo` when the flag is truthy, in addition to the existing root / sudo-missing cases, so the Proxy Agent runs without `sudo` (the operator trusts the CA manually, e.g. via `NODE_EXTRA_CA_CERTS`). Argv-array `spawn` preserved — no shell interpolation (Hard Rule #13). ([#6122](https://github.com/diegosouzapw/OmniRoute/issues/6122)) Regression guard: `tests/unit/mitm-systemCommands-no-sudo.test.ts`. (thanks @powellnorma)
+- **feat(providers):** add **Requesty** as an OpenAI-compatible gateway provider (BYOK, base `https://router.requesty.ai/v1`, ~200 free requests/day) — wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/requesty/`, `src/shared/constants/providers/apikey/gateways.ts`). ([#6120](https://github.com/diegosouzapw/OmniRoute/issues/6120)) Regression guard: `tests/unit/requesty-provider.test.ts`. (thanks @chirag127)
+- **feat(dashboard):** add **configured-only / available-only filters** to the Free Provider Rankings page ([#6150](https://github.com/diegosouzapw/OmniRoute/issues/6150)) — hide providers you haven't configured, or whose connections are all rate-limited / out of quota, via server-side query params (`?configuredOnly` / `?availableOnly` on `GET /api/free-provider-rankings`) backed by a testable lib helper reusing the in-process connection state (no Redis). Both filters default off, so the default view is unchanged; this supersedes the earlier client-side "Configured Only" toggle (#6245) with an available-only dimension and unit-tested logic. Regression guard: `tests/unit/freeProviderRankings-filters.test.ts`.
+
+### 🔧 Bug Fixes
+
+- **fix(dashboard):** adding a second API key connection for the same provider no longer silently overwrites the first — the Add-API-key modal now derives a unique default connection name (`main`, then `main-2`, `main-3`, …) so the backend name-based upsert can't collide (#6499 — thanks @dilneiss).
+- **fix(compression):** the session-dedup engine now also deduplicates a large multi-line block repeated **within a single message** (intra-message dedup), not just across turns; the compression-preview API surfaces a `fallbackReason`, and the fusion panel reports how many models were rate-limited vs failed on total-panel failure (#6501 — thanks @chirag127).
+- **fix(compression):** a stacked-pipeline step naming an unregistered engine now surfaces a `validationErrors` entry instead of silently no-op'ing, so misconfigured pipelines are visible in the preview API (#6506 — thanks @chirag127).
+- **feat(usage):** add a **Codex reset-credit redemption** flow to the Provider Limits UI ([#6361](https://github.com/diegosouzapw/OmniRoute/pull/6361)) — a `useCodexResetCreditRedemption` hook + `/api/usage/codex-reset-credit` route + `codexResetCredits` lib let you redeem banked Codex reset credits from the quota card. Regression guard: `tests/unit/codex-reset-credits.test.ts`. (thanks @JxnLexn)
+- **feat(glm):** add **team-plan quota settings** for `glm-cn` connections ([#6351](https://github.com/diegosouzapw/OmniRoute/pull/6351)) — a dedicated `GlmTeamQuotaFields` form section (team quota id / limits) threaded through the Add/Edit connection modals, persisted via `providerSpecificData`, with the GLM usage service reading the team quota. Regression guards: `tests/unit/glm-team-quota.test.ts`, `provider-specific-data-schema.test.ts`. (thanks @hao3039032)
+- **feat(providers):** add **TinyFish** web-fetch/search support ([#6349](https://github.com/diegosouzapw/OmniRoute/pull/6349)) — a `tinyfish-fetch` executor + `/v1/web/fetch` route + MCP web-fetch tool, registered as a specialty-media provider with request-validation and a search-provider catalog entry. Regression guards: `tests/unit/executor-tinyfish-fetch.test.ts`, `web-fetch-handler.test.ts`, `mcp-web-fetch-tool.test.ts`, `provider-validation-tinyfish.test.ts`. (thanks @dtybnrj)
+- **fix(cli):** `omniroute launch-codex` now spawns `codex.cmd` through a shell on Windows (the npm `.cmd` shim is unresolvable by bare `spawn` → ENOENT), mirroring the qodercli Windows fix (#6263) ([#6312](https://github.com/diegosouzapw/OmniRoute/pull/6312)). Regression guard: `tests/unit/launch-codex-windows-spawn-6312.test.ts`. (thanks @swingtempo)
+- **fix(codex):** isolate the **Spark** quota from the shared Codex quota and stabilize the quota UI ordering / hydration so per-scope limits render consistently ([#6336](https://github.com/diegosouzapw/OmniRoute/pull/6336)). Regression guards: `tests/unit/codex-quota-selection-hydration.test.ts`, `provider-limits-ui.test.ts` + 3 more. (thanks @xz-dev)
+- **feat(api):** add a `hidePaidModels` setting that filters paid-only models out of the `/v1/models` catalog. Regression guard: `tests/unit/models-catalog-hide-paid.test.ts`. (thanks @chirag127)
+- **fix(api-manager):** the fallback model picker now preserves combos instead of dropping them when a primary model is unavailable ([#6443](https://github.com/diegosouzapw/OmniRoute/pull/6443)). Regression guard: `tests/unit/api-manager-page-static.test.ts`. (thanks @jmengit)
+- **fix(providers):** recoverable Antigravity / Cloud-Code (Gemini Code Assist) `403` responses ([#6452](https://github.com/diegosouzapw/OmniRoute/pull/6452)) are now classified as a retryable project-config error instead of a terminal account ban, so a fixable project/API-disabled 403 no longer forces a ~1-year cooldown / full OAuth reconnect. Regression guard: `tests/unit/errorclassifier-antigravity-403.test.ts`. (thanks @developerjillur)
+- **fix(mitm):** `sanitizeHeaders` now redacts `Set-Cookie` response headers so upstream session cookies never leak into logs / diagnostics ([#6451](https://github.com/diegosouzapw/OmniRoute/pull/6451)). Regression guard: `tests/unit/mitm-sanitize-headers.test.ts`. (thanks @developerjillur)
+- **fix(api):** `/api/compression/preview` now accepts `mode: "caveman"` and correctly handles stacked / zero-compression previews ([#6425](https://github.com/diegosouzapw/OmniRoute/issues/6425)). Regression guard: `tests/unit/api/compression-preview-caveman-and-stacked-6425.test.ts`. (thanks @chirag127)
+- **feat(providers):** add **Zed** hosted LLM aggregator as a native-app provider ([#6118](https://github.com/diegosouzapw/OmniRoute/pull/6118)) — OAuth sign-in via the Zed hosted flow, registered through the shared provider registry + executor. Regression guards: `tests/unit/zed-oauth-provider.test.ts`, `zed-import-utils.test.ts`, `zed-docker-detect.test.ts`, `mitm-handler-zed.test.ts`. VPS-validated via live operator login (Hard Rule #18).
+- **fix(oauth):** the Kiro SSO-cache auto-import now **preserves the IDC region** — cross-region Amazon Q / Kiro profiles imported from the SSO cache are no longer collapsed to the default region ([#6113](https://github.com/diegosouzapw/OmniRoute/pull/6113)). Regression guard: `tests/unit/kiro-auto-import-idc-2059.test.ts`. VPS-validated via live operator login (Hard Rule #18).
+- **fix(dashboard):** passthrough model aliases no longer collide when two namespaced model ids share a last segment (port from 9router#1850, [#6431](https://github.com/diegosouzapw/OmniRoute/pull/6431)). `enx/gpt-5.5` and `enx/codebuddy/gpt-5.5` both auto-generated the alias `gpt-5.5`, so the second model could never be added (the UI just alerted "alias already exists"). Aliases are now disambiguated deterministically — bare last segment when free, then parent-qualified (`codebuddy-gpt-5.5`), then a numeric suffix — while re-adding the exact same model id is still blocked. Regression guard: `tests/unit/passthrough-alias-1850.test.ts`. (thanks @arpicato)
+- **fix(translator):** preserve a Gemini `functionResponse` co-located with other parts (another `functionCall`, or trailing `text`) in the same content when translating **Gemini → OpenAI** ([#6376](https://github.com/diegosouzapw/OmniRoute/pull/6376)). `convertGeminiContent()` early-returned the tool message on the first `functionResponse` part, dropping any co-located parts; such contents are now pre-split (one tool message per `functionResponse`, emitted first, plus one message for the remaining parts). Regression guard: `tests/unit/gemini-to-openai-function-response.test.ts`. (thanks @warelik)
+- **fix(headroom):** detect a python interpreter managed by **mise / pyenv / asdf / conda** (port from 9router#2353, [#6382](https://github.com/diegosouzapw/OmniRoute/pull/6382)). Headroom's python probe (`src/lib/headroom/detect.ts`) searched a hardcoded `PATH`, but version managers expose their interpreters via shim dirs that only join `PATH` through interactive-shell activation — which the non-interactive server never runs, so a managed python (≥3.10) was invisible and Headroom reported it missing. The search path now prepends the well-known shim/bin dirs (`~/.local/share/mise/shims`, `~/.pyenv/shims`, `~/.asdf/shims`, `$CONDA_PREFIX/bin`, `~/.local/bin`, respecting `MISE_DATA_DIR`/`PYENV_ROOT`/`ASDF_DATA_DIR` when set), and a new `HEADROOM_PYTHON` env override lets operators point straight at their interpreter (mirroring `HEADROOM_URL`). Still shell-free (`execFileSync`). Regression guard: `tests/unit/headroom-detect.test.ts` (5). (thanks @loopyd)
+- **fix(executors):** strip the OpenAI-Codex/Claude-CLI `client_metadata` passthrough field for **NVIDIA** requests (port from 9router#1887, [#6411](https://github.com/diegosouzapw/OmniRoute/pull/6411)). NVIDIA's OpenAI-compatible wrapper rejects it with `400 Unsupported parameter`, the same class already handled for `cerebras`/`mistral`; `nvidia` (executor `default`) was missing from the strip allowlist so Codex/Claude-Code passthrough requests 400'd. Regression guard: `tests/unit/executor-default-strip-client-metadata.test.ts` (+nvidia case). (thanks @phidinhmanh)
+- **fix(translator):** strip the Claude-style `thinking` field for **NVIDIA `z-ai/glm-5.2`** (port from 9router#2023, [#6413](https://github.com/diegosouzapw/OmniRoute/pull/6413)). NVIDIA's OpenAI-compatible wrapper 400s on `thinking` (a Claude-format client routed here leaves a `thinking:{type:"adaptive"}`); the existing strip rule only dropped `reasoning`. Same class already handled for `minimax-m2.7`. Regression guard: `tests/unit/nvidia-minimax-thinking-strip.test.ts` (+glm-5.2 case). (thanks @phidinhmanh)
+- **fix(translator):** suppress the streamed `</think>` close marker for the **Antigravity IDE** client (port from 9router#1061, [#6415](https://github.com/diegosouzapw/OmniRoute/pull/6415)). On thinking-only turns Antigravity rendered a bare `</think>` as the sole visible content, tripping its loop-detection and wasting requests. Antigravity's UA (`vscode/<v> (Antigravity/<v>)`) is added to the marker-suppress allowlist (alongside OpenCode); Claude Code / Cursor still get the marker, and `x-omniroute-thinking-marker: on` force-restores it. Regression guard: `tests/unit/think-close-marker-suppress-5245.test.ts`. (thanks @abdofallah)
+- **fix(executors):** strip nested `reasoning_content` from messages for **Mistral** (port from 9router#1649, [#6417](https://github.com/diegosouzapw/OmniRoute/pull/6417)). Mistral's API returns `422 extra_forbidden` when an assistant message carries `reasoning_content` (replayed thinking from a prior turn, e.g. via the Codex `/responses` path); the generic top-level 400 field-downgrade retry never covered the nested per-message field. `DefaultExecutor` now strips it for provider `mistral` only, so DeepSeek (which requires replayed `reasoning_content`) is unaffected. Regression guard: `tests/unit/mistral-strip-reasoning-content-1649.test.ts`. (thanks @xxy9468615)
+- **fix(executors):** strip the `client_metadata` passthrough field on the **OpenCode** path (port from 9router#1442, [#6418](https://github.com/diegosouzapw/OmniRoute/pull/6418)). OpenCode upstreams (e.g. `kimi-k2.6` via opencode-go) reject it with `400 "Extra inputs are not permitted, field: 'client_metadata'"`; the DefaultExecutor strip only covered cerebras/mistral and `OpencodeExecutor` extends `BaseExecutor` directly, so nothing removed it there. Regression guard: `tests/unit/opencode-strip-client-metadata-1442.test.ts`. (thanks @yanpaing007)
+- **fix(executors):** inject the `reasoning_content` echo for the native **Moonshot Kimi** provider (port from 9router#1480, [#6419](https://github.com/diegosouzapw/OmniRoute/pull/6419)). Kimi (executor `default`) is a thinking-mode upstream that 400s with "reasoning_content must be passed back" when a prior assistant turn lacks it; the placeholder injection was only wired into the OpenCode meta-provider, so direct multi-turn Kimi conversations failed. Scoped to `kimi` (gateway-served models matching the thinking-model name pattern are unaffected). Regression guard: `tests/unit/kimi-native-reasoning-injected-1480.test.ts`. (thanks @2220258345)
+- **fix(executors):** recover from a strict gateway's `context_management: Extra inputs are not permitted` 400 (port from 9router#1468, [#6420](https://github.com/diegosouzapw/OmniRoute/pull/6420)). **Claude Code** always sends a top-level `context_management` field; strict anthropic-compatible gateways reject it. The dedicated context-editing 400-fallback only fired when OmniRoute's own `contextEditing` feature was enabled (default off), so a client-sent field passed through untouched and 400'd. `context_management` is now in the generic reactive field-strip list, so it's stripped-and-retried once regardless of the feature flag (with correct request re-signing for claude-compatible relays). Regression guard: `tests/unit/provider-field-strips.test.ts`. (thanks @ohahe52-dot)
+- **fix(network):** enable **RFC 8305 Happy Eyeballs** (`autoSelectFamily`) on the direct-egress undici dispatcher (port from 9router#1237, [#6423](https://github.com/diegosouzapw/OmniRoute/pull/6423)). When DNS returns both IPv6 (AAAA) and IPv4 (A) and the IPv6 route is broken (e.g. a NAT64 `64:ff9b::` prefix without routing), undici tried IPv6 first and hung until `ETIMEDOUT` (then a 502 + account lockout), even though `curl` reached the same host. The direct dispatcher now races both families and uses whichever connects first. Proxy paths pin family via `proxyTls` and are unaffected. Regression guard: `tests/unit/direct-dispatcher-pipelining-4580.test.ts`. (thanks @adentdk)
+- **fix(combo):** round-robin now advances the rotation pointer past the model that **actually served**, not the eagerly-scheduled one (port from 9router#948, [#6428](https://github.com/diegosouzapw/OmniRoute/pull/6428)). With `stickyLimit: 1` (true round-robin), when the scheduled model failed and a _different_ model served via fallback, the counter had already advanced +1 from the scheduled index — so the next request reused the fallback-served model, degrading round-robin into hot-spotting on whichever model was healthy. The pointer now advances to the served index + 1 (mirroring the sticky-limit>1 path). Session-stickiness (#3825) and distribution are preserved. Regression guard: `tests/unit/combo-rr-fallback-advance-948.test.ts`. (thanks @binsarjr)
+- **fix(sse):** a non-string `model` field is now rejected with a `400` before the resolver, instead of crashing downstream `.toLowerCase()`/`.split()` calls into an empty-body `500` that escapes the error sanitizer ([#6407](https://github.com/diegosouzapw/OmniRoute/issues/6407)). Regression guard: `tests/unit/chat-non-string-model-6407.test.ts`. (thanks @chirag127)
+- **fix(api):** unknown `/api/*` routes now return a JSON `404` (instead of the dashboard HTML shell) and scalar chat params (`model`/`temperature`/etc.) are validated **before** the provider lookup so malformed requests fail fast with a clear `400` ([#6424](https://github.com/diegosouzapw/OmniRoute/issues/6424), [#6412](https://github.com/diegosouzapw/OmniRoute/issues/6412)). Regression guards: `tests/unit/api/api-catchall-json-404.test.ts`, `tests/unit/chat-early-schema-validation-6412.test.ts`. (thanks @chirag127)
+- **fix(api):** `/v1/chat/completions` now rejects a non-JSON `Content-Type` with a `400` before parsing the body ([#6414](https://github.com/diegosouzapw/OmniRoute/issues/6414)). Regression guard: `tests/unit/v1-chat-completions-content-type-6414.test.ts`. (thanks @chirag127)
+- **fix(api):** the `X-OmniRoute-Compression` response header is now echoed on `/v1/chat/completions` and `/v1/completions` ([#6422](https://github.com/diegosouzapw/OmniRoute/issues/6422)). Regression guard: `tests/unit/compression-header-echo-6422.test.ts`. (thanks @chirag127)
+- **fix(api):** concurrent `GET /v1/models` requests are coalesced into a single catalog build ([#6408](https://github.com/diegosouzapw/OmniRoute/issues/6408)). Regression guard: `tests/unit/v1-models-concurrent-6408.test.ts`. (thanks @chirag127)
+- **fix(api):** `/v1/completions` now echoes the requested `body.model` in its JSON + streamed responses ([#6429](https://github.com/diegosouzapw/OmniRoute/pull/6429)). Regression guard: `tests/unit/completions-body-model-echo.test.ts`. (thanks @chirag127)
+- **fix(api):** env-var master keys now see the full `/v1/models` catalog ([#6406](https://github.com/diegosouzapw/OmniRoute/issues/6406)). Regression guard: `tests/unit/models-catalog-envkey-6406.test.ts`. (thanks @chirag127)
+- **fix(api):** non-streaming `/v1/completions` responses now echo `body.model` aligned with the `X-OmniRoute-Model` header ([#6426](https://github.com/diegosouzapw/OmniRoute/issues/6426)). Regression guard: `tests/unit/v1-completions-model-header-match-6426.test.ts`. (thanks @chirag127)
+- **fix(api):** unknown `/v1/*` routes now return a JSON `404 not_found` instead of the Next.js dashboard HTML shell ([#6405](https://github.com/diegosouzapw/OmniRoute/issues/6405)). Regression guard: `tests/unit/api/v1-catchall-json-404.test.ts`. (thanks @chirag127)
+- **fix(api):** the per-connection provider models route now degrades to the shipped catalog when a provider's `/models` endpoint answers with a redirect ([#6267](https://github.com/diegosouzapw/OmniRoute/issues/6267)) — a `qwen-web` import failed with a raw `Redirect blocked … (307)` 503. `safeOutboundFetch` throws `REDIRECT_BLOCKED` on the 307, `getSafeOutboundFetchErrorStatus` maps it to 503, and `buildDiscoveryErrorFallbackResponse` treated every 503 as a hard error — so the non-empty `getModelsByProviderId("qwen-web")` catalog was never surfaced. A models-endpoint redirect is not a fixable-config error (unlike `URL_GUARD_BLOCKED`/`INVALID_URL`, which stay hard errors), so it now falls back to the local/cached catalog before the 503 short-circuit. General fix — covers any config-driven provider that 307s. Regression guard: `tests/unit/provider-models-qwen-web-redirect-6267.test.ts`. (thanks @chirag127)
+- **fix(api):** the per-connection provider models route (MCP `list_models_catalog` + the dashboard import view) now merges USER-ADDED custom models into its response ([#6247](https://github.com/diegosouzapw/OmniRoute/issues/6247)) — custom models live in the `key_value` namespace `customModels`, which the live REST `/api/v1/models` already merges, but `src/app/api/providers/[id]/models/route.ts` never read `getCustomModels`, so custom models were dropped on both the discovery-success and local_catalog paths. They are now folded into the returned model list (deduped by id, stamped `owned_by: provider`), fixing MCP + the dashboard import view in one place. Regression guard: `tests/unit/provider-models-custom-merge-6247.test.ts`. (thanks @RCrushMe)
+
+- **fix(providers):** GitLab Duo tool-calling follow-up turns no longer fail upstream with `422 {"detail":"Validation error"}` (tokens 0/0, rejected pre-inference). The #6234 tool-result-feedback fix serialized the **entire** multi-turn conversation into GitLab's single-file `code_suggestions` (`small_file`) generation endpoint — folded history that turn-N sent as an oversized `current_file.content_above_cursor` **and** duplicated verbatim into `user_instruction`, tripping the AI-Gateway's `small_file` validation guard. The executor now **bounds** that prompt: it keeps system + latest user message + the most-recent tool round (dropping older turns), caps oversized tool results, and stops duplicating the full prompt into `user_instruction` (which now carries only the short latest user message) — while still feeding the most-recent tool result back so the agent continues (`open-sse/executors/gitlab.ts`, [#6220](https://github.com/diegosouzapw/OmniRoute/issues/6220)). The unit test covers the bounding logic; the upstream 422→200 clearing is VPS-only (Hard Rule #18). Regression guard: `tests/unit/gitlab-tool-exchange-bounded-6220.test.ts`.
+
+- **fix(i18n):** the provider-detail (`/dashboard/providers/[id]`) connection-status filter labels no longer render as `__MISSING__:All` / `__MISSING__:Active` / `__MISSING__:Error` / `__MISSING__:Banned` / `__MISSING__:CreditsExhausted` in non-English locales (notably pt-BR) ([#6290](https://github.com/diegosouzapw/OmniRoute/issues/6290)). Root cause was **not** the namespace mismatch the issue guessed — the `providers.filter*` keys resolve correctly in `en.json`; the debt lived in the locale mirrors (`src/i18n/messages/*.json`), where these five keys carried the `__MISSING__:` sync sentinel in ~15 locales and were absent entirely in ~26 others, so next-intl found the key and echoed the sentinel verbatim. All 40 non-English/-Chinese mirrors now ship real translations for the five `providers.filter*` labels. Regression guard: `tests/unit/i18n-provider-filter-keys-6290.test.ts`. (thanks @diegosouzapw)
+
+- **fix(providers):** the `copilot-m365-web` streaming executor now emits `debug`-level WebSocket diagnostics ([#6210](https://github.com/diegosouzapw/OmniRoute/issues/6210)) — the outbound WS URL (with the `access_token` **redacted** via `redactWsUrl()`), handshake success/failure, and each received SignalR frame's `type`/`target`. Previously the streaming path logged nothing, so an empty `content:null` response (the M365 Education / Starter tier symptom fixed in #6234) was undiagnosable even at `APP_LOG_LEVEL=debug`. The change is debug-level and side-effect-free — it does not alter streaming behavior or the frame parser, and the token never reaches the logs. Regression guard: `tests/unit/copilot-m365-web-logging-6210.test.ts` (thanks @qpeyba)
+
+- **fix(resilience):** a round-robin combo no longer returns `503 all upstream accounts are unavailable` when a compatibility-rejected target is actually healthy ([#6238](https://github.com/diegosouzapw/OmniRoute/issues/6238)). `filterTargetsByRequestCompatibility` drops request-incompatible targets (tool/vision/structured-output unsupported, or below the required context window) **before** any availability check runs, and its `compatible.length === 0` safety net only fired when _all_ targets were filtered — not when the kept targets later all turned out runtime-unavailable (circuit-open / cooldown / no credentials). So a combo could 503 while a compat-rejected-but-healthy provider sat unused. `handleRoundRobinCombo` now keeps the compat-rejected set and, when every compat-kept target was skipped without a single real attempt, probes those rejected targets as a **last-resort fallback tier** (via the new pure `open-sse/services/combo/comboCompatFallback.ts`) before crystallizing the 503. Regression guard: `tests/unit/combo-roundrobin-compat-fallback-6238.test.ts`. (thanks @ThongAccount)
+
+- **fix(startup):** best-effort self-heal for a corrupted Turbopack dev cache on Windows ([#6289](https://github.com/diegosouzapw/OmniRoute/issues/6289)). On Windows, `pnpm dev` can fail at startup when Turbopack `mmap`s a persistent-cache SST file and the OS refuses the mapping (`os error 1455` — "paging file too small"), which Turbopack surfaces as a misleading `Module not found: Can't resolve '@/shared/utils/machine'`. This is a **known upstream Turbopack cache-corruption bug — not our code**. The dev launcher (`scripts/dev/run-next.mjs`) now wraps `nextApp.prepare()` and, when it rejects with that signature (`isTurbopackCacheCorruption` in the new `scripts/dev/turbopackCacheHeal.mjs`), purges `.build/next/**/cache/turbopack` and retries **once** with a clear log. **Caveat — best-effort only:** the corruption often surfaces as a runtime overlay rather than a `prepare()` rejection, so this cannot always intercept it; the reliable remedy remains manually deleting the Turbopack cache dir. Regression guard: `tests/unit/turbopack-cache-heal-6289.test.ts`. (thanks @chirag127)
+
+- **fix(providers):** qodercli PAT auth no longer fails with `spawn qodercli ENOENT` on Windows ([#6263](https://github.com/diegosouzapw/OmniRoute/issues/6263)) — `spawnQoderCli` spawned the bare `qodercli` name with `shell:false` and an unenriched env, so the npm `.cmd` wrapper under `%APPDATA%\npm` (a user-PATH directory) was never resolved. It now resolves the absolute `.cmd`/`.exe` path through the existing `getCliRuntimeStatus("qoder")` resolver in `src/shared/services/cliRuntime.ts` (memoized), spawns with `shell` when the target is a `.cmd`/`.bat`, and uses the cliRuntime-enriched env (PATH + PATHEXT + APPDATA); the ENOENT error now lists the searched paths plus the `CLI_QODER_BIN` override. End-to-end spawn on a real Windows host is host-only (Hard Rule #18); the path-resolution logic is unit-tested. Regression guard: `tests/unit/qodercli-windows-resolve-6263.test.ts`. (thanks @chirag127)
+
+- **fix(sse):** the reasoning-token buffer no longer inflates **probe-sized `max_tokens`** ([#6274](https://github.com/diegosouzapw/OmniRoute/issues/6274)) — Claude Code's `/model` capability check sends `max_tokens: 1`, but for a thinking-capable model with a large output cap (e.g. `glm-5.2`) the #3587 headroom heuristic (`max(current + 1000, ceil(current * 1.5))`) rewrote it to `1001` and forwarded that upstream, wasting tokens on a request that was never a genuine reasoning budget. `resolveReasoningBufferedMaxTokens()` (`open-sse/services/reasoningTokenBuffer.ts`) now short-circuits and returns the caller's value verbatim when it is below the new `REASONING_BUFFER_MIN_TRIGGER` (256) threshold — a tiny explicit limit is a probe, not a reasoning request. Real budgets still receive the #3587 headroom unchanged, and the guard runs after the existing capability checks so unknown / non-reasoning models keep returning `null`. Regression guard: `tests/unit/reasoning-token-buffer-6274.test.ts`. (thanks @brightfiscalband)
+
+- **fix(cli):** `omniroute reset-password` now works as a real subcommand, and password resets over piped (non-TTY) stdin actually apply ([#6261](https://github.com/diegosouzapw/OmniRoute/issues/6261), [#6258](https://github.com/diegosouzapw/OmniRoute/issues/6258)). Two coupled defects: (1) **#6261** — `bin/omniroute.mjs` routed everything through Commander with only two pre-Commander bypasses (`--mcp`, `reset-encrypted-columns`), so `omniroute reset-password` was rejected as an unknown command; only the separate `omniroute-reset-password` bin worked, while the docs falsely advertised the subcommand (incl. a bogus "legacy alias still works"). A pre-Commander bypass mirroring `reset-encrypted-columns` now dynamically imports `bin/reset-password.mjs` (which self-executes) before Commander parses; the three doc lines were corrected. (2) **#6258** — `bin/reset-password.mjs` issued two sequential `rl.question` prompts; under piped stdin the second read never settled at EOF, so `main()` never reached `resetManagementPassword` and the reset was a silent no-op (both prompts printed, no success, password unchanged). The CLI now detects non-TTY stdin and reads it once (first line = password, second line = confirm if present, else reused), adds a `--password-stdin` flag (entire stdin is the password, no confirmation), and exits `0` explicitly so the success line always flushes; interactive TTY behavior is unchanged. Regression guard: `tests/unit/reset-password-cli-6261-6258.test.ts` (3). (thanks @chirag127)
+
+- **fix(db):** the mass-migration **safety abort** now tells the operator how to bypass it and stops flooding the log ([#6260](https://github.com/diegosouzapw/OmniRoute/issues/6260)) — after restoring a backup that wiped the migration tracking table, `runMigrations()` threw the abort on every downstream `ensureDbInitialized()`, re-logging the full banner 11+ times, and the message never mentioned the existing `OMNIROUTE_MAX_PENDING_MIGRATIONS` escape hatch. The abort text now appends a bypass hint (set `OMNIROUTE_MAX_PENDING_MIGRATIONS=0` in `server.env` / `DATA_DIR/.env`), and a new `MigrationSafetyAbortError` is memoized so repeated calls in the same process throw the same instance and emit a single concise line instead of the full cascade. Regression guard: `tests/unit/migration-safety-abort-6260.test.ts`. (thanks @chirag127)
+
+- **fix(auth):** importing a **distinct** Codex/ChatGPT OAuth `auth.json` is no longer falsely rejected as "already exists" when it belongs to a different user in the same workspace ([#6301](https://github.com/diegosouzapw/OmniRoute/issues/6301)). `findExistingCodexConnection` (in `src/lib/oauth/utils/codexAuthImport.ts`) deduped **only** on `providerSpecificData.workspaceId === accountId`, where `accountId` is the shared `chatgpt_account_id`/`tokens.account_id` — so two members of the same ChatGPT Team collapsed onto a single connection (409 `duplicate_account`). The id_token's `https://api.openai.com/auth` claim carries a per-user `chatgpt_user_id` alongside the workspace id (the device-flow path already persisted it as `chatgptUserId`, but the import path did not). Now `parseAndValidateCodexAuth` extracts `userId` (`chatgpt_user_id` → `user_id` → JWT `sub`) into `ParsedCodexAuth`, the create/update paths persist `chatgptUserId` in `providerSpecificData` (mirroring `codex.ts`), and dedup keys on `workspaceId` **AND** `chatgptUserId` — with a backward-compat fallback to legacy accountId-only matching when no stored connection for that workspace records a `chatgptUserId`, so genuinely-same accounts still dedup. Regression guard: `tests/unit/codex-auth-import-userid-dedup-6301.test.ts` (4). (thanks @anungma)
+
+- **fix(providers):** importing models for the **venice-web** provider no longer fails with a red "Provider venice-web does not support models listing" ([#6269](https://github.com/diegosouzapw/OmniRoute/issues/6269)). `venice-web` is a web-cookie provider with an executor but no upstream `/v1/models` endpoint and no registry `models`, so the models route fell through to the tail `400`. Mirroring the `jules`/`linkup-search`/`ollama-search` fix (#5569), it now ships a static local catalog entry in `src/lib/providers/staticModels.ts` — seeding the current Venice lineup (`venice-uncensored`, `llama-3.3-70b`, `qwen3-235b`, `qwen3-4b`, `deepseek-r1-671b`; Venice rotates its catalog, see docs.venice.ai/models/overview) — so the route returns `200` with `source:"local_catalog"`, `intentional:true`. Regression guard: `tests/unit/static-models-venice-web-6269.test.ts`. (thanks @chirag127)
+
+- **fix(api):** the specialty model catalogs (`/v1/embeddings`, `/v1/images`, `/v1/music`, `/v1/videos` model lists) are now derived from the **unified catalog filtered by a predicate** (`getSpecialtyModelsResponse`) instead of ad-hoc per-route logic, so they consistently respect active-credential visibility and stay in sync with the main catalog ([#6303](https://github.com/diegosouzapw/OmniRoute/pull/6303)). Regression guard: `tests/unit/specialty-model-catalog-routes.test.ts`. (thanks @makcimbx)
+- **fix(api):** the agent-bridge server route now resolves the MITM manager via a **dynamic `import("@/mitm/manager.runtime")`** so Turbopack does not statically pull the stub (or over-bundle the manager), and the agent-skills generator anchors its output base path with `path.join(process.cwd(), …)` so Turbopack's static analyzer stops tracing the whole project root ([#6329](https://github.com/diegosouzapw/OmniRoute/issues/6329), [#6366](https://github.com/diegosouzapw/OmniRoute/pull/6366)). Regression guard: `tests/unit/agent-bridge-server-route-dynamic-import.test.ts`. (thanks @Iammilansoni)
+- **fix(api):** internal probes (combo-test, cloud-sync verify) now pick a **management-scoped / allow-all API key** instead of naively grabbing `getApiKeys()[0]` — a restricted `self:usage` first row made the probe fail with "Model X is not allowed for this API key" even when the combo path was healthy (`pickApiKeyForInternalUse` in `src/lib/db/apiKeys.ts`). The API-manager model editor also falls back to `/api/models?all=true` when `/v1/models` is catalog-protected ([#6372](https://github.com/diegosouzapw/OmniRoute/pull/6372)). Regression guard: `tests/unit/pick-internal-api-key-6372.test.ts`. (thanks @jmengit)
+- **fix(live-ws):** the Live Dashboard WebSocket server now **rejects on bind failure** (e.g. `EADDRINUSE` when the API bridge already holds the port) instead of letting the error surface as an unhandled `error` event that crash-loops the process — the `error` listener is attached to `wss` (not `server`) and releases the EventBus subscription on a failed start ([#6324](https://github.com/diegosouzapw/OmniRoute/issues/6324)). Regression guard: `tests/unit/live-ws-eaddrinuse-6324.test.ts`. (thanks @vinayakkulkarni)
+- **fix(dashboard):** the Home provider-topology widget now trusts the live provider-metrics snapshot — it uses `topology.errorProvider` and live `activeRequests` directly instead of re-deriving state from a stale `lastErrorAt` or applying a frontend timeout filter, so the topology reflects real-time provider health ([#6322](https://github.com/diegosouzapw/OmniRoute/pull/6322)). Regression guard: `tests/unit/home-provider-topology-live-state.test.ts`. (thanks @xz-dev)
+- **fix(sse):** strip zero-width markers from streamed **tool-call arguments** — a follow-up to [#5857](https://github.com/diegosouzapw/OmniRoute/pull/5857). That PR removed injected zero-width joiners (U+200D) from streamed assistant text/reasoning but deliberately left tool-call argument JSON byte-exact. The request-side obfuscation (`open-sse/services/claudeCodeObfuscation.ts`) injects ZWJ into agent words — including the temp path inside the Bash tool description — and Claude models copy that verbatim into generated commands, which are delivered as tool-call arguments rather than assistant text. As a result the ZWJ survived and corrupted code blocks (e.g. a temp path rendered with an invisible joiner). Now `open-sse/handlers/responseSanitizer.ts` strips zero-width code points from tool-call argument strings at every emit site (OpenAI non-stream/stream chat `tool_calls` + legacy `function_call`, native Responses `function_call` items, the OpenAI→Responses conversion, and the native Responses streaming `response.function_call_arguments.delta/.done` events). Only zero-width code points are removed; JSON structure and all other bytes stay identical (no parse/restringify), so normal arguments remain byte-exact. Regression guard: 6 new cases in `tests/unit/response-sanitizer.test.ts` (suite 50/50).
+
+- **fix(nodejs):** the default app log path now resolves under `DATA_DIR` (`~/.omniroute/logs/application/app.log`) instead of `process.cwd()` ([#6197](https://github.com/diegosouzapw/OmniRoute/issues/6197)) — the globally-installed CLI runs from an arbitrary working directory, so anchoring the default to cwd made file logging silently write to (or no-op under) an unrelated directory, contradicting the documented `.env.example` default. `getAppLogFilePath()` now computes the default lazily via the pure `resolveDataDir()` resolver (honours a per-process `DATA_DIR`, no directory-creation side effect); an explicit `APP_LOG_FILE_PATH` still wins. Regression guard: `tests/unit/logenv-datadir-path-6197.test.ts` (3).
+
+- **fix(docker):** AgentBridge/`startMitm` no longer aborts in containers/headless when the Antigravity-default DNS step can't write `/etc/hosts` ([#6127](https://github.com/diegosouzapw/OmniRoute/issues/6127)), and the privileged command's stderr now reaches `app.log` instead of only a bare exit code hitting the toast ([#6198](https://github.com/diegosouzapw/OmniRoute/issues/6198)). The default DNS step (`addDNSEntry`) was called unguarded while cert install and the two sibling DNS steps were each best-effort — in the runtime Docker image (`USER node`, no `sudo`, read-only `/etc/hosts`) it threw `Command failed with code 1` out of `startMitmInternal` and killed the whole start, discarding the stderr. The three DNS steps are extracted into a best-effort `provisionDnsEntries()` where each failure is logged with the full `err` (stderr included, folded in by `systemCommands.ts`) and never aborts the start. Regression guard: `tests/unit/mitm-dns-graceful-degrade-6127.test.ts` (4).
+
+- **fix(providers):** copilot-m365-web now supports the M365 Education "Starter / OfficeWebIncludedCopilot" tier and no longer returns an empty `content:null` stream ([#6210](https://github.com/diegosouzapw/OmniRoute/issues/6210)). Two gaps: (1) `buildWsUrl()` hardcoded the individual-consumer scenario (`OfficeWebPaidConsumerCopilot`, `isEdu=false`) — the EDU tier is now opt-in via `providerSpecificData.tier="edu"`, emitting `scenario=OfficeWebIncludedCopilot`/`isEdu=true` (the individual path is unchanged); (2) the EDU/GPT-5.5 path streams deltas via `arguments[0].writeAtCursor` (incremental) instead of only `messages[].text` (accumulated snapshots), which the parser dropped — a new `accumulateBotContent()` folds both formats, with `type:2 item.result.message` as a last-resort fallback. Regression guard: `tests/unit/copilot-m365-edu-writeatcursor-6210.test.ts` (10). (thanks @qpeyba)
+
+- **fix(providers):** GitLab Duo executor now feeds tool results back into the prompt instead of looping ([#6220](https://github.com/diegosouzapw/OmniRoute/issues/6220)) — `buildPrompt()` branched only on `system`/`user` and took `userParts.at(-1)`, silently dropping the `assistant{tool_calls}` + `tool{result}` turns the client appended, so the reconstructed prompt was byte-identical to turn 1 and the model re-emitted the same `<tool>` call forever. When a tool exchange is present the full conversation is now serialized, folding each tool result back keyed by its `tool_call_id`; simple conversations keep the legacy shape. Complements the tool_call emission from [#6051](https://github.com/diegosouzapw/OmniRoute/issues/6051) (the `kilo-duplicate` label was a false positive — different, sequential defect). Regression guard: `tests/unit/gitlab-tool-result-feedback-6220.test.ts` (4).
+
+- **fix(providers):** opencode-go/opencode-zen can now synthesize the OpenCode CLI identity headers Cloudflare requires on VPS egress ([#5997](https://github.com/diegosouzapw/OmniRoute/issues/5997)) — on a datacenter VPS, `opencode.ai/zen/go/v1/chat/completions` 403s (HTML challenge) requests lacking CLI identity, while the reporter's control curl proved that `User-Agent: opencode-cli/1.0.0` + `x-opencode-client: cli` + `x-opencode-project: default` + fresh request/session UUIDs succeed. Opt-in via `OPENCODE_SYNTHESIZE_CLI_HEADERS=true` (values overridable via `OPENCODE_GO_USER_AGENT`/`OPENCODE_USER_AGENT`/`OPENCODE_CLIENT`/`OPENCODE_PROJECT`); it fills only headers the client did not already send. Kept **off by default** — the forward-only path is deliberate (fabricating a wrong value risks upstream rejection; a prior dedup regressed with `opencode/local`), so this replaces the fragile local header-injection shim without changing default behavior. Regression guard: `tests/unit/opencode-cli-headers-synthesis-5997.test.ts` (6). (thanks @aleksesipenko)
+- fix(resilience): sticky session affinity now evicts and fails over to another account when the pinned account is exhausted/unavailable (#6219)
+
+- fix(sse): Responses API passthrough now drops internal commentary-phase output before forwarding to clients (gated by RESPONSES_PASSTHROUGH_DROP_COMMENTARY, default on) (#6199)
+
+- **fix(sse):** tool-call function schemas with a root `type: null` are now coerced to `type: "object"` before dispatch ([#6359](https://github.com/diegosouzapw/OmniRoute/issues/6359)) — clients like the Codex app emit `parameters: { type: null, ... }` for some tools, which OpenAI-compatible upstreams reject with `400 Invalid schema for function '...': schema must be a JSON Schema of 'type: "object"', got 'type: null'`, failing the whole request. `toolSchemaSanitizer` already stripped the null; it now re-adds the mandatory root `"object"` type (and empty `properties`/open `additionalProperties` when absent). Combinator roots (`anyOf`/`oneOf`/`allOf`) and explicit root types are left untouched. Regression guard: 5 new cases in `tests/unit/tool-schema-sanitizer.test.mjs`.
+
+- **fix(docker):** AgentBridge no longer fails to start on npm/Electron/VPS installs with "MITM manager stub reached at runtime" ([#6344](https://github.com/diegosouzapw/OmniRoute/issues/6344)) — v3.8.45 flipped the production bundler default to Turbopack, but `next.config.mjs` aliased `@/mitm/manager` to its Docker-only degraded stub **unconditionally**. That was harmless while Docker (which sets the alias intentionally for #3390 graceful degradation) was the sole Turbopack consumer, but once every artifact built with Turbopack the stub shipped to all non-Docker users and `startMitm` threw on the first Agent-Bridge start. The alias is now opt-in via `OMNIROUTE_MITM_STUB=1` (set only by the Dockerfile) through the shared `scripts/build/mitm-stub-flag.mjs` helper; default builds bundle the real manager. Regression guard: `tests/unit/mitm-stub-alias-6344.test.mjs` (4).
+
+- **fix(proxy):** stop the v3.8.44 proxy regression that leaked the real IP and disabled healthy proxies ([#6246](https://github.com/diegosouzapw/OmniRoute/issues/6246)). Two coupled defects from the new health scheduler: (1) **IP leak** — when a proxy assigned to a connection was marked `inactive`, resolution fell through to a **direct** egress instead of blocking, exposing the operator's real IP; (2) **over-deactivation** — the sweep flipped a proxy to `inactive` on the **first** failed probe and counted our own 5s timeout / a probe-target `5xx` as the proxy's fault, so healthy paid proxies vanished from egress selection ("my proxies are not being used anymore"). Fix: the sweep decision is extracted into a pure, network-free `decideProxyHealthAction` (`src/lib/proxyHealth/decision.ts`) — by default the health check now **only counts/logs and never downgrades status** (a proxy is downgraded/removed only with `PROXY_AUTO_REMOVE=true`, after `PROXY_AUTO_REMOVE_AFTER` **consecutive** conclusive failures); probes are classified tri-state so an inconclusive result (our timeout, or a `5xx` from the probe target) never penalizes the proxy, and the probe timeout is raised 5s→15s. Separately, `safeResolveProxy` now **fails closed** via the existing policy: a connection whose assigned proxy is dead is blocked instead of leaking direct (`hasBlockingProxyAssignment`), honoring the explicit `proxy off` toggles and the `PROXY_FAIL_OPEN=true` opt-out. Existing proxies stuck `inactive` by the old behavior need a one-time manual re-activate (the operator owns proxy status). Regression guards: `tests/unit/proxy-health-decide-action-6246.test.ts`, `tests/unit/proxy-assigned-unavailable-6246.test.ts`.
+
+- **fix(proxy):** make "Test All" read-only and add bulk enable/disable ([#6246](https://github.com/diegosouzapw/OmniRoute/issues/6246)). Complements the core fail-closed / scheduler fix (#6296) with the two remaining reporter asks. (1) The **"Test All" button** (`POST /api/settings/proxies/auto-test`) used to flip a proxy to `inactive` on a failed reachability probe; since the egress selector excludes `inactive` proxies, a flaky probe (an unreachable `httpbin.org`, a proxy that blocks `HEAD`, or a slow paid proxy) silently disabled every proxy that failed — "Test All" is now **read-only by default** (only the operator sets a proxy active/inactive; opt back into the legacy test-and-set with `PROXY_HEALTH_AUTO_DEACTIVATE=true`). (2) Adds a **bulk enable/disable** proxies endpoint + toolbar action (`POST /api/settings/proxies/batch-activate`) so an operator can re-activate proxies in one click. Regression guard: `tests/unit/proxy-health-6246.test.ts`. (thanks @tenshiak)
+
+- **chatcore (tools): stop the default 128-tool cap from silently dropping opencode's `task`/MCP tools.** opencode (used as an MCP/agent host) sends a large tool list; when it exceeds the speculative `MAX_TOOLS_LIMIT` (128) default, `truncateToolList` did a blind `tools.slice(0, 128)`, dropping every tool past index 128 — including opencode's built-in `task` tool (subagent launch) and many MCP tools, so models routed through OmniRoute could no longer spawn subagents or reach part of their tools. The cap exists to avoid upstream `400`s for providers with real hard limits (e.g. grok-cli 200), so it is kept for those: detection of the opencode client (`isOpencodeClient` — any `x-opencode-*` header, or `opencode` in the user-agent) now only bypasses the **speculative 128 default**, never a known provider ceiling. Precedence is explicit — a proactive/detected provider limit always truncates (even for opencode); otherwise opencode forwards its full tool list; otherwise the unchanged 128 default applies to every other client. Refactors `getEffectiveToolLimit` into `getKnownToolLimit(provider) ?? DEFAULT_LIMIT` (byte-identical for existing callers) and fixes a cosmetic debug-log that reported the truncated count instead of the original. Regression guard: `tests/unit/tool-limit-detector.test.ts`.
+
+- **fix(mitm):** the macOS MITM-cert install check now matches the system keychain again. `security find-certificate -a -Z` prints the SHA-1 as a colon-less hex string, but the installed-check compared it against `getCertFingerprint()`'s colon-separated form, so the substring match never hit — the cert was reported as not-installed and re-prompted for the sudo install on every run. Fingerprints are now normalized (colons stripped, upper-cased) on both sides via the extracted `macCertOutputHasFingerprint` helper. Regression guard: `tests/unit/mitm-cert-mac-fingerprint.test.ts`. ([#6204](https://github.com/diegosouzapw/OmniRoute/pull/6204), closes [#6134](https://github.com/diegosouzapw/OmniRoute/issues/6134) — thanks @rianonehub)
+
+- **fix(api):** `/v1/messages/count_tokens` now counts `tool_use`, `tool_result` and `thinking` content blocks (and array-form `system` prompts) in the local-estimation path, instead of only `text`. Real agentic conversations keep ~95% of their tokens inside tool results; the previous estimate returned near-zero for them, which silently broke Claude Code's auto-compaction (context grew past the window with no compaction until the upstream API rejected the request). The real provider-side count path is unchanged. Regression guard: `tests/unit/messages-count-tokens-route.test.ts`. ([#6221](https://github.com/diegosouzapw/OmniRoute/pull/6221) — thanks @luweiCN)
+
+- **fix(antigravity):** strip a trailing assistant prefill turn for Vertex Claude models to avoid upstream 400s ([#6114](https://github.com/diegosouzapw/OmniRoute/pull/6114)). Regression guard: `tests/unit/antigravity-claude-prefill-strip.test.ts`. (thanks @anki1kr)
+
+- **fix(security):** the mutable cloud-agent routes (`/api/cloud/credentials/update`, `/api/cloud/models/alias`) now require management auth instead of being treated as public. They were classified as public API routes, so a request without management credentials could update stored cloud-agent credentials and model aliases. They are removed from the public-route set, classified as management routes in the authz pipeline, and gated by `requireManagementAuth`; cloud **read**/auth routes stay public. Regression guards: `tests/unit/cloud-write-auth.test.ts`, `tests/unit/authz/classify.test.ts`, `tests/unit/public-api-routes.test.ts`. ([#6233](https://github.com/diegosouzapw/OmniRoute/pull/6233) — thanks @vittoroliveira-dev)
+
+- **refactor(dashboard):** extract the onboarding-wizard "Open provider details" link target into a pure, unit-tested `buildProviderDetailsHref(connection)` helper. The wizard already routes by `connection.id` (the node UUID) rather than the provider category slug (#6144/#6145); this hardens that behavior behind a tested helper that guards a missing id/connection. Regression guard: `tests/unit/provider-onboarding-href.test.ts`. ([#6166](https://github.com/diegosouzapw/OmniRoute/pull/6166) — thanks @KooshaPari)
+
+- **fix(security):** the doubao synthetic device-id generator now derives its digits via an unbiased crypto-random draw (rejection sampling over `crypto.randomBytes()`) instead of a `% 10` reduction, closing a CodeQL `js/biased-cryptographic-random` finding.
+- **fix(agentSkills):** the GitHub-skills generator now resolves `outputDir` to an absolute path before writing, fixing a regression introduced by #6366 (relative-to-cwd base path) that could write generated skill files to the wrong directory.
+- **fix(security):** `/api/keys/{id}/devices` now checks the HTTP method before auth/validation, returning a `405` for non-GET/DELETE verbs instead of a misleading `401`/`500` (closes a `dast-smoke` QUERY-method finding).
+- **fix(quality):** clear the last 2 heavy quality-gate reds on the release tip (cycle pre-flight).
+- **fix(mitm):** the test suite and CI can never mutate the OS trust store — `OMNIROUTE_SKIP_SYSTEM_TRUST=1` is set globally for tests/CI so `installCert`/`uninstallCert`/`installTproxyCa` skip the privileged OS dispatch ([#6310](https://github.com/diegosouzapw/OmniRoute/pull/6310); full detail is under the [3.8.45] section below — this branch received it via the parallel-cycle sync-back).
+- **fix(api):** `POST /api/github-skills` now Zod-validates its request body; documented the new quality-gate env vars and pinned the merge-integrity GitHub Actions to a commit SHA.
+- **fix(skills):** generate the missing `omni-github-skills` registry entry and align the agent-skills catalog-count tests (follow-up to #6186).
+- **fix(quality):** clear the cycle's 11 net-new ESLint errors and make `validate-release-green` suppressions-aware.
+
+### 📝 Maintenance
+
+- **i18n(it):** add 118 missing Italian (`it`) translations (net-additive — no existing keys dropped, valid JSON), improving Italian UI coverage. ([#6212](https://github.com/diegosouzapw/OmniRoute/pull/6212) — thanks @serverless83)
+- **chore(providers):** remove deprecated **MiMo V2** model entries from the catalogs (xiaomi-mimo, opencode-go, zenmux-free, audio TTS) — the upstream V2 line is superseded by MiMo V2.5; drops `mimo-v2-tts`, `mimo-v2-pro`, `mimo-v2-omni`, `mimo-v2-flash`, `mimo-v2-flash-free` and realigns the provider-catalog tests. ([#6248](https://github.com/diegosouzapw/OmniRoute/pull/6248) — thanks @backryun)
+- **chore(release):** ~50 commits on this branch are v3.8.45 pre-flight/hardening fixes and CI-perf work that landed here via the parallel-cycle sync-back (`sync-next-cycle.mjs`, Hard Rule #21) **after** the `v3.8.45` git tag was cut, and are already fully documented under the **[3.8.45]** section below — listed here only so the per-cycle commit-coverage check (`npm run release:uncovered`) doesn't flag them as gaps. Provider/catalog/UX/backend fixes: #6041, #6078, #6108, #6135, #6148, #6149, #6154, #6158, #6161, #6162, #6163, #6164, #6165, #6170, #6177, #6178, #6181, #6186, #6187, #6191, #6193, #6194, #6195, #6200, #6205, #6208, #6209, #6211, #6213, #6223, #6224, #6225, #6226, #6227, #6228, #6229, #6230, #6235, #6291, #6292. CI/release-pipeline work: #6167, #6203, #6214, #6215, #6218, #6273, #6275, #6283, #6284, #6285, #6300, #6305.
+- **chore(release):** additional zero-ref release-cycle plumbing on this branch, kept out of `release:uncovered` on purpose (no `#N` in the commit subject to cite): opening the v3.8.46 cycle, opening/closing the v3.8.45 cycle, the finalized [3.8.45] CHANGELOG i18n sync-back to 42 mirrors, the v3.8.45 cognitive/cyclomatic and file-size drift rebaselines, ESLint stale-suppression pruning (4,273 → 4,233), and clearing test-masking/docs-all pre-flight reds for v3.8.45.
+
+### ⚡ Performance & Infrastructure
+
+- **perf(release-green):** the pre-flight validator (`scripts/quality/validate-release-green.mjs`) now runs its 4 slow suites (unit / vitest / integration / pack-artifact) **concurrently** via `Promise.all` — pre-flight wall time drops from ~the sum of the suites to ~the slowest one (~30min saved per round; Phase 0 was the nº1 bottleneck of the v3.8.45 release benchmark, 2h54 of 6h34 e2e). Guard: `tests/unit/validate-release-green.test.ts` ("runs the slow suites CONCURRENTLY"). ([#6319](https://github.com/diegosouzapw/OmniRoute/pull/6319))
+- **fix(ci):** `scripts/release/sync-next-cycle.mjs` — two defects found live in its first production run (v3.8.45 Phase 5): (1) the `git()` helper's default 1 MiB `maxBuffer` crashed with `ENOBUFS` on `git show origin/main:CHANGELOG.md` (the CHANGELOG alone is >1 MiB) — widened to 64 MiB; (2) the i18n resync only propagated the `[NEXT]` (TBD) section, leaving the just-shipped finalized section as "— TBD" in all 42 mirrors — it now also syncs `[prevVersion]` bounded by the heading below it (new exported pure helper `versionAfter`). Guards: +5 tests in `tests/unit/sync-next-cycle.test.ts` (8/8). ([#6327](https://github.com/diegosouzapw/OmniRoute/pull/6327))
+- **test(ci):** concurrency-sensitive flaky tests are **quarantined into a serial pass** (`tests/unit/serial/`, `--test-concurrency=1`, appended to every unit runner incl. sharded variants — the serial pass is sharded too so concurrent shard jobs never self-collide). Initial set: `glm-coding-plan-monthly-3580`, `quota-division-blocks`, `provider-health-autopilot`, `combo-health-autopilot` — the class behind the ~28min CI wedges/re-runs (two live 1h+ wedges cancelled during this PR's own validation). Discovery + TIA gates track the new glob; systemic root cause (async logger writing after teardown) tracked in [#6360](https://github.com/diegosouzapw/OmniRoute/issues/6360). Guard: `tests/unit/test-serial-quarantine.test.ts` (4). ([#6347](https://github.com/diegosouzapw/OmniRoute/pull/6347))
+
+### 🙌 Contributors
+
+Thanks to everyone whose work landed in v3.8.46:
+
+| Contributor                                                    | PRs / Issues           |
+| -------------------------------------------------------------- | ---------------------- |
+| [@2220258345](https://github.com/2220258345)                   | direct commit / report |
+| [@abdofallah](https://github.com/abdofallah)                   | direct commit / report |
+| [@adentdk](https://github.com/adentdk)                         | direct commit / report |
+| [@aleksesipenko](https://github.com/aleksesipenko)             | direct commit / report |
+| [@anki1kr](https://github.com/anki1kr)                         | direct commit / report |
+| [@anungma](https://github.com/anungma)                         | direct commit / report |
+| [@arpicato](https://github.com/arpicato)                       | direct commit / report |
+| [@backryun](https://github.com/backryun)                       | #6248                  |
+| [@binsarjr](https://github.com/binsarjr)                       | direct commit / report |
+| [@brightfiscalband](https://github.com/brightfiscalband)       | direct commit / report |
+| [@chirag127](https://github.com/chirag127)                     | #6501, #6506           |
+| [@developerjillur](https://github.com/developerjillur)         | direct commit / report |
+| [@dilneiss](https://github.com/dilneiss)                       | #6499                  |
+| [@dtybnrj](https://github.com/dtybnrj)                         | direct commit / report |
+| [@Forcerecon](https://github.com/Forcerecon)                   | direct commit / report |
+| [@hao3039032](https://github.com/hao3039032)                   | direct commit / report |
+| [@Iammilansoni](https://github.com/Iammilansoni)               | #6150, #6245           |
+| [@jmengit](https://github.com/jmengit)                         | direct commit / report |
+| [@jordansilly77-stack](https://github.com/jordansilly77-stack) | direct commit / report |
+| [@JxnLexn](https://github.com/JxnLexn)                         | direct commit / report |
+| [@KooshaPari](https://github.com/KooshaPari)                   | #6166                  |
+| [@loopyd](https://github.com/loopyd)                           | direct commit / report |
+| [@luweiCN](https://github.com/luweiCN)                         | #6221                  |
+| [@makcimbx](https://github.com/makcimbx)                       | direct commit / report |
+| [@muflifadla38](https://github.com/muflifadla38)               | direct commit / report |
+| [@newnol](https://github.com/newnol)                           | direct commit / report |
+| [@ofekbetzalel](https://github.com/ofekbetzalel)               | direct commit / report |
+| [@ohahe52-dot](https://github.com/ohahe52-dot)                 | direct commit / report |
+| [@phidinhmanh](https://github.com/phidinhmanh)                 | direct commit / report |
+| [@powellnorma](https://github.com/powellnorma)                 | direct commit / report |
+| [@qpeyba](https://github.com/qpeyba)                           | direct commit / report |
+| [@RaviTharuma](https://github.com/RaviTharuma)                 | direct commit / report |
+| [@RCrushMe](https://github.com/RCrushMe)                       | direct commit / report |
+| [@rianonehub](https://github.com/rianonehub)                   | #6134, #6204           |
+| [@serverless83](https://github.com/serverless83)               | #6212                  |
+| [@swingtempo](https://github.com/swingtempo)                   | direct commit / report |
+| [@tenshiak](https://github.com/tenshiak)                       | direct commit / report |
+| [@ThongAccount](https://github.com/ThongAccount)               | direct commit / report |
+| [@UnrealAryan](https://github.com/UnrealAryan)                 | direct commit / report |
+| [@vinayakkulkarni](https://github.com/vinayakkulkarni)         | direct commit / report |
+| [@vittoroliveira-dev](https://github.com/vittoroliveira-dev)   | #6233                  |
+| [@warelik](https://github.com/warelik)                         | direct commit / report |
+| [@xxy9468615](https://github.com/xxy9468615)                   | direct commit / report |
+| [@xz-dev](https://github.com/xz-dev)                           | direct commit / report |
+| [@yanpaing007](https://github.com/yanpaing007)                 | direct commit / report |
+| [@diegosouzapw](https://github.com/diegosouzapw)               | maintainer             |
+
+---
+
 ## [3.8.45] — 2026-07-06
 
 ### ✨ New Features

docs/i18n/de/CHANGELOG.md

@@ -6,6 +6,213 @@
 
 ## [3.8.31] — 2026-06-20
 
+## [3.8.46] — TBD
+
+### ✨ New Features
+
+- **feat(sse):** **hide paid-only models from `auto/*` routing** when `hidePaidModels` is on ([#6512](https://github.com/diegosouzapw/OmniRoute/issues/6512)) — follow-up to #6328/#6495. PR #6495 hid paid-only models from the `GET /v1/models` listing, but `auto/*` combos (`auto/best-coding`, `auto/glm`, …) could still pick a paid-only backend into their candidate pool → a 402/403 at request time. `createVirtualAutoCombo` now filters the candidate pool through the new pure `open-sse/services/autoCombo/paidModelFilter.ts` (`filterPaidOnlyCandidates`), applying the same free-model predicate #6495 uses in `catalog.ts` (`providerHasFreeModels(provider) && isFreeModel(provider, {id})`) whenever `settings.hidePaidModels === true`. Applied before the category/tier/family narrowing, so it covers every `auto/*` combo; an all-paid pool degrades to the existing graceful empty-pool path. **Opt-in — default OFF leaves the pool unchanged** (identity). Regression guard: `tests/unit/autoCombo/paid-model-filter-6512.test.ts` (4, incl. the default-off identity guard).
+- **feat(sse):** **provider-family auto combos** — `auto/glm`, `auto/minimax`, `auto/mimo`, `auto/zai`, `auto/gemma`, `auto/llama`, `auto/gemini` ([#6453](https://github.com/diegosouzapw/OmniRoute/issues/6453)) — new routable ids that materialize an on-demand virtual combo spanning whatever installed backends currently expose that model family, degrading gracefully as backends rotate. A new pure `open-sse/services/autoCombo/modelFamily.ts` (`detectModelFamily`) classifies by model-id prefix for six families; `zai` is instead resolved by provider id (z.ai's hosted API serves the same `glm-*` model ids as every other GLM backend, so `auto/zai` means "route to my z.ai backend specifically" vs `auto/glm`'s "any connected GLM backend"). Reuses the existing `createVirtualAutoCombo` on-demand materialization path (no DB writes) and the `/v1/models` catalog advertising loop. Regression guard: `tests/unit/autoCombo/provider-family-combos.test.ts` (11).
+- **feat(proxy):** native **proxy-pool round-robin / egress IP rotation** ([#6365](https://github.com/diegosouzapw/OmniRoute/issues/6365)) — a scope (global / provider / account) can now hold **multiple** proxies as a pool with a rotation strategy, so outbound requests cycle their egress IP instead of pinning one proxy per scope. Migration `117_proxy_pool_rotation.sql` lifts the `UNIQUE(scope, scope_id)` constraint (rebuild via the canonical rename/copy/drop; existing single assignments become 1-element pools) and adds a `proxy_scope_rotation` companion table holding the per-scope strategy + a persisted monotonic round-robin cursor. Strategies: `round-robin` (default, monotonic cursor — never `Math.random`), `random`, and `sticky-per-N-min`. Resolution (`resolveProxyForScopeFromRegistry` / `resolveProxyForConnectionFromRegistry`) now fetches the alive, position-ordered candidate set (unchanged `PROXY_ALIVE_PREDICATE`) and applies the strategy; an empty / all-dead pool still returns `null` — the #6246 fail-closed guard is **untouched** (never falls through to direct egress). Backend + DB only; dashboard pool-builder UI is a follow-up. Regression guard: `tests/unit/proxy-pool-rotation-6365.test.ts` (8, incl. fail-closed + backward-compat).
+- **feat(providers):** end-to-end **tool/function calling on the native Gemini `/v1beta` endpoint** ([#6222](https://github.com/diegosouzapw/OmniRoute/issues/6222)) — both directions of the Gemini↔OpenAI conversion now preserve tool data (previously silently dropped). Request side: `convertGeminiToInternal` (extracted to its own testable module) maps `tools[].functionDeclarations` → OpenAI `tools`, prior `functionCall` parts → assistant `tool_calls`, and `functionResponse` parts → `tool`-role messages. Response side: `convertOpenAIResponseToGemini` emits `parts[].functionCall {name,args}` from `message.tool_calls`, and the streaming `openAIChunkToGeminiChunk` accumulates fragmented `tool_calls` deltas by index into complete `functionCall` parts. The non-Gemini client paths (Claude, OpenAI-Responses) already preserved tool calls — this closes the gap specific to the native Gemini surface. Regression guard: `tests/unit/v1beta-gemini-tool-calling-6222.test.ts` (6, incl. a streaming SSE round-trip).
+- **feat(providers):** copilot-m365-web **enterprise / work tier** support ([#6334](https://github.com/diegosouzapw/OmniRoute/issues/6334)) — mirrors the EDU-tier pattern (#6210): `M365ConnectionParams` gains an `agent` field, a new opt-in `M365_ENTERPRISE_OVERRIDES` preset (`agent=work`, `scenario=officeweb`, `licenseType=Premium`) applies via `providerSpecificData.tier="enterprise"` (alias `"work"`), and `agent` is also overridable directly via `providerSpecificData.agent`. `buildWsUrl` was hardcoding `agent="web"` (the one enterprise-distinguishing param with no override path), so a Premium work account handshook then returned an empty stream. The individual and EDU paths are untouched. Kilo's dup flag vs #6210 (EDU tier) was a false positive — different tier. Regression guard: `tests/unit/copilot-m365-enterprise-6334.test.ts` (7). End-to-end confirmation on a real Premium work account is a live-VPS validation follow-up (Hard Rule #18). (thanks @Forcerecon)
+- **feat(api):** standardized, provider-agnostic **`effort` + `thinking` request params** ([#6241](https://github.com/diegosouzapw/OmniRoute/issues/6241)) — a thin standardization layer over the existing mature per-provider reasoning plumbing (no provider mapper touched). `providerChatCompletionSchema` gains a canonical `effort` (reusing the shared `none/low/medium/high/xhigh` vocabulary — the UI tiers `extra`/`max` collapse onto `xhigh`) and a boolean `thinking`. A pure `normalizeReasoningRequest` (wired once in `src/sse/handlers/chat.ts`, before any reasoning field is read) folds them onto the fields the translators already consume (`reasoning_effort` / `reasoning.effort` / `thinking`), so they fan out to Anthropic / Gemini / xAI / Responses — an explicit client `reasoning_effort` / object-shaped `thinking` always wins (backward-compatible). `/models` additively exposes `supportsThinking` + `effort_tiers` so the frontend can render the toggles (UI component is a follow-up). Regression guard: `tests/unit/effort-thinking-standardization-6241.test.ts` (12). (thanks @Iammilansoni, @shabeer)
+- **feat(combo):** new **`pipeline` (sequential) combo strategy** ([#6297](https://github.com/diegosouzapw/OmniRoute/issues/6297)) — the 18th routing strategy runs targets **in order**, threading each step's output into the next step's input, with an optional per-step `prompt` (system instruction); only the final step's response is returned. Distinct from `fusion` (parallel fan-out + judge). Implemented as a self-contained `open-sse/services/pipeline.ts` (sibling to `fusion.ts`), dispatched from `combo.ts`; the step list reuses `combo.models` order and reads an optional `prompt` off each target (backward-compatible — ignored by every other strategy). Intermediate steps run non-streaming with tools stripped (complete prose to thread forward); the final step keeps the client's `stream` flag + tools. A failing/empty/unparseable intermediate step fails the whole pipeline explicitly via a sanitized error (never silently swallowed). Kilo's dup flag vs #563 was a false positive (that's model→chain selection; this is a sequential chain). Regression guard: `tests/unit/combo-pipeline-strategy.test.ts` (5). (thanks @ofekbetzalel)
+- **feat(ci):** `check:test-masking` now flags **inline-reimplemented prod conditions** ([#6348](https://github.com/diegosouzapw/OmniRoute/issues/6348)) — a new **report-only** subcheck (v2, 6A.10 family) catches the wrong-shape contract test: a test that recomputes the condition under test inline instead of importing/exercising the real function (the #6216 class, where `=== 500` → `>= 500` stayed green because the test re-implemented the branch). For each added/modified test file it warns when the file textually duplicates a ≥3-token conditional from a production file touched in the same PR **and** does not import the symbol/module owning it, via a pure, fixture-tested `findReimplementedConditions()` with an allowlist mirroring `assertReductionAllowlist`. Report-only for now (does not fail the gate) — to be promoted to blocking after a triage cycle. Regression guard: `tests/unit/check-test-masking.test.ts` (45).
+- **feat(sse):** per-connection routing override (native vs CLIProxyAPI) ([#6339](https://github.com/diegosouzapw/OmniRoute/issues/6339)) — the previously-dead `isCliproxyapiDeepModeEnabled` helper is now wired into `resolveExecutorWithProxy`: a single connection can opt itself into the CLIProxyAPI passthrough executor via `providerSpecificData.cliproxyapiMode="claude-native"`, with precedence **connection override > provider `upstream_proxy_config` mode > default**. `resolveExecutorWithProxy` now receives the resolved connection's `providerSpecificData` (threaded from `chatCore.ts`), so one connection can deep-route while the provider's other connections stay native — no DB schema change (the toggle rides in `providerSpecificData`). Also resolves the same-provider mixing ask in #6340. Regression guard: `tests/unit/chatcore-executor-proxy.test.ts` (9). (thanks @RaviTharuma)
+- **feat(dashboard):** "Add session cookie" modal now shows a prominent **"Open ‹host› →"** link to the provider's own site ([#6268](https://github.com/diegosouzapw/OmniRoute/issues/6268)) — every `-web` cookie-session provider (chatgpt-web, claude-web, gemini-web, kimi-web, lmarena, qwen-web, m365-copilot-web, …) renders a one-click external link (opening the provider's login/home page in a new tab) so operators no longer tab away to retype the URL mid-setup. The host resolves from a pure, unit-tested `resolveWebProviderHost()` (prefers `WEB_COOKIE_PROVIDERS[id].website`, falls back to the registry `baseUrl` origin); non-web providers render exactly as before. Kilo's dup flag vs #6265 (modal-too-small-on-1080p) was a false positive — distinct concern. Regression guard: `tests/unit/resolve-web-provider-host.test.ts` (5). (thanks @chirag127)
+- **feat(providers):** add **DigitalOcean** AI (serverless inference) as an OpenAI-compatible API-key provider ([#6373](https://github.com/diegosouzapw/OmniRoute/pull/6373)) — base `https://inference.do-ai.run/v1`, wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/digitalocean/`, `src/shared/constants/providers/apikey/inference-hosts.ts`). Regression guard: `tests/unit/digitalocean-provider.test.ts`. (thanks @newnol)
+- **feat(providers):** add **Huancheng Public API** (`hcnsec`) as an OpenAI-compatible regional provider ([#6410](https://github.com/diegosouzapw/OmniRoute/pull/6410)) — Xinjiang Huancheng Cybersecurity's public LLM platform (base `https://api.hcnsec.cn/v1`, free credits via daily check-ins), wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/hcnsec/`, `src/shared/constants/providers/apikey/regional.ts`). Regression guard: `tests/unit/hcnsec-provider.test.ts`. (thanks @UnrealAryan)
+- **feat(dashboard):** the web-session credential guide now shows an **"Open {host}" link** ([#6316](https://github.com/diegosouzapw/OmniRoute/pull/6316)) to the provider's sign-in site (derived from the provider `website` via `getProviderWebsiteHost`), so you can jump straight to the page where the cookie/session must be captured. Regression guard: `tests/unit/web-session-provider-link-6316.test.ts`. (thanks @jordansilly77-stack)
+- **feat(cerebras):** add the **Gemma 4 31B** model (`gemma-4-31b`) to the Cerebras registry + pricing table ([#6331](https://github.com/diegosouzapw/OmniRoute/pull/6331)). Regression guard extends `tests/unit/t28-model-catalog-updates.test.ts`. (thanks @backryun)
+- **feat(providers):** add **Yuanbao (web)** as a cookie-session provider ([#6196](https://github.com/diegosouzapw/OmniRoute/issues/6196)) — `yuanbao-web` (Tencent Yuanbao, `yuanbao.tencent.com`) with cookie-only auth (`hy_user`/`hy_token` + public agent id), SSE→OpenAI translation incl. `reasoning_content`, exposing DeepSeek V3/R1 + Hunyuan / Hunyuan-T1. Regression guard: `tests/unit/providers-yuanbao-web.test.ts`. `together-web` was **deferred** (no verifiable web-session endpoint — needs a captured request) and `huggingchat-web` **dropped** (the existing `huggingchat` already is a web-cookie provider). (thanks @chirag127)
+- **feat(providers):** route the built-in **agentrouter** through the dynamic Claude-Code wire image ([#6056](https://github.com/diegosouzapw/OmniRoute/issues/6056)) — a small static allow-set (`CC_WIRE_IMAGE_BUILTINS` in `open-sse/services/ccWireImageBuiltins.ts`), consulted by `isClaudeCodeCompatible` / `isClaudeCodeCompatibleProvider` / `applyFingerprint`, makes agentrouter adopt the CC wire-image headers + fingerprint **while guarding the CC baseUrl/auth branches** so it keeps its own registry `baseUrl` and `x-api-key` auth. Regression guard: `tests/unit/agentrouter-cc-wire-image.test.ts` (asserts the wire image is applied AND agentrouter's baseUrl/auth are preserved). Live WAF-acceptance against agentrouter.org is a VPS validation follow-up (Hard Rule #18).
+- **feat(providers):** **bulk-add API keys for Cloudflare Workers AI** ([#6174](https://github.com/diegosouzapw/OmniRoute/issues/6174)) — `cloudflare-ai` is removed from the bulk-add exclusion list and the bulk parser gains a 3-field `name|accountId|apiKey` mode; the bulk route now builds a **per-entry** `providerSpecificData` so each key carries its own `accountId` (fixing the previous shared-object reuse), and both the create + key-validation paths receive it. Regression guard: `tests/unit/bulk-api-key-parser-cloudflare.test.ts`. (thanks @muflifadla38)
+- **feat(dashboard):** routing/settings UX clarity ([#6147](https://github.com/diegosouzapw/OmniRoute/issues/6147)) — (1) weighted combos show the **effective routing share %** next to each weight when weights don't sum to 100 (`WeightTotalBar.tsx`); (2) the status widget's user-facing **"Cloud Sync" label is renamed** to "Remote Settings Sync" (`CloudSyncStatus.tsx`; internal ids/state untouched); (3) built-in providers gain an **opt-in advanced base-URL override** (`isBaseUrlOverrideEligibleProvider`, hidden behind an "Advanced" toggle, reusing the existing `providerSpecificData.baseUrl` persistence — not globally widened). Regression guard: `tests/unit/routing-settings-ux-6147.test.ts`.
+- **feat(combo):** add an option to **disable session stickiness**, per-combo or globally — round-robin / random combos can rotate to a different connection on every request instead of pinning a whole conversation to one connection by its first-message hash. Resolution precedence per-combo `config.disableSessionStickiness` → global `settings.disableSessionStickiness` → default `false` (preserves the #3825 prompt-cache/504 fix); gates **both** stickiness call sites in `open-sse/services/combo.ts`. Exposed as a global toggle (Combo Defaults) and a per-combo Inherit/on/off control. ([#6168](https://github.com/diegosouzapw/OmniRoute/issues/6168)) Regression guard: `tests/unit/combo-disable-session-stickiness.test.ts`. (thanks @RCrushMe)
+- **feat(docker):** add the `OMNIROUTE_NO_SUDO` env flag for root-less / user-namespaced deployments — the MITM cert-trust command path (`resolveSudoSpawn` in `src/mitm/systemCommands.ts`) now strips the leading `sudo` when the flag is truthy, in addition to the existing root / sudo-missing cases, so the Proxy Agent runs without `sudo` (the operator trusts the CA manually, e.g. via `NODE_EXTRA_CA_CERTS`). Argv-array `spawn` preserved — no shell interpolation (Hard Rule #13). ([#6122](https://github.com/diegosouzapw/OmniRoute/issues/6122)) Regression guard: `tests/unit/mitm-systemCommands-no-sudo.test.ts`. (thanks @powellnorma)
+- **feat(providers):** add **Requesty** as an OpenAI-compatible gateway provider (BYOK, base `https://router.requesty.ai/v1`, ~200 free requests/day) — wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/requesty/`, `src/shared/constants/providers/apikey/gateways.ts`). ([#6120](https://github.com/diegosouzapw/OmniRoute/issues/6120)) Regression guard: `tests/unit/requesty-provider.test.ts`. (thanks @chirag127)
+- **feat(dashboard):** add **configured-only / available-only filters** to the Free Provider Rankings page ([#6150](https://github.com/diegosouzapw/OmniRoute/issues/6150)) — hide providers you haven't configured, or whose connections are all rate-limited / out of quota, via server-side query params (`?configuredOnly` / `?availableOnly` on `GET /api/free-provider-rankings`) backed by a testable lib helper reusing the in-process connection state (no Redis). Both filters default off, so the default view is unchanged; this supersedes the earlier client-side "Configured Only" toggle (#6245) with an available-only dimension and unit-tested logic. Regression guard: `tests/unit/freeProviderRankings-filters.test.ts`.
+
+### 🔧 Bug Fixes
+
+- **fix(dashboard):** adding a second API key connection for the same provider no longer silently overwrites the first — the Add-API-key modal now derives a unique default connection name (`main`, then `main-2`, `main-3`, …) so the backend name-based upsert can't collide (#6499 — thanks @dilneiss).
+- **fix(compression):** the session-dedup engine now also deduplicates a large multi-line block repeated **within a single message** (intra-message dedup), not just across turns; the compression-preview API surfaces a `fallbackReason`, and the fusion panel reports how many models were rate-limited vs failed on total-panel failure (#6501 — thanks @chirag127).
+- **fix(compression):** a stacked-pipeline step naming an unregistered engine now surfaces a `validationErrors` entry instead of silently no-op'ing, so misconfigured pipelines are visible in the preview API (#6506 — thanks @chirag127).
+- **feat(usage):** add a **Codex reset-credit redemption** flow to the Provider Limits UI ([#6361](https://github.com/diegosouzapw/OmniRoute/pull/6361)) — a `useCodexResetCreditRedemption` hook + `/api/usage/codex-reset-credit` route + `codexResetCredits` lib let you redeem banked Codex reset credits from the quota card. Regression guard: `tests/unit/codex-reset-credits.test.ts`. (thanks @JxnLexn)
+- **feat(glm):** add **team-plan quota settings** for `glm-cn` connections ([#6351](https://github.com/diegosouzapw/OmniRoute/pull/6351)) — a dedicated `GlmTeamQuotaFields` form section (team quota id / limits) threaded through the Add/Edit connection modals, persisted via `providerSpecificData`, with the GLM usage service reading the team quota. Regression guards: `tests/unit/glm-team-quota.test.ts`, `provider-specific-data-schema.test.ts`. (thanks @hao3039032)
+- **feat(providers):** add **TinyFish** web-fetch/search support ([#6349](https://github.com/diegosouzapw/OmniRoute/pull/6349)) — a `tinyfish-fetch` executor + `/v1/web/fetch` route + MCP web-fetch tool, registered as a specialty-media provider with request-validation and a search-provider catalog entry. Regression guards: `tests/unit/executor-tinyfish-fetch.test.ts`, `web-fetch-handler.test.ts`, `mcp-web-fetch-tool.test.ts`, `provider-validation-tinyfish.test.ts`. (thanks @dtybnrj)
+- **fix(cli):** `omniroute launch-codex` now spawns `codex.cmd` through a shell on Windows (the npm `.cmd` shim is unresolvable by bare `spawn` → ENOENT), mirroring the qodercli Windows fix (#6263) ([#6312](https://github.com/diegosouzapw/OmniRoute/pull/6312)). Regression guard: `tests/unit/launch-codex-windows-spawn-6312.test.ts`. (thanks @swingtempo)
+- **fix(codex):** isolate the **Spark** quota from the shared Codex quota and stabilize the quota UI ordering / hydration so per-scope limits render consistently ([#6336](https://github.com/diegosouzapw/OmniRoute/pull/6336)). Regression guards: `tests/unit/codex-quota-selection-hydration.test.ts`, `provider-limits-ui.test.ts` + 3 more. (thanks @xz-dev)
+- **feat(api):** add a `hidePaidModels` setting that filters paid-only models out of the `/v1/models` catalog. Regression guard: `tests/unit/models-catalog-hide-paid.test.ts`. (thanks @chirag127)
+- **fix(api-manager):** the fallback model picker now preserves combos instead of dropping them when a primary model is unavailable ([#6443](https://github.com/diegosouzapw/OmniRoute/pull/6443)). Regression guard: `tests/unit/api-manager-page-static.test.ts`. (thanks @jmengit)
+- **fix(providers):** recoverable Antigravity / Cloud-Code (Gemini Code Assist) `403` responses ([#6452](https://github.com/diegosouzapw/OmniRoute/pull/6452)) are now classified as a retryable project-config error instead of a terminal account ban, so a fixable project/API-disabled 403 no longer forces a ~1-year cooldown / full OAuth reconnect. Regression guard: `tests/unit/errorclassifier-antigravity-403.test.ts`. (thanks @developerjillur)
+- **fix(mitm):** `sanitizeHeaders` now redacts `Set-Cookie` response headers so upstream session cookies never leak into logs / diagnostics ([#6451](https://github.com/diegosouzapw/OmniRoute/pull/6451)). Regression guard: `tests/unit/mitm-sanitize-headers.test.ts`. (thanks @developerjillur)
+- **fix(api):** `/api/compression/preview` now accepts `mode: "caveman"` and correctly handles stacked / zero-compression previews ([#6425](https://github.com/diegosouzapw/OmniRoute/issues/6425)). Regression guard: `tests/unit/api/compression-preview-caveman-and-stacked-6425.test.ts`. (thanks @chirag127)
+- **feat(providers):** add **Zed** hosted LLM aggregator as a native-app provider ([#6118](https://github.com/diegosouzapw/OmniRoute/pull/6118)) — OAuth sign-in via the Zed hosted flow, registered through the shared provider registry + executor. Regression guards: `tests/unit/zed-oauth-provider.test.ts`, `zed-import-utils.test.ts`, `zed-docker-detect.test.ts`, `mitm-handler-zed.test.ts`. VPS-validated via live operator login (Hard Rule #18).
+- **fix(oauth):** the Kiro SSO-cache auto-import now **preserves the IDC region** — cross-region Amazon Q / Kiro profiles imported from the SSO cache are no longer collapsed to the default region ([#6113](https://github.com/diegosouzapw/OmniRoute/pull/6113)). Regression guard: `tests/unit/kiro-auto-import-idc-2059.test.ts`. VPS-validated via live operator login (Hard Rule #18).
+- **fix(dashboard):** passthrough model aliases no longer collide when two namespaced model ids share a last segment (port from 9router#1850, [#6431](https://github.com/diegosouzapw/OmniRoute/pull/6431)). `enx/gpt-5.5` and `enx/codebuddy/gpt-5.5` both auto-generated the alias `gpt-5.5`, so the second model could never be added (the UI just alerted "alias already exists"). Aliases are now disambiguated deterministically — bare last segment when free, then parent-qualified (`codebuddy-gpt-5.5`), then a numeric suffix — while re-adding the exact same model id is still blocked. Regression guard: `tests/unit/passthrough-alias-1850.test.ts`. (thanks @arpicato)
+- **fix(translator):** preserve a Gemini `functionResponse` co-located with other parts (another `functionCall`, or trailing `text`) in the same content when translating **Gemini → OpenAI** ([#6376](https://github.com/diegosouzapw/OmniRoute/pull/6376)). `convertGeminiContent()` early-returned the tool message on the first `functionResponse` part, dropping any co-located parts; such contents are now pre-split (one tool message per `functionResponse`, emitted first, plus one message for the remaining parts). Regression guard: `tests/unit/gemini-to-openai-function-response.test.ts`. (thanks @warelik)
+- **fix(headroom):** detect a python interpreter managed by **mise / pyenv / asdf / conda** (port from 9router#2353, [#6382](https://github.com/diegosouzapw/OmniRoute/pull/6382)). Headroom's python probe (`src/lib/headroom/detect.ts`) searched a hardcoded `PATH`, but version managers expose their interpreters via shim dirs that only join `PATH` through interactive-shell activation — which the non-interactive server never runs, so a managed python (≥3.10) was invisible and Headroom reported it missing. The search path now prepends the well-known shim/bin dirs (`~/.local/share/mise/shims`, `~/.pyenv/shims`, `~/.asdf/shims`, `$CONDA_PREFIX/bin`, `~/.local/bin`, respecting `MISE_DATA_DIR`/`PYENV_ROOT`/`ASDF_DATA_DIR` when set), and a new `HEADROOM_PYTHON` env override lets operators point straight at their interpreter (mirroring `HEADROOM_URL`). Still shell-free (`execFileSync`). Regression guard: `tests/unit/headroom-detect.test.ts` (5). (thanks @loopyd)
+- **fix(executors):** strip the OpenAI-Codex/Claude-CLI `client_metadata` passthrough field for **NVIDIA** requests (port from 9router#1887, [#6411](https://github.com/diegosouzapw/OmniRoute/pull/6411)). NVIDIA's OpenAI-compatible wrapper rejects it with `400 Unsupported parameter`, the same class already handled for `cerebras`/`mistral`; `nvidia` (executor `default`) was missing from the strip allowlist so Codex/Claude-Code passthrough requests 400'd. Regression guard: `tests/unit/executor-default-strip-client-metadata.test.ts` (+nvidia case). (thanks @phidinhmanh)
+- **fix(translator):** strip the Claude-style `thinking` field for **NVIDIA `z-ai/glm-5.2`** (port from 9router#2023, [#6413](https://github.com/diegosouzapw/OmniRoute/pull/6413)). NVIDIA's OpenAI-compatible wrapper 400s on `thinking` (a Claude-format client routed here leaves a `thinking:{type:"adaptive"}`); the existing strip rule only dropped `reasoning`. Same class already handled for `minimax-m2.7`. Regression guard: `tests/unit/nvidia-minimax-thinking-strip.test.ts` (+glm-5.2 case). (thanks @phidinhmanh)
+- **fix(translator):** suppress the streamed `</think>` close marker for the **Antigravity IDE** client (port from 9router#1061, [#6415](https://github.com/diegosouzapw/OmniRoute/pull/6415)). On thinking-only turns Antigravity rendered a bare `</think>` as the sole visible content, tripping its loop-detection and wasting requests. Antigravity's UA (`vscode/<v> (Antigravity/<v>)`) is added to the marker-suppress allowlist (alongside OpenCode); Claude Code / Cursor still get the marker, and `x-omniroute-thinking-marker: on` force-restores it. Regression guard: `tests/unit/think-close-marker-suppress-5245.test.ts`. (thanks @abdofallah)
+- **fix(executors):** strip nested `reasoning_content` from messages for **Mistral** (port from 9router#1649, [#6417](https://github.com/diegosouzapw/OmniRoute/pull/6417)). Mistral's API returns `422 extra_forbidden` when an assistant message carries `reasoning_content` (replayed thinking from a prior turn, e.g. via the Codex `/responses` path); the generic top-level 400 field-downgrade retry never covered the nested per-message field. `DefaultExecutor` now strips it for provider `mistral` only, so DeepSeek (which requires replayed `reasoning_content`) is unaffected. Regression guard: `tests/unit/mistral-strip-reasoning-content-1649.test.ts`. (thanks @xxy9468615)
+- **fix(executors):** strip the `client_metadata` passthrough field on the **OpenCode** path (port from 9router#1442, [#6418](https://github.com/diegosouzapw/OmniRoute/pull/6418)). OpenCode upstreams (e.g. `kimi-k2.6` via opencode-go) reject it with `400 "Extra inputs are not permitted, field: 'client_metadata'"`; the DefaultExecutor strip only covered cerebras/mistral and `OpencodeExecutor` extends `BaseExecutor` directly, so nothing removed it there. Regression guard: `tests/unit/opencode-strip-client-metadata-1442.test.ts`. (thanks @yanpaing007)
+- **fix(executors):** inject the `reasoning_content` echo for the native **Moonshot Kimi** provider (port from 9router#1480, [#6419](https://github.com/diegosouzapw/OmniRoute/pull/6419)). Kimi (executor `default`) is a thinking-mode upstream that 400s with "reasoning_content must be passed back" when a prior assistant turn lacks it; the placeholder injection was only wired into the OpenCode meta-provider, so direct multi-turn Kimi conversations failed. Scoped to `kimi` (gateway-served models matching the thinking-model name pattern are unaffected). Regression guard: `tests/unit/kimi-native-reasoning-injected-1480.test.ts`. (thanks @2220258345)
+- **fix(executors):** recover from a strict gateway's `context_management: Extra inputs are not permitted` 400 (port from 9router#1468, [#6420](https://github.com/diegosouzapw/OmniRoute/pull/6420)). **Claude Code** always sends a top-level `context_management` field; strict anthropic-compatible gateways reject it. The dedicated context-editing 400-fallback only fired when OmniRoute's own `contextEditing` feature was enabled (default off), so a client-sent field passed through untouched and 400'd. `context_management` is now in the generic reactive field-strip list, so it's stripped-and-retried once regardless of the feature flag (with correct request re-signing for claude-compatible relays). Regression guard: `tests/unit/provider-field-strips.test.ts`. (thanks @ohahe52-dot)
+- **fix(network):** enable **RFC 8305 Happy Eyeballs** (`autoSelectFamily`) on the direct-egress undici dispatcher (port from 9router#1237, [#6423](https://github.com/diegosouzapw/OmniRoute/pull/6423)). When DNS returns both IPv6 (AAAA) and IPv4 (A) and the IPv6 route is broken (e.g. a NAT64 `64:ff9b::` prefix without routing), undici tried IPv6 first and hung until `ETIMEDOUT` (then a 502 + account lockout), even though `curl` reached the same host. The direct dispatcher now races both families and uses whichever connects first. Proxy paths pin family via `proxyTls` and are unaffected. Regression guard: `tests/unit/direct-dispatcher-pipelining-4580.test.ts`. (thanks @adentdk)
+- **fix(combo):** round-robin now advances the rotation pointer past the model that **actually served**, not the eagerly-scheduled one (port from 9router#948, [#6428](https://github.com/diegosouzapw/OmniRoute/pull/6428)). With `stickyLimit: 1` (true round-robin), when the scheduled model failed and a _different_ model served via fallback, the counter had already advanced +1 from the scheduled index — so the next request reused the fallback-served model, degrading round-robin into hot-spotting on whichever model was healthy. The pointer now advances to the served index + 1 (mirroring the sticky-limit>1 path). Session-stickiness (#3825) and distribution are preserved. Regression guard: `tests/unit/combo-rr-fallback-advance-948.test.ts`. (thanks @binsarjr)
+- **fix(sse):** a non-string `model` field is now rejected with a `400` before the resolver, instead of crashing downstream `.toLowerCase()`/`.split()` calls into an empty-body `500` that escapes the error sanitizer ([#6407](https://github.com/diegosouzapw/OmniRoute/issues/6407)). Regression guard: `tests/unit/chat-non-string-model-6407.test.ts`. (thanks @chirag127)
+- **fix(api):** unknown `/api/*` routes now return a JSON `404` (instead of the dashboard HTML shell) and scalar chat params (`model`/`temperature`/etc.) are validated **before** the provider lookup so malformed requests fail fast with a clear `400` ([#6424](https://github.com/diegosouzapw/OmniRoute/issues/6424), [#6412](https://github.com/diegosouzapw/OmniRoute/issues/6412)). Regression guards: `tests/unit/api/api-catchall-json-404.test.ts`, `tests/unit/chat-early-schema-validation-6412.test.ts`. (thanks @chirag127)
+- **fix(api):** `/v1/chat/completions` now rejects a non-JSON `Content-Type` with a `400` before parsing the body ([#6414](https://github.com/diegosouzapw/OmniRoute/issues/6414)). Regression guard: `tests/unit/v1-chat-completions-content-type-6414.test.ts`. (thanks @chirag127)
+- **fix(api):** the `X-OmniRoute-Compression` response header is now echoed on `/v1/chat/completions` and `/v1/completions` ([#6422](https://github.com/diegosouzapw/OmniRoute/issues/6422)). Regression guard: `tests/unit/compression-header-echo-6422.test.ts`. (thanks @chirag127)
+- **fix(api):** concurrent `GET /v1/models` requests are coalesced into a single catalog build ([#6408](https://github.com/diegosouzapw/OmniRoute/issues/6408)). Regression guard: `tests/unit/v1-models-concurrent-6408.test.ts`. (thanks @chirag127)
+- **fix(api):** `/v1/completions` now echoes the requested `body.model` in its JSON + streamed responses ([#6429](https://github.com/diegosouzapw/OmniRoute/pull/6429)). Regression guard: `tests/unit/completions-body-model-echo.test.ts`. (thanks @chirag127)
+- **fix(api):** env-var master keys now see the full `/v1/models` catalog ([#6406](https://github.com/diegosouzapw/OmniRoute/issues/6406)). Regression guard: `tests/unit/models-catalog-envkey-6406.test.ts`. (thanks @chirag127)
+- **fix(api):** non-streaming `/v1/completions` responses now echo `body.model` aligned with the `X-OmniRoute-Model` header ([#6426](https://github.com/diegosouzapw/OmniRoute/issues/6426)). Regression guard: `tests/unit/v1-completions-model-header-match-6426.test.ts`. (thanks @chirag127)
+- **fix(api):** unknown `/v1/*` routes now return a JSON `404 not_found` instead of the Next.js dashboard HTML shell ([#6405](https://github.com/diegosouzapw/OmniRoute/issues/6405)). Regression guard: `tests/unit/api/v1-catchall-json-404.test.ts`. (thanks @chirag127)
+- **fix(api):** the per-connection provider models route now degrades to the shipped catalog when a provider's `/models` endpoint answers with a redirect ([#6267](https://github.com/diegosouzapw/OmniRoute/issues/6267)) — a `qwen-web` import failed with a raw `Redirect blocked … (307)` 503. `safeOutboundFetch` throws `REDIRECT_BLOCKED` on the 307, `getSafeOutboundFetchErrorStatus` maps it to 503, and `buildDiscoveryErrorFallbackResponse` treated every 503 as a hard error — so the non-empty `getModelsByProviderId("qwen-web")` catalog was never surfaced. A models-endpoint redirect is not a fixable-config error (unlike `URL_GUARD_BLOCKED`/`INVALID_URL`, which stay hard errors), so it now falls back to the local/cached catalog before the 503 short-circuit. General fix — covers any config-driven provider that 307s. Regression guard: `tests/unit/provider-models-qwen-web-redirect-6267.test.ts`. (thanks @chirag127)
+- **fix(api):** the per-connection provider models route (MCP `list_models_catalog` + the dashboard import view) now merges USER-ADDED custom models into its response ([#6247](https://github.com/diegosouzapw/OmniRoute/issues/6247)) — custom models live in the `key_value` namespace `customModels`, which the live REST `/api/v1/models` already merges, but `src/app/api/providers/[id]/models/route.ts` never read `getCustomModels`, so custom models were dropped on both the discovery-success and local_catalog paths. They are now folded into the returned model list (deduped by id, stamped `owned_by: provider`), fixing MCP + the dashboard import view in one place. Regression guard: `tests/unit/provider-models-custom-merge-6247.test.ts`. (thanks @RCrushMe)
+
+- **fix(providers):** GitLab Duo tool-calling follow-up turns no longer fail upstream with `422 {"detail":"Validation error"}` (tokens 0/0, rejected pre-inference). The #6234 tool-result-feedback fix serialized the **entire** multi-turn conversation into GitLab's single-file `code_suggestions` (`small_file`) generation endpoint — folded history that turn-N sent as an oversized `current_file.content_above_cursor` **and** duplicated verbatim into `user_instruction`, tripping the AI-Gateway's `small_file` validation guard. The executor now **bounds** that prompt: it keeps system + latest user message + the most-recent tool round (dropping older turns), caps oversized tool results, and stops duplicating the full prompt into `user_instruction` (which now carries only the short latest user message) — while still feeding the most-recent tool result back so the agent continues (`open-sse/executors/gitlab.ts`, [#6220](https://github.com/diegosouzapw/OmniRoute/issues/6220)). The unit test covers the bounding logic; the upstream 422→200 clearing is VPS-only (Hard Rule #18). Regression guard: `tests/unit/gitlab-tool-exchange-bounded-6220.test.ts`.
+
+- **fix(i18n):** the provider-detail (`/dashboard/providers/[id]`) connection-status filter labels no longer render as `__MISSING__:All` / `__MISSING__:Active` / `__MISSING__:Error` / `__MISSING__:Banned` / `__MISSING__:CreditsExhausted` in non-English locales (notably pt-BR) ([#6290](https://github.com/diegosouzapw/OmniRoute/issues/6290)). Root cause was **not** the namespace mismatch the issue guessed — the `providers.filter*` keys resolve correctly in `en.json`; the debt lived in the locale mirrors (`src/i18n/messages/*.json`), where these five keys carried the `__MISSING__:` sync sentinel in ~15 locales and were absent entirely in ~26 others, so next-intl found the key and echoed the sentinel verbatim. All 40 non-English/-Chinese mirrors now ship real translations for the five `providers.filter*` labels. Regression guard: `tests/unit/i18n-provider-filter-keys-6290.test.ts`. (thanks @diegosouzapw)
+
+- **fix(providers):** the `copilot-m365-web` streaming executor now emits `debug`-level WebSocket diagnostics ([#6210](https://github.com/diegosouzapw/OmniRoute/issues/6210)) — the outbound WS URL (with the `access_token` **redacted** via `redactWsUrl()`), handshake success/failure, and each received SignalR frame's `type`/`target`. Previously the streaming path logged nothing, so an empty `content:null` response (the M365 Education / Starter tier symptom fixed in #6234) was undiagnosable even at `APP_LOG_LEVEL=debug`. The change is debug-level and side-effect-free — it does not alter streaming behavior or the frame parser, and the token never reaches the logs. Regression guard: `tests/unit/copilot-m365-web-logging-6210.test.ts` (thanks @qpeyba)
+
+- **fix(resilience):** a round-robin combo no longer returns `503 all upstream accounts are unavailable` when a compatibility-rejected target is actually healthy ([#6238](https://github.com/diegosouzapw/OmniRoute/issues/6238)). `filterTargetsByRequestCompatibility` drops request-incompatible targets (tool/vision/structured-output unsupported, or below the required context window) **before** any availability check runs, and its `compatible.length === 0` safety net only fired when _all_ targets were filtered — not when the kept targets later all turned out runtime-unavailable (circuit-open / cooldown / no credentials). So a combo could 503 while a compat-rejected-but-healthy provider sat unused. `handleRoundRobinCombo` now keeps the compat-rejected set and, when every compat-kept target was skipped without a single real attempt, probes those rejected targets as a **last-resort fallback tier** (via the new pure `open-sse/services/combo/comboCompatFallback.ts`) before crystallizing the 503. Regression guard: `tests/unit/combo-roundrobin-compat-fallback-6238.test.ts`. (thanks @ThongAccount)
+
+- **fix(startup):** best-effort self-heal for a corrupted Turbopack dev cache on Windows ([#6289](https://github.com/diegosouzapw/OmniRoute/issues/6289)). On Windows, `pnpm dev` can fail at startup when Turbopack `mmap`s a persistent-cache SST file and the OS refuses the mapping (`os error 1455` — "paging file too small"), which Turbopack surfaces as a misleading `Module not found: Can't resolve '@/shared/utils/machine'`. This is a **known upstream Turbopack cache-corruption bug — not our code**. The dev launcher (`scripts/dev/run-next.mjs`) now wraps `nextApp.prepare()` and, when it rejects with that signature (`isTurbopackCacheCorruption` in the new `scripts/dev/turbopackCacheHeal.mjs`), purges `.build/next/**/cache/turbopack` and retries **once** with a clear log. **Caveat — best-effort only:** the corruption often surfaces as a runtime overlay rather than a `prepare()` rejection, so this cannot always intercept it; the reliable remedy remains manually deleting the Turbopack cache dir. Regression guard: `tests/unit/turbopack-cache-heal-6289.test.ts`. (thanks @chirag127)
+
+- **fix(providers):** qodercli PAT auth no longer fails with `spawn qodercli ENOENT` on Windows ([#6263](https://github.com/diegosouzapw/OmniRoute/issues/6263)) — `spawnQoderCli` spawned the bare `qodercli` name with `shell:false` and an unenriched env, so the npm `.cmd` wrapper under `%APPDATA%\npm` (a user-PATH directory) was never resolved. It now resolves the absolute `.cmd`/`.exe` path through the existing `getCliRuntimeStatus("qoder")` resolver in `src/shared/services/cliRuntime.ts` (memoized), spawns with `shell` when the target is a `.cmd`/`.bat`, and uses the cliRuntime-enriched env (PATH + PATHEXT + APPDATA); the ENOENT error now lists the searched paths plus the `CLI_QODER_BIN` override. End-to-end spawn on a real Windows host is host-only (Hard Rule #18); the path-resolution logic is unit-tested. Regression guard: `tests/unit/qodercli-windows-resolve-6263.test.ts`. (thanks @chirag127)
+
+- **fix(sse):** the reasoning-token buffer no longer inflates **probe-sized `max_tokens`** ([#6274](https://github.com/diegosouzapw/OmniRoute/issues/6274)) — Claude Code's `/model` capability check sends `max_tokens: 1`, but for a thinking-capable model with a large output cap (e.g. `glm-5.2`) the #3587 headroom heuristic (`max(current + 1000, ceil(current * 1.5))`) rewrote it to `1001` and forwarded that upstream, wasting tokens on a request that was never a genuine reasoning budget. `resolveReasoningBufferedMaxTokens()` (`open-sse/services/reasoningTokenBuffer.ts`) now short-circuits and returns the caller's value verbatim when it is below the new `REASONING_BUFFER_MIN_TRIGGER` (256) threshold — a tiny explicit limit is a probe, not a reasoning request. Real budgets still receive the #3587 headroom unchanged, and the guard runs after the existing capability checks so unknown / non-reasoning models keep returning `null`. Regression guard: `tests/unit/reasoning-token-buffer-6274.test.ts`. (thanks @brightfiscalband)
+
+- **fix(cli):** `omniroute reset-password` now works as a real subcommand, and password resets over piped (non-TTY) stdin actually apply ([#6261](https://github.com/diegosouzapw/OmniRoute/issues/6261), [#6258](https://github.com/diegosouzapw/OmniRoute/issues/6258)). Two coupled defects: (1) **#6261** — `bin/omniroute.mjs` routed everything through Commander with only two pre-Commander bypasses (`--mcp`, `reset-encrypted-columns`), so `omniroute reset-password` was rejected as an unknown command; only the separate `omniroute-reset-password` bin worked, while the docs falsely advertised the subcommand (incl. a bogus "legacy alias still works"). A pre-Commander bypass mirroring `reset-encrypted-columns` now dynamically imports `bin/reset-password.mjs` (which self-executes) before Commander parses; the three doc lines were corrected. (2) **#6258** — `bin/reset-password.mjs` issued two sequential `rl.question` prompts; under piped stdin the second read never settled at EOF, so `main()` never reached `resetManagementPassword` and the reset was a silent no-op (both prompts printed, no success, password unchanged). The CLI now detects non-TTY stdin and reads it once (first line = password, second line = confirm if present, else reused), adds a `--password-stdin` flag (entire stdin is the password, no confirmation), and exits `0` explicitly so the success line always flushes; interactive TTY behavior is unchanged. Regression guard: `tests/unit/reset-password-cli-6261-6258.test.ts` (3). (thanks @chirag127)
+
+- **fix(db):** the mass-migration **safety abort** now tells the operator how to bypass it and stops flooding the log ([#6260](https://github.com/diegosouzapw/OmniRoute/issues/6260)) — after restoring a backup that wiped the migration tracking table, `runMigrations()` threw the abort on every downstream `ensureDbInitialized()`, re-logging the full banner 11+ times, and the message never mentioned the existing `OMNIROUTE_MAX_PENDING_MIGRATIONS` escape hatch. The abort text now appends a bypass hint (set `OMNIROUTE_MAX_PENDING_MIGRATIONS=0` in `server.env` / `DATA_DIR/.env`), and a new `MigrationSafetyAbortError` is memoized so repeated calls in the same process throw the same instance and emit a single concise line instead of the full cascade. Regression guard: `tests/unit/migration-safety-abort-6260.test.ts`. (thanks @chirag127)
+
+- **fix(auth):** importing a **distinct** Codex/ChatGPT OAuth `auth.json` is no longer falsely rejected as "already exists" when it belongs to a different user in the same workspace ([#6301](https://github.com/diegosouzapw/OmniRoute/issues/6301)). `findExistingCodexConnection` (in `src/lib/oauth/utils/codexAuthImport.ts`) deduped **only** on `providerSpecificData.workspaceId === accountId`, where `accountId` is the shared `chatgpt_account_id`/`tokens.account_id` — so two members of the same ChatGPT Team collapsed onto a single connection (409 `duplicate_account`). The id_token's `https://api.openai.com/auth` claim carries a per-user `chatgpt_user_id` alongside the workspace id (the device-flow path already persisted it as `chatgptUserId`, but the import path did not). Now `parseAndValidateCodexAuth` extracts `userId` (`chatgpt_user_id` → `user_id` → JWT `sub`) into `ParsedCodexAuth`, the create/update paths persist `chatgptUserId` in `providerSpecificData` (mirroring `codex.ts`), and dedup keys on `workspaceId` **AND** `chatgptUserId` — with a backward-compat fallback to legacy accountId-only matching when no stored connection for that workspace records a `chatgptUserId`, so genuinely-same accounts still dedup. Regression guard: `tests/unit/codex-auth-import-userid-dedup-6301.test.ts` (4). (thanks @anungma)
+
+- **fix(providers):** importing models for the **venice-web** provider no longer fails with a red "Provider venice-web does not support models listing" ([#6269](https://github.com/diegosouzapw/OmniRoute/issues/6269)). `venice-web` is a web-cookie provider with an executor but no upstream `/v1/models` endpoint and no registry `models`, so the models route fell through to the tail `400`. Mirroring the `jules`/`linkup-search`/`ollama-search` fix (#5569), it now ships a static local catalog entry in `src/lib/providers/staticModels.ts` — seeding the current Venice lineup (`venice-uncensored`, `llama-3.3-70b`, `qwen3-235b`, `qwen3-4b`, `deepseek-r1-671b`; Venice rotates its catalog, see docs.venice.ai/models/overview) — so the route returns `200` with `source:"local_catalog"`, `intentional:true`. Regression guard: `tests/unit/static-models-venice-web-6269.test.ts`. (thanks @chirag127)
+
+- **fix(api):** the specialty model catalogs (`/v1/embeddings`, `/v1/images`, `/v1/music`, `/v1/videos` model lists) are now derived from the **unified catalog filtered by a predicate** (`getSpecialtyModelsResponse`) instead of ad-hoc per-route logic, so they consistently respect active-credential visibility and stay in sync with the main catalog ([#6303](https://github.com/diegosouzapw/OmniRoute/pull/6303)). Regression guard: `tests/unit/specialty-model-catalog-routes.test.ts`. (thanks @makcimbx)
+- **fix(api):** the agent-bridge server route now resolves the MITM manager via a **dynamic `import("@/mitm/manager.runtime")`** so Turbopack does not statically pull the stub (or over-bundle the manager), and the agent-skills generator anchors its output base path with `path.join(process.cwd(), …)` so Turbopack's static analyzer stops tracing the whole project root ([#6329](https://github.com/diegosouzapw/OmniRoute/issues/6329), [#6366](https://github.com/diegosouzapw/OmniRoute/pull/6366)). Regression guard: `tests/unit/agent-bridge-server-route-dynamic-import.test.ts`. (thanks @Iammilansoni)
+- **fix(api):** internal probes (combo-test, cloud-sync verify) now pick a **management-scoped / allow-all API key** instead of naively grabbing `getApiKeys()[0]` — a restricted `self:usage` first row made the probe fail with "Model X is not allowed for this API key" even when the combo path was healthy (`pickApiKeyForInternalUse` in `src/lib/db/apiKeys.ts`). The API-manager model editor also falls back to `/api/models?all=true` when `/v1/models` is catalog-protected ([#6372](https://github.com/diegosouzapw/OmniRoute/pull/6372)). Regression guard: `tests/unit/pick-internal-api-key-6372.test.ts`. (thanks @jmengit)
+- **fix(live-ws):** the Live Dashboard WebSocket server now **rejects on bind failure** (e.g. `EADDRINUSE` when the API bridge already holds the port) instead of letting the error surface as an unhandled `error` event that crash-loops the process — the `error` listener is attached to `wss` (not `server`) and releases the EventBus subscription on a failed start ([#6324](https://github.com/diegosouzapw/OmniRoute/issues/6324)). Regression guard: `tests/unit/live-ws-eaddrinuse-6324.test.ts`. (thanks @vinayakkulkarni)
+- **fix(dashboard):** the Home provider-topology widget now trusts the live provider-metrics snapshot — it uses `topology.errorProvider` and live `activeRequests` directly instead of re-deriving state from a stale `lastErrorAt` or applying a frontend timeout filter, so the topology reflects real-time provider health ([#6322](https://github.com/diegosouzapw/OmniRoute/pull/6322)). Regression guard: `tests/unit/home-provider-topology-live-state.test.ts`. (thanks @xz-dev)
+- **fix(sse):** strip zero-width markers from streamed **tool-call arguments** — a follow-up to [#5857](https://github.com/diegosouzapw/OmniRoute/pull/5857). That PR removed injected zero-width joiners (U+200D) from streamed assistant text/reasoning but deliberately left tool-call argument JSON byte-exact. The request-side obfuscation (`open-sse/services/claudeCodeObfuscation.ts`) injects ZWJ into agent words — including the temp path inside the Bash tool description — and Claude models copy that verbatim into generated commands, which are delivered as tool-call arguments rather than assistant text. As a result the ZWJ survived and corrupted code blocks (e.g. a temp path rendered with an invisible joiner). Now `open-sse/handlers/responseSanitizer.ts` strips zero-width code points from tool-call argument strings at every emit site (OpenAI non-stream/stream chat `tool_calls` + legacy `function_call`, native Responses `function_call` items, the OpenAI→Responses conversion, and the native Responses streaming `response.function_call_arguments.delta/.done` events). Only zero-width code points are removed; JSON structure and all other bytes stay identical (no parse/restringify), so normal arguments remain byte-exact. Regression guard: 6 new cases in `tests/unit/response-sanitizer.test.ts` (suite 50/50).
+
+- **fix(nodejs):** the default app log path now resolves under `DATA_DIR` (`~/.omniroute/logs/application/app.log`) instead of `process.cwd()` ([#6197](https://github.com/diegosouzapw/OmniRoute/issues/6197)) — the globally-installed CLI runs from an arbitrary working directory, so anchoring the default to cwd made file logging silently write to (or no-op under) an unrelated directory, contradicting the documented `.env.example` default. `getAppLogFilePath()` now computes the default lazily via the pure `resolveDataDir()` resolver (honours a per-process `DATA_DIR`, no directory-creation side effect); an explicit `APP_LOG_FILE_PATH` still wins. Regression guard: `tests/unit/logenv-datadir-path-6197.test.ts` (3).
+
+- **fix(docker):** AgentBridge/`startMitm` no longer aborts in containers/headless when the Antigravity-default DNS step can't write `/etc/hosts` ([#6127](https://github.com/diegosouzapw/OmniRoute/issues/6127)), and the privileged command's stderr now reaches `app.log` instead of only a bare exit code hitting the toast ([#6198](https://github.com/diegosouzapw/OmniRoute/issues/6198)). The default DNS step (`addDNSEntry`) was called unguarded while cert install and the two sibling DNS steps were each best-effort — in the runtime Docker image (`USER node`, no `sudo`, read-only `/etc/hosts`) it threw `Command failed with code 1` out of `startMitmInternal` and killed the whole start, discarding the stderr. The three DNS steps are extracted into a best-effort `provisionDnsEntries()` where each failure is logged with the full `err` (stderr included, folded in by `systemCommands.ts`) and never aborts the start. Regression guard: `tests/unit/mitm-dns-graceful-degrade-6127.test.ts` (4).
+
+- **fix(providers):** copilot-m365-web now supports the M365 Education "Starter / OfficeWebIncludedCopilot" tier and no longer returns an empty `content:null` stream ([#6210](https://github.com/diegosouzapw/OmniRoute/issues/6210)). Two gaps: (1) `buildWsUrl()` hardcoded the individual-consumer scenario (`OfficeWebPaidConsumerCopilot`, `isEdu=false`) — the EDU tier is now opt-in via `providerSpecificData.tier="edu"`, emitting `scenario=OfficeWebIncludedCopilot`/`isEdu=true` (the individual path is unchanged); (2) the EDU/GPT-5.5 path streams deltas via `arguments[0].writeAtCursor` (incremental) instead of only `messages[].text` (accumulated snapshots), which the parser dropped — a new `accumulateBotContent()` folds both formats, with `type:2 item.result.message` as a last-resort fallback. Regression guard: `tests/unit/copilot-m365-edu-writeatcursor-6210.test.ts` (10). (thanks @qpeyba)
+
+- **fix(providers):** GitLab Duo executor now feeds tool results back into the prompt instead of looping ([#6220](https://github.com/diegosouzapw/OmniRoute/issues/6220)) — `buildPrompt()` branched only on `system`/`user` and took `userParts.at(-1)`, silently dropping the `assistant{tool_calls}` + `tool{result}` turns the client appended, so the reconstructed prompt was byte-identical to turn 1 and the model re-emitted the same `<tool>` call forever. When a tool exchange is present the full conversation is now serialized, folding each tool result back keyed by its `tool_call_id`; simple conversations keep the legacy shape. Complements the tool_call emission from [#6051](https://github.com/diegosouzapw/OmniRoute/issues/6051) (the `kilo-duplicate` label was a false positive — different, sequential defect). Regression guard: `tests/unit/gitlab-tool-result-feedback-6220.test.ts` (4).
+
+- **fix(providers):** opencode-go/opencode-zen can now synthesize the OpenCode CLI identity headers Cloudflare requires on VPS egress ([#5997](https://github.com/diegosouzapw/OmniRoute/issues/5997)) — on a datacenter VPS, `opencode.ai/zen/go/v1/chat/completions` 403s (HTML challenge) requests lacking CLI identity, while the reporter's control curl proved that `User-Agent: opencode-cli/1.0.0` + `x-opencode-client: cli` + `x-opencode-project: default` + fresh request/session UUIDs succeed. Opt-in via `OPENCODE_SYNTHESIZE_CLI_HEADERS=true` (values overridable via `OPENCODE_GO_USER_AGENT`/`OPENCODE_USER_AGENT`/`OPENCODE_CLIENT`/`OPENCODE_PROJECT`); it fills only headers the client did not already send. Kept **off by default** — the forward-only path is deliberate (fabricating a wrong value risks upstream rejection; a prior dedup regressed with `opencode/local`), so this replaces the fragile local header-injection shim without changing default behavior. Regression guard: `tests/unit/opencode-cli-headers-synthesis-5997.test.ts` (6). (thanks @aleksesipenko)
+- fix(resilience): sticky session affinity now evicts and fails over to another account when the pinned account is exhausted/unavailable (#6219)
+
+- fix(sse): Responses API passthrough now drops internal commentary-phase output before forwarding to clients (gated by RESPONSES_PASSTHROUGH_DROP_COMMENTARY, default on) (#6199)
+
+- **fix(sse):** tool-call function schemas with a root `type: null` are now coerced to `type: "object"` before dispatch ([#6359](https://github.com/diegosouzapw/OmniRoute/issues/6359)) — clients like the Codex app emit `parameters: { type: null, ... }` for some tools, which OpenAI-compatible upstreams reject with `400 Invalid schema for function '...': schema must be a JSON Schema of 'type: "object"', got 'type: null'`, failing the whole request. `toolSchemaSanitizer` already stripped the null; it now re-adds the mandatory root `"object"` type (and empty `properties`/open `additionalProperties` when absent). Combinator roots (`anyOf`/`oneOf`/`allOf`) and explicit root types are left untouched. Regression guard: 5 new cases in `tests/unit/tool-schema-sanitizer.test.mjs`.
+
+- **fix(docker):** AgentBridge no longer fails to start on npm/Electron/VPS installs with "MITM manager stub reached at runtime" ([#6344](https://github.com/diegosouzapw/OmniRoute/issues/6344)) — v3.8.45 flipped the production bundler default to Turbopack, but `next.config.mjs` aliased `@/mitm/manager` to its Docker-only degraded stub **unconditionally**. That was harmless while Docker (which sets the alias intentionally for #3390 graceful degradation) was the sole Turbopack consumer, but once every artifact built with Turbopack the stub shipped to all non-Docker users and `startMitm` threw on the first Agent-Bridge start. The alias is now opt-in via `OMNIROUTE_MITM_STUB=1` (set only by the Dockerfile) through the shared `scripts/build/mitm-stub-flag.mjs` helper; default builds bundle the real manager. Regression guard: `tests/unit/mitm-stub-alias-6344.test.mjs` (4).
+
+- **fix(proxy):** stop the v3.8.44 proxy regression that leaked the real IP and disabled healthy proxies ([#6246](https://github.com/diegosouzapw/OmniRoute/issues/6246)). Two coupled defects from the new health scheduler: (1) **IP leak** — when a proxy assigned to a connection was marked `inactive`, resolution fell through to a **direct** egress instead of blocking, exposing the operator's real IP; (2) **over-deactivation** — the sweep flipped a proxy to `inactive` on the **first** failed probe and counted our own 5s timeout / a probe-target `5xx` as the proxy's fault, so healthy paid proxies vanished from egress selection ("my proxies are not being used anymore"). Fix: the sweep decision is extracted into a pure, network-free `decideProxyHealthAction` (`src/lib/proxyHealth/decision.ts`) — by default the health check now **only counts/logs and never downgrades status** (a proxy is downgraded/removed only with `PROXY_AUTO_REMOVE=true`, after `PROXY_AUTO_REMOVE_AFTER` **consecutive** conclusive failures); probes are classified tri-state so an inconclusive result (our timeout, or a `5xx` from the probe target) never penalizes the proxy, and the probe timeout is raised 5s→15s. Separately, `safeResolveProxy` now **fails closed** via the existing policy: a connection whose assigned proxy is dead is blocked instead of leaking direct (`hasBlockingProxyAssignment`), honoring the explicit `proxy off` toggles and the `PROXY_FAIL_OPEN=true` opt-out. Existing proxies stuck `inactive` by the old behavior need a one-time manual re-activate (the operator owns proxy status). Regression guards: `tests/unit/proxy-health-decide-action-6246.test.ts`, `tests/unit/proxy-assigned-unavailable-6246.test.ts`.
+
+- **fix(proxy):** make "Test All" read-only and add bulk enable/disable ([#6246](https://github.com/diegosouzapw/OmniRoute/issues/6246)). Complements the core fail-closed / scheduler fix (#6296) with the two remaining reporter asks. (1) The **"Test All" button** (`POST /api/settings/proxies/auto-test`) used to flip a proxy to `inactive` on a failed reachability probe; since the egress selector excludes `inactive` proxies, a flaky probe (an unreachable `httpbin.org`, a proxy that blocks `HEAD`, or a slow paid proxy) silently disabled every proxy that failed — "Test All" is now **read-only by default** (only the operator sets a proxy active/inactive; opt back into the legacy test-and-set with `PROXY_HEALTH_AUTO_DEACTIVATE=true`). (2) Adds a **bulk enable/disable** proxies endpoint + toolbar action (`POST /api/settings/proxies/batch-activate`) so an operator can re-activate proxies in one click. Regression guard: `tests/unit/proxy-health-6246.test.ts`. (thanks @tenshiak)
+
+- **chatcore (tools): stop the default 128-tool cap from silently dropping opencode's `task`/MCP tools.** opencode (used as an MCP/agent host) sends a large tool list; when it exceeds the speculative `MAX_TOOLS_LIMIT` (128) default, `truncateToolList` did a blind `tools.slice(0, 128)`, dropping every tool past index 128 — including opencode's built-in `task` tool (subagent launch) and many MCP tools, so models routed through OmniRoute could no longer spawn subagents or reach part of their tools. The cap exists to avoid upstream `400`s for providers with real hard limits (e.g. grok-cli 200), so it is kept for those: detection of the opencode client (`isOpencodeClient` — any `x-opencode-*` header, or `opencode` in the user-agent) now only bypasses the **speculative 128 default**, never a known provider ceiling. Precedence is explicit — a proactive/detected provider limit always truncates (even for opencode); otherwise opencode forwards its full tool list; otherwise the unchanged 128 default applies to every other client. Refactors `getEffectiveToolLimit` into `getKnownToolLimit(provider) ?? DEFAULT_LIMIT` (byte-identical for existing callers) and fixes a cosmetic debug-log that reported the truncated count instead of the original. Regression guard: `tests/unit/tool-limit-detector.test.ts`.
+
+- **fix(mitm):** the macOS MITM-cert install check now matches the system keychain again. `security find-certificate -a -Z` prints the SHA-1 as a colon-less hex string, but the installed-check compared it against `getCertFingerprint()`'s colon-separated form, so the substring match never hit — the cert was reported as not-installed and re-prompted for the sudo install on every run. Fingerprints are now normalized (colons stripped, upper-cased) on both sides via the extracted `macCertOutputHasFingerprint` helper. Regression guard: `tests/unit/mitm-cert-mac-fingerprint.test.ts`. ([#6204](https://github.com/diegosouzapw/OmniRoute/pull/6204), closes [#6134](https://github.com/diegosouzapw/OmniRoute/issues/6134) — thanks @rianonehub)
+
+- **fix(api):** `/v1/messages/count_tokens` now counts `tool_use`, `tool_result` and `thinking` content blocks (and array-form `system` prompts) in the local-estimation path, instead of only `text`. Real agentic conversations keep ~95% of their tokens inside tool results; the previous estimate returned near-zero for them, which silently broke Claude Code's auto-compaction (context grew past the window with no compaction until the upstream API rejected the request). The real provider-side count path is unchanged. Regression guard: `tests/unit/messages-count-tokens-route.test.ts`. ([#6221](https://github.com/diegosouzapw/OmniRoute/pull/6221) — thanks @luweiCN)
+
+- **fix(antigravity):** strip a trailing assistant prefill turn for Vertex Claude models to avoid upstream 400s ([#6114](https://github.com/diegosouzapw/OmniRoute/pull/6114)). Regression guard: `tests/unit/antigravity-claude-prefill-strip.test.ts`. (thanks @anki1kr)
+
+- **fix(security):** the mutable cloud-agent routes (`/api/cloud/credentials/update`, `/api/cloud/models/alias`) now require management auth instead of being treated as public. They were classified as public API routes, so a request without management credentials could update stored cloud-agent credentials and model aliases. They are removed from the public-route set, classified as management routes in the authz pipeline, and gated by `requireManagementAuth`; cloud **read**/auth routes stay public. Regression guards: `tests/unit/cloud-write-auth.test.ts`, `tests/unit/authz/classify.test.ts`, `tests/unit/public-api-routes.test.ts`. ([#6233](https://github.com/diegosouzapw/OmniRoute/pull/6233) — thanks @vittoroliveira-dev)
+
+- **refactor(dashboard):** extract the onboarding-wizard "Open provider details" link target into a pure, unit-tested `buildProviderDetailsHref(connection)` helper. The wizard already routes by `connection.id` (the node UUID) rather than the provider category slug (#6144/#6145); this hardens that behavior behind a tested helper that guards a missing id/connection. Regression guard: `tests/unit/provider-onboarding-href.test.ts`. ([#6166](https://github.com/diegosouzapw/OmniRoute/pull/6166) — thanks @KooshaPari)
+
+- **fix(security):** the doubao synthetic device-id generator now derives its digits via an unbiased crypto-random draw (rejection sampling over `crypto.randomBytes()`) instead of a `% 10` reduction, closing a CodeQL `js/biased-cryptographic-random` finding.
+- **fix(agentSkills):** the GitHub-skills generator now resolves `outputDir` to an absolute path before writing, fixing a regression introduced by #6366 (relative-to-cwd base path) that could write generated skill files to the wrong directory.
+- **fix(security):** `/api/keys/{id}/devices` now checks the HTTP method before auth/validation, returning a `405` for non-GET/DELETE verbs instead of a misleading `401`/`500` (closes a `dast-smoke` QUERY-method finding).
+- **fix(quality):** clear the last 2 heavy quality-gate reds on the release tip (cycle pre-flight).
+- **fix(mitm):** the test suite and CI can never mutate the OS trust store — `OMNIROUTE_SKIP_SYSTEM_TRUST=1` is set globally for tests/CI so `installCert`/`uninstallCert`/`installTproxyCa` skip the privileged OS dispatch ([#6310](https://github.com/diegosouzapw/OmniRoute/pull/6310); full detail is under the [3.8.45] section below — this branch received it via the parallel-cycle sync-back).
+- **fix(api):** `POST /api/github-skills` now Zod-validates its request body; documented the new quality-gate env vars and pinned the merge-integrity GitHub Actions to a commit SHA.
+- **fix(skills):** generate the missing `omni-github-skills` registry entry and align the agent-skills catalog-count tests (follow-up to #6186).
+- **fix(quality):** clear the cycle's 11 net-new ESLint errors and make `validate-release-green` suppressions-aware.
+
+### 📝 Maintenance
+
+- **i18n(it):** add 118 missing Italian (`it`) translations (net-additive — no existing keys dropped, valid JSON), improving Italian UI coverage. ([#6212](https://github.com/diegosouzapw/OmniRoute/pull/6212) — thanks @serverless83)
+- **chore(providers):** remove deprecated **MiMo V2** model entries from the catalogs (xiaomi-mimo, opencode-go, zenmux-free, audio TTS) — the upstream V2 line is superseded by MiMo V2.5; drops `mimo-v2-tts`, `mimo-v2-pro`, `mimo-v2-omni`, `mimo-v2-flash`, `mimo-v2-flash-free` and realigns the provider-catalog tests. ([#6248](https://github.com/diegosouzapw/OmniRoute/pull/6248) — thanks @backryun)
+- **chore(release):** ~50 commits on this branch are v3.8.45 pre-flight/hardening fixes and CI-perf work that landed here via the parallel-cycle sync-back (`sync-next-cycle.mjs`, Hard Rule #21) **after** the `v3.8.45` git tag was cut, and are already fully documented under the **[3.8.45]** section below — listed here only so the per-cycle commit-coverage check (`npm run release:uncovered`) doesn't flag them as gaps. Provider/catalog/UX/backend fixes: #6041, #6078, #6108, #6135, #6148, #6149, #6154, #6158, #6161, #6162, #6163, #6164, #6165, #6170, #6177, #6178, #6181, #6186, #6187, #6191, #6193, #6194, #6195, #6200, #6205, #6208, #6209, #6211, #6213, #6223, #6224, #6225, #6226, #6227, #6228, #6229, #6230, #6235, #6291, #6292. CI/release-pipeline work: #6167, #6203, #6214, #6215, #6218, #6273, #6275, #6283, #6284, #6285, #6300, #6305.
+- **chore(release):** additional zero-ref release-cycle plumbing on this branch, kept out of `release:uncovered` on purpose (no `#N` in the commit subject to cite): opening the v3.8.46 cycle, opening/closing the v3.8.45 cycle, the finalized [3.8.45] CHANGELOG i18n sync-back to 42 mirrors, the v3.8.45 cognitive/cyclomatic and file-size drift rebaselines, ESLint stale-suppression pruning (4,273 → 4,233), and clearing test-masking/docs-all pre-flight reds for v3.8.45.
+
+### ⚡ Performance & Infrastructure
+
+- **perf(release-green):** the pre-flight validator (`scripts/quality/validate-release-green.mjs`) now runs its 4 slow suites (unit / vitest / integration / pack-artifact) **concurrently** via `Promise.all` — pre-flight wall time drops from ~the sum of the suites to ~the slowest one (~30min saved per round; Phase 0 was the nº1 bottleneck of the v3.8.45 release benchmark, 2h54 of 6h34 e2e). Guard: `tests/unit/validate-release-green.test.ts` ("runs the slow suites CONCURRENTLY"). ([#6319](https://github.com/diegosouzapw/OmniRoute/pull/6319))
+- **fix(ci):** `scripts/release/sync-next-cycle.mjs` — two defects found live in its first production run (v3.8.45 Phase 5): (1) the `git()` helper's default 1 MiB `maxBuffer` crashed with `ENOBUFS` on `git show origin/main:CHANGELOG.md` (the CHANGELOG alone is >1 MiB) — widened to 64 MiB; (2) the i18n resync only propagated the `[NEXT]` (TBD) section, leaving the just-shipped finalized section as "— TBD" in all 42 mirrors — it now also syncs `[prevVersion]` bounded by the heading below it (new exported pure helper `versionAfter`). Guards: +5 tests in `tests/unit/sync-next-cycle.test.ts` (8/8). ([#6327](https://github.com/diegosouzapw/OmniRoute/pull/6327))
+- **test(ci):** concurrency-sensitive flaky tests are **quarantined into a serial pass** (`tests/unit/serial/`, `--test-concurrency=1`, appended to every unit runner incl. sharded variants — the serial pass is sharded too so concurrent shard jobs never self-collide). Initial set: `glm-coding-plan-monthly-3580`, `quota-division-blocks`, `provider-health-autopilot`, `combo-health-autopilot` — the class behind the ~28min CI wedges/re-runs (two live 1h+ wedges cancelled during this PR's own validation). Discovery + TIA gates track the new glob; systemic root cause (async logger writing after teardown) tracked in [#6360](https://github.com/diegosouzapw/OmniRoute/issues/6360). Guard: `tests/unit/test-serial-quarantine.test.ts` (4). ([#6347](https://github.com/diegosouzapw/OmniRoute/pull/6347))
+
+### 🙌 Contributors
+
+Thanks to everyone whose work landed in v3.8.46:
+
+| Contributor                                                    | PRs / Issues           |
+| -------------------------------------------------------------- | ---------------------- |
+| [@2220258345](https://github.com/2220258345)                   | direct commit / report |
+| [@abdofallah](https://github.com/abdofallah)                   | direct commit / report |
+| [@adentdk](https://github.com/adentdk)                         | direct commit / report |
+| [@aleksesipenko](https://github.com/aleksesipenko)             | direct commit / report |
+| [@anki1kr](https://github.com/anki1kr)                         | direct commit / report |
+| [@anungma](https://github.com/anungma)                         | direct commit / report |
+| [@arpicato](https://github.com/arpicato)                       | direct commit / report |
+| [@backryun](https://github.com/backryun)                       | #6248                  |
+| [@binsarjr](https://github.com/binsarjr)                       | direct commit / report |
+| [@brightfiscalband](https://github.com/brightfiscalband)       | direct commit / report |
+| [@chirag127](https://github.com/chirag127)                     | #6501, #6506           |
+| [@developerjillur](https://github.com/developerjillur)         | direct commit / report |
+| [@dilneiss](https://github.com/dilneiss)                       | #6499                  |
+| [@dtybnrj](https://github.com/dtybnrj)                         | direct commit / report |
+| [@Forcerecon](https://github.com/Forcerecon)                   | direct commit / report |
+| [@hao3039032](https://github.com/hao3039032)                   | direct commit / report |
+| [@Iammilansoni](https://github.com/Iammilansoni)               | #6150, #6245           |
+| [@jmengit](https://github.com/jmengit)                         | direct commit / report |
+| [@jordansilly77-stack](https://github.com/jordansilly77-stack) | direct commit / report |
+| [@JxnLexn](https://github.com/JxnLexn)                         | direct commit / report |
+| [@KooshaPari](https://github.com/KooshaPari)                   | #6166                  |
+| [@loopyd](https://github.com/loopyd)                           | direct commit / report |
+| [@luweiCN](https://github.com/luweiCN)                         | #6221                  |
+| [@makcimbx](https://github.com/makcimbx)                       | direct commit / report |
+| [@muflifadla38](https://github.com/muflifadla38)               | direct commit / report |
+| [@newnol](https://github.com/newnol)                           | direct commit / report |
+| [@ofekbetzalel](https://github.com/ofekbetzalel)               | direct commit / report |
+| [@ohahe52-dot](https://github.com/ohahe52-dot)                 | direct commit / report |
+| [@phidinhmanh](https://github.com/phidinhmanh)                 | direct commit / report |
+| [@powellnorma](https://github.com/powellnorma)                 | direct commit / report |
+| [@qpeyba](https://github.com/qpeyba)                           | direct commit / report |
+| [@RaviTharuma](https://github.com/RaviTharuma)                 | direct commit / report |
+| [@RCrushMe](https://github.com/RCrushMe)                       | direct commit / report |
+| [@rianonehub](https://github.com/rianonehub)                   | #6134, #6204           |
+| [@serverless83](https://github.com/serverless83)               | #6212                  |
+| [@swingtempo](https://github.com/swingtempo)                   | direct commit / report |
+| [@tenshiak](https://github.com/tenshiak)                       | direct commit / report |
+| [@ThongAccount](https://github.com/ThongAccount)               | direct commit / report |
+| [@UnrealAryan](https://github.com/UnrealAryan)                 | direct commit / report |
+| [@vinayakkulkarni](https://github.com/vinayakkulkarni)         | direct commit / report |
+| [@vittoroliveira-dev](https://github.com/vittoroliveira-dev)   | #6233                  |
+| [@warelik](https://github.com/warelik)                         | direct commit / report |
+| [@xxy9468615](https://github.com/xxy9468615)                   | direct commit / report |
+| [@xz-dev](https://github.com/xz-dev)                           | direct commit / report |
+| [@yanpaing007](https://github.com/yanpaing007)                 | direct commit / report |
+| [@diegosouzapw](https://github.com/diegosouzapw)               | maintainer             |
+
+---
+
 ## [3.8.45] — 2026-07-06
 
 ### ✨ New Features

docs/i18n/es/CHANGELOG.md

@@ -6,6 +6,213 @@
 
 ## [3.8.31] — 2026-06-20
 
+## [3.8.46] — TBD
+
+### ✨ New Features
+
+- **feat(sse):** **hide paid-only models from `auto/*` routing** when `hidePaidModels` is on ([#6512](https://github.com/diegosouzapw/OmniRoute/issues/6512)) — follow-up to #6328/#6495. PR #6495 hid paid-only models from the `GET /v1/models` listing, but `auto/*` combos (`auto/best-coding`, `auto/glm`, …) could still pick a paid-only backend into their candidate pool → a 402/403 at request time. `createVirtualAutoCombo` now filters the candidate pool through the new pure `open-sse/services/autoCombo/paidModelFilter.ts` (`filterPaidOnlyCandidates`), applying the same free-model predicate #6495 uses in `catalog.ts` (`providerHasFreeModels(provider) && isFreeModel(provider, {id})`) whenever `settings.hidePaidModels === true`. Applied before the category/tier/family narrowing, so it covers every `auto/*` combo; an all-paid pool degrades to the existing graceful empty-pool path. **Opt-in — default OFF leaves the pool unchanged** (identity). Regression guard: `tests/unit/autoCombo/paid-model-filter-6512.test.ts` (4, incl. the default-off identity guard).
+- **feat(sse):** **provider-family auto combos** — `auto/glm`, `auto/minimax`, `auto/mimo`, `auto/zai`, `auto/gemma`, `auto/llama`, `auto/gemini` ([#6453](https://github.com/diegosouzapw/OmniRoute/issues/6453)) — new routable ids that materialize an on-demand virtual combo spanning whatever installed backends currently expose that model family, degrading gracefully as backends rotate. A new pure `open-sse/services/autoCombo/modelFamily.ts` (`detectModelFamily`) classifies by model-id prefix for six families; `zai` is instead resolved by provider id (z.ai's hosted API serves the same `glm-*` model ids as every other GLM backend, so `auto/zai` means "route to my z.ai backend specifically" vs `auto/glm`'s "any connected GLM backend"). Reuses the existing `createVirtualAutoCombo` on-demand materialization path (no DB writes) and the `/v1/models` catalog advertising loop. Regression guard: `tests/unit/autoCombo/provider-family-combos.test.ts` (11).
+- **feat(proxy):** native **proxy-pool round-robin / egress IP rotation** ([#6365](https://github.com/diegosouzapw/OmniRoute/issues/6365)) — a scope (global / provider / account) can now hold **multiple** proxies as a pool with a rotation strategy, so outbound requests cycle their egress IP instead of pinning one proxy per scope. Migration `117_proxy_pool_rotation.sql` lifts the `UNIQUE(scope, scope_id)` constraint (rebuild via the canonical rename/copy/drop; existing single assignments become 1-element pools) and adds a `proxy_scope_rotation` companion table holding the per-scope strategy + a persisted monotonic round-robin cursor. Strategies: `round-robin` (default, monotonic cursor — never `Math.random`), `random`, and `sticky-per-N-min`. Resolution (`resolveProxyForScopeFromRegistry` / `resolveProxyForConnectionFromRegistry`) now fetches the alive, position-ordered candidate set (unchanged `PROXY_ALIVE_PREDICATE`) and applies the strategy; an empty / all-dead pool still returns `null` — the #6246 fail-closed guard is **untouched** (never falls through to direct egress). Backend + DB only; dashboard pool-builder UI is a follow-up. Regression guard: `tests/unit/proxy-pool-rotation-6365.test.ts` (8, incl. fail-closed + backward-compat).
+- **feat(providers):** end-to-end **tool/function calling on the native Gemini `/v1beta` endpoint** ([#6222](https://github.com/diegosouzapw/OmniRoute/issues/6222)) — both directions of the Gemini↔OpenAI conversion now preserve tool data (previously silently dropped). Request side: `convertGeminiToInternal` (extracted to its own testable module) maps `tools[].functionDeclarations` → OpenAI `tools`, prior `functionCall` parts → assistant `tool_calls`, and `functionResponse` parts → `tool`-role messages. Response side: `convertOpenAIResponseToGemini` emits `parts[].functionCall {name,args}` from `message.tool_calls`, and the streaming `openAIChunkToGeminiChunk` accumulates fragmented `tool_calls` deltas by index into complete `functionCall` parts. The non-Gemini client paths (Claude, OpenAI-Responses) already preserved tool calls — this closes the gap specific to the native Gemini surface. Regression guard: `tests/unit/v1beta-gemini-tool-calling-6222.test.ts` (6, incl. a streaming SSE round-trip).
+- **feat(providers):** copilot-m365-web **enterprise / work tier** support ([#6334](https://github.com/diegosouzapw/OmniRoute/issues/6334)) — mirrors the EDU-tier pattern (#6210): `M365ConnectionParams` gains an `agent` field, a new opt-in `M365_ENTERPRISE_OVERRIDES` preset (`agent=work`, `scenario=officeweb`, `licenseType=Premium`) applies via `providerSpecificData.tier="enterprise"` (alias `"work"`), and `agent` is also overridable directly via `providerSpecificData.agent`. `buildWsUrl` was hardcoding `agent="web"` (the one enterprise-distinguishing param with no override path), so a Premium work account handshook then returned an empty stream. The individual and EDU paths are untouched. Kilo's dup flag vs #6210 (EDU tier) was a false positive — different tier. Regression guard: `tests/unit/copilot-m365-enterprise-6334.test.ts` (7). End-to-end confirmation on a real Premium work account is a live-VPS validation follow-up (Hard Rule #18). (thanks @Forcerecon)
+- **feat(api):** standardized, provider-agnostic **`effort` + `thinking` request params** ([#6241](https://github.com/diegosouzapw/OmniRoute/issues/6241)) — a thin standardization layer over the existing mature per-provider reasoning plumbing (no provider mapper touched). `providerChatCompletionSchema` gains a canonical `effort` (reusing the shared `none/low/medium/high/xhigh` vocabulary — the UI tiers `extra`/`max` collapse onto `xhigh`) and a boolean `thinking`. A pure `normalizeReasoningRequest` (wired once in `src/sse/handlers/chat.ts`, before any reasoning field is read) folds them onto the fields the translators already consume (`reasoning_effort` / `reasoning.effort` / `thinking`), so they fan out to Anthropic / Gemini / xAI / Responses — an explicit client `reasoning_effort` / object-shaped `thinking` always wins (backward-compatible). `/models` additively exposes `supportsThinking` + `effort_tiers` so the frontend can render the toggles (UI component is a follow-up). Regression guard: `tests/unit/effort-thinking-standardization-6241.test.ts` (12). (thanks @Iammilansoni, @shabeer)
+- **feat(combo):** new **`pipeline` (sequential) combo strategy** ([#6297](https://github.com/diegosouzapw/OmniRoute/issues/6297)) — the 18th routing strategy runs targets **in order**, threading each step's output into the next step's input, with an optional per-step `prompt` (system instruction); only the final step's response is returned. Distinct from `fusion` (parallel fan-out + judge). Implemented as a self-contained `open-sse/services/pipeline.ts` (sibling to `fusion.ts`), dispatched from `combo.ts`; the step list reuses `combo.models` order and reads an optional `prompt` off each target (backward-compatible — ignored by every other strategy). Intermediate steps run non-streaming with tools stripped (complete prose to thread forward); the final step keeps the client's `stream` flag + tools. A failing/empty/unparseable intermediate step fails the whole pipeline explicitly via a sanitized error (never silently swallowed). Kilo's dup flag vs #563 was a false positive (that's model→chain selection; this is a sequential chain). Regression guard: `tests/unit/combo-pipeline-strategy.test.ts` (5). (thanks @ofekbetzalel)
+- **feat(ci):** `check:test-masking` now flags **inline-reimplemented prod conditions** ([#6348](https://github.com/diegosouzapw/OmniRoute/issues/6348)) — a new **report-only** subcheck (v2, 6A.10 family) catches the wrong-shape contract test: a test that recomputes the condition under test inline instead of importing/exercising the real function (the #6216 class, where `=== 500` → `>= 500` stayed green because the test re-implemented the branch). For each added/modified test file it warns when the file textually duplicates a ≥3-token conditional from a production file touched in the same PR **and** does not import the symbol/module owning it, via a pure, fixture-tested `findReimplementedConditions()` with an allowlist mirroring `assertReductionAllowlist`. Report-only for now (does not fail the gate) — to be promoted to blocking after a triage cycle. Regression guard: `tests/unit/check-test-masking.test.ts` (45).
+- **feat(sse):** per-connection routing override (native vs CLIProxyAPI) ([#6339](https://github.com/diegosouzapw/OmniRoute/issues/6339)) — the previously-dead `isCliproxyapiDeepModeEnabled` helper is now wired into `resolveExecutorWithProxy`: a single connection can opt itself into the CLIProxyAPI passthrough executor via `providerSpecificData.cliproxyapiMode="claude-native"`, with precedence **connection override > provider `upstream_proxy_config` mode > default**. `resolveExecutorWithProxy` now receives the resolved connection's `providerSpecificData` (threaded from `chatCore.ts`), so one connection can deep-route while the provider's other connections stay native — no DB schema change (the toggle rides in `providerSpecificData`). Also resolves the same-provider mixing ask in #6340. Regression guard: `tests/unit/chatcore-executor-proxy.test.ts` (9). (thanks @RaviTharuma)
+- **feat(dashboard):** "Add session cookie" modal now shows a prominent **"Open ‹host› →"** link to the provider's own site ([#6268](https://github.com/diegosouzapw/OmniRoute/issues/6268)) — every `-web` cookie-session provider (chatgpt-web, claude-web, gemini-web, kimi-web, lmarena, qwen-web, m365-copilot-web, …) renders a one-click external link (opening the provider's login/home page in a new tab) so operators no longer tab away to retype the URL mid-setup. The host resolves from a pure, unit-tested `resolveWebProviderHost()` (prefers `WEB_COOKIE_PROVIDERS[id].website`, falls back to the registry `baseUrl` origin); non-web providers render exactly as before. Kilo's dup flag vs #6265 (modal-too-small-on-1080p) was a false positive — distinct concern. Regression guard: `tests/unit/resolve-web-provider-host.test.ts` (5). (thanks @chirag127)
+- **feat(providers):** add **DigitalOcean** AI (serverless inference) as an OpenAI-compatible API-key provider ([#6373](https://github.com/diegosouzapw/OmniRoute/pull/6373)) — base `https://inference.do-ai.run/v1`, wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/digitalocean/`, `src/shared/constants/providers/apikey/inference-hosts.ts`). Regression guard: `tests/unit/digitalocean-provider.test.ts`. (thanks @newnol)
+- **feat(providers):** add **Huancheng Public API** (`hcnsec`) as an OpenAI-compatible regional provider ([#6410](https://github.com/diegosouzapw/OmniRoute/pull/6410)) — Xinjiang Huancheng Cybersecurity's public LLM platform (base `https://api.hcnsec.cn/v1`, free credits via daily check-ins), wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/hcnsec/`, `src/shared/constants/providers/apikey/regional.ts`). Regression guard: `tests/unit/hcnsec-provider.test.ts`. (thanks @UnrealAryan)
+- **feat(dashboard):** the web-session credential guide now shows an **"Open {host}" link** ([#6316](https://github.com/diegosouzapw/OmniRoute/pull/6316)) to the provider's sign-in site (derived from the provider `website` via `getProviderWebsiteHost`), so you can jump straight to the page where the cookie/session must be captured. Regression guard: `tests/unit/web-session-provider-link-6316.test.ts`. (thanks @jordansilly77-stack)
+- **feat(cerebras):** add the **Gemma 4 31B** model (`gemma-4-31b`) to the Cerebras registry + pricing table ([#6331](https://github.com/diegosouzapw/OmniRoute/pull/6331)). Regression guard extends `tests/unit/t28-model-catalog-updates.test.ts`. (thanks @backryun)
+- **feat(providers):** add **Yuanbao (web)** as a cookie-session provider ([#6196](https://github.com/diegosouzapw/OmniRoute/issues/6196)) — `yuanbao-web` (Tencent Yuanbao, `yuanbao.tencent.com`) with cookie-only auth (`hy_user`/`hy_token` + public agent id), SSE→OpenAI translation incl. `reasoning_content`, exposing DeepSeek V3/R1 + Hunyuan / Hunyuan-T1. Regression guard: `tests/unit/providers-yuanbao-web.test.ts`. `together-web` was **deferred** (no verifiable web-session endpoint — needs a captured request) and `huggingchat-web` **dropped** (the existing `huggingchat` already is a web-cookie provider). (thanks @chirag127)
+- **feat(providers):** route the built-in **agentrouter** through the dynamic Claude-Code wire image ([#6056](https://github.com/diegosouzapw/OmniRoute/issues/6056)) — a small static allow-set (`CC_WIRE_IMAGE_BUILTINS` in `open-sse/services/ccWireImageBuiltins.ts`), consulted by `isClaudeCodeCompatible` / `isClaudeCodeCompatibleProvider` / `applyFingerprint`, makes agentrouter adopt the CC wire-image headers + fingerprint **while guarding the CC baseUrl/auth branches** so it keeps its own registry `baseUrl` and `x-api-key` auth. Regression guard: `tests/unit/agentrouter-cc-wire-image.test.ts` (asserts the wire image is applied AND agentrouter's baseUrl/auth are preserved). Live WAF-acceptance against agentrouter.org is a VPS validation follow-up (Hard Rule #18).
+- **feat(providers):** **bulk-add API keys for Cloudflare Workers AI** ([#6174](https://github.com/diegosouzapw/OmniRoute/issues/6174)) — `cloudflare-ai` is removed from the bulk-add exclusion list and the bulk parser gains a 3-field `name|accountId|apiKey` mode; the bulk route now builds a **per-entry** `providerSpecificData` so each key carries its own `accountId` (fixing the previous shared-object reuse), and both the create + key-validation paths receive it. Regression guard: `tests/unit/bulk-api-key-parser-cloudflare.test.ts`. (thanks @muflifadla38)
+- **feat(dashboard):** routing/settings UX clarity ([#6147](https://github.com/diegosouzapw/OmniRoute/issues/6147)) — (1) weighted combos show the **effective routing share %** next to each weight when weights don't sum to 100 (`WeightTotalBar.tsx`); (2) the status widget's user-facing **"Cloud Sync" label is renamed** to "Remote Settings Sync" (`CloudSyncStatus.tsx`; internal ids/state untouched); (3) built-in providers gain an **opt-in advanced base-URL override** (`isBaseUrlOverrideEligibleProvider`, hidden behind an "Advanced" toggle, reusing the existing `providerSpecificData.baseUrl` persistence — not globally widened). Regression guard: `tests/unit/routing-settings-ux-6147.test.ts`.
+- **feat(combo):** add an option to **disable session stickiness**, per-combo or globally — round-robin / random combos can rotate to a different connection on every request instead of pinning a whole conversation to one connection by its first-message hash. Resolution precedence per-combo `config.disableSessionStickiness` → global `settings.disableSessionStickiness` → default `false` (preserves the #3825 prompt-cache/504 fix); gates **both** stickiness call sites in `open-sse/services/combo.ts`. Exposed as a global toggle (Combo Defaults) and a per-combo Inherit/on/off control. ([#6168](https://github.com/diegosouzapw/OmniRoute/issues/6168)) Regression guard: `tests/unit/combo-disable-session-stickiness.test.ts`. (thanks @RCrushMe)
+- **feat(docker):** add the `OMNIROUTE_NO_SUDO` env flag for root-less / user-namespaced deployments — the MITM cert-trust command path (`resolveSudoSpawn` in `src/mitm/systemCommands.ts`) now strips the leading `sudo` when the flag is truthy, in addition to the existing root / sudo-missing cases, so the Proxy Agent runs without `sudo` (the operator trusts the CA manually, e.g. via `NODE_EXTRA_CA_CERTS`). Argv-array `spawn` preserved — no shell interpolation (Hard Rule #13). ([#6122](https://github.com/diegosouzapw/OmniRoute/issues/6122)) Regression guard: `tests/unit/mitm-systemCommands-no-sudo.test.ts`. (thanks @powellnorma)
+- **feat(providers):** add **Requesty** as an OpenAI-compatible gateway provider (BYOK, base `https://router.requesty.ai/v1`, ~200 free requests/day) — wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/requesty/`, `src/shared/constants/providers/apikey/gateways.ts`). ([#6120](https://github.com/diegosouzapw/OmniRoute/issues/6120)) Regression guard: `tests/unit/requesty-provider.test.ts`. (thanks @chirag127)
+- **feat(dashboard):** add **configured-only / available-only filters** to the Free Provider Rankings page ([#6150](https://github.com/diegosouzapw/OmniRoute/issues/6150)) — hide providers you haven't configured, or whose connections are all rate-limited / out of quota, via server-side query params (`?configuredOnly` / `?availableOnly` on `GET /api/free-provider-rankings`) backed by a testable lib helper reusing the in-process connection state (no Redis). Both filters default off, so the default view is unchanged; this supersedes the earlier client-side "Configured Only" toggle (#6245) with an available-only dimension and unit-tested logic. Regression guard: `tests/unit/freeProviderRankings-filters.test.ts`.
+
+### 🔧 Bug Fixes
+
+- **fix(dashboard):** adding a second API key connection for the same provider no longer silently overwrites the first — the Add-API-key modal now derives a unique default connection name (`main`, then `main-2`, `main-3`, …) so the backend name-based upsert can't collide (#6499 — thanks @dilneiss).
+- **fix(compression):** the session-dedup engine now also deduplicates a large multi-line block repeated **within a single message** (intra-message dedup), not just across turns; the compression-preview API surfaces a `fallbackReason`, and the fusion panel reports how many models were rate-limited vs failed on total-panel failure (#6501 — thanks @chirag127).
+- **fix(compression):** a stacked-pipeline step naming an unregistered engine now surfaces a `validationErrors` entry instead of silently no-op'ing, so misconfigured pipelines are visible in the preview API (#6506 — thanks @chirag127).
+- **feat(usage):** add a **Codex reset-credit redemption** flow to the Provider Limits UI ([#6361](https://github.com/diegosouzapw/OmniRoute/pull/6361)) — a `useCodexResetCreditRedemption` hook + `/api/usage/codex-reset-credit` route + `codexResetCredits` lib let you redeem banked Codex reset credits from the quota card. Regression guard: `tests/unit/codex-reset-credits.test.ts`. (thanks @JxnLexn)
+- **feat(glm):** add **team-plan quota settings** for `glm-cn` connections ([#6351](https://github.com/diegosouzapw/OmniRoute/pull/6351)) — a dedicated `GlmTeamQuotaFields` form section (team quota id / limits) threaded through the Add/Edit connection modals, persisted via `providerSpecificData`, with the GLM usage service reading the team quota. Regression guards: `tests/unit/glm-team-quota.test.ts`, `provider-specific-data-schema.test.ts`. (thanks @hao3039032)
+- **feat(providers):** add **TinyFish** web-fetch/search support ([#6349](https://github.com/diegosouzapw/OmniRoute/pull/6349)) — a `tinyfish-fetch` executor + `/v1/web/fetch` route + MCP web-fetch tool, registered as a specialty-media provider with request-validation and a search-provider catalog entry. Regression guards: `tests/unit/executor-tinyfish-fetch.test.ts`, `web-fetch-handler.test.ts`, `mcp-web-fetch-tool.test.ts`, `provider-validation-tinyfish.test.ts`. (thanks @dtybnrj)
+- **fix(cli):** `omniroute launch-codex` now spawns `codex.cmd` through a shell on Windows (the npm `.cmd` shim is unresolvable by bare `spawn` → ENOENT), mirroring the qodercli Windows fix (#6263) ([#6312](https://github.com/diegosouzapw/OmniRoute/pull/6312)). Regression guard: `tests/unit/launch-codex-windows-spawn-6312.test.ts`. (thanks @swingtempo)
+- **fix(codex):** isolate the **Spark** quota from the shared Codex quota and stabilize the quota UI ordering / hydration so per-scope limits render consistently ([#6336](https://github.com/diegosouzapw/OmniRoute/pull/6336)). Regression guards: `tests/unit/codex-quota-selection-hydration.test.ts`, `provider-limits-ui.test.ts` + 3 more. (thanks @xz-dev)
+- **feat(api):** add a `hidePaidModels` setting that filters paid-only models out of the `/v1/models` catalog. Regression guard: `tests/unit/models-catalog-hide-paid.test.ts`. (thanks @chirag127)
+- **fix(api-manager):** the fallback model picker now preserves combos instead of dropping them when a primary model is unavailable ([#6443](https://github.com/diegosouzapw/OmniRoute/pull/6443)). Regression guard: `tests/unit/api-manager-page-static.test.ts`. (thanks @jmengit)
+- **fix(providers):** recoverable Antigravity / Cloud-Code (Gemini Code Assist) `403` responses ([#6452](https://github.com/diegosouzapw/OmniRoute/pull/6452)) are now classified as a retryable project-config error instead of a terminal account ban, so a fixable project/API-disabled 403 no longer forces a ~1-year cooldown / full OAuth reconnect. Regression guard: `tests/unit/errorclassifier-antigravity-403.test.ts`. (thanks @developerjillur)
+- **fix(mitm):** `sanitizeHeaders` now redacts `Set-Cookie` response headers so upstream session cookies never leak into logs / diagnostics ([#6451](https://github.com/diegosouzapw/OmniRoute/pull/6451)). Regression guard: `tests/unit/mitm-sanitize-headers.test.ts`. (thanks @developerjillur)
+- **fix(api):** `/api/compression/preview` now accepts `mode: "caveman"` and correctly handles stacked / zero-compression previews ([#6425](https://github.com/diegosouzapw/OmniRoute/issues/6425)). Regression guard: `tests/unit/api/compression-preview-caveman-and-stacked-6425.test.ts`. (thanks @chirag127)
+- **feat(providers):** add **Zed** hosted LLM aggregator as a native-app provider ([#6118](https://github.com/diegosouzapw/OmniRoute/pull/6118)) — OAuth sign-in via the Zed hosted flow, registered through the shared provider registry + executor. Regression guards: `tests/unit/zed-oauth-provider.test.ts`, `zed-import-utils.test.ts`, `zed-docker-detect.test.ts`, `mitm-handler-zed.test.ts`. VPS-validated via live operator login (Hard Rule #18).
+- **fix(oauth):** the Kiro SSO-cache auto-import now **preserves the IDC region** — cross-region Amazon Q / Kiro profiles imported from the SSO cache are no longer collapsed to the default region ([#6113](https://github.com/diegosouzapw/OmniRoute/pull/6113)). Regression guard: `tests/unit/kiro-auto-import-idc-2059.test.ts`. VPS-validated via live operator login (Hard Rule #18).
+- **fix(dashboard):** passthrough model aliases no longer collide when two namespaced model ids share a last segment (port from 9router#1850, [#6431](https://github.com/diegosouzapw/OmniRoute/pull/6431)). `enx/gpt-5.5` and `enx/codebuddy/gpt-5.5` both auto-generated the alias `gpt-5.5`, so the second model could never be added (the UI just alerted "alias already exists"). Aliases are now disambiguated deterministically — bare last segment when free, then parent-qualified (`codebuddy-gpt-5.5`), then a numeric suffix — while re-adding the exact same model id is still blocked. Regression guard: `tests/unit/passthrough-alias-1850.test.ts`. (thanks @arpicato)
+- **fix(translator):** preserve a Gemini `functionResponse` co-located with other parts (another `functionCall`, or trailing `text`) in the same content when translating **Gemini → OpenAI** ([#6376](https://github.com/diegosouzapw/OmniRoute/pull/6376)). `convertGeminiContent()` early-returned the tool message on the first `functionResponse` part, dropping any co-located parts; such contents are now pre-split (one tool message per `functionResponse`, emitted first, plus one message for the remaining parts). Regression guard: `tests/unit/gemini-to-openai-function-response.test.ts`. (thanks @warelik)
+- **fix(headroom):** detect a python interpreter managed by **mise / pyenv / asdf / conda** (port from 9router#2353, [#6382](https://github.com/diegosouzapw/OmniRoute/pull/6382)). Headroom's python probe (`src/lib/headroom/detect.ts`) searched a hardcoded `PATH`, but version managers expose their interpreters via shim dirs that only join `PATH` through interactive-shell activation — which the non-interactive server never runs, so a managed python (≥3.10) was invisible and Headroom reported it missing. The search path now prepends the well-known shim/bin dirs (`~/.local/share/mise/shims`, `~/.pyenv/shims`, `~/.asdf/shims`, `$CONDA_PREFIX/bin`, `~/.local/bin`, respecting `MISE_DATA_DIR`/`PYENV_ROOT`/`ASDF_DATA_DIR` when set), and a new `HEADROOM_PYTHON` env override lets operators point straight at their interpreter (mirroring `HEADROOM_URL`). Still shell-free (`execFileSync`). Regression guard: `tests/unit/headroom-detect.test.ts` (5). (thanks @loopyd)
+- **fix(executors):** strip the OpenAI-Codex/Claude-CLI `client_metadata` passthrough field for **NVIDIA** requests (port from 9router#1887, [#6411](https://github.com/diegosouzapw/OmniRoute/pull/6411)). NVIDIA's OpenAI-compatible wrapper rejects it with `400 Unsupported parameter`, the same class already handled for `cerebras`/`mistral`; `nvidia` (executor `default`) was missing from the strip allowlist so Codex/Claude-Code passthrough requests 400'd. Regression guard: `tests/unit/executor-default-strip-client-metadata.test.ts` (+nvidia case). (thanks @phidinhmanh)
+- **fix(translator):** strip the Claude-style `thinking` field for **NVIDIA `z-ai/glm-5.2`** (port from 9router#2023, [#6413](https://github.com/diegosouzapw/OmniRoute/pull/6413)). NVIDIA's OpenAI-compatible wrapper 400s on `thinking` (a Claude-format client routed here leaves a `thinking:{type:"adaptive"}`); the existing strip rule only dropped `reasoning`. Same class already handled for `minimax-m2.7`. Regression guard: `tests/unit/nvidia-minimax-thinking-strip.test.ts` (+glm-5.2 case). (thanks @phidinhmanh)
+- **fix(translator):** suppress the streamed `</think>` close marker for the **Antigravity IDE** client (port from 9router#1061, [#6415](https://github.com/diegosouzapw/OmniRoute/pull/6415)). On thinking-only turns Antigravity rendered a bare `</think>` as the sole visible content, tripping its loop-detection and wasting requests. Antigravity's UA (`vscode/<v> (Antigravity/<v>)`) is added to the marker-suppress allowlist (alongside OpenCode); Claude Code / Cursor still get the marker, and `x-omniroute-thinking-marker: on` force-restores it. Regression guard: `tests/unit/think-close-marker-suppress-5245.test.ts`. (thanks @abdofallah)
+- **fix(executors):** strip nested `reasoning_content` from messages for **Mistral** (port from 9router#1649, [#6417](https://github.com/diegosouzapw/OmniRoute/pull/6417)). Mistral's API returns `422 extra_forbidden` when an assistant message carries `reasoning_content` (replayed thinking from a prior turn, e.g. via the Codex `/responses` path); the generic top-level 400 field-downgrade retry never covered the nested per-message field. `DefaultExecutor` now strips it for provider `mistral` only, so DeepSeek (which requires replayed `reasoning_content`) is unaffected. Regression guard: `tests/unit/mistral-strip-reasoning-content-1649.test.ts`. (thanks @xxy9468615)
+- **fix(executors):** strip the `client_metadata` passthrough field on the **OpenCode** path (port from 9router#1442, [#6418](https://github.com/diegosouzapw/OmniRoute/pull/6418)). OpenCode upstreams (e.g. `kimi-k2.6` via opencode-go) reject it with `400 "Extra inputs are not permitted, field: 'client_metadata'"`; the DefaultExecutor strip only covered cerebras/mistral and `OpencodeExecutor` extends `BaseExecutor` directly, so nothing removed it there. Regression guard: `tests/unit/opencode-strip-client-metadata-1442.test.ts`. (thanks @yanpaing007)
+- **fix(executors):** inject the `reasoning_content` echo for the native **Moonshot Kimi** provider (port from 9router#1480, [#6419](https://github.com/diegosouzapw/OmniRoute/pull/6419)). Kimi (executor `default`) is a thinking-mode upstream that 400s with "reasoning_content must be passed back" when a prior assistant turn lacks it; the placeholder injection was only wired into the OpenCode meta-provider, so direct multi-turn Kimi conversations failed. Scoped to `kimi` (gateway-served models matching the thinking-model name pattern are unaffected). Regression guard: `tests/unit/kimi-native-reasoning-injected-1480.test.ts`. (thanks @2220258345)
+- **fix(executors):** recover from a strict gateway's `context_management: Extra inputs are not permitted` 400 (port from 9router#1468, [#6420](https://github.com/diegosouzapw/OmniRoute/pull/6420)). **Claude Code** always sends a top-level `context_management` field; strict anthropic-compatible gateways reject it. The dedicated context-editing 400-fallback only fired when OmniRoute's own `contextEditing` feature was enabled (default off), so a client-sent field passed through untouched and 400'd. `context_management` is now in the generic reactive field-strip list, so it's stripped-and-retried once regardless of the feature flag (with correct request re-signing for claude-compatible relays). Regression guard: `tests/unit/provider-field-strips.test.ts`. (thanks @ohahe52-dot)
+- **fix(network):** enable **RFC 8305 Happy Eyeballs** (`autoSelectFamily`) on the direct-egress undici dispatcher (port from 9router#1237, [#6423](https://github.com/diegosouzapw/OmniRoute/pull/6423)). When DNS returns both IPv6 (AAAA) and IPv4 (A) and the IPv6 route is broken (e.g. a NAT64 `64:ff9b::` prefix without routing), undici tried IPv6 first and hung until `ETIMEDOUT` (then a 502 + account lockout), even though `curl` reached the same host. The direct dispatcher now races both families and uses whichever connects first. Proxy paths pin family via `proxyTls` and are unaffected. Regression guard: `tests/unit/direct-dispatcher-pipelining-4580.test.ts`. (thanks @adentdk)
+- **fix(combo):** round-robin now advances the rotation pointer past the model that **actually served**, not the eagerly-scheduled one (port from 9router#948, [#6428](https://github.com/diegosouzapw/OmniRoute/pull/6428)). With `stickyLimit: 1` (true round-robin), when the scheduled model failed and a _different_ model served via fallback, the counter had already advanced +1 from the scheduled index — so the next request reused the fallback-served model, degrading round-robin into hot-spotting on whichever model was healthy. The pointer now advances to the served index + 1 (mirroring the sticky-limit>1 path). Session-stickiness (#3825) and distribution are preserved. Regression guard: `tests/unit/combo-rr-fallback-advance-948.test.ts`. (thanks @binsarjr)
+- **fix(sse):** a non-string `model` field is now rejected with a `400` before the resolver, instead of crashing downstream `.toLowerCase()`/`.split()` calls into an empty-body `500` that escapes the error sanitizer ([#6407](https://github.com/diegosouzapw/OmniRoute/issues/6407)). Regression guard: `tests/unit/chat-non-string-model-6407.test.ts`. (thanks @chirag127)
+- **fix(api):** unknown `/api/*` routes now return a JSON `404` (instead of the dashboard HTML shell) and scalar chat params (`model`/`temperature`/etc.) are validated **before** the provider lookup so malformed requests fail fast with a clear `400` ([#6424](https://github.com/diegosouzapw/OmniRoute/issues/6424), [#6412](https://github.com/diegosouzapw/OmniRoute/issues/6412)). Regression guards: `tests/unit/api/api-catchall-json-404.test.ts`, `tests/unit/chat-early-schema-validation-6412.test.ts`. (thanks @chirag127)
+- **fix(api):** `/v1/chat/completions` now rejects a non-JSON `Content-Type` with a `400` before parsing the body ([#6414](https://github.com/diegosouzapw/OmniRoute/issues/6414)). Regression guard: `tests/unit/v1-chat-completions-content-type-6414.test.ts`. (thanks @chirag127)
+- **fix(api):** the `X-OmniRoute-Compression` response header is now echoed on `/v1/chat/completions` and `/v1/completions` ([#6422](https://github.com/diegosouzapw/OmniRoute/issues/6422)). Regression guard: `tests/unit/compression-header-echo-6422.test.ts`. (thanks @chirag127)
+- **fix(api):** concurrent `GET /v1/models` requests are coalesced into a single catalog build ([#6408](https://github.com/diegosouzapw/OmniRoute/issues/6408)). Regression guard: `tests/unit/v1-models-concurrent-6408.test.ts`. (thanks @chirag127)
+- **fix(api):** `/v1/completions` now echoes the requested `body.model` in its JSON + streamed responses ([#6429](https://github.com/diegosouzapw/OmniRoute/pull/6429)). Regression guard: `tests/unit/completions-body-model-echo.test.ts`. (thanks @chirag127)
+- **fix(api):** env-var master keys now see the full `/v1/models` catalog ([#6406](https://github.com/diegosouzapw/OmniRoute/issues/6406)). Regression guard: `tests/unit/models-catalog-envkey-6406.test.ts`. (thanks @chirag127)
+- **fix(api):** non-streaming `/v1/completions` responses now echo `body.model` aligned with the `X-OmniRoute-Model` header ([#6426](https://github.com/diegosouzapw/OmniRoute/issues/6426)). Regression guard: `tests/unit/v1-completions-model-header-match-6426.test.ts`. (thanks @chirag127)
+- **fix(api):** unknown `/v1/*` routes now return a JSON `404 not_found` instead of the Next.js dashboard HTML shell ([#6405](https://github.com/diegosouzapw/OmniRoute/issues/6405)). Regression guard: `tests/unit/api/v1-catchall-json-404.test.ts`. (thanks @chirag127)
+- **fix(api):** the per-connection provider models route now degrades to the shipped catalog when a provider's `/models` endpoint answers with a redirect ([#6267](https://github.com/diegosouzapw/OmniRoute/issues/6267)) — a `qwen-web` import failed with a raw `Redirect blocked … (307)` 503. `safeOutboundFetch` throws `REDIRECT_BLOCKED` on the 307, `getSafeOutboundFetchErrorStatus` maps it to 503, and `buildDiscoveryErrorFallbackResponse` treated every 503 as a hard error — so the non-empty `getModelsByProviderId("qwen-web")` catalog was never surfaced. A models-endpoint redirect is not a fixable-config error (unlike `URL_GUARD_BLOCKED`/`INVALID_URL`, which stay hard errors), so it now falls back to the local/cached catalog before the 503 short-circuit. General fix — covers any config-driven provider that 307s. Regression guard: `tests/unit/provider-models-qwen-web-redirect-6267.test.ts`. (thanks @chirag127)
+- **fix(api):** the per-connection provider models route (MCP `list_models_catalog` + the dashboard import view) now merges USER-ADDED custom models into its response ([#6247](https://github.com/diegosouzapw/OmniRoute/issues/6247)) — custom models live in the `key_value` namespace `customModels`, which the live REST `/api/v1/models` already merges, but `src/app/api/providers/[id]/models/route.ts` never read `getCustomModels`, so custom models were dropped on both the discovery-success and local_catalog paths. They are now folded into the returned model list (deduped by id, stamped `owned_by: provider`), fixing MCP + the dashboard import view in one place. Regression guard: `tests/unit/provider-models-custom-merge-6247.test.ts`. (thanks @RCrushMe)
+
+- **fix(providers):** GitLab Duo tool-calling follow-up turns no longer fail upstream with `422 {"detail":"Validation error"}` (tokens 0/0, rejected pre-inference). The #6234 tool-result-feedback fix serialized the **entire** multi-turn conversation into GitLab's single-file `code_suggestions` (`small_file`) generation endpoint — folded history that turn-N sent as an oversized `current_file.content_above_cursor` **and** duplicated verbatim into `user_instruction`, tripping the AI-Gateway's `small_file` validation guard. The executor now **bounds** that prompt: it keeps system + latest user message + the most-recent tool round (dropping older turns), caps oversized tool results, and stops duplicating the full prompt into `user_instruction` (which now carries only the short latest user message) — while still feeding the most-recent tool result back so the agent continues (`open-sse/executors/gitlab.ts`, [#6220](https://github.com/diegosouzapw/OmniRoute/issues/6220)). The unit test covers the bounding logic; the upstream 422→200 clearing is VPS-only (Hard Rule #18). Regression guard: `tests/unit/gitlab-tool-exchange-bounded-6220.test.ts`.
+
+- **fix(i18n):** the provider-detail (`/dashboard/providers/[id]`) connection-status filter labels no longer render as `__MISSING__:All` / `__MISSING__:Active` / `__MISSING__:Error` / `__MISSING__:Banned` / `__MISSING__:CreditsExhausted` in non-English locales (notably pt-BR) ([#6290](https://github.com/diegosouzapw/OmniRoute/issues/6290)). Root cause was **not** the namespace mismatch the issue guessed — the `providers.filter*` keys resolve correctly in `en.json`; the debt lived in the locale mirrors (`src/i18n/messages/*.json`), where these five keys carried the `__MISSING__:` sync sentinel in ~15 locales and were absent entirely in ~26 others, so next-intl found the key and echoed the sentinel verbatim. All 40 non-English/-Chinese mirrors now ship real translations for the five `providers.filter*` labels. Regression guard: `tests/unit/i18n-provider-filter-keys-6290.test.ts`. (thanks @diegosouzapw)
+
+- **fix(providers):** the `copilot-m365-web` streaming executor now emits `debug`-level WebSocket diagnostics ([#6210](https://github.com/diegosouzapw/OmniRoute/issues/6210)) — the outbound WS URL (with the `access_token` **redacted** via `redactWsUrl()`), handshake success/failure, and each received SignalR frame's `type`/`target`. Previously the streaming path logged nothing, so an empty `content:null` response (the M365 Education / Starter tier symptom fixed in #6234) was undiagnosable even at `APP_LOG_LEVEL=debug`. The change is debug-level and side-effect-free — it does not alter streaming behavior or the frame parser, and the token never reaches the logs. Regression guard: `tests/unit/copilot-m365-web-logging-6210.test.ts` (thanks @qpeyba)
+
+- **fix(resilience):** a round-robin combo no longer returns `503 all upstream accounts are unavailable` when a compatibility-rejected target is actually healthy ([#6238](https://github.com/diegosouzapw/OmniRoute/issues/6238)). `filterTargetsByRequestCompatibility` drops request-incompatible targets (tool/vision/structured-output unsupported, or below the required context window) **before** any availability check runs, and its `compatible.length === 0` safety net only fired when _all_ targets were filtered — not when the kept targets later all turned out runtime-unavailable (circuit-open / cooldown / no credentials). So a combo could 503 while a compat-rejected-but-healthy provider sat unused. `handleRoundRobinCombo` now keeps the compat-rejected set and, when every compat-kept target was skipped without a single real attempt, probes those rejected targets as a **last-resort fallback tier** (via the new pure `open-sse/services/combo/comboCompatFallback.ts`) before crystallizing the 503. Regression guard: `tests/unit/combo-roundrobin-compat-fallback-6238.test.ts`. (thanks @ThongAccount)
+
+- **fix(startup):** best-effort self-heal for a corrupted Turbopack dev cache on Windows ([#6289](https://github.com/diegosouzapw/OmniRoute/issues/6289)). On Windows, `pnpm dev` can fail at startup when Turbopack `mmap`s a persistent-cache SST file and the OS refuses the mapping (`os error 1455` — "paging file too small"), which Turbopack surfaces as a misleading `Module not found: Can't resolve '@/shared/utils/machine'`. This is a **known upstream Turbopack cache-corruption bug — not our code**. The dev launcher (`scripts/dev/run-next.mjs`) now wraps `nextApp.prepare()` and, when it rejects with that signature (`isTurbopackCacheCorruption` in the new `scripts/dev/turbopackCacheHeal.mjs`), purges `.build/next/**/cache/turbopack` and retries **once** with a clear log. **Caveat — best-effort only:** the corruption often surfaces as a runtime overlay rather than a `prepare()` rejection, so this cannot always intercept it; the reliable remedy remains manually deleting the Turbopack cache dir. Regression guard: `tests/unit/turbopack-cache-heal-6289.test.ts`. (thanks @chirag127)
+
+- **fix(providers):** qodercli PAT auth no longer fails with `spawn qodercli ENOENT` on Windows ([#6263](https://github.com/diegosouzapw/OmniRoute/issues/6263)) — `spawnQoderCli` spawned the bare `qodercli` name with `shell:false` and an unenriched env, so the npm `.cmd` wrapper under `%APPDATA%\npm` (a user-PATH directory) was never resolved. It now resolves the absolute `.cmd`/`.exe` path through the existing `getCliRuntimeStatus("qoder")` resolver in `src/shared/services/cliRuntime.ts` (memoized), spawns with `shell` when the target is a `.cmd`/`.bat`, and uses the cliRuntime-enriched env (PATH + PATHEXT + APPDATA); the ENOENT error now lists the searched paths plus the `CLI_QODER_BIN` override. End-to-end spawn on a real Windows host is host-only (Hard Rule #18); the path-resolution logic is unit-tested. Regression guard: `tests/unit/qodercli-windows-resolve-6263.test.ts`. (thanks @chirag127)
+
+- **fix(sse):** the reasoning-token buffer no longer inflates **probe-sized `max_tokens`** ([#6274](https://github.com/diegosouzapw/OmniRoute/issues/6274)) — Claude Code's `/model` capability check sends `max_tokens: 1`, but for a thinking-capable model with a large output cap (e.g. `glm-5.2`) the #3587 headroom heuristic (`max(current + 1000, ceil(current * 1.5))`) rewrote it to `1001` and forwarded that upstream, wasting tokens on a request that was never a genuine reasoning budget. `resolveReasoningBufferedMaxTokens()` (`open-sse/services/reasoningTokenBuffer.ts`) now short-circuits and returns the caller's value verbatim when it is below the new `REASONING_BUFFER_MIN_TRIGGER` (256) threshold — a tiny explicit limit is a probe, not a reasoning request. Real budgets still receive the #3587 headroom unchanged, and the guard runs after the existing capability checks so unknown / non-reasoning models keep returning `null`. Regression guard: `tests/unit/reasoning-token-buffer-6274.test.ts`. (thanks @brightfiscalband)
+
+- **fix(cli):** `omniroute reset-password` now works as a real subcommand, and password resets over piped (non-TTY) stdin actually apply ([#6261](https://github.com/diegosouzapw/OmniRoute/issues/6261), [#6258](https://github.com/diegosouzapw/OmniRoute/issues/6258)). Two coupled defects: (1) **#6261** — `bin/omniroute.mjs` routed everything through Commander with only two pre-Commander bypasses (`--mcp`, `reset-encrypted-columns`), so `omniroute reset-password` was rejected as an unknown command; only the separate `omniroute-reset-password` bin worked, while the docs falsely advertised the subcommand (incl. a bogus "legacy alias still works"). A pre-Commander bypass mirroring `reset-encrypted-columns` now dynamically imports `bin/reset-password.mjs` (which self-executes) before Commander parses; the three doc lines were corrected. (2) **#6258** — `bin/reset-password.mjs` issued two sequential `rl.question` prompts; under piped stdin the second read never settled at EOF, so `main()` never reached `resetManagementPassword` and the reset was a silent no-op (both prompts printed, no success, password unchanged). The CLI now detects non-TTY stdin and reads it once (first line = password, second line = confirm if present, else reused), adds a `--password-stdin` flag (entire stdin is the password, no confirmation), and exits `0` explicitly so the success line always flushes; interactive TTY behavior is unchanged. Regression guard: `tests/unit/reset-password-cli-6261-6258.test.ts` (3). (thanks @chirag127)
+
+- **fix(db):** the mass-migration **safety abort** now tells the operator how to bypass it and stops flooding the log ([#6260](https://github.com/diegosouzapw/OmniRoute/issues/6260)) — after restoring a backup that wiped the migration tracking table, `runMigrations()` threw the abort on every downstream `ensureDbInitialized()`, re-logging the full banner 11+ times, and the message never mentioned the existing `OMNIROUTE_MAX_PENDING_MIGRATIONS` escape hatch. The abort text now appends a bypass hint (set `OMNIROUTE_MAX_PENDING_MIGRATIONS=0` in `server.env` / `DATA_DIR/.env`), and a new `MigrationSafetyAbortError` is memoized so repeated calls in the same process throw the same instance and emit a single concise line instead of the full cascade. Regression guard: `tests/unit/migration-safety-abort-6260.test.ts`. (thanks @chirag127)
+
+- **fix(auth):** importing a **distinct** Codex/ChatGPT OAuth `auth.json` is no longer falsely rejected as "already exists" when it belongs to a different user in the same workspace ([#6301](https://github.com/diegosouzapw/OmniRoute/issues/6301)). `findExistingCodexConnection` (in `src/lib/oauth/utils/codexAuthImport.ts`) deduped **only** on `providerSpecificData.workspaceId === accountId`, where `accountId` is the shared `chatgpt_account_id`/`tokens.account_id` — so two members of the same ChatGPT Team collapsed onto a single connection (409 `duplicate_account`). The id_token's `https://api.openai.com/auth` claim carries a per-user `chatgpt_user_id` alongside the workspace id (the device-flow path already persisted it as `chatgptUserId`, but the import path did not). Now `parseAndValidateCodexAuth` extracts `userId` (`chatgpt_user_id` → `user_id` → JWT `sub`) into `ParsedCodexAuth`, the create/update paths persist `chatgptUserId` in `providerSpecificData` (mirroring `codex.ts`), and dedup keys on `workspaceId` **AND** `chatgptUserId` — with a backward-compat fallback to legacy accountId-only matching when no stored connection for that workspace records a `chatgptUserId`, so genuinely-same accounts still dedup. Regression guard: `tests/unit/codex-auth-import-userid-dedup-6301.test.ts` (4). (thanks @anungma)
+
+- **fix(providers):** importing models for the **venice-web** provider no longer fails with a red "Provider venice-web does not support models listing" ([#6269](https://github.com/diegosouzapw/OmniRoute/issues/6269)). `venice-web` is a web-cookie provider with an executor but no upstream `/v1/models` endpoint and no registry `models`, so the models route fell through to the tail `400`. Mirroring the `jules`/`linkup-search`/`ollama-search` fix (#5569), it now ships a static local catalog entry in `src/lib/providers/staticModels.ts` — seeding the current Venice lineup (`venice-uncensored`, `llama-3.3-70b`, `qwen3-235b`, `qwen3-4b`, `deepseek-r1-671b`; Venice rotates its catalog, see docs.venice.ai/models/overview) — so the route returns `200` with `source:"local_catalog"`, `intentional:true`. Regression guard: `tests/unit/static-models-venice-web-6269.test.ts`. (thanks @chirag127)
+
+- **fix(api):** the specialty model catalogs (`/v1/embeddings`, `/v1/images`, `/v1/music`, `/v1/videos` model lists) are now derived from the **unified catalog filtered by a predicate** (`getSpecialtyModelsResponse`) instead of ad-hoc per-route logic, so they consistently respect active-credential visibility and stay in sync with the main catalog ([#6303](https://github.com/diegosouzapw/OmniRoute/pull/6303)). Regression guard: `tests/unit/specialty-model-catalog-routes.test.ts`. (thanks @makcimbx)
+- **fix(api):** the agent-bridge server route now resolves the MITM manager via a **dynamic `import("@/mitm/manager.runtime")`** so Turbopack does not statically pull the stub (or over-bundle the manager), and the agent-skills generator anchors its output base path with `path.join(process.cwd(), …)` so Turbopack's static analyzer stops tracing the whole project root ([#6329](https://github.com/diegosouzapw/OmniRoute/issues/6329), [#6366](https://github.com/diegosouzapw/OmniRoute/pull/6366)). Regression guard: `tests/unit/agent-bridge-server-route-dynamic-import.test.ts`. (thanks @Iammilansoni)
+- **fix(api):** internal probes (combo-test, cloud-sync verify) now pick a **management-scoped / allow-all API key** instead of naively grabbing `getApiKeys()[0]` — a restricted `self:usage` first row made the probe fail with "Model X is not allowed for this API key" even when the combo path was healthy (`pickApiKeyForInternalUse` in `src/lib/db/apiKeys.ts`). The API-manager model editor also falls back to `/api/models?all=true` when `/v1/models` is catalog-protected ([#6372](https://github.com/diegosouzapw/OmniRoute/pull/6372)). Regression guard: `tests/unit/pick-internal-api-key-6372.test.ts`. (thanks @jmengit)
+- **fix(live-ws):** the Live Dashboard WebSocket server now **rejects on bind failure** (e.g. `EADDRINUSE` when the API bridge already holds the port) instead of letting the error surface as an unhandled `error` event that crash-loops the process — the `error` listener is attached to `wss` (not `server`) and releases the EventBus subscription on a failed start ([#6324](https://github.com/diegosouzapw/OmniRoute/issues/6324)). Regression guard: `tests/unit/live-ws-eaddrinuse-6324.test.ts`. (thanks @vinayakkulkarni)
+- **fix(dashboard):** the Home provider-topology widget now trusts the live provider-metrics snapshot — it uses `topology.errorProvider` and live `activeRequests` directly instead of re-deriving state from a stale `lastErrorAt` or applying a frontend timeout filter, so the topology reflects real-time provider health ([#6322](https://github.com/diegosouzapw/OmniRoute/pull/6322)). Regression guard: `tests/unit/home-provider-topology-live-state.test.ts`. (thanks @xz-dev)
+- **fix(sse):** strip zero-width markers from streamed **tool-call arguments** — a follow-up to [#5857](https://github.com/diegosouzapw/OmniRoute/pull/5857). That PR removed injected zero-width joiners (U+200D) from streamed assistant text/reasoning but deliberately left tool-call argument JSON byte-exact. The request-side obfuscation (`open-sse/services/claudeCodeObfuscation.ts`) injects ZWJ into agent words — including the temp path inside the Bash tool description — and Claude models copy that verbatim into generated commands, which are delivered as tool-call arguments rather than assistant text. As a result the ZWJ survived and corrupted code blocks (e.g. a temp path rendered with an invisible joiner). Now `open-sse/handlers/responseSanitizer.ts` strips zero-width code points from tool-call argument strings at every emit site (OpenAI non-stream/stream chat `tool_calls` + legacy `function_call`, native Responses `function_call` items, the OpenAI→Responses conversion, and the native Responses streaming `response.function_call_arguments.delta/.done` events). Only zero-width code points are removed; JSON structure and all other bytes stay identical (no parse/restringify), so normal arguments remain byte-exact. Regression guard: 6 new cases in `tests/unit/response-sanitizer.test.ts` (suite 50/50).
+
+- **fix(nodejs):** the default app log path now resolves under `DATA_DIR` (`~/.omniroute/logs/application/app.log`) instead of `process.cwd()` ([#6197](https://github.com/diegosouzapw/OmniRoute/issues/6197)) — the globally-installed CLI runs from an arbitrary working directory, so anchoring the default to cwd made file logging silently write to (or no-op under) an unrelated directory, contradicting the documented `.env.example` default. `getAppLogFilePath()` now computes the default lazily via the pure `resolveDataDir()` resolver (honours a per-process `DATA_DIR`, no directory-creation side effect); an explicit `APP_LOG_FILE_PATH` still wins. Regression guard: `tests/unit/logenv-datadir-path-6197.test.ts` (3).
+
+- **fix(docker):** AgentBridge/`startMitm` no longer aborts in containers/headless when the Antigravity-default DNS step can't write `/etc/hosts` ([#6127](https://github.com/diegosouzapw/OmniRoute/issues/6127)), and the privileged command's stderr now reaches `app.log` instead of only a bare exit code hitting the toast ([#6198](https://github.com/diegosouzapw/OmniRoute/issues/6198)). The default DNS step (`addDNSEntry`) was called unguarded while cert install and the two sibling DNS steps were each best-effort — in the runtime Docker image (`USER node`, no `sudo`, read-only `/etc/hosts`) it threw `Command failed with code 1` out of `startMitmInternal` and killed the whole start, discarding the stderr. The three DNS steps are extracted into a best-effort `provisionDnsEntries()` where each failure is logged with the full `err` (stderr included, folded in by `systemCommands.ts`) and never aborts the start. Regression guard: `tests/unit/mitm-dns-graceful-degrade-6127.test.ts` (4).
+
+- **fix(providers):** copilot-m365-web now supports the M365 Education "Starter / OfficeWebIncludedCopilot" tier and no longer returns an empty `content:null` stream ([#6210](https://github.com/diegosouzapw/OmniRoute/issues/6210)). Two gaps: (1) `buildWsUrl()` hardcoded the individual-consumer scenario (`OfficeWebPaidConsumerCopilot`, `isEdu=false`) — the EDU tier is now opt-in via `providerSpecificData.tier="edu"`, emitting `scenario=OfficeWebIncludedCopilot`/`isEdu=true` (the individual path is unchanged); (2) the EDU/GPT-5.5 path streams deltas via `arguments[0].writeAtCursor` (incremental) instead of only `messages[].text` (accumulated snapshots), which the parser dropped — a new `accumulateBotContent()` folds both formats, with `type:2 item.result.message` as a last-resort fallback. Regression guard: `tests/unit/copilot-m365-edu-writeatcursor-6210.test.ts` (10). (thanks @qpeyba)
+
+- **fix(providers):** GitLab Duo executor now feeds tool results back into the prompt instead of looping ([#6220](https://github.com/diegosouzapw/OmniRoute/issues/6220)) — `buildPrompt()` branched only on `system`/`user` and took `userParts.at(-1)`, silently dropping the `assistant{tool_calls}` + `tool{result}` turns the client appended, so the reconstructed prompt was byte-identical to turn 1 and the model re-emitted the same `<tool>` call forever. When a tool exchange is present the full conversation is now serialized, folding each tool result back keyed by its `tool_call_id`; simple conversations keep the legacy shape. Complements the tool_call emission from [#6051](https://github.com/diegosouzapw/OmniRoute/issues/6051) (the `kilo-duplicate` label was a false positive — different, sequential defect). Regression guard: `tests/unit/gitlab-tool-result-feedback-6220.test.ts` (4).
+
+- **fix(providers):** opencode-go/opencode-zen can now synthesize the OpenCode CLI identity headers Cloudflare requires on VPS egress ([#5997](https://github.com/diegosouzapw/OmniRoute/issues/5997)) — on a datacenter VPS, `opencode.ai/zen/go/v1/chat/completions` 403s (HTML challenge) requests lacking CLI identity, while the reporter's control curl proved that `User-Agent: opencode-cli/1.0.0` + `x-opencode-client: cli` + `x-opencode-project: default` + fresh request/session UUIDs succeed. Opt-in via `OPENCODE_SYNTHESIZE_CLI_HEADERS=true` (values overridable via `OPENCODE_GO_USER_AGENT`/`OPENCODE_USER_AGENT`/`OPENCODE_CLIENT`/`OPENCODE_PROJECT`); it fills only headers the client did not already send. Kept **off by default** — the forward-only path is deliberate (fabricating a wrong value risks upstream rejection; a prior dedup regressed with `opencode/local`), so this replaces the fragile local header-injection shim without changing default behavior. Regression guard: `tests/unit/opencode-cli-headers-synthesis-5997.test.ts` (6). (thanks @aleksesipenko)
+- fix(resilience): sticky session affinity now evicts and fails over to another account when the pinned account is exhausted/unavailable (#6219)
+
+- fix(sse): Responses API passthrough now drops internal commentary-phase output before forwarding to clients (gated by RESPONSES_PASSTHROUGH_DROP_COMMENTARY, default on) (#6199)
+
+- **fix(sse):** tool-call function schemas with a root `type: null` are now coerced to `type: "object"` before dispatch ([#6359](https://github.com/diegosouzapw/OmniRoute/issues/6359)) — clients like the Codex app emit `parameters: { type: null, ... }` for some tools, which OpenAI-compatible upstreams reject with `400 Invalid schema for function '...': schema must be a JSON Schema of 'type: "object"', got 'type: null'`, failing the whole request. `toolSchemaSanitizer` already stripped the null; it now re-adds the mandatory root `"object"` type (and empty `properties`/open `additionalProperties` when absent). Combinator roots (`anyOf`/`oneOf`/`allOf`) and explicit root types are left untouched. Regression guard: 5 new cases in `tests/unit/tool-schema-sanitizer.test.mjs`.
+
+- **fix(docker):** AgentBridge no longer fails to start on npm/Electron/VPS installs with "MITM manager stub reached at runtime" ([#6344](https://github.com/diegosouzapw/OmniRoute/issues/6344)) — v3.8.45 flipped the production bundler default to Turbopack, but `next.config.mjs` aliased `@/mitm/manager` to its Docker-only degraded stub **unconditionally**. That was harmless while Docker (which sets the alias intentionally for #3390 graceful degradation) was the sole Turbopack consumer, but once every artifact built with Turbopack the stub shipped to all non-Docker users and `startMitm` threw on the first Agent-Bridge start. The alias is now opt-in via `OMNIROUTE_MITM_STUB=1` (set only by the Dockerfile) through the shared `scripts/build/mitm-stub-flag.mjs` helper; default builds bundle the real manager. Regression guard: `tests/unit/mitm-stub-alias-6344.test.mjs` (4).
+
+- **fix(proxy):** stop the v3.8.44 proxy regression that leaked the real IP and disabled healthy proxies ([#6246](https://github.com/diegosouzapw/OmniRoute/issues/6246)). Two coupled defects from the new health scheduler: (1) **IP leak** — when a proxy assigned to a connection was marked `inactive`, resolution fell through to a **direct** egress instead of blocking, exposing the operator's real IP; (2) **over-deactivation** — the sweep flipped a proxy to `inactive` on the **first** failed probe and counted our own 5s timeout / a probe-target `5xx` as the proxy's fault, so healthy paid proxies vanished from egress selection ("my proxies are not being used anymore"). Fix: the sweep decision is extracted into a pure, network-free `decideProxyHealthAction` (`src/lib/proxyHealth/decision.ts`) — by default the health check now **only counts/logs and never downgrades status** (a proxy is downgraded/removed only with `PROXY_AUTO_REMOVE=true`, after `PROXY_AUTO_REMOVE_AFTER` **consecutive** conclusive failures); probes are classified tri-state so an inconclusive result (our timeout, or a `5xx` from the probe target) never penalizes the proxy, and the probe timeout is raised 5s→15s. Separately, `safeResolveProxy` now **fails closed** via the existing policy: a connection whose assigned proxy is dead is blocked instead of leaking direct (`hasBlockingProxyAssignment`), honoring the explicit `proxy off` toggles and the `PROXY_FAIL_OPEN=true` opt-out. Existing proxies stuck `inactive` by the old behavior need a one-time manual re-activate (the operator owns proxy status). Regression guards: `tests/unit/proxy-health-decide-action-6246.test.ts`, `tests/unit/proxy-assigned-unavailable-6246.test.ts`.
+
+- **fix(proxy):** make "Test All" read-only and add bulk enable/disable ([#6246](https://github.com/diegosouzapw/OmniRoute/issues/6246)). Complements the core fail-closed / scheduler fix (#6296) with the two remaining reporter asks. (1) The **"Test All" button** (`POST /api/settings/proxies/auto-test`) used to flip a proxy to `inactive` on a failed reachability probe; since the egress selector excludes `inactive` proxies, a flaky probe (an unreachable `httpbin.org`, a proxy that blocks `HEAD`, or a slow paid proxy) silently disabled every proxy that failed — "Test All" is now **read-only by default** (only the operator sets a proxy active/inactive; opt back into the legacy test-and-set with `PROXY_HEALTH_AUTO_DEACTIVATE=true`). (2) Adds a **bulk enable/disable** proxies endpoint + toolbar action (`POST /api/settings/proxies/batch-activate`) so an operator can re-activate proxies in one click. Regression guard: `tests/unit/proxy-health-6246.test.ts`. (thanks @tenshiak)
+
+- **chatcore (tools): stop the default 128-tool cap from silently dropping opencode's `task`/MCP tools.** opencode (used as an MCP/agent host) sends a large tool list; when it exceeds the speculative `MAX_TOOLS_LIMIT` (128) default, `truncateToolList` did a blind `tools.slice(0, 128)`, dropping every tool past index 128 — including opencode's built-in `task` tool (subagent launch) and many MCP tools, so models routed through OmniRoute could no longer spawn subagents or reach part of their tools. The cap exists to avoid upstream `400`s for providers with real hard limits (e.g. grok-cli 200), so it is kept for those: detection of the opencode client (`isOpencodeClient` — any `x-opencode-*` header, or `opencode` in the user-agent) now only bypasses the **speculative 128 default**, never a known provider ceiling. Precedence is explicit — a proactive/detected provider limit always truncates (even for opencode); otherwise opencode forwards its full tool list; otherwise the unchanged 128 default applies to every other client. Refactors `getEffectiveToolLimit` into `getKnownToolLimit(provider) ?? DEFAULT_LIMIT` (byte-identical for existing callers) and fixes a cosmetic debug-log that reported the truncated count instead of the original. Regression guard: `tests/unit/tool-limit-detector.test.ts`.
+
+- **fix(mitm):** the macOS MITM-cert install check now matches the system keychain again. `security find-certificate -a -Z` prints the SHA-1 as a colon-less hex string, but the installed-check compared it against `getCertFingerprint()`'s colon-separated form, so the substring match never hit — the cert was reported as not-installed and re-prompted for the sudo install on every run. Fingerprints are now normalized (colons stripped, upper-cased) on both sides via the extracted `macCertOutputHasFingerprint` helper. Regression guard: `tests/unit/mitm-cert-mac-fingerprint.test.ts`. ([#6204](https://github.com/diegosouzapw/OmniRoute/pull/6204), closes [#6134](https://github.com/diegosouzapw/OmniRoute/issues/6134) — thanks @rianonehub)
+
+- **fix(api):** `/v1/messages/count_tokens` now counts `tool_use`, `tool_result` and `thinking` content blocks (and array-form `system` prompts) in the local-estimation path, instead of only `text`. Real agentic conversations keep ~95% of their tokens inside tool results; the previous estimate returned near-zero for them, which silently broke Claude Code's auto-compaction (context grew past the window with no compaction until the upstream API rejected the request). The real provider-side count path is unchanged. Regression guard: `tests/unit/messages-count-tokens-route.test.ts`. ([#6221](https://github.com/diegosouzapw/OmniRoute/pull/6221) — thanks @luweiCN)
+
+- **fix(antigravity):** strip a trailing assistant prefill turn for Vertex Claude models to avoid upstream 400s ([#6114](https://github.com/diegosouzapw/OmniRoute/pull/6114)). Regression guard: `tests/unit/antigravity-claude-prefill-strip.test.ts`. (thanks @anki1kr)
+
+- **fix(security):** the mutable cloud-agent routes (`/api/cloud/credentials/update`, `/api/cloud/models/alias`) now require management auth instead of being treated as public. They were classified as public API routes, so a request without management credentials could update stored cloud-agent credentials and model aliases. They are removed from the public-route set, classified as management routes in the authz pipeline, and gated by `requireManagementAuth`; cloud **read**/auth routes stay public. Regression guards: `tests/unit/cloud-write-auth.test.ts`, `tests/unit/authz/classify.test.ts`, `tests/unit/public-api-routes.test.ts`. ([#6233](https://github.com/diegosouzapw/OmniRoute/pull/6233) — thanks @vittoroliveira-dev)
+
+- **refactor(dashboard):** extract the onboarding-wizard "Open provider details" link target into a pure, unit-tested `buildProviderDetailsHref(connection)` helper. The wizard already routes by `connection.id` (the node UUID) rather than the provider category slug (#6144/#6145); this hardens that behavior behind a tested helper that guards a missing id/connection. Regression guard: `tests/unit/provider-onboarding-href.test.ts`. ([#6166](https://github.com/diegosouzapw/OmniRoute/pull/6166) — thanks @KooshaPari)
+
+- **fix(security):** the doubao synthetic device-id generator now derives its digits via an unbiased crypto-random draw (rejection sampling over `crypto.randomBytes()`) instead of a `% 10` reduction, closing a CodeQL `js/biased-cryptographic-random` finding.
+- **fix(agentSkills):** the GitHub-skills generator now resolves `outputDir` to an absolute path before writing, fixing a regression introduced by #6366 (relative-to-cwd base path) that could write generated skill files to the wrong directory.
+- **fix(security):** `/api/keys/{id}/devices` now checks the HTTP method before auth/validation, returning a `405` for non-GET/DELETE verbs instead of a misleading `401`/`500` (closes a `dast-smoke` QUERY-method finding).
+- **fix(quality):** clear the last 2 heavy quality-gate reds on the release tip (cycle pre-flight).
+- **fix(mitm):** the test suite and CI can never mutate the OS trust store — `OMNIROUTE_SKIP_SYSTEM_TRUST=1` is set globally for tests/CI so `installCert`/`uninstallCert`/`installTproxyCa` skip the privileged OS dispatch ([#6310](https://github.com/diegosouzapw/OmniRoute/pull/6310); full detail is under the [3.8.45] section below — this branch received it via the parallel-cycle sync-back).
+- **fix(api):** `POST /api/github-skills` now Zod-validates its request body; documented the new quality-gate env vars and pinned the merge-integrity GitHub Actions to a commit SHA.
+- **fix(skills):** generate the missing `omni-github-skills` registry entry and align the agent-skills catalog-count tests (follow-up to #6186).
+- **fix(quality):** clear the cycle's 11 net-new ESLint errors and make `validate-release-green` suppressions-aware.
+
+### 📝 Maintenance
+
+- **i18n(it):** add 118 missing Italian (`it`) translations (net-additive — no existing keys dropped, valid JSON), improving Italian UI coverage. ([#6212](https://github.com/diegosouzapw/OmniRoute/pull/6212) — thanks @serverless83)
+- **chore(providers):** remove deprecated **MiMo V2** model entries from the catalogs (xiaomi-mimo, opencode-go, zenmux-free, audio TTS) — the upstream V2 line is superseded by MiMo V2.5; drops `mimo-v2-tts`, `mimo-v2-pro`, `mimo-v2-omni`, `mimo-v2-flash`, `mimo-v2-flash-free` and realigns the provider-catalog tests. ([#6248](https://github.com/diegosouzapw/OmniRoute/pull/6248) — thanks @backryun)
+- **chore(release):** ~50 commits on this branch are v3.8.45 pre-flight/hardening fixes and CI-perf work that landed here via the parallel-cycle sync-back (`sync-next-cycle.mjs`, Hard Rule #21) **after** the `v3.8.45` git tag was cut, and are already fully documented under the **[3.8.45]** section below — listed here only so the per-cycle commit-coverage check (`npm run release:uncovered`) doesn't flag them as gaps. Provider/catalog/UX/backend fixes: #6041, #6078, #6108, #6135, #6148, #6149, #6154, #6158, #6161, #6162, #6163, #6164, #6165, #6170, #6177, #6178, #6181, #6186, #6187, #6191, #6193, #6194, #6195, #6200, #6205, #6208, #6209, #6211, #6213, #6223, #6224, #6225, #6226, #6227, #6228, #6229, #6230, #6235, #6291, #6292. CI/release-pipeline work: #6167, #6203, #6214, #6215, #6218, #6273, #6275, #6283, #6284, #6285, #6300, #6305.
+- **chore(release):** additional zero-ref release-cycle plumbing on this branch, kept out of `release:uncovered` on purpose (no `#N` in the commit subject to cite): opening the v3.8.46 cycle, opening/closing the v3.8.45 cycle, the finalized [3.8.45] CHANGELOG i18n sync-back to 42 mirrors, the v3.8.45 cognitive/cyclomatic and file-size drift rebaselines, ESLint stale-suppression pruning (4,273 → 4,233), and clearing test-masking/docs-all pre-flight reds for v3.8.45.
+
+### ⚡ Performance & Infrastructure
+
+- **perf(release-green):** the pre-flight validator (`scripts/quality/validate-release-green.mjs`) now runs its 4 slow suites (unit / vitest / integration / pack-artifact) **concurrently** via `Promise.all` — pre-flight wall time drops from ~the sum of the suites to ~the slowest one (~30min saved per round; Phase 0 was the nº1 bottleneck of the v3.8.45 release benchmark, 2h54 of 6h34 e2e). Guard: `tests/unit/validate-release-green.test.ts` ("runs the slow suites CONCURRENTLY"). ([#6319](https://github.com/diegosouzapw/OmniRoute/pull/6319))
+- **fix(ci):** `scripts/release/sync-next-cycle.mjs` — two defects found live in its first production run (v3.8.45 Phase 5): (1) the `git()` helper's default 1 MiB `maxBuffer` crashed with `ENOBUFS` on `git show origin/main:CHANGELOG.md` (the CHANGELOG alone is >1 MiB) — widened to 64 MiB; (2) the i18n resync only propagated the `[NEXT]` (TBD) section, leaving the just-shipped finalized section as "— TBD" in all 42 mirrors — it now also syncs `[prevVersion]` bounded by the heading below it (new exported pure helper `versionAfter`). Guards: +5 tests in `tests/unit/sync-next-cycle.test.ts` (8/8). ([#6327](https://github.com/diegosouzapw/OmniRoute/pull/6327))
+- **test(ci):** concurrency-sensitive flaky tests are **quarantined into a serial pass** (`tests/unit/serial/`, `--test-concurrency=1`, appended to every unit runner incl. sharded variants — the serial pass is sharded too so concurrent shard jobs never self-collide). Initial set: `glm-coding-plan-monthly-3580`, `quota-division-blocks`, `provider-health-autopilot`, `combo-health-autopilot` — the class behind the ~28min CI wedges/re-runs (two live 1h+ wedges cancelled during this PR's own validation). Discovery + TIA gates track the new glob; systemic root cause (async logger writing after teardown) tracked in [#6360](https://github.com/diegosouzapw/OmniRoute/issues/6360). Guard: `tests/unit/test-serial-quarantine.test.ts` (4). ([#6347](https://github.com/diegosouzapw/OmniRoute/pull/6347))
+
+### 🙌 Contributors
+
+Thanks to everyone whose work landed in v3.8.46:
+
+| Contributor                                                    | PRs / Issues           |
+| -------------------------------------------------------------- | ---------------------- |
+| [@2220258345](https://github.com/2220258345)                   | direct commit / report |
+| [@abdofallah](https://github.com/abdofallah)                   | direct commit / report |
+| [@adentdk](https://github.com/adentdk)                         | direct commit / report |
+| [@aleksesipenko](https://github.com/aleksesipenko)             | direct commit / report |
+| [@anki1kr](https://github.com/anki1kr)                         | direct commit / report |
+| [@anungma](https://github.com/anungma)                         | direct commit / report |
+| [@arpicato](https://github.com/arpicato)                       | direct commit / report |
+| [@backryun](https://github.com/backryun)                       | #6248                  |
+| [@binsarjr](https://github.com/binsarjr)                       | direct commit / report |
+| [@brightfiscalband](https://github.com/brightfiscalband)       | direct commit / report |
+| [@chirag127](https://github.com/chirag127)                     | #6501, #6506           |
+| [@developerjillur](https://github.com/developerjillur)         | direct commit / report |
+| [@dilneiss](https://github.com/dilneiss)                       | #6499                  |
+| [@dtybnrj](https://github.com/dtybnrj)                         | direct commit / report |
+| [@Forcerecon](https://github.com/Forcerecon)                   | direct commit / report |
+| [@hao3039032](https://github.com/hao3039032)                   | direct commit / report |
+| [@Iammilansoni](https://github.com/Iammilansoni)               | #6150, #6245           |
+| [@jmengit](https://github.com/jmengit)                         | direct commit / report |
+| [@jordansilly77-stack](https://github.com/jordansilly77-stack) | direct commit / report |
+| [@JxnLexn](https://github.com/JxnLexn)                         | direct commit / report |
+| [@KooshaPari](https://github.com/KooshaPari)                   | #6166                  |
+| [@loopyd](https://github.com/loopyd)                           | direct commit / report |
+| [@luweiCN](https://github.com/luweiCN)                         | #6221                  |
+| [@makcimbx](https://github.com/makcimbx)                       | direct commit / report |
+| [@muflifadla38](https://github.com/muflifadla38)               | direct commit / report |
+| [@newnol](https://github.com/newnol)                           | direct commit / report |
+| [@ofekbetzalel](https://github.com/ofekbetzalel)               | direct commit / report |
+| [@ohahe52-dot](https://github.com/ohahe52-dot)                 | direct commit / report |
+| [@phidinhmanh](https://github.com/phidinhmanh)                 | direct commit / report |
+| [@powellnorma](https://github.com/powellnorma)                 | direct commit / report |
+| [@qpeyba](https://github.com/qpeyba)                           | direct commit / report |
+| [@RaviTharuma](https://github.com/RaviTharuma)                 | direct commit / report |
+| [@RCrushMe](https://github.com/RCrushMe)                       | direct commit / report |
+| [@rianonehub](https://github.com/rianonehub)                   | #6134, #6204           |
+| [@serverless83](https://github.com/serverless83)               | #6212                  |
+| [@swingtempo](https://github.com/swingtempo)                   | direct commit / report |
+| [@tenshiak](https://github.com/tenshiak)                       | direct commit / report |
+| [@ThongAccount](https://github.com/ThongAccount)               | direct commit / report |
+| [@UnrealAryan](https://github.com/UnrealAryan)                 | direct commit / report |
+| [@vinayakkulkarni](https://github.com/vinayakkulkarni)         | direct commit / report |
+| [@vittoroliveira-dev](https://github.com/vittoroliveira-dev)   | #6233                  |
+| [@warelik](https://github.com/warelik)                         | direct commit / report |
+| [@xxy9468615](https://github.com/xxy9468615)                   | direct commit / report |
+| [@xz-dev](https://github.com/xz-dev)                           | direct commit / report |
+| [@yanpaing007](https://github.com/yanpaing007)                 | direct commit / report |
+| [@diegosouzapw](https://github.com/diegosouzapw)               | maintainer             |
+
+---
+
 ## [3.8.45] — 2026-07-06
 
 ### ✨ New Features

docs/i18n/fa/CHANGELOG.md

@@ -6,6 +6,213 @@
 
 ## [3.8.31] — 2026-06-20
 
+## [3.8.46] — TBD
+
+### ✨ New Features
+
+- **feat(sse):** **hide paid-only models from `auto/*` routing** when `hidePaidModels` is on ([#6512](https://github.com/diegosouzapw/OmniRoute/issues/6512)) — follow-up to #6328/#6495. PR #6495 hid paid-only models from the `GET /v1/models` listing, but `auto/*` combos (`auto/best-coding`, `auto/glm`, …) could still pick a paid-only backend into their candidate pool → a 402/403 at request time. `createVirtualAutoCombo` now filters the candidate pool through the new pure `open-sse/services/autoCombo/paidModelFilter.ts` (`filterPaidOnlyCandidates`), applying the same free-model predicate #6495 uses in `catalog.ts` (`providerHasFreeModels(provider) && isFreeModel(provider, {id})`) whenever `settings.hidePaidModels === true`. Applied before the category/tier/family narrowing, so it covers every `auto/*` combo; an all-paid pool degrades to the existing graceful empty-pool path. **Opt-in — default OFF leaves the pool unchanged** (identity). Regression guard: `tests/unit/autoCombo/paid-model-filter-6512.test.ts` (4, incl. the default-off identity guard).
+- **feat(sse):** **provider-family auto combos** — `auto/glm`, `auto/minimax`, `auto/mimo`, `auto/zai`, `auto/gemma`, `auto/llama`, `auto/gemini` ([#6453](https://github.com/diegosouzapw/OmniRoute/issues/6453)) — new routable ids that materialize an on-demand virtual combo spanning whatever installed backends currently expose that model family, degrading gracefully as backends rotate. A new pure `open-sse/services/autoCombo/modelFamily.ts` (`detectModelFamily`) classifies by model-id prefix for six families; `zai` is instead resolved by provider id (z.ai's hosted API serves the same `glm-*` model ids as every other GLM backend, so `auto/zai` means "route to my z.ai backend specifically" vs `auto/glm`'s "any connected GLM backend"). Reuses the existing `createVirtualAutoCombo` on-demand materialization path (no DB writes) and the `/v1/models` catalog advertising loop. Regression guard: `tests/unit/autoCombo/provider-family-combos.test.ts` (11).
+- **feat(proxy):** native **proxy-pool round-robin / egress IP rotation** ([#6365](https://github.com/diegosouzapw/OmniRoute/issues/6365)) — a scope (global / provider / account) can now hold **multiple** proxies as a pool with a rotation strategy, so outbound requests cycle their egress IP instead of pinning one proxy per scope. Migration `117_proxy_pool_rotation.sql` lifts the `UNIQUE(scope, scope_id)` constraint (rebuild via the canonical rename/copy/drop; existing single assignments become 1-element pools) and adds a `proxy_scope_rotation` companion table holding the per-scope strategy + a persisted monotonic round-robin cursor. Strategies: `round-robin` (default, monotonic cursor — never `Math.random`), `random`, and `sticky-per-N-min`. Resolution (`resolveProxyForScopeFromRegistry` / `resolveProxyForConnectionFromRegistry`) now fetches the alive, position-ordered candidate set (unchanged `PROXY_ALIVE_PREDICATE`) and applies the strategy; an empty / all-dead pool still returns `null` — the #6246 fail-closed guard is **untouched** (never falls through to direct egress). Backend + DB only; dashboard pool-builder UI is a follow-up. Regression guard: `tests/unit/proxy-pool-rotation-6365.test.ts` (8, incl. fail-closed + backward-compat).
+- **feat(providers):** end-to-end **tool/function calling on the native Gemini `/v1beta` endpoint** ([#6222](https://github.com/diegosouzapw/OmniRoute/issues/6222)) — both directions of the Gemini↔OpenAI conversion now preserve tool data (previously silently dropped). Request side: `convertGeminiToInternal` (extracted to its own testable module) maps `tools[].functionDeclarations` → OpenAI `tools`, prior `functionCall` parts → assistant `tool_calls`, and `functionResponse` parts → `tool`-role messages. Response side: `convertOpenAIResponseToGemini` emits `parts[].functionCall {name,args}` from `message.tool_calls`, and the streaming `openAIChunkToGeminiChunk` accumulates fragmented `tool_calls` deltas by index into complete `functionCall` parts. The non-Gemini client paths (Claude, OpenAI-Responses) already preserved tool calls — this closes the gap specific to the native Gemini surface. Regression guard: `tests/unit/v1beta-gemini-tool-calling-6222.test.ts` (6, incl. a streaming SSE round-trip).
+- **feat(providers):** copilot-m365-web **enterprise / work tier** support ([#6334](https://github.com/diegosouzapw/OmniRoute/issues/6334)) — mirrors the EDU-tier pattern (#6210): `M365ConnectionParams` gains an `agent` field, a new opt-in `M365_ENTERPRISE_OVERRIDES` preset (`agent=work`, `scenario=officeweb`, `licenseType=Premium`) applies via `providerSpecificData.tier="enterprise"` (alias `"work"`), and `agent` is also overridable directly via `providerSpecificData.agent`. `buildWsUrl` was hardcoding `agent="web"` (the one enterprise-distinguishing param with no override path), so a Premium work account handshook then returned an empty stream. The individual and EDU paths are untouched. Kilo's dup flag vs #6210 (EDU tier) was a false positive — different tier. Regression guard: `tests/unit/copilot-m365-enterprise-6334.test.ts` (7). End-to-end confirmation on a real Premium work account is a live-VPS validation follow-up (Hard Rule #18). (thanks @Forcerecon)
+- **feat(api):** standardized, provider-agnostic **`effort` + `thinking` request params** ([#6241](https://github.com/diegosouzapw/OmniRoute/issues/6241)) — a thin standardization layer over the existing mature per-provider reasoning plumbing (no provider mapper touched). `providerChatCompletionSchema` gains a canonical `effort` (reusing the shared `none/low/medium/high/xhigh` vocabulary — the UI tiers `extra`/`max` collapse onto `xhigh`) and a boolean `thinking`. A pure `normalizeReasoningRequest` (wired once in `src/sse/handlers/chat.ts`, before any reasoning field is read) folds them onto the fields the translators already consume (`reasoning_effort` / `reasoning.effort` / `thinking`), so they fan out to Anthropic / Gemini / xAI / Responses — an explicit client `reasoning_effort` / object-shaped `thinking` always wins (backward-compatible). `/models` additively exposes `supportsThinking` + `effort_tiers` so the frontend can render the toggles (UI component is a follow-up). Regression guard: `tests/unit/effort-thinking-standardization-6241.test.ts` (12). (thanks @Iammilansoni, @shabeer)
+- **feat(combo):** new **`pipeline` (sequential) combo strategy** ([#6297](https://github.com/diegosouzapw/OmniRoute/issues/6297)) — the 18th routing strategy runs targets **in order**, threading each step's output into the next step's input, with an optional per-step `prompt` (system instruction); only the final step's response is returned. Distinct from `fusion` (parallel fan-out + judge). Implemented as a self-contained `open-sse/services/pipeline.ts` (sibling to `fusion.ts`), dispatched from `combo.ts`; the step list reuses `combo.models` order and reads an optional `prompt` off each target (backward-compatible — ignored by every other strategy). Intermediate steps run non-streaming with tools stripped (complete prose to thread forward); the final step keeps the client's `stream` flag + tools. A failing/empty/unparseable intermediate step fails the whole pipeline explicitly via a sanitized error (never silently swallowed). Kilo's dup flag vs #563 was a false positive (that's model→chain selection; this is a sequential chain). Regression guard: `tests/unit/combo-pipeline-strategy.test.ts` (5). (thanks @ofekbetzalel)
+- **feat(ci):** `check:test-masking` now flags **inline-reimplemented prod conditions** ([#6348](https://github.com/diegosouzapw/OmniRoute/issues/6348)) — a new **report-only** subcheck (v2, 6A.10 family) catches the wrong-shape contract test: a test that recomputes the condition under test inline instead of importing/exercising the real function (the #6216 class, where `=== 500` → `>= 500` stayed green because the test re-implemented the branch). For each added/modified test file it warns when the file textually duplicates a ≥3-token conditional from a production file touched in the same PR **and** does not import the symbol/module owning it, via a pure, fixture-tested `findReimplementedConditions()` with an allowlist mirroring `assertReductionAllowlist`. Report-only for now (does not fail the gate) — to be promoted to blocking after a triage cycle. Regression guard: `tests/unit/check-test-masking.test.ts` (45).
+- **feat(sse):** per-connection routing override (native vs CLIProxyAPI) ([#6339](https://github.com/diegosouzapw/OmniRoute/issues/6339)) — the previously-dead `isCliproxyapiDeepModeEnabled` helper is now wired into `resolveExecutorWithProxy`: a single connection can opt itself into the CLIProxyAPI passthrough executor via `providerSpecificData.cliproxyapiMode="claude-native"`, with precedence **connection override > provider `upstream_proxy_config` mode > default**. `resolveExecutorWithProxy` now receives the resolved connection's `providerSpecificData` (threaded from `chatCore.ts`), so one connection can deep-route while the provider's other connections stay native — no DB schema change (the toggle rides in `providerSpecificData`). Also resolves the same-provider mixing ask in #6340. Regression guard: `tests/unit/chatcore-executor-proxy.test.ts` (9). (thanks @RaviTharuma)
+- **feat(dashboard):** "Add session cookie" modal now shows a prominent **"Open ‹host› →"** link to the provider's own site ([#6268](https://github.com/diegosouzapw/OmniRoute/issues/6268)) — every `-web` cookie-session provider (chatgpt-web, claude-web, gemini-web, kimi-web, lmarena, qwen-web, m365-copilot-web, …) renders a one-click external link (opening the provider's login/home page in a new tab) so operators no longer tab away to retype the URL mid-setup. The host resolves from a pure, unit-tested `resolveWebProviderHost()` (prefers `WEB_COOKIE_PROVIDERS[id].website`, falls back to the registry `baseUrl` origin); non-web providers render exactly as before. Kilo's dup flag vs #6265 (modal-too-small-on-1080p) was a false positive — distinct concern. Regression guard: `tests/unit/resolve-web-provider-host.test.ts` (5). (thanks @chirag127)
+- **feat(providers):** add **DigitalOcean** AI (serverless inference) as an OpenAI-compatible API-key provider ([#6373](https://github.com/diegosouzapw/OmniRoute/pull/6373)) — base `https://inference.do-ai.run/v1`, wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/digitalocean/`, `src/shared/constants/providers/apikey/inference-hosts.ts`). Regression guard: `tests/unit/digitalocean-provider.test.ts`. (thanks @newnol)
+- **feat(providers):** add **Huancheng Public API** (`hcnsec`) as an OpenAI-compatible regional provider ([#6410](https://github.com/diegosouzapw/OmniRoute/pull/6410)) — Xinjiang Huancheng Cybersecurity's public LLM platform (base `https://api.hcnsec.cn/v1`, free credits via daily check-ins), wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/hcnsec/`, `src/shared/constants/providers/apikey/regional.ts`). Regression guard: `tests/unit/hcnsec-provider.test.ts`. (thanks @UnrealAryan)
+- **feat(dashboard):** the web-session credential guide now shows an **"Open {host}" link** ([#6316](https://github.com/diegosouzapw/OmniRoute/pull/6316)) to the provider's sign-in site (derived from the provider `website` via `getProviderWebsiteHost`), so you can jump straight to the page where the cookie/session must be captured. Regression guard: `tests/unit/web-session-provider-link-6316.test.ts`. (thanks @jordansilly77-stack)
+- **feat(cerebras):** add the **Gemma 4 31B** model (`gemma-4-31b`) to the Cerebras registry + pricing table ([#6331](https://github.com/diegosouzapw/OmniRoute/pull/6331)). Regression guard extends `tests/unit/t28-model-catalog-updates.test.ts`. (thanks @backryun)
+- **feat(providers):** add **Yuanbao (web)** as a cookie-session provider ([#6196](https://github.com/diegosouzapw/OmniRoute/issues/6196)) — `yuanbao-web` (Tencent Yuanbao, `yuanbao.tencent.com`) with cookie-only auth (`hy_user`/`hy_token` + public agent id), SSE→OpenAI translation incl. `reasoning_content`, exposing DeepSeek V3/R1 + Hunyuan / Hunyuan-T1. Regression guard: `tests/unit/providers-yuanbao-web.test.ts`. `together-web` was **deferred** (no verifiable web-session endpoint — needs a captured request) and `huggingchat-web` **dropped** (the existing `huggingchat` already is a web-cookie provider). (thanks @chirag127)
+- **feat(providers):** route the built-in **agentrouter** through the dynamic Claude-Code wire image ([#6056](https://github.com/diegosouzapw/OmniRoute/issues/6056)) — a small static allow-set (`CC_WIRE_IMAGE_BUILTINS` in `open-sse/services/ccWireImageBuiltins.ts`), consulted by `isClaudeCodeCompatible` / `isClaudeCodeCompatibleProvider` / `applyFingerprint`, makes agentrouter adopt the CC wire-image headers + fingerprint **while guarding the CC baseUrl/auth branches** so it keeps its own registry `baseUrl` and `x-api-key` auth. Regression guard: `tests/unit/agentrouter-cc-wire-image.test.ts` (asserts the wire image is applied AND agentrouter's baseUrl/auth are preserved). Live WAF-acceptance against agentrouter.org is a VPS validation follow-up (Hard Rule #18).
+- **feat(providers):** **bulk-add API keys for Cloudflare Workers AI** ([#6174](https://github.com/diegosouzapw/OmniRoute/issues/6174)) — `cloudflare-ai` is removed from the bulk-add exclusion list and the bulk parser gains a 3-field `name|accountId|apiKey` mode; the bulk route now builds a **per-entry** `providerSpecificData` so each key carries its own `accountId` (fixing the previous shared-object reuse), and both the create + key-validation paths receive it. Regression guard: `tests/unit/bulk-api-key-parser-cloudflare.test.ts`. (thanks @muflifadla38)
+- **feat(dashboard):** routing/settings UX clarity ([#6147](https://github.com/diegosouzapw/OmniRoute/issues/6147)) — (1) weighted combos show the **effective routing share %** next to each weight when weights don't sum to 100 (`WeightTotalBar.tsx`); (2) the status widget's user-facing **"Cloud Sync" label is renamed** to "Remote Settings Sync" (`CloudSyncStatus.tsx`; internal ids/state untouched); (3) built-in providers gain an **opt-in advanced base-URL override** (`isBaseUrlOverrideEligibleProvider`, hidden behind an "Advanced" toggle, reusing the existing `providerSpecificData.baseUrl` persistence — not globally widened). Regression guard: `tests/unit/routing-settings-ux-6147.test.ts`.
+- **feat(combo):** add an option to **disable session stickiness**, per-combo or globally — round-robin / random combos can rotate to a different connection on every request instead of pinning a whole conversation to one connection by its first-message hash. Resolution precedence per-combo `config.disableSessionStickiness` → global `settings.disableSessionStickiness` → default `false` (preserves the #3825 prompt-cache/504 fix); gates **both** stickiness call sites in `open-sse/services/combo.ts`. Exposed as a global toggle (Combo Defaults) and a per-combo Inherit/on/off control. ([#6168](https://github.com/diegosouzapw/OmniRoute/issues/6168)) Regression guard: `tests/unit/combo-disable-session-stickiness.test.ts`. (thanks @RCrushMe)
+- **feat(docker):** add the `OMNIROUTE_NO_SUDO` env flag for root-less / user-namespaced deployments — the MITM cert-trust command path (`resolveSudoSpawn` in `src/mitm/systemCommands.ts`) now strips the leading `sudo` when the flag is truthy, in addition to the existing root / sudo-missing cases, so the Proxy Agent runs without `sudo` (the operator trusts the CA manually, e.g. via `NODE_EXTRA_CA_CERTS`). Argv-array `spawn` preserved — no shell interpolation (Hard Rule #13). ([#6122](https://github.com/diegosouzapw/OmniRoute/issues/6122)) Regression guard: `tests/unit/mitm-systemCommands-no-sudo.test.ts`. (thanks @powellnorma)
+- **feat(providers):** add **Requesty** as an OpenAI-compatible gateway provider (BYOK, base `https://router.requesty.ai/v1`, ~200 free requests/day) — wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/requesty/`, `src/shared/constants/providers/apikey/gateways.ts`). ([#6120](https://github.com/diegosouzapw/OmniRoute/issues/6120)) Regression guard: `tests/unit/requesty-provider.test.ts`. (thanks @chirag127)
+- **feat(dashboard):** add **configured-only / available-only filters** to the Free Provider Rankings page ([#6150](https://github.com/diegosouzapw/OmniRoute/issues/6150)) — hide providers you haven't configured, or whose connections are all rate-limited / out of quota, via server-side query params (`?configuredOnly` / `?availableOnly` on `GET /api/free-provider-rankings`) backed by a testable lib helper reusing the in-process connection state (no Redis). Both filters default off, so the default view is unchanged; this supersedes the earlier client-side "Configured Only" toggle (#6245) with an available-only dimension and unit-tested logic. Regression guard: `tests/unit/freeProviderRankings-filters.test.ts`.
+
+### 🔧 Bug Fixes
+
+- **fix(dashboard):** adding a second API key connection for the same provider no longer silently overwrites the first — the Add-API-key modal now derives a unique default connection name (`main`, then `main-2`, `main-3`, …) so the backend name-based upsert can't collide (#6499 — thanks @dilneiss).
+- **fix(compression):** the session-dedup engine now also deduplicates a large multi-line block repeated **within a single message** (intra-message dedup), not just across turns; the compression-preview API surfaces a `fallbackReason`, and the fusion panel reports how many models were rate-limited vs failed on total-panel failure (#6501 — thanks @chirag127).
+- **fix(compression):** a stacked-pipeline step naming an unregistered engine now surfaces a `validationErrors` entry instead of silently no-op'ing, so misconfigured pipelines are visible in the preview API (#6506 — thanks @chirag127).
+- **feat(usage):** add a **Codex reset-credit redemption** flow to the Provider Limits UI ([#6361](https://github.com/diegosouzapw/OmniRoute/pull/6361)) — a `useCodexResetCreditRedemption` hook + `/api/usage/codex-reset-credit` route + `codexResetCredits` lib let you redeem banked Codex reset credits from the quota card. Regression guard: `tests/unit/codex-reset-credits.test.ts`. (thanks @JxnLexn)
+- **feat(glm):** add **team-plan quota settings** for `glm-cn` connections ([#6351](https://github.com/diegosouzapw/OmniRoute/pull/6351)) — a dedicated `GlmTeamQuotaFields` form section (team quota id / limits) threaded through the Add/Edit connection modals, persisted via `providerSpecificData`, with the GLM usage service reading the team quota. Regression guards: `tests/unit/glm-team-quota.test.ts`, `provider-specific-data-schema.test.ts`. (thanks @hao3039032)
+- **feat(providers):** add **TinyFish** web-fetch/search support ([#6349](https://github.com/diegosouzapw/OmniRoute/pull/6349)) — a `tinyfish-fetch` executor + `/v1/web/fetch` route + MCP web-fetch tool, registered as a specialty-media provider with request-validation and a search-provider catalog entry. Regression guards: `tests/unit/executor-tinyfish-fetch.test.ts`, `web-fetch-handler.test.ts`, `mcp-web-fetch-tool.test.ts`, `provider-validation-tinyfish.test.ts`. (thanks @dtybnrj)
+- **fix(cli):** `omniroute launch-codex` now spawns `codex.cmd` through a shell on Windows (the npm `.cmd` shim is unresolvable by bare `spawn` → ENOENT), mirroring the qodercli Windows fix (#6263) ([#6312](https://github.com/diegosouzapw/OmniRoute/pull/6312)). Regression guard: `tests/unit/launch-codex-windows-spawn-6312.test.ts`. (thanks @swingtempo)
+- **fix(codex):** isolate the **Spark** quota from the shared Codex quota and stabilize the quota UI ordering / hydration so per-scope limits render consistently ([#6336](https://github.com/diegosouzapw/OmniRoute/pull/6336)). Regression guards: `tests/unit/codex-quota-selection-hydration.test.ts`, `provider-limits-ui.test.ts` + 3 more. (thanks @xz-dev)
+- **feat(api):** add a `hidePaidModels` setting that filters paid-only models out of the `/v1/models` catalog. Regression guard: `tests/unit/models-catalog-hide-paid.test.ts`. (thanks @chirag127)
+- **fix(api-manager):** the fallback model picker now preserves combos instead of dropping them when a primary model is unavailable ([#6443](https://github.com/diegosouzapw/OmniRoute/pull/6443)). Regression guard: `tests/unit/api-manager-page-static.test.ts`. (thanks @jmengit)
+- **fix(providers):** recoverable Antigravity / Cloud-Code (Gemini Code Assist) `403` responses ([#6452](https://github.com/diegosouzapw/OmniRoute/pull/6452)) are now classified as a retryable project-config error instead of a terminal account ban, so a fixable project/API-disabled 403 no longer forces a ~1-year cooldown / full OAuth reconnect. Regression guard: `tests/unit/errorclassifier-antigravity-403.test.ts`. (thanks @developerjillur)
+- **fix(mitm):** `sanitizeHeaders` now redacts `Set-Cookie` response headers so upstream session cookies never leak into logs / diagnostics ([#6451](https://github.com/diegosouzapw/OmniRoute/pull/6451)). Regression guard: `tests/unit/mitm-sanitize-headers.test.ts`. (thanks @developerjillur)
+- **fix(api):** `/api/compression/preview` now accepts `mode: "caveman"` and correctly handles stacked / zero-compression previews ([#6425](https://github.com/diegosouzapw/OmniRoute/issues/6425)). Regression guard: `tests/unit/api/compression-preview-caveman-and-stacked-6425.test.ts`. (thanks @chirag127)
+- **feat(providers):** add **Zed** hosted LLM aggregator as a native-app provider ([#6118](https://github.com/diegosouzapw/OmniRoute/pull/6118)) — OAuth sign-in via the Zed hosted flow, registered through the shared provider registry + executor. Regression guards: `tests/unit/zed-oauth-provider.test.ts`, `zed-import-utils.test.ts`, `zed-docker-detect.test.ts`, `mitm-handler-zed.test.ts`. VPS-validated via live operator login (Hard Rule #18).
+- **fix(oauth):** the Kiro SSO-cache auto-import now **preserves the IDC region** — cross-region Amazon Q / Kiro profiles imported from the SSO cache are no longer collapsed to the default region ([#6113](https://github.com/diegosouzapw/OmniRoute/pull/6113)). Regression guard: `tests/unit/kiro-auto-import-idc-2059.test.ts`. VPS-validated via live operator login (Hard Rule #18).
+- **fix(dashboard):** passthrough model aliases no longer collide when two namespaced model ids share a last segment (port from 9router#1850, [#6431](https://github.com/diegosouzapw/OmniRoute/pull/6431)). `enx/gpt-5.5` and `enx/codebuddy/gpt-5.5` both auto-generated the alias `gpt-5.5`, so the second model could never be added (the UI just alerted "alias already exists"). Aliases are now disambiguated deterministically — bare last segment when free, then parent-qualified (`codebuddy-gpt-5.5`), then a numeric suffix — while re-adding the exact same model id is still blocked. Regression guard: `tests/unit/passthrough-alias-1850.test.ts`. (thanks @arpicato)
+- **fix(translator):** preserve a Gemini `functionResponse` co-located with other parts (another `functionCall`, or trailing `text`) in the same content when translating **Gemini → OpenAI** ([#6376](https://github.com/diegosouzapw/OmniRoute/pull/6376)). `convertGeminiContent()` early-returned the tool message on the first `functionResponse` part, dropping any co-located parts; such contents are now pre-split (one tool message per `functionResponse`, emitted first, plus one message for the remaining parts). Regression guard: `tests/unit/gemini-to-openai-function-response.test.ts`. (thanks @warelik)
+- **fix(headroom):** detect a python interpreter managed by **mise / pyenv / asdf / conda** (port from 9router#2353, [#6382](https://github.com/diegosouzapw/OmniRoute/pull/6382)). Headroom's python probe (`src/lib/headroom/detect.ts`) searched a hardcoded `PATH`, but version managers expose their interpreters via shim dirs that only join `PATH` through interactive-shell activation — which the non-interactive server never runs, so a managed python (≥3.10) was invisible and Headroom reported it missing. The search path now prepends the well-known shim/bin dirs (`~/.local/share/mise/shims`, `~/.pyenv/shims`, `~/.asdf/shims`, `$CONDA_PREFIX/bin`, `~/.local/bin`, respecting `MISE_DATA_DIR`/`PYENV_ROOT`/`ASDF_DATA_DIR` when set), and a new `HEADROOM_PYTHON` env override lets operators point straight at their interpreter (mirroring `HEADROOM_URL`). Still shell-free (`execFileSync`). Regression guard: `tests/unit/headroom-detect.test.ts` (5). (thanks @loopyd)
+- **fix(executors):** strip the OpenAI-Codex/Claude-CLI `client_metadata` passthrough field for **NVIDIA** requests (port from 9router#1887, [#6411](https://github.com/diegosouzapw/OmniRoute/pull/6411)). NVIDIA's OpenAI-compatible wrapper rejects it with `400 Unsupported parameter`, the same class already handled for `cerebras`/`mistral`; `nvidia` (executor `default`) was missing from the strip allowlist so Codex/Claude-Code passthrough requests 400'd. Regression guard: `tests/unit/executor-default-strip-client-metadata.test.ts` (+nvidia case). (thanks @phidinhmanh)
+- **fix(translator):** strip the Claude-style `thinking` field for **NVIDIA `z-ai/glm-5.2`** (port from 9router#2023, [#6413](https://github.com/diegosouzapw/OmniRoute/pull/6413)). NVIDIA's OpenAI-compatible wrapper 400s on `thinking` (a Claude-format client routed here leaves a `thinking:{type:"adaptive"}`); the existing strip rule only dropped `reasoning`. Same class already handled for `minimax-m2.7`. Regression guard: `tests/unit/nvidia-minimax-thinking-strip.test.ts` (+glm-5.2 case). (thanks @phidinhmanh)
+- **fix(translator):** suppress the streamed `</think>` close marker for the **Antigravity IDE** client (port from 9router#1061, [#6415](https://github.com/diegosouzapw/OmniRoute/pull/6415)). On thinking-only turns Antigravity rendered a bare `</think>` as the sole visible content, tripping its loop-detection and wasting requests. Antigravity's UA (`vscode/<v> (Antigravity/<v>)`) is added to the marker-suppress allowlist (alongside OpenCode); Claude Code / Cursor still get the marker, and `x-omniroute-thinking-marker: on` force-restores it. Regression guard: `tests/unit/think-close-marker-suppress-5245.test.ts`. (thanks @abdofallah)
+- **fix(executors):** strip nested `reasoning_content` from messages for **Mistral** (port from 9router#1649, [#6417](https://github.com/diegosouzapw/OmniRoute/pull/6417)). Mistral's API returns `422 extra_forbidden` when an assistant message carries `reasoning_content` (replayed thinking from a prior turn, e.g. via the Codex `/responses` path); the generic top-level 400 field-downgrade retry never covered the nested per-message field. `DefaultExecutor` now strips it for provider `mistral` only, so DeepSeek (which requires replayed `reasoning_content`) is unaffected. Regression guard: `tests/unit/mistral-strip-reasoning-content-1649.test.ts`. (thanks @xxy9468615)
+- **fix(executors):** strip the `client_metadata` passthrough field on the **OpenCode** path (port from 9router#1442, [#6418](https://github.com/diegosouzapw/OmniRoute/pull/6418)). OpenCode upstreams (e.g. `kimi-k2.6` via opencode-go) reject it with `400 "Extra inputs are not permitted, field: 'client_metadata'"`; the DefaultExecutor strip only covered cerebras/mistral and `OpencodeExecutor` extends `BaseExecutor` directly, so nothing removed it there. Regression guard: `tests/unit/opencode-strip-client-metadata-1442.test.ts`. (thanks @yanpaing007)
+- **fix(executors):** inject the `reasoning_content` echo for the native **Moonshot Kimi** provider (port from 9router#1480, [#6419](https://github.com/diegosouzapw/OmniRoute/pull/6419)). Kimi (executor `default`) is a thinking-mode upstream that 400s with "reasoning_content must be passed back" when a prior assistant turn lacks it; the placeholder injection was only wired into the OpenCode meta-provider, so direct multi-turn Kimi conversations failed. Scoped to `kimi` (gateway-served models matching the thinking-model name pattern are unaffected). Regression guard: `tests/unit/kimi-native-reasoning-injected-1480.test.ts`. (thanks @2220258345)
+- **fix(executors):** recover from a strict gateway's `context_management: Extra inputs are not permitted` 400 (port from 9router#1468, [#6420](https://github.com/diegosouzapw/OmniRoute/pull/6420)). **Claude Code** always sends a top-level `context_management` field; strict anthropic-compatible gateways reject it. The dedicated context-editing 400-fallback only fired when OmniRoute's own `contextEditing` feature was enabled (default off), so a client-sent field passed through untouched and 400'd. `context_management` is now in the generic reactive field-strip list, so it's stripped-and-retried once regardless of the feature flag (with correct request re-signing for claude-compatible relays). Regression guard: `tests/unit/provider-field-strips.test.ts`. (thanks @ohahe52-dot)
+- **fix(network):** enable **RFC 8305 Happy Eyeballs** (`autoSelectFamily`) on the direct-egress undici dispatcher (port from 9router#1237, [#6423](https://github.com/diegosouzapw/OmniRoute/pull/6423)). When DNS returns both IPv6 (AAAA) and IPv4 (A) and the IPv6 route is broken (e.g. a NAT64 `64:ff9b::` prefix without routing), undici tried IPv6 first and hung until `ETIMEDOUT` (then a 502 + account lockout), even though `curl` reached the same host. The direct dispatcher now races both families and uses whichever connects first. Proxy paths pin family via `proxyTls` and are unaffected. Regression guard: `tests/unit/direct-dispatcher-pipelining-4580.test.ts`. (thanks @adentdk)
+- **fix(combo):** round-robin now advances the rotation pointer past the model that **actually served**, not the eagerly-scheduled one (port from 9router#948, [#6428](https://github.com/diegosouzapw/OmniRoute/pull/6428)). With `stickyLimit: 1` (true round-robin), when the scheduled model failed and a _different_ model served via fallback, the counter had already advanced +1 from the scheduled index — so the next request reused the fallback-served model, degrading round-robin into hot-spotting on whichever model was healthy. The pointer now advances to the served index + 1 (mirroring the sticky-limit>1 path). Session-stickiness (#3825) and distribution are preserved. Regression guard: `tests/unit/combo-rr-fallback-advance-948.test.ts`. (thanks @binsarjr)
+- **fix(sse):** a non-string `model` field is now rejected with a `400` before the resolver, instead of crashing downstream `.toLowerCase()`/`.split()` calls into an empty-body `500` that escapes the error sanitizer ([#6407](https://github.com/diegosouzapw/OmniRoute/issues/6407)). Regression guard: `tests/unit/chat-non-string-model-6407.test.ts`. (thanks @chirag127)
+- **fix(api):** unknown `/api/*` routes now return a JSON `404` (instead of the dashboard HTML shell) and scalar chat params (`model`/`temperature`/etc.) are validated **before** the provider lookup so malformed requests fail fast with a clear `400` ([#6424](https://github.com/diegosouzapw/OmniRoute/issues/6424), [#6412](https://github.com/diegosouzapw/OmniRoute/issues/6412)). Regression guards: `tests/unit/api/api-catchall-json-404.test.ts`, `tests/unit/chat-early-schema-validation-6412.test.ts`. (thanks @chirag127)
+- **fix(api):** `/v1/chat/completions` now rejects a non-JSON `Content-Type` with a `400` before parsing the body ([#6414](https://github.com/diegosouzapw/OmniRoute/issues/6414)). Regression guard: `tests/unit/v1-chat-completions-content-type-6414.test.ts`. (thanks @chirag127)
+- **fix(api):** the `X-OmniRoute-Compression` response header is now echoed on `/v1/chat/completions` and `/v1/completions` ([#6422](https://github.com/diegosouzapw/OmniRoute/issues/6422)). Regression guard: `tests/unit/compression-header-echo-6422.test.ts`. (thanks @chirag127)
+- **fix(api):** concurrent `GET /v1/models` requests are coalesced into a single catalog build ([#6408](https://github.com/diegosouzapw/OmniRoute/issues/6408)). Regression guard: `tests/unit/v1-models-concurrent-6408.test.ts`. (thanks @chirag127)
+- **fix(api):** `/v1/completions` now echoes the requested `body.model` in its JSON + streamed responses ([#6429](https://github.com/diegosouzapw/OmniRoute/pull/6429)). Regression guard: `tests/unit/completions-body-model-echo.test.ts`. (thanks @chirag127)
+- **fix(api):** env-var master keys now see the full `/v1/models` catalog ([#6406](https://github.com/diegosouzapw/OmniRoute/issues/6406)). Regression guard: `tests/unit/models-catalog-envkey-6406.test.ts`. (thanks @chirag127)
+- **fix(api):** non-streaming `/v1/completions` responses now echo `body.model` aligned with the `X-OmniRoute-Model` header ([#6426](https://github.com/diegosouzapw/OmniRoute/issues/6426)). Regression guard: `tests/unit/v1-completions-model-header-match-6426.test.ts`. (thanks @chirag127)
+- **fix(api):** unknown `/v1/*` routes now return a JSON `404 not_found` instead of the Next.js dashboard HTML shell ([#6405](https://github.com/diegosouzapw/OmniRoute/issues/6405)). Regression guard: `tests/unit/api/v1-catchall-json-404.test.ts`. (thanks @chirag127)
+- **fix(api):** the per-connection provider models route now degrades to the shipped catalog when a provider's `/models` endpoint answers with a redirect ([#6267](https://github.com/diegosouzapw/OmniRoute/issues/6267)) — a `qwen-web` import failed with a raw `Redirect blocked … (307)` 503. `safeOutboundFetch` throws `REDIRECT_BLOCKED` on the 307, `getSafeOutboundFetchErrorStatus` maps it to 503, and `buildDiscoveryErrorFallbackResponse` treated every 503 as a hard error — so the non-empty `getModelsByProviderId("qwen-web")` catalog was never surfaced. A models-endpoint redirect is not a fixable-config error (unlike `URL_GUARD_BLOCKED`/`INVALID_URL`, which stay hard errors), so it now falls back to the local/cached catalog before the 503 short-circuit. General fix — covers any config-driven provider that 307s. Regression guard: `tests/unit/provider-models-qwen-web-redirect-6267.test.ts`. (thanks @chirag127)
+- **fix(api):** the per-connection provider models route (MCP `list_models_catalog` + the dashboard import view) now merges USER-ADDED custom models into its response ([#6247](https://github.com/diegosouzapw/OmniRoute/issues/6247)) — custom models live in the `key_value` namespace `customModels`, which the live REST `/api/v1/models` already merges, but `src/app/api/providers/[id]/models/route.ts` never read `getCustomModels`, so custom models were dropped on both the discovery-success and local_catalog paths. They are now folded into the returned model list (deduped by id, stamped `owned_by: provider`), fixing MCP + the dashboard import view in one place. Regression guard: `tests/unit/provider-models-custom-merge-6247.test.ts`. (thanks @RCrushMe)
+
+- **fix(providers):** GitLab Duo tool-calling follow-up turns no longer fail upstream with `422 {"detail":"Validation error"}` (tokens 0/0, rejected pre-inference). The #6234 tool-result-feedback fix serialized the **entire** multi-turn conversation into GitLab's single-file `code_suggestions` (`small_file`) generation endpoint — folded history that turn-N sent as an oversized `current_file.content_above_cursor` **and** duplicated verbatim into `user_instruction`, tripping the AI-Gateway's `small_file` validation guard. The executor now **bounds** that prompt: it keeps system + latest user message + the most-recent tool round (dropping older turns), caps oversized tool results, and stops duplicating the full prompt into `user_instruction` (which now carries only the short latest user message) — while still feeding the most-recent tool result back so the agent continues (`open-sse/executors/gitlab.ts`, [#6220](https://github.com/diegosouzapw/OmniRoute/issues/6220)). The unit test covers the bounding logic; the upstream 422→200 clearing is VPS-only (Hard Rule #18). Regression guard: `tests/unit/gitlab-tool-exchange-bounded-6220.test.ts`.
+
+- **fix(i18n):** the provider-detail (`/dashboard/providers/[id]`) connection-status filter labels no longer render as `__MISSING__:All` / `__MISSING__:Active` / `__MISSING__:Error` / `__MISSING__:Banned` / `__MISSING__:CreditsExhausted` in non-English locales (notably pt-BR) ([#6290](https://github.com/diegosouzapw/OmniRoute/issues/6290)). Root cause was **not** the namespace mismatch the issue guessed — the `providers.filter*` keys resolve correctly in `en.json`; the debt lived in the locale mirrors (`src/i18n/messages/*.json`), where these five keys carried the `__MISSING__:` sync sentinel in ~15 locales and were absent entirely in ~26 others, so next-intl found the key and echoed the sentinel verbatim. All 40 non-English/-Chinese mirrors now ship real translations for the five `providers.filter*` labels. Regression guard: `tests/unit/i18n-provider-filter-keys-6290.test.ts`. (thanks @diegosouzapw)
+
+- **fix(providers):** the `copilot-m365-web` streaming executor now emits `debug`-level WebSocket diagnostics ([#6210](https://github.com/diegosouzapw/OmniRoute/issues/6210)) — the outbound WS URL (with the `access_token` **redacted** via `redactWsUrl()`), handshake success/failure, and each received SignalR frame's `type`/`target`. Previously the streaming path logged nothing, so an empty `content:null` response (the M365 Education / Starter tier symptom fixed in #6234) was undiagnosable even at `APP_LOG_LEVEL=debug`. The change is debug-level and side-effect-free — it does not alter streaming behavior or the frame parser, and the token never reaches the logs. Regression guard: `tests/unit/copilot-m365-web-logging-6210.test.ts` (thanks @qpeyba)
+
+- **fix(resilience):** a round-robin combo no longer returns `503 all upstream accounts are unavailable` when a compatibility-rejected target is actually healthy ([#6238](https://github.com/diegosouzapw/OmniRoute/issues/6238)). `filterTargetsByRequestCompatibility` drops request-incompatible targets (tool/vision/structured-output unsupported, or below the required context window) **before** any availability check runs, and its `compatible.length === 0` safety net only fired when _all_ targets were filtered — not when the kept targets later all turned out runtime-unavailable (circuit-open / cooldown / no credentials). So a combo could 503 while a compat-rejected-but-healthy provider sat unused. `handleRoundRobinCombo` now keeps the compat-rejected set and, when every compat-kept target was skipped without a single real attempt, probes those rejected targets as a **last-resort fallback tier** (via the new pure `open-sse/services/combo/comboCompatFallback.ts`) before crystallizing the 503. Regression guard: `tests/unit/combo-roundrobin-compat-fallback-6238.test.ts`. (thanks @ThongAccount)
+
+- **fix(startup):** best-effort self-heal for a corrupted Turbopack dev cache on Windows ([#6289](https://github.com/diegosouzapw/OmniRoute/issues/6289)). On Windows, `pnpm dev` can fail at startup when Turbopack `mmap`s a persistent-cache SST file and the OS refuses the mapping (`os error 1455` — "paging file too small"), which Turbopack surfaces as a misleading `Module not found: Can't resolve '@/shared/utils/machine'`. This is a **known upstream Turbopack cache-corruption bug — not our code**. The dev launcher (`scripts/dev/run-next.mjs`) now wraps `nextApp.prepare()` and, when it rejects with that signature (`isTurbopackCacheCorruption` in the new `scripts/dev/turbopackCacheHeal.mjs`), purges `.build/next/**/cache/turbopack` and retries **once** with a clear log. **Caveat — best-effort only:** the corruption often surfaces as a runtime overlay rather than a `prepare()` rejection, so this cannot always intercept it; the reliable remedy remains manually deleting the Turbopack cache dir. Regression guard: `tests/unit/turbopack-cache-heal-6289.test.ts`. (thanks @chirag127)
+
+- **fix(providers):** qodercli PAT auth no longer fails with `spawn qodercli ENOENT` on Windows ([#6263](https://github.com/diegosouzapw/OmniRoute/issues/6263)) — `spawnQoderCli` spawned the bare `qodercli` name with `shell:false` and an unenriched env, so the npm `.cmd` wrapper under `%APPDATA%\npm` (a user-PATH directory) was never resolved. It now resolves the absolute `.cmd`/`.exe` path through the existing `getCliRuntimeStatus("qoder")` resolver in `src/shared/services/cliRuntime.ts` (memoized), spawns with `shell` when the target is a `.cmd`/`.bat`, and uses the cliRuntime-enriched env (PATH + PATHEXT + APPDATA); the ENOENT error now lists the searched paths plus the `CLI_QODER_BIN` override. End-to-end spawn on a real Windows host is host-only (Hard Rule #18); the path-resolution logic is unit-tested. Regression guard: `tests/unit/qodercli-windows-resolve-6263.test.ts`. (thanks @chirag127)
+
+- **fix(sse):** the reasoning-token buffer no longer inflates **probe-sized `max_tokens`** ([#6274](https://github.com/diegosouzapw/OmniRoute/issues/6274)) — Claude Code's `/model` capability check sends `max_tokens: 1`, but for a thinking-capable model with a large output cap (e.g. `glm-5.2`) the #3587 headroom heuristic (`max(current + 1000, ceil(current * 1.5))`) rewrote it to `1001` and forwarded that upstream, wasting tokens on a request that was never a genuine reasoning budget. `resolveReasoningBufferedMaxTokens()` (`open-sse/services/reasoningTokenBuffer.ts`) now short-circuits and returns the caller's value verbatim when it is below the new `REASONING_BUFFER_MIN_TRIGGER` (256) threshold — a tiny explicit limit is a probe, not a reasoning request. Real budgets still receive the #3587 headroom unchanged, and the guard runs after the existing capability checks so unknown / non-reasoning models keep returning `null`. Regression guard: `tests/unit/reasoning-token-buffer-6274.test.ts`. (thanks @brightfiscalband)
+
+- **fix(cli):** `omniroute reset-password` now works as a real subcommand, and password resets over piped (non-TTY) stdin actually apply ([#6261](https://github.com/diegosouzapw/OmniRoute/issues/6261), [#6258](https://github.com/diegosouzapw/OmniRoute/issues/6258)). Two coupled defects: (1) **#6261** — `bin/omniroute.mjs` routed everything through Commander with only two pre-Commander bypasses (`--mcp`, `reset-encrypted-columns`), so `omniroute reset-password` was rejected as an unknown command; only the separate `omniroute-reset-password` bin worked, while the docs falsely advertised the subcommand (incl. a bogus "legacy alias still works"). A pre-Commander bypass mirroring `reset-encrypted-columns` now dynamically imports `bin/reset-password.mjs` (which self-executes) before Commander parses; the three doc lines were corrected. (2) **#6258** — `bin/reset-password.mjs` issued two sequential `rl.question` prompts; under piped stdin the second read never settled at EOF, so `main()` never reached `resetManagementPassword` and the reset was a silent no-op (both prompts printed, no success, password unchanged). The CLI now detects non-TTY stdin and reads it once (first line = password, second line = confirm if present, else reused), adds a `--password-stdin` flag (entire stdin is the password, no confirmation), and exits `0` explicitly so the success line always flushes; interactive TTY behavior is unchanged. Regression guard: `tests/unit/reset-password-cli-6261-6258.test.ts` (3). (thanks @chirag127)
+
+- **fix(db):** the mass-migration **safety abort** now tells the operator how to bypass it and stops flooding the log ([#6260](https://github.com/diegosouzapw/OmniRoute/issues/6260)) — after restoring a backup that wiped the migration tracking table, `runMigrations()` threw the abort on every downstream `ensureDbInitialized()`, re-logging the full banner 11+ times, and the message never mentioned the existing `OMNIROUTE_MAX_PENDING_MIGRATIONS` escape hatch. The abort text now appends a bypass hint (set `OMNIROUTE_MAX_PENDING_MIGRATIONS=0` in `server.env` / `DATA_DIR/.env`), and a new `MigrationSafetyAbortError` is memoized so repeated calls in the same process throw the same instance and emit a single concise line instead of the full cascade. Regression guard: `tests/unit/migration-safety-abort-6260.test.ts`. (thanks @chirag127)
+
+- **fix(auth):** importing a **distinct** Codex/ChatGPT OAuth `auth.json` is no longer falsely rejected as "already exists" when it belongs to a different user in the same workspace ([#6301](https://github.com/diegosouzapw/OmniRoute/issues/6301)). `findExistingCodexConnection` (in `src/lib/oauth/utils/codexAuthImport.ts`) deduped **only** on `providerSpecificData.workspaceId === accountId`, where `accountId` is the shared `chatgpt_account_id`/`tokens.account_id` — so two members of the same ChatGPT Team collapsed onto a single connection (409 `duplicate_account`). The id_token's `https://api.openai.com/auth` claim carries a per-user `chatgpt_user_id` alongside the workspace id (the device-flow path already persisted it as `chatgptUserId`, but the import path did not). Now `parseAndValidateCodexAuth` extracts `userId` (`chatgpt_user_id` → `user_id` → JWT `sub`) into `ParsedCodexAuth`, the create/update paths persist `chatgptUserId` in `providerSpecificData` (mirroring `codex.ts`), and dedup keys on `workspaceId` **AND** `chatgptUserId` — with a backward-compat fallback to legacy accountId-only matching when no stored connection for that workspace records a `chatgptUserId`, so genuinely-same accounts still dedup. Regression guard: `tests/unit/codex-auth-import-userid-dedup-6301.test.ts` (4). (thanks @anungma)
+
+- **fix(providers):** importing models for the **venice-web** provider no longer fails with a red "Provider venice-web does not support models listing" ([#6269](https://github.com/diegosouzapw/OmniRoute/issues/6269)). `venice-web` is a web-cookie provider with an executor but no upstream `/v1/models` endpoint and no registry `models`, so the models route fell through to the tail `400`. Mirroring the `jules`/`linkup-search`/`ollama-search` fix (#5569), it now ships a static local catalog entry in `src/lib/providers/staticModels.ts` — seeding the current Venice lineup (`venice-uncensored`, `llama-3.3-70b`, `qwen3-235b`, `qwen3-4b`, `deepseek-r1-671b`; Venice rotates its catalog, see docs.venice.ai/models/overview) — so the route returns `200` with `source:"local_catalog"`, `intentional:true`. Regression guard: `tests/unit/static-models-venice-web-6269.test.ts`. (thanks @chirag127)
+
+- **fix(api):** the specialty model catalogs (`/v1/embeddings`, `/v1/images`, `/v1/music`, `/v1/videos` model lists) are now derived from the **unified catalog filtered by a predicate** (`getSpecialtyModelsResponse`) instead of ad-hoc per-route logic, so they consistently respect active-credential visibility and stay in sync with the main catalog ([#6303](https://github.com/diegosouzapw/OmniRoute/pull/6303)). Regression guard: `tests/unit/specialty-model-catalog-routes.test.ts`. (thanks @makcimbx)
+- **fix(api):** the agent-bridge server route now resolves the MITM manager via a **dynamic `import("@/mitm/manager.runtime")`** so Turbopack does not statically pull the stub (or over-bundle the manager), and the agent-skills generator anchors its output base path with `path.join(process.cwd(), …)` so Turbopack's static analyzer stops tracing the whole project root ([#6329](https://github.com/diegosouzapw/OmniRoute/issues/6329), [#6366](https://github.com/diegosouzapw/OmniRoute/pull/6366)). Regression guard: `tests/unit/agent-bridge-server-route-dynamic-import.test.ts`. (thanks @Iammilansoni)
+- **fix(api):** internal probes (combo-test, cloud-sync verify) now pick a **management-scoped / allow-all API key** instead of naively grabbing `getApiKeys()[0]` — a restricted `self:usage` first row made the probe fail with "Model X is not allowed for this API key" even when the combo path was healthy (`pickApiKeyForInternalUse` in `src/lib/db/apiKeys.ts`). The API-manager model editor also falls back to `/api/models?all=true` when `/v1/models` is catalog-protected ([#6372](https://github.com/diegosouzapw/OmniRoute/pull/6372)). Regression guard: `tests/unit/pick-internal-api-key-6372.test.ts`. (thanks @jmengit)
+- **fix(live-ws):** the Live Dashboard WebSocket server now **rejects on bind failure** (e.g. `EADDRINUSE` when the API bridge already holds the port) instead of letting the error surface as an unhandled `error` event that crash-loops the process — the `error` listener is attached to `wss` (not `server`) and releases the EventBus subscription on a failed start ([#6324](https://github.com/diegosouzapw/OmniRoute/issues/6324)). Regression guard: `tests/unit/live-ws-eaddrinuse-6324.test.ts`. (thanks @vinayakkulkarni)
+- **fix(dashboard):** the Home provider-topology widget now trusts the live provider-metrics snapshot — it uses `topology.errorProvider` and live `activeRequests` directly instead of re-deriving state from a stale `lastErrorAt` or applying a frontend timeout filter, so the topology reflects real-time provider health ([#6322](https://github.com/diegosouzapw/OmniRoute/pull/6322)). Regression guard: `tests/unit/home-provider-topology-live-state.test.ts`. (thanks @xz-dev)
+- **fix(sse):** strip zero-width markers from streamed **tool-call arguments** — a follow-up to [#5857](https://github.com/diegosouzapw/OmniRoute/pull/5857). That PR removed injected zero-width joiners (U+200D) from streamed assistant text/reasoning but deliberately left tool-call argument JSON byte-exact. The request-side obfuscation (`open-sse/services/claudeCodeObfuscation.ts`) injects ZWJ into agent words — including the temp path inside the Bash tool description — and Claude models copy that verbatim into generated commands, which are delivered as tool-call arguments rather than assistant text. As a result the ZWJ survived and corrupted code blocks (e.g. a temp path rendered with an invisible joiner). Now `open-sse/handlers/responseSanitizer.ts` strips zero-width code points from tool-call argument strings at every emit site (OpenAI non-stream/stream chat `tool_calls` + legacy `function_call`, native Responses `function_call` items, the OpenAI→Responses conversion, and the native Responses streaming `response.function_call_arguments.delta/.done` events). Only zero-width code points are removed; JSON structure and all other bytes stay identical (no parse/restringify), so normal arguments remain byte-exact. Regression guard: 6 new cases in `tests/unit/response-sanitizer.test.ts` (suite 50/50).
+
+- **fix(nodejs):** the default app log path now resolves under `DATA_DIR` (`~/.omniroute/logs/application/app.log`) instead of `process.cwd()` ([#6197](https://github.com/diegosouzapw/OmniRoute/issues/6197)) — the globally-installed CLI runs from an arbitrary working directory, so anchoring the default to cwd made file logging silently write to (or no-op under) an unrelated directory, contradicting the documented `.env.example` default. `getAppLogFilePath()` now computes the default lazily via the pure `resolveDataDir()` resolver (honours a per-process `DATA_DIR`, no directory-creation side effect); an explicit `APP_LOG_FILE_PATH` still wins. Regression guard: `tests/unit/logenv-datadir-path-6197.test.ts` (3).
+
+- **fix(docker):** AgentBridge/`startMitm` no longer aborts in containers/headless when the Antigravity-default DNS step can't write `/etc/hosts` ([#6127](https://github.com/diegosouzapw/OmniRoute/issues/6127)), and the privileged command's stderr now reaches `app.log` instead of only a bare exit code hitting the toast ([#6198](https://github.com/diegosouzapw/OmniRoute/issues/6198)). The default DNS step (`addDNSEntry`) was called unguarded while cert install and the two sibling DNS steps were each best-effort — in the runtime Docker image (`USER node`, no `sudo`, read-only `/etc/hosts`) it threw `Command failed with code 1` out of `startMitmInternal` and killed the whole start, discarding the stderr. The three DNS steps are extracted into a best-effort `provisionDnsEntries()` where each failure is logged with the full `err` (stderr included, folded in by `systemCommands.ts`) and never aborts the start. Regression guard: `tests/unit/mitm-dns-graceful-degrade-6127.test.ts` (4).
+
+- **fix(providers):** copilot-m365-web now supports the M365 Education "Starter / OfficeWebIncludedCopilot" tier and no longer returns an empty `content:null` stream ([#6210](https://github.com/diegosouzapw/OmniRoute/issues/6210)). Two gaps: (1) `buildWsUrl()` hardcoded the individual-consumer scenario (`OfficeWebPaidConsumerCopilot`, `isEdu=false`) — the EDU tier is now opt-in via `providerSpecificData.tier="edu"`, emitting `scenario=OfficeWebIncludedCopilot`/`isEdu=true` (the individual path is unchanged); (2) the EDU/GPT-5.5 path streams deltas via `arguments[0].writeAtCursor` (incremental) instead of only `messages[].text` (accumulated snapshots), which the parser dropped — a new `accumulateBotContent()` folds both formats, with `type:2 item.result.message` as a last-resort fallback. Regression guard: `tests/unit/copilot-m365-edu-writeatcursor-6210.test.ts` (10). (thanks @qpeyba)
+
+- **fix(providers):** GitLab Duo executor now feeds tool results back into the prompt instead of looping ([#6220](https://github.com/diegosouzapw/OmniRoute/issues/6220)) — `buildPrompt()` branched only on `system`/`user` and took `userParts.at(-1)`, silently dropping the `assistant{tool_calls}` + `tool{result}` turns the client appended, so the reconstructed prompt was byte-identical to turn 1 and the model re-emitted the same `<tool>` call forever. When a tool exchange is present the full conversation is now serialized, folding each tool result back keyed by its `tool_call_id`; simple conversations keep the legacy shape. Complements the tool_call emission from [#6051](https://github.com/diegosouzapw/OmniRoute/issues/6051) (the `kilo-duplicate` label was a false positive — different, sequential defect). Regression guard: `tests/unit/gitlab-tool-result-feedback-6220.test.ts` (4).
+
+- **fix(providers):** opencode-go/opencode-zen can now synthesize the OpenCode CLI identity headers Cloudflare requires on VPS egress ([#5997](https://github.com/diegosouzapw/OmniRoute/issues/5997)) — on a datacenter VPS, `opencode.ai/zen/go/v1/chat/completions` 403s (HTML challenge) requests lacking CLI identity, while the reporter's control curl proved that `User-Agent: opencode-cli/1.0.0` + `x-opencode-client: cli` + `x-opencode-project: default` + fresh request/session UUIDs succeed. Opt-in via `OPENCODE_SYNTHESIZE_CLI_HEADERS=true` (values overridable via `OPENCODE_GO_USER_AGENT`/`OPENCODE_USER_AGENT`/`OPENCODE_CLIENT`/`OPENCODE_PROJECT`); it fills only headers the client did not already send. Kept **off by default** — the forward-only path is deliberate (fabricating a wrong value risks upstream rejection; a prior dedup regressed with `opencode/local`), so this replaces the fragile local header-injection shim without changing default behavior. Regression guard: `tests/unit/opencode-cli-headers-synthesis-5997.test.ts` (6). (thanks @aleksesipenko)
+- fix(resilience): sticky session affinity now evicts and fails over to another account when the pinned account is exhausted/unavailable (#6219)
+
+- fix(sse): Responses API passthrough now drops internal commentary-phase output before forwarding to clients (gated by RESPONSES_PASSTHROUGH_DROP_COMMENTARY, default on) (#6199)
+
+- **fix(sse):** tool-call function schemas with a root `type: null` are now coerced to `type: "object"` before dispatch ([#6359](https://github.com/diegosouzapw/OmniRoute/issues/6359)) — clients like the Codex app emit `parameters: { type: null, ... }` for some tools, which OpenAI-compatible upstreams reject with `400 Invalid schema for function '...': schema must be a JSON Schema of 'type: "object"', got 'type: null'`, failing the whole request. `toolSchemaSanitizer` already stripped the null; it now re-adds the mandatory root `"object"` type (and empty `properties`/open `additionalProperties` when absent). Combinator roots (`anyOf`/`oneOf`/`allOf`) and explicit root types are left untouched. Regression guard: 5 new cases in `tests/unit/tool-schema-sanitizer.test.mjs`.
+
+- **fix(docker):** AgentBridge no longer fails to start on npm/Electron/VPS installs with "MITM manager stub reached at runtime" ([#6344](https://github.com/diegosouzapw/OmniRoute/issues/6344)) — v3.8.45 flipped the production bundler default to Turbopack, but `next.config.mjs` aliased `@/mitm/manager` to its Docker-only degraded stub **unconditionally**. That was harmless while Docker (which sets the alias intentionally for #3390 graceful degradation) was the sole Turbopack consumer, but once every artifact built with Turbopack the stub shipped to all non-Docker users and `startMitm` threw on the first Agent-Bridge start. The alias is now opt-in via `OMNIROUTE_MITM_STUB=1` (set only by the Dockerfile) through the shared `scripts/build/mitm-stub-flag.mjs` helper; default builds bundle the real manager. Regression guard: `tests/unit/mitm-stub-alias-6344.test.mjs` (4).
+
+- **fix(proxy):** stop the v3.8.44 proxy regression that leaked the real IP and disabled healthy proxies ([#6246](https://github.com/diegosouzapw/OmniRoute/issues/6246)). Two coupled defects from the new health scheduler: (1) **IP leak** — when a proxy assigned to a connection was marked `inactive`, resolution fell through to a **direct** egress instead of blocking, exposing the operator's real IP; (2) **over-deactivation** — the sweep flipped a proxy to `inactive` on the **first** failed probe and counted our own 5s timeout / a probe-target `5xx` as the proxy's fault, so healthy paid proxies vanished from egress selection ("my proxies are not being used anymore"). Fix: the sweep decision is extracted into a pure, network-free `decideProxyHealthAction` (`src/lib/proxyHealth/decision.ts`) — by default the health check now **only counts/logs and never downgrades status** (a proxy is downgraded/removed only with `PROXY_AUTO_REMOVE=true`, after `PROXY_AUTO_REMOVE_AFTER` **consecutive** conclusive failures); probes are classified tri-state so an inconclusive result (our timeout, or a `5xx` from the probe target) never penalizes the proxy, and the probe timeout is raised 5s→15s. Separately, `safeResolveProxy` now **fails closed** via the existing policy: a connection whose assigned proxy is dead is blocked instead of leaking direct (`hasBlockingProxyAssignment`), honoring the explicit `proxy off` toggles and the `PROXY_FAIL_OPEN=true` opt-out. Existing proxies stuck `inactive` by the old behavior need a one-time manual re-activate (the operator owns proxy status). Regression guards: `tests/unit/proxy-health-decide-action-6246.test.ts`, `tests/unit/proxy-assigned-unavailable-6246.test.ts`.
+
+- **fix(proxy):** make "Test All" read-only and add bulk enable/disable ([#6246](https://github.com/diegosouzapw/OmniRoute/issues/6246)). Complements the core fail-closed / scheduler fix (#6296) with the two remaining reporter asks. (1) The **"Test All" button** (`POST /api/settings/proxies/auto-test`) used to flip a proxy to `inactive` on a failed reachability probe; since the egress selector excludes `inactive` proxies, a flaky probe (an unreachable `httpbin.org`, a proxy that blocks `HEAD`, or a slow paid proxy) silently disabled every proxy that failed — "Test All" is now **read-only by default** (only the operator sets a proxy active/inactive; opt back into the legacy test-and-set with `PROXY_HEALTH_AUTO_DEACTIVATE=true`). (2) Adds a **bulk enable/disable** proxies endpoint + toolbar action (`POST /api/settings/proxies/batch-activate`) so an operator can re-activate proxies in one click. Regression guard: `tests/unit/proxy-health-6246.test.ts`. (thanks @tenshiak)
+
+- **chatcore (tools): stop the default 128-tool cap from silently dropping opencode's `task`/MCP tools.** opencode (used as an MCP/agent host) sends a large tool list; when it exceeds the speculative `MAX_TOOLS_LIMIT` (128) default, `truncateToolList` did a blind `tools.slice(0, 128)`, dropping every tool past index 128 — including opencode's built-in `task` tool (subagent launch) and many MCP tools, so models routed through OmniRoute could no longer spawn subagents or reach part of their tools. The cap exists to avoid upstream `400`s for providers with real hard limits (e.g. grok-cli 200), so it is kept for those: detection of the opencode client (`isOpencodeClient` — any `x-opencode-*` header, or `opencode` in the user-agent) now only bypasses the **speculative 128 default**, never a known provider ceiling. Precedence is explicit — a proactive/detected provider limit always truncates (even for opencode); otherwise opencode forwards its full tool list; otherwise the unchanged 128 default applies to every other client. Refactors `getEffectiveToolLimit` into `getKnownToolLimit(provider) ?? DEFAULT_LIMIT` (byte-identical for existing callers) and fixes a cosmetic debug-log that reported the truncated count instead of the original. Regression guard: `tests/unit/tool-limit-detector.test.ts`.
+
+- **fix(mitm):** the macOS MITM-cert install check now matches the system keychain again. `security find-certificate -a -Z` prints the SHA-1 as a colon-less hex string, but the installed-check compared it against `getCertFingerprint()`'s colon-separated form, so the substring match never hit — the cert was reported as not-installed and re-prompted for the sudo install on every run. Fingerprints are now normalized (colons stripped, upper-cased) on both sides via the extracted `macCertOutputHasFingerprint` helper. Regression guard: `tests/unit/mitm-cert-mac-fingerprint.test.ts`. ([#6204](https://github.com/diegosouzapw/OmniRoute/pull/6204), closes [#6134](https://github.com/diegosouzapw/OmniRoute/issues/6134) — thanks @rianonehub)
+
+- **fix(api):** `/v1/messages/count_tokens` now counts `tool_use`, `tool_result` and `thinking` content blocks (and array-form `system` prompts) in the local-estimation path, instead of only `text`. Real agentic conversations keep ~95% of their tokens inside tool results; the previous estimate returned near-zero for them, which silently broke Claude Code's auto-compaction (context grew past the window with no compaction until the upstream API rejected the request). The real provider-side count path is unchanged. Regression guard: `tests/unit/messages-count-tokens-route.test.ts`. ([#6221](https://github.com/diegosouzapw/OmniRoute/pull/6221) — thanks @luweiCN)
+
+- **fix(antigravity):** strip a trailing assistant prefill turn for Vertex Claude models to avoid upstream 400s ([#6114](https://github.com/diegosouzapw/OmniRoute/pull/6114)). Regression guard: `tests/unit/antigravity-claude-prefill-strip.test.ts`. (thanks @anki1kr)
+
+- **fix(security):** the mutable cloud-agent routes (`/api/cloud/credentials/update`, `/api/cloud/models/alias`) now require management auth instead of being treated as public. They were classified as public API routes, so a request without management credentials could update stored cloud-agent credentials and model aliases. They are removed from the public-route set, classified as management routes in the authz pipeline, and gated by `requireManagementAuth`; cloud **read**/auth routes stay public. Regression guards: `tests/unit/cloud-write-auth.test.ts`, `tests/unit/authz/classify.test.ts`, `tests/unit/public-api-routes.test.ts`. ([#6233](https://github.com/diegosouzapw/OmniRoute/pull/6233) — thanks @vittoroliveira-dev)
+
+- **refactor(dashboard):** extract the onboarding-wizard "Open provider details" link target into a pure, unit-tested `buildProviderDetailsHref(connection)` helper. The wizard already routes by `connection.id` (the node UUID) rather than the provider category slug (#6144/#6145); this hardens that behavior behind a tested helper that guards a missing id/connection. Regression guard: `tests/unit/provider-onboarding-href.test.ts`. ([#6166](https://github.com/diegosouzapw/OmniRoute/pull/6166) — thanks @KooshaPari)
+
+- **fix(security):** the doubao synthetic device-id generator now derives its digits via an unbiased crypto-random draw (rejection sampling over `crypto.randomBytes()`) instead of a `% 10` reduction, closing a CodeQL `js/biased-cryptographic-random` finding.
+- **fix(agentSkills):** the GitHub-skills generator now resolves `outputDir` to an absolute path before writing, fixing a regression introduced by #6366 (relative-to-cwd base path) that could write generated skill files to the wrong directory.
+- **fix(security):** `/api/keys/{id}/devices` now checks the HTTP method before auth/validation, returning a `405` for non-GET/DELETE verbs instead of a misleading `401`/`500` (closes a `dast-smoke` QUERY-method finding).
+- **fix(quality):** clear the last 2 heavy quality-gate reds on the release tip (cycle pre-flight).
+- **fix(mitm):** the test suite and CI can never mutate the OS trust store — `OMNIROUTE_SKIP_SYSTEM_TRUST=1` is set globally for tests/CI so `installCert`/`uninstallCert`/`installTproxyCa` skip the privileged OS dispatch ([#6310](https://github.com/diegosouzapw/OmniRoute/pull/6310); full detail is under the [3.8.45] section below — this branch received it via the parallel-cycle sync-back).
+- **fix(api):** `POST /api/github-skills` now Zod-validates its request body; documented the new quality-gate env vars and pinned the merge-integrity GitHub Actions to a commit SHA.
+- **fix(skills):** generate the missing `omni-github-skills` registry entry and align the agent-skills catalog-count tests (follow-up to #6186).
+- **fix(quality):** clear the cycle's 11 net-new ESLint errors and make `validate-release-green` suppressions-aware.
+
+### 📝 Maintenance
+
+- **i18n(it):** add 118 missing Italian (`it`) translations (net-additive — no existing keys dropped, valid JSON), improving Italian UI coverage. ([#6212](https://github.com/diegosouzapw/OmniRoute/pull/6212) — thanks @serverless83)
+- **chore(providers):** remove deprecated **MiMo V2** model entries from the catalogs (xiaomi-mimo, opencode-go, zenmux-free, audio TTS) — the upstream V2 line is superseded by MiMo V2.5; drops `mimo-v2-tts`, `mimo-v2-pro`, `mimo-v2-omni`, `mimo-v2-flash`, `mimo-v2-flash-free` and realigns the provider-catalog tests. ([#6248](https://github.com/diegosouzapw/OmniRoute/pull/6248) — thanks @backryun)
+- **chore(release):** ~50 commits on this branch are v3.8.45 pre-flight/hardening fixes and CI-perf work that landed here via the parallel-cycle sync-back (`sync-next-cycle.mjs`, Hard Rule #21) **after** the `v3.8.45` git tag was cut, and are already fully documented under the **[3.8.45]** section below — listed here only so the per-cycle commit-coverage check (`npm run release:uncovered`) doesn't flag them as gaps. Provider/catalog/UX/backend fixes: #6041, #6078, #6108, #6135, #6148, #6149, #6154, #6158, #6161, #6162, #6163, #6164, #6165, #6170, #6177, #6178, #6181, #6186, #6187, #6191, #6193, #6194, #6195, #6200, #6205, #6208, #6209, #6211, #6213, #6223, #6224, #6225, #6226, #6227, #6228, #6229, #6230, #6235, #6291, #6292. CI/release-pipeline work: #6167, #6203, #6214, #6215, #6218, #6273, #6275, #6283, #6284, #6285, #6300, #6305.
+- **chore(release):** additional zero-ref release-cycle plumbing on this branch, kept out of `release:uncovered` on purpose (no `#N` in the commit subject to cite): opening the v3.8.46 cycle, opening/closing the v3.8.45 cycle, the finalized [3.8.45] CHANGELOG i18n sync-back to 42 mirrors, the v3.8.45 cognitive/cyclomatic and file-size drift rebaselines, ESLint stale-suppression pruning (4,273 → 4,233), and clearing test-masking/docs-all pre-flight reds for v3.8.45.
+
+### ⚡ Performance & Infrastructure
+
+- **perf(release-green):** the pre-flight validator (`scripts/quality/validate-release-green.mjs`) now runs its 4 slow suites (unit / vitest / integration / pack-artifact) **concurrently** via `Promise.all` — pre-flight wall time drops from ~the sum of the suites to ~the slowest one (~30min saved per round; Phase 0 was the nº1 bottleneck of the v3.8.45 release benchmark, 2h54 of 6h34 e2e). Guard: `tests/unit/validate-release-green.test.ts` ("runs the slow suites CONCURRENTLY"). ([#6319](https://github.com/diegosouzapw/OmniRoute/pull/6319))
+- **fix(ci):** `scripts/release/sync-next-cycle.mjs` — two defects found live in its first production run (v3.8.45 Phase 5): (1) the `git()` helper's default 1 MiB `maxBuffer` crashed with `ENOBUFS` on `git show origin/main:CHANGELOG.md` (the CHANGELOG alone is >1 MiB) — widened to 64 MiB; (2) the i18n resync only propagated the `[NEXT]` (TBD) section, leaving the just-shipped finalized section as "— TBD" in all 42 mirrors — it now also syncs `[prevVersion]` bounded by the heading below it (new exported pure helper `versionAfter`). Guards: +5 tests in `tests/unit/sync-next-cycle.test.ts` (8/8). ([#6327](https://github.com/diegosouzapw/OmniRoute/pull/6327))
+- **test(ci):** concurrency-sensitive flaky tests are **quarantined into a serial pass** (`tests/unit/serial/`, `--test-concurrency=1`, appended to every unit runner incl. sharded variants — the serial pass is sharded too so concurrent shard jobs never self-collide). Initial set: `glm-coding-plan-monthly-3580`, `quota-division-blocks`, `provider-health-autopilot`, `combo-health-autopilot` — the class behind the ~28min CI wedges/re-runs (two live 1h+ wedges cancelled during this PR's own validation). Discovery + TIA gates track the new glob; systemic root cause (async logger writing after teardown) tracked in [#6360](https://github.com/diegosouzapw/OmniRoute/issues/6360). Guard: `tests/unit/test-serial-quarantine.test.ts` (4). ([#6347](https://github.com/diegosouzapw/OmniRoute/pull/6347))
+
+### 🙌 Contributors
+
+Thanks to everyone whose work landed in v3.8.46:
+
+| Contributor                                                    | PRs / Issues           |
+| -------------------------------------------------------------- | ---------------------- |
+| [@2220258345](https://github.com/2220258345)                   | direct commit / report |
+| [@abdofallah](https://github.com/abdofallah)                   | direct commit / report |
+| [@adentdk](https://github.com/adentdk)                         | direct commit / report |
+| [@aleksesipenko](https://github.com/aleksesipenko)             | direct commit / report |
+| [@anki1kr](https://github.com/anki1kr)                         | direct commit / report |
+| [@anungma](https://github.com/anungma)                         | direct commit / report |
+| [@arpicato](https://github.com/arpicato)                       | direct commit / report |
+| [@backryun](https://github.com/backryun)                       | #6248                  |
+| [@binsarjr](https://github.com/binsarjr)                       | direct commit / report |
+| [@brightfiscalband](https://github.com/brightfiscalband)       | direct commit / report |
+| [@chirag127](https://github.com/chirag127)                     | #6501, #6506           |
+| [@developerjillur](https://github.com/developerjillur)         | direct commit / report |
+| [@dilneiss](https://github.com/dilneiss)                       | #6499                  |
+| [@dtybnrj](https://github.com/dtybnrj)                         | direct commit / report |
+| [@Forcerecon](https://github.com/Forcerecon)                   | direct commit / report |
+| [@hao3039032](https://github.com/hao3039032)                   | direct commit / report |
+| [@Iammilansoni](https://github.com/Iammilansoni)               | #6150, #6245           |
+| [@jmengit](https://github.com/jmengit)                         | direct commit / report |
+| [@jordansilly77-stack](https://github.com/jordansilly77-stack) | direct commit / report |
+| [@JxnLexn](https://github.com/JxnLexn)                         | direct commit / report |
+| [@KooshaPari](https://github.com/KooshaPari)                   | #6166                  |
+| [@loopyd](https://github.com/loopyd)                           | direct commit / report |
+| [@luweiCN](https://github.com/luweiCN)                         | #6221                  |
+| [@makcimbx](https://github.com/makcimbx)                       | direct commit / report |
+| [@muflifadla38](https://github.com/muflifadla38)               | direct commit / report |
+| [@newnol](https://github.com/newnol)                           | direct commit / report |
+| [@ofekbetzalel](https://github.com/ofekbetzalel)               | direct commit / report |
+| [@ohahe52-dot](https://github.com/ohahe52-dot)                 | direct commit / report |
+| [@phidinhmanh](https://github.com/phidinhmanh)                 | direct commit / report |
+| [@powellnorma](https://github.com/powellnorma)                 | direct commit / report |
+| [@qpeyba](https://github.com/qpeyba)                           | direct commit / report |
+| [@RaviTharuma](https://github.com/RaviTharuma)                 | direct commit / report |
+| [@RCrushMe](https://github.com/RCrushMe)                       | direct commit / report |
+| [@rianonehub](https://github.com/rianonehub)                   | #6134, #6204           |
+| [@serverless83](https://github.com/serverless83)               | #6212                  |
+| [@swingtempo](https://github.com/swingtempo)                   | direct commit / report |
+| [@tenshiak](https://github.com/tenshiak)                       | direct commit / report |
+| [@ThongAccount](https://github.com/ThongAccount)               | direct commit / report |
+| [@UnrealAryan](https://github.com/UnrealAryan)                 | direct commit / report |
+| [@vinayakkulkarni](https://github.com/vinayakkulkarni)         | direct commit / report |
+| [@vittoroliveira-dev](https://github.com/vittoroliveira-dev)   | #6233                  |
+| [@warelik](https://github.com/warelik)                         | direct commit / report |
+| [@xxy9468615](https://github.com/xxy9468615)                   | direct commit / report |
+| [@xz-dev](https://github.com/xz-dev)                           | direct commit / report |
+| [@yanpaing007](https://github.com/yanpaing007)                 | direct commit / report |
+| [@diegosouzapw](https://github.com/diegosouzapw)               | maintainer             |
+
+---
+
 ## [3.8.45] — 2026-07-06
 
 ### ✨ New Features

docs/i18n/fi/CHANGELOG.md

@@ -6,6 +6,213 @@
 
 ## [3.8.31] — 2026-06-20
 
+## [3.8.46] — TBD
+
+### ✨ New Features
+
+- **feat(sse):** **hide paid-only models from `auto/*` routing** when `hidePaidModels` is on ([#6512](https://github.com/diegosouzapw/OmniRoute/issues/6512)) — follow-up to #6328/#6495. PR #6495 hid paid-only models from the `GET /v1/models` listing, but `auto/*` combos (`auto/best-coding`, `auto/glm`, …) could still pick a paid-only backend into their candidate pool → a 402/403 at request time. `createVirtualAutoCombo` now filters the candidate pool through the new pure `open-sse/services/autoCombo/paidModelFilter.ts` (`filterPaidOnlyCandidates`), applying the same free-model predicate #6495 uses in `catalog.ts` (`providerHasFreeModels(provider) && isFreeModel(provider, {id})`) whenever `settings.hidePaidModels === true`. Applied before the category/tier/family narrowing, so it covers every `auto/*` combo; an all-paid pool degrades to the existing graceful empty-pool path. **Opt-in — default OFF leaves the pool unchanged** (identity). Regression guard: `tests/unit/autoCombo/paid-model-filter-6512.test.ts` (4, incl. the default-off identity guard).
+- **feat(sse):** **provider-family auto combos** — `auto/glm`, `auto/minimax`, `auto/mimo`, `auto/zai`, `auto/gemma`, `auto/llama`, `auto/gemini` ([#6453](https://github.com/diegosouzapw/OmniRoute/issues/6453)) — new routable ids that materialize an on-demand virtual combo spanning whatever installed backends currently expose that model family, degrading gracefully as backends rotate. A new pure `open-sse/services/autoCombo/modelFamily.ts` (`detectModelFamily`) classifies by model-id prefix for six families; `zai` is instead resolved by provider id (z.ai's hosted API serves the same `glm-*` model ids as every other GLM backend, so `auto/zai` means "route to my z.ai backend specifically" vs `auto/glm`'s "any connected GLM backend"). Reuses the existing `createVirtualAutoCombo` on-demand materialization path (no DB writes) and the `/v1/models` catalog advertising loop. Regression guard: `tests/unit/autoCombo/provider-family-combos.test.ts` (11).
+- **feat(proxy):** native **proxy-pool round-robin / egress IP rotation** ([#6365](https://github.com/diegosouzapw/OmniRoute/issues/6365)) — a scope (global / provider / account) can now hold **multiple** proxies as a pool with a rotation strategy, so outbound requests cycle their egress IP instead of pinning one proxy per scope. Migration `117_proxy_pool_rotation.sql` lifts the `UNIQUE(scope, scope_id)` constraint (rebuild via the canonical rename/copy/drop; existing single assignments become 1-element pools) and adds a `proxy_scope_rotation` companion table holding the per-scope strategy + a persisted monotonic round-robin cursor. Strategies: `round-robin` (default, monotonic cursor — never `Math.random`), `random`, and `sticky-per-N-min`. Resolution (`resolveProxyForScopeFromRegistry` / `resolveProxyForConnectionFromRegistry`) now fetches the alive, position-ordered candidate set (unchanged `PROXY_ALIVE_PREDICATE`) and applies the strategy; an empty / all-dead pool still returns `null` — the #6246 fail-closed guard is **untouched** (never falls through to direct egress). Backend + DB only; dashboard pool-builder UI is a follow-up. Regression guard: `tests/unit/proxy-pool-rotation-6365.test.ts` (8, incl. fail-closed + backward-compat).
+- **feat(providers):** end-to-end **tool/function calling on the native Gemini `/v1beta` endpoint** ([#6222](https://github.com/diegosouzapw/OmniRoute/issues/6222)) — both directions of the Gemini↔OpenAI conversion now preserve tool data (previously silently dropped). Request side: `convertGeminiToInternal` (extracted to its own testable module) maps `tools[].functionDeclarations` → OpenAI `tools`, prior `functionCall` parts → assistant `tool_calls`, and `functionResponse` parts → `tool`-role messages. Response side: `convertOpenAIResponseToGemini` emits `parts[].functionCall {name,args}` from `message.tool_calls`, and the streaming `openAIChunkToGeminiChunk` accumulates fragmented `tool_calls` deltas by index into complete `functionCall` parts. The non-Gemini client paths (Claude, OpenAI-Responses) already preserved tool calls — this closes the gap specific to the native Gemini surface. Regression guard: `tests/unit/v1beta-gemini-tool-calling-6222.test.ts` (6, incl. a streaming SSE round-trip).
+- **feat(providers):** copilot-m365-web **enterprise / work tier** support ([#6334](https://github.com/diegosouzapw/OmniRoute/issues/6334)) — mirrors the EDU-tier pattern (#6210): `M365ConnectionParams` gains an `agent` field, a new opt-in `M365_ENTERPRISE_OVERRIDES` preset (`agent=work`, `scenario=officeweb`, `licenseType=Premium`) applies via `providerSpecificData.tier="enterprise"` (alias `"work"`), and `agent` is also overridable directly via `providerSpecificData.agent`. `buildWsUrl` was hardcoding `agent="web"` (the one enterprise-distinguishing param with no override path), so a Premium work account handshook then returned an empty stream. The individual and EDU paths are untouched. Kilo's dup flag vs #6210 (EDU tier) was a false positive — different tier. Regression guard: `tests/unit/copilot-m365-enterprise-6334.test.ts` (7). End-to-end confirmation on a real Premium work account is a live-VPS validation follow-up (Hard Rule #18). (thanks @Forcerecon)
+- **feat(api):** standardized, provider-agnostic **`effort` + `thinking` request params** ([#6241](https://github.com/diegosouzapw/OmniRoute/issues/6241)) — a thin standardization layer over the existing mature per-provider reasoning plumbing (no provider mapper touched). `providerChatCompletionSchema` gains a canonical `effort` (reusing the shared `none/low/medium/high/xhigh` vocabulary — the UI tiers `extra`/`max` collapse onto `xhigh`) and a boolean `thinking`. A pure `normalizeReasoningRequest` (wired once in `src/sse/handlers/chat.ts`, before any reasoning field is read) folds them onto the fields the translators already consume (`reasoning_effort` / `reasoning.effort` / `thinking`), so they fan out to Anthropic / Gemini / xAI / Responses — an explicit client `reasoning_effort` / object-shaped `thinking` always wins (backward-compatible). `/models` additively exposes `supportsThinking` + `effort_tiers` so the frontend can render the toggles (UI component is a follow-up). Regression guard: `tests/unit/effort-thinking-standardization-6241.test.ts` (12). (thanks @Iammilansoni, @shabeer)
+- **feat(combo):** new **`pipeline` (sequential) combo strategy** ([#6297](https://github.com/diegosouzapw/OmniRoute/issues/6297)) — the 18th routing strategy runs targets **in order**, threading each step's output into the next step's input, with an optional per-step `prompt` (system instruction); only the final step's response is returned. Distinct from `fusion` (parallel fan-out + judge). Implemented as a self-contained `open-sse/services/pipeline.ts` (sibling to `fusion.ts`), dispatched from `combo.ts`; the step list reuses `combo.models` order and reads an optional `prompt` off each target (backward-compatible — ignored by every other strategy). Intermediate steps run non-streaming with tools stripped (complete prose to thread forward); the final step keeps the client's `stream` flag + tools. A failing/empty/unparseable intermediate step fails the whole pipeline explicitly via a sanitized error (never silently swallowed). Kilo's dup flag vs #563 was a false positive (that's model→chain selection; this is a sequential chain). Regression guard: `tests/unit/combo-pipeline-strategy.test.ts` (5). (thanks @ofekbetzalel)
+- **feat(ci):** `check:test-masking` now flags **inline-reimplemented prod conditions** ([#6348](https://github.com/diegosouzapw/OmniRoute/issues/6348)) — a new **report-only** subcheck (v2, 6A.10 family) catches the wrong-shape contract test: a test that recomputes the condition under test inline instead of importing/exercising the real function (the #6216 class, where `=== 500` → `>= 500` stayed green because the test re-implemented the branch). For each added/modified test file it warns when the file textually duplicates a ≥3-token conditional from a production file touched in the same PR **and** does not import the symbol/module owning it, via a pure, fixture-tested `findReimplementedConditions()` with an allowlist mirroring `assertReductionAllowlist`. Report-only for now (does not fail the gate) — to be promoted to blocking after a triage cycle. Regression guard: `tests/unit/check-test-masking.test.ts` (45).
+- **feat(sse):** per-connection routing override (native vs CLIProxyAPI) ([#6339](https://github.com/diegosouzapw/OmniRoute/issues/6339)) — the previously-dead `isCliproxyapiDeepModeEnabled` helper is now wired into `resolveExecutorWithProxy`: a single connection can opt itself into the CLIProxyAPI passthrough executor via `providerSpecificData.cliproxyapiMode="claude-native"`, with precedence **connection override > provider `upstream_proxy_config` mode > default**. `resolveExecutorWithProxy` now receives the resolved connection's `providerSpecificData` (threaded from `chatCore.ts`), so one connection can deep-route while the provider's other connections stay native — no DB schema change (the toggle rides in `providerSpecificData`). Also resolves the same-provider mixing ask in #6340. Regression guard: `tests/unit/chatcore-executor-proxy.test.ts` (9). (thanks @RaviTharuma)
+- **feat(dashboard):** "Add session cookie" modal now shows a prominent **"Open ‹host› →"** link to the provider's own site ([#6268](https://github.com/diegosouzapw/OmniRoute/issues/6268)) — every `-web` cookie-session provider (chatgpt-web, claude-web, gemini-web, kimi-web, lmarena, qwen-web, m365-copilot-web, …) renders a one-click external link (opening the provider's login/home page in a new tab) so operators no longer tab away to retype the URL mid-setup. The host resolves from a pure, unit-tested `resolveWebProviderHost()` (prefers `WEB_COOKIE_PROVIDERS[id].website`, falls back to the registry `baseUrl` origin); non-web providers render exactly as before. Kilo's dup flag vs #6265 (modal-too-small-on-1080p) was a false positive — distinct concern. Regression guard: `tests/unit/resolve-web-provider-host.test.ts` (5). (thanks @chirag127)
+- **feat(providers):** add **DigitalOcean** AI (serverless inference) as an OpenAI-compatible API-key provider ([#6373](https://github.com/diegosouzapw/OmniRoute/pull/6373)) — base `https://inference.do-ai.run/v1`, wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/digitalocean/`, `src/shared/constants/providers/apikey/inference-hosts.ts`). Regression guard: `tests/unit/digitalocean-provider.test.ts`. (thanks @newnol)
+- **feat(providers):** add **Huancheng Public API** (`hcnsec`) as an OpenAI-compatible regional provider ([#6410](https://github.com/diegosouzapw/OmniRoute/pull/6410)) — Xinjiang Huancheng Cybersecurity's public LLM platform (base `https://api.hcnsec.cn/v1`, free credits via daily check-ins), wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/hcnsec/`, `src/shared/constants/providers/apikey/regional.ts`). Regression guard: `tests/unit/hcnsec-provider.test.ts`. (thanks @UnrealAryan)
+- **feat(dashboard):** the web-session credential guide now shows an **"Open {host}" link** ([#6316](https://github.com/diegosouzapw/OmniRoute/pull/6316)) to the provider's sign-in site (derived from the provider `website` via `getProviderWebsiteHost`), so you can jump straight to the page where the cookie/session must be captured. Regression guard: `tests/unit/web-session-provider-link-6316.test.ts`. (thanks @jordansilly77-stack)
+- **feat(cerebras):** add the **Gemma 4 31B** model (`gemma-4-31b`) to the Cerebras registry + pricing table ([#6331](https://github.com/diegosouzapw/OmniRoute/pull/6331)). Regression guard extends `tests/unit/t28-model-catalog-updates.test.ts`. (thanks @backryun)
+- **feat(providers):** add **Yuanbao (web)** as a cookie-session provider ([#6196](https://github.com/diegosouzapw/OmniRoute/issues/6196)) — `yuanbao-web` (Tencent Yuanbao, `yuanbao.tencent.com`) with cookie-only auth (`hy_user`/`hy_token` + public agent id), SSE→OpenAI translation incl. `reasoning_content`, exposing DeepSeek V3/R1 + Hunyuan / Hunyuan-T1. Regression guard: `tests/unit/providers-yuanbao-web.test.ts`. `together-web` was **deferred** (no verifiable web-session endpoint — needs a captured request) and `huggingchat-web` **dropped** (the existing `huggingchat` already is a web-cookie provider). (thanks @chirag127)
+- **feat(providers):** route the built-in **agentrouter** through the dynamic Claude-Code wire image ([#6056](https://github.com/diegosouzapw/OmniRoute/issues/6056)) — a small static allow-set (`CC_WIRE_IMAGE_BUILTINS` in `open-sse/services/ccWireImageBuiltins.ts`), consulted by `isClaudeCodeCompatible` / `isClaudeCodeCompatibleProvider` / `applyFingerprint`, makes agentrouter adopt the CC wire-image headers + fingerprint **while guarding the CC baseUrl/auth branches** so it keeps its own registry `baseUrl` and `x-api-key` auth. Regression guard: `tests/unit/agentrouter-cc-wire-image.test.ts` (asserts the wire image is applied AND agentrouter's baseUrl/auth are preserved). Live WAF-acceptance against agentrouter.org is a VPS validation follow-up (Hard Rule #18).
+- **feat(providers):** **bulk-add API keys for Cloudflare Workers AI** ([#6174](https://github.com/diegosouzapw/OmniRoute/issues/6174)) — `cloudflare-ai` is removed from the bulk-add exclusion list and the bulk parser gains a 3-field `name|accountId|apiKey` mode; the bulk route now builds a **per-entry** `providerSpecificData` so each key carries its own `accountId` (fixing the previous shared-object reuse), and both the create + key-validation paths receive it. Regression guard: `tests/unit/bulk-api-key-parser-cloudflare.test.ts`. (thanks @muflifadla38)
+- **feat(dashboard):** routing/settings UX clarity ([#6147](https://github.com/diegosouzapw/OmniRoute/issues/6147)) — (1) weighted combos show the **effective routing share %** next to each weight when weights don't sum to 100 (`WeightTotalBar.tsx`); (2) the status widget's user-facing **"Cloud Sync" label is renamed** to "Remote Settings Sync" (`CloudSyncStatus.tsx`; internal ids/state untouched); (3) built-in providers gain an **opt-in advanced base-URL override** (`isBaseUrlOverrideEligibleProvider`, hidden behind an "Advanced" toggle, reusing the existing `providerSpecificData.baseUrl` persistence — not globally widened). Regression guard: `tests/unit/routing-settings-ux-6147.test.ts`.
+- **feat(combo):** add an option to **disable session stickiness**, per-combo or globally — round-robin / random combos can rotate to a different connection on every request instead of pinning a whole conversation to one connection by its first-message hash. Resolution precedence per-combo `config.disableSessionStickiness` → global `settings.disableSessionStickiness` → default `false` (preserves the #3825 prompt-cache/504 fix); gates **both** stickiness call sites in `open-sse/services/combo.ts`. Exposed as a global toggle (Combo Defaults) and a per-combo Inherit/on/off control. ([#6168](https://github.com/diegosouzapw/OmniRoute/issues/6168)) Regression guard: `tests/unit/combo-disable-session-stickiness.test.ts`. (thanks @RCrushMe)
+- **feat(docker):** add the `OMNIROUTE_NO_SUDO` env flag for root-less / user-namespaced deployments — the MITM cert-trust command path (`resolveSudoSpawn` in `src/mitm/systemCommands.ts`) now strips the leading `sudo` when the flag is truthy, in addition to the existing root / sudo-missing cases, so the Proxy Agent runs without `sudo` (the operator trusts the CA manually, e.g. via `NODE_EXTRA_CA_CERTS`). Argv-array `spawn` preserved — no shell interpolation (Hard Rule #13). ([#6122](https://github.com/diegosouzapw/OmniRoute/issues/6122)) Regression guard: `tests/unit/mitm-systemCommands-no-sudo.test.ts`. (thanks @powellnorma)
+- **feat(providers):** add **Requesty** as an OpenAI-compatible gateway provider (BYOK, base `https://router.requesty.ai/v1`, ~200 free requests/day) — wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/requesty/`, `src/shared/constants/providers/apikey/gateways.ts`). ([#6120](https://github.com/diegosouzapw/OmniRoute/issues/6120)) Regression guard: `tests/unit/requesty-provider.test.ts`. (thanks @chirag127)
+- **feat(dashboard):** add **configured-only / available-only filters** to the Free Provider Rankings page ([#6150](https://github.com/diegosouzapw/OmniRoute/issues/6150)) — hide providers you haven't configured, or whose connections are all rate-limited / out of quota, via server-side query params (`?configuredOnly` / `?availableOnly` on `GET /api/free-provider-rankings`) backed by a testable lib helper reusing the in-process connection state (no Redis). Both filters default off, so the default view is unchanged; this supersedes the earlier client-side "Configured Only" toggle (#6245) with an available-only dimension and unit-tested logic. Regression guard: `tests/unit/freeProviderRankings-filters.test.ts`.
+
+### 🔧 Bug Fixes
+
+- **fix(dashboard):** adding a second API key connection for the same provider no longer silently overwrites the first — the Add-API-key modal now derives a unique default connection name (`main`, then `main-2`, `main-3`, …) so the backend name-based upsert can't collide (#6499 — thanks @dilneiss).
+- **fix(compression):** the session-dedup engine now also deduplicates a large multi-line block repeated **within a single message** (intra-message dedup), not just across turns; the compression-preview API surfaces a `fallbackReason`, and the fusion panel reports how many models were rate-limited vs failed on total-panel failure (#6501 — thanks @chirag127).
+- **fix(compression):** a stacked-pipeline step naming an unregistered engine now surfaces a `validationErrors` entry instead of silently no-op'ing, so misconfigured pipelines are visible in the preview API (#6506 — thanks @chirag127).
+- **feat(usage):** add a **Codex reset-credit redemption** flow to the Provider Limits UI ([#6361](https://github.com/diegosouzapw/OmniRoute/pull/6361)) — a `useCodexResetCreditRedemption` hook + `/api/usage/codex-reset-credit` route + `codexResetCredits` lib let you redeem banked Codex reset credits from the quota card. Regression guard: `tests/unit/codex-reset-credits.test.ts`. (thanks @JxnLexn)
+- **feat(glm):** add **team-plan quota settings** for `glm-cn` connections ([#6351](https://github.com/diegosouzapw/OmniRoute/pull/6351)) — a dedicated `GlmTeamQuotaFields` form section (team quota id / limits) threaded through the Add/Edit connection modals, persisted via `providerSpecificData`, with the GLM usage service reading the team quota. Regression guards: `tests/unit/glm-team-quota.test.ts`, `provider-specific-data-schema.test.ts`. (thanks @hao3039032)
+- **feat(providers):** add **TinyFish** web-fetch/search support ([#6349](https://github.com/diegosouzapw/OmniRoute/pull/6349)) — a `tinyfish-fetch` executor + `/v1/web/fetch` route + MCP web-fetch tool, registered as a specialty-media provider with request-validation and a search-provider catalog entry. Regression guards: `tests/unit/executor-tinyfish-fetch.test.ts`, `web-fetch-handler.test.ts`, `mcp-web-fetch-tool.test.ts`, `provider-validation-tinyfish.test.ts`. (thanks @dtybnrj)
+- **fix(cli):** `omniroute launch-codex` now spawns `codex.cmd` through a shell on Windows (the npm `.cmd` shim is unresolvable by bare `spawn` → ENOENT), mirroring the qodercli Windows fix (#6263) ([#6312](https://github.com/diegosouzapw/OmniRoute/pull/6312)). Regression guard: `tests/unit/launch-codex-windows-spawn-6312.test.ts`. (thanks @swingtempo)
+- **fix(codex):** isolate the **Spark** quota from the shared Codex quota and stabilize the quota UI ordering / hydration so per-scope limits render consistently ([#6336](https://github.com/diegosouzapw/OmniRoute/pull/6336)). Regression guards: `tests/unit/codex-quota-selection-hydration.test.ts`, `provider-limits-ui.test.ts` + 3 more. (thanks @xz-dev)
+- **feat(api):** add a `hidePaidModels` setting that filters paid-only models out of the `/v1/models` catalog. Regression guard: `tests/unit/models-catalog-hide-paid.test.ts`. (thanks @chirag127)
+- **fix(api-manager):** the fallback model picker now preserves combos instead of dropping them when a primary model is unavailable ([#6443](https://github.com/diegosouzapw/OmniRoute/pull/6443)). Regression guard: `tests/unit/api-manager-page-static.test.ts`. (thanks @jmengit)
+- **fix(providers):** recoverable Antigravity / Cloud-Code (Gemini Code Assist) `403` responses ([#6452](https://github.com/diegosouzapw/OmniRoute/pull/6452)) are now classified as a retryable project-config error instead of a terminal account ban, so a fixable project/API-disabled 403 no longer forces a ~1-year cooldown / full OAuth reconnect. Regression guard: `tests/unit/errorclassifier-antigravity-403.test.ts`. (thanks @developerjillur)
+- **fix(mitm):** `sanitizeHeaders` now redacts `Set-Cookie` response headers so upstream session cookies never leak into logs / diagnostics ([#6451](https://github.com/diegosouzapw/OmniRoute/pull/6451)). Regression guard: `tests/unit/mitm-sanitize-headers.test.ts`. (thanks @developerjillur)
+- **fix(api):** `/api/compression/preview` now accepts `mode: "caveman"` and correctly handles stacked / zero-compression previews ([#6425](https://github.com/diegosouzapw/OmniRoute/issues/6425)). Regression guard: `tests/unit/api/compression-preview-caveman-and-stacked-6425.test.ts`. (thanks @chirag127)
+- **feat(providers):** add **Zed** hosted LLM aggregator as a native-app provider ([#6118](https://github.com/diegosouzapw/OmniRoute/pull/6118)) — OAuth sign-in via the Zed hosted flow, registered through the shared provider registry + executor. Regression guards: `tests/unit/zed-oauth-provider.test.ts`, `zed-import-utils.test.ts`, `zed-docker-detect.test.ts`, `mitm-handler-zed.test.ts`. VPS-validated via live operator login (Hard Rule #18).
+- **fix(oauth):** the Kiro SSO-cache auto-import now **preserves the IDC region** — cross-region Amazon Q / Kiro profiles imported from the SSO cache are no longer collapsed to the default region ([#6113](https://github.com/diegosouzapw/OmniRoute/pull/6113)). Regression guard: `tests/unit/kiro-auto-import-idc-2059.test.ts`. VPS-validated via live operator login (Hard Rule #18).
+- **fix(dashboard):** passthrough model aliases no longer collide when two namespaced model ids share a last segment (port from 9router#1850, [#6431](https://github.com/diegosouzapw/OmniRoute/pull/6431)). `enx/gpt-5.5` and `enx/codebuddy/gpt-5.5` both auto-generated the alias `gpt-5.5`, so the second model could never be added (the UI just alerted "alias already exists"). Aliases are now disambiguated deterministically — bare last segment when free, then parent-qualified (`codebuddy-gpt-5.5`), then a numeric suffix — while re-adding the exact same model id is still blocked. Regression guard: `tests/unit/passthrough-alias-1850.test.ts`. (thanks @arpicato)
+- **fix(translator):** preserve a Gemini `functionResponse` co-located with other parts (another `functionCall`, or trailing `text`) in the same content when translating **Gemini → OpenAI** ([#6376](https://github.com/diegosouzapw/OmniRoute/pull/6376)). `convertGeminiContent()` early-returned the tool message on the first `functionResponse` part, dropping any co-located parts; such contents are now pre-split (one tool message per `functionResponse`, emitted first, plus one message for the remaining parts). Regression guard: `tests/unit/gemini-to-openai-function-response.test.ts`. (thanks @warelik)
+- **fix(headroom):** detect a python interpreter managed by **mise / pyenv / asdf / conda** (port from 9router#2353, [#6382](https://github.com/diegosouzapw/OmniRoute/pull/6382)). Headroom's python probe (`src/lib/headroom/detect.ts`) searched a hardcoded `PATH`, but version managers expose their interpreters via shim dirs that only join `PATH` through interactive-shell activation — which the non-interactive server never runs, so a managed python (≥3.10) was invisible and Headroom reported it missing. The search path now prepends the well-known shim/bin dirs (`~/.local/share/mise/shims`, `~/.pyenv/shims`, `~/.asdf/shims`, `$CONDA_PREFIX/bin`, `~/.local/bin`, respecting `MISE_DATA_DIR`/`PYENV_ROOT`/`ASDF_DATA_DIR` when set), and a new `HEADROOM_PYTHON` env override lets operators point straight at their interpreter (mirroring `HEADROOM_URL`). Still shell-free (`execFileSync`). Regression guard: `tests/unit/headroom-detect.test.ts` (5). (thanks @loopyd)
+- **fix(executors):** strip the OpenAI-Codex/Claude-CLI `client_metadata` passthrough field for **NVIDIA** requests (port from 9router#1887, [#6411](https://github.com/diegosouzapw/OmniRoute/pull/6411)). NVIDIA's OpenAI-compatible wrapper rejects it with `400 Unsupported parameter`, the same class already handled for `cerebras`/`mistral`; `nvidia` (executor `default`) was missing from the strip allowlist so Codex/Claude-Code passthrough requests 400'd. Regression guard: `tests/unit/executor-default-strip-client-metadata.test.ts` (+nvidia case). (thanks @phidinhmanh)
+- **fix(translator):** strip the Claude-style `thinking` field for **NVIDIA `z-ai/glm-5.2`** (port from 9router#2023, [#6413](https://github.com/diegosouzapw/OmniRoute/pull/6413)). NVIDIA's OpenAI-compatible wrapper 400s on `thinking` (a Claude-format client routed here leaves a `thinking:{type:"adaptive"}`); the existing strip rule only dropped `reasoning`. Same class already handled for `minimax-m2.7`. Regression guard: `tests/unit/nvidia-minimax-thinking-strip.test.ts` (+glm-5.2 case). (thanks @phidinhmanh)
+- **fix(translator):** suppress the streamed `</think>` close marker for the **Antigravity IDE** client (port from 9router#1061, [#6415](https://github.com/diegosouzapw/OmniRoute/pull/6415)). On thinking-only turns Antigravity rendered a bare `</think>` as the sole visible content, tripping its loop-detection and wasting requests. Antigravity's UA (`vscode/<v> (Antigravity/<v>)`) is added to the marker-suppress allowlist (alongside OpenCode); Claude Code / Cursor still get the marker, and `x-omniroute-thinking-marker: on` force-restores it. Regression guard: `tests/unit/think-close-marker-suppress-5245.test.ts`. (thanks @abdofallah)
+- **fix(executors):** strip nested `reasoning_content` from messages for **Mistral** (port from 9router#1649, [#6417](https://github.com/diegosouzapw/OmniRoute/pull/6417)). Mistral's API returns `422 extra_forbidden` when an assistant message carries `reasoning_content` (replayed thinking from a prior turn, e.g. via the Codex `/responses` path); the generic top-level 400 field-downgrade retry never covered the nested per-message field. `DefaultExecutor` now strips it for provider `mistral` only, so DeepSeek (which requires replayed `reasoning_content`) is unaffected. Regression guard: `tests/unit/mistral-strip-reasoning-content-1649.test.ts`. (thanks @xxy9468615)
+- **fix(executors):** strip the `client_metadata` passthrough field on the **OpenCode** path (port from 9router#1442, [#6418](https://github.com/diegosouzapw/OmniRoute/pull/6418)). OpenCode upstreams (e.g. `kimi-k2.6` via opencode-go) reject it with `400 "Extra inputs are not permitted, field: 'client_metadata'"`; the DefaultExecutor strip only covered cerebras/mistral and `OpencodeExecutor` extends `BaseExecutor` directly, so nothing removed it there. Regression guard: `tests/unit/opencode-strip-client-metadata-1442.test.ts`. (thanks @yanpaing007)
+- **fix(executors):** inject the `reasoning_content` echo for the native **Moonshot Kimi** provider (port from 9router#1480, [#6419](https://github.com/diegosouzapw/OmniRoute/pull/6419)). Kimi (executor `default`) is a thinking-mode upstream that 400s with "reasoning_content must be passed back" when a prior assistant turn lacks it; the placeholder injection was only wired into the OpenCode meta-provider, so direct multi-turn Kimi conversations failed. Scoped to `kimi` (gateway-served models matching the thinking-model name pattern are unaffected). Regression guard: `tests/unit/kimi-native-reasoning-injected-1480.test.ts`. (thanks @2220258345)
+- **fix(executors):** recover from a strict gateway's `context_management: Extra inputs are not permitted` 400 (port from 9router#1468, [#6420](https://github.com/diegosouzapw/OmniRoute/pull/6420)). **Claude Code** always sends a top-level `context_management` field; strict anthropic-compatible gateways reject it. The dedicated context-editing 400-fallback only fired when OmniRoute's own `contextEditing` feature was enabled (default off), so a client-sent field passed through untouched and 400'd. `context_management` is now in the generic reactive field-strip list, so it's stripped-and-retried once regardless of the feature flag (with correct request re-signing for claude-compatible relays). Regression guard: `tests/unit/provider-field-strips.test.ts`. (thanks @ohahe52-dot)
+- **fix(network):** enable **RFC 8305 Happy Eyeballs** (`autoSelectFamily`) on the direct-egress undici dispatcher (port from 9router#1237, [#6423](https://github.com/diegosouzapw/OmniRoute/pull/6423)). When DNS returns both IPv6 (AAAA) and IPv4 (A) and the IPv6 route is broken (e.g. a NAT64 `64:ff9b::` prefix without routing), undici tried IPv6 first and hung until `ETIMEDOUT` (then a 502 + account lockout), even though `curl` reached the same host. The direct dispatcher now races both families and uses whichever connects first. Proxy paths pin family via `proxyTls` and are unaffected. Regression guard: `tests/unit/direct-dispatcher-pipelining-4580.test.ts`. (thanks @adentdk)
+- **fix(combo):** round-robin now advances the rotation pointer past the model that **actually served**, not the eagerly-scheduled one (port from 9router#948, [#6428](https://github.com/diegosouzapw/OmniRoute/pull/6428)). With `stickyLimit: 1` (true round-robin), when the scheduled model failed and a _different_ model served via fallback, the counter had already advanced +1 from the scheduled index — so the next request reused the fallback-served model, degrading round-robin into hot-spotting on whichever model was healthy. The pointer now advances to the served index + 1 (mirroring the sticky-limit>1 path). Session-stickiness (#3825) and distribution are preserved. Regression guard: `tests/unit/combo-rr-fallback-advance-948.test.ts`. (thanks @binsarjr)
+- **fix(sse):** a non-string `model` field is now rejected with a `400` before the resolver, instead of crashing downstream `.toLowerCase()`/`.split()` calls into an empty-body `500` that escapes the error sanitizer ([#6407](https://github.com/diegosouzapw/OmniRoute/issues/6407)). Regression guard: `tests/unit/chat-non-string-model-6407.test.ts`. (thanks @chirag127)
+- **fix(api):** unknown `/api/*` routes now return a JSON `404` (instead of the dashboard HTML shell) and scalar chat params (`model`/`temperature`/etc.) are validated **before** the provider lookup so malformed requests fail fast with a clear `400` ([#6424](https://github.com/diegosouzapw/OmniRoute/issues/6424), [#6412](https://github.com/diegosouzapw/OmniRoute/issues/6412)). Regression guards: `tests/unit/api/api-catchall-json-404.test.ts`, `tests/unit/chat-early-schema-validation-6412.test.ts`. (thanks @chirag127)
+- **fix(api):** `/v1/chat/completions` now rejects a non-JSON `Content-Type` with a `400` before parsing the body ([#6414](https://github.com/diegosouzapw/OmniRoute/issues/6414)). Regression guard: `tests/unit/v1-chat-completions-content-type-6414.test.ts`. (thanks @chirag127)
+- **fix(api):** the `X-OmniRoute-Compression` response header is now echoed on `/v1/chat/completions` and `/v1/completions` ([#6422](https://github.com/diegosouzapw/OmniRoute/issues/6422)). Regression guard: `tests/unit/compression-header-echo-6422.test.ts`. (thanks @chirag127)
+- **fix(api):** concurrent `GET /v1/models` requests are coalesced into a single catalog build ([#6408](https://github.com/diegosouzapw/OmniRoute/issues/6408)). Regression guard: `tests/unit/v1-models-concurrent-6408.test.ts`. (thanks @chirag127)
+- **fix(api):** `/v1/completions` now echoes the requested `body.model` in its JSON + streamed responses ([#6429](https://github.com/diegosouzapw/OmniRoute/pull/6429)). Regression guard: `tests/unit/completions-body-model-echo.test.ts`. (thanks @chirag127)
+- **fix(api):** env-var master keys now see the full `/v1/models` catalog ([#6406](https://github.com/diegosouzapw/OmniRoute/issues/6406)). Regression guard: `tests/unit/models-catalog-envkey-6406.test.ts`. (thanks @chirag127)
+- **fix(api):** non-streaming `/v1/completions` responses now echo `body.model` aligned with the `X-OmniRoute-Model` header ([#6426](https://github.com/diegosouzapw/OmniRoute/issues/6426)). Regression guard: `tests/unit/v1-completions-model-header-match-6426.test.ts`. (thanks @chirag127)
+- **fix(api):** unknown `/v1/*` routes now return a JSON `404 not_found` instead of the Next.js dashboard HTML shell ([#6405](https://github.com/diegosouzapw/OmniRoute/issues/6405)). Regression guard: `tests/unit/api/v1-catchall-json-404.test.ts`. (thanks @chirag127)
+- **fix(api):** the per-connection provider models route now degrades to the shipped catalog when a provider's `/models` endpoint answers with a redirect ([#6267](https://github.com/diegosouzapw/OmniRoute/issues/6267)) — a `qwen-web` import failed with a raw `Redirect blocked … (307)` 503. `safeOutboundFetch` throws `REDIRECT_BLOCKED` on the 307, `getSafeOutboundFetchErrorStatus` maps it to 503, and `buildDiscoveryErrorFallbackResponse` treated every 503 as a hard error — so the non-empty `getModelsByProviderId("qwen-web")` catalog was never surfaced. A models-endpoint redirect is not a fixable-config error (unlike `URL_GUARD_BLOCKED`/`INVALID_URL`, which stay hard errors), so it now falls back to the local/cached catalog before the 503 short-circuit. General fix — covers any config-driven provider that 307s. Regression guard: `tests/unit/provider-models-qwen-web-redirect-6267.test.ts`. (thanks @chirag127)
+- **fix(api):** the per-connection provider models route (MCP `list_models_catalog` + the dashboard import view) now merges USER-ADDED custom models into its response ([#6247](https://github.com/diegosouzapw/OmniRoute/issues/6247)) — custom models live in the `key_value` namespace `customModels`, which the live REST `/api/v1/models` already merges, but `src/app/api/providers/[id]/models/route.ts` never read `getCustomModels`, so custom models were dropped on both the discovery-success and local_catalog paths. They are now folded into the returned model list (deduped by id, stamped `owned_by: provider`), fixing MCP + the dashboard import view in one place. Regression guard: `tests/unit/provider-models-custom-merge-6247.test.ts`. (thanks @RCrushMe)
+
+- **fix(providers):** GitLab Duo tool-calling follow-up turns no longer fail upstream with `422 {"detail":"Validation error"}` (tokens 0/0, rejected pre-inference). The #6234 tool-result-feedback fix serialized the **entire** multi-turn conversation into GitLab's single-file `code_suggestions` (`small_file`) generation endpoint — folded history that turn-N sent as an oversized `current_file.content_above_cursor` **and** duplicated verbatim into `user_instruction`, tripping the AI-Gateway's `small_file` validation guard. The executor now **bounds** that prompt: it keeps system + latest user message + the most-recent tool round (dropping older turns), caps oversized tool results, and stops duplicating the full prompt into `user_instruction` (which now carries only the short latest user message) — while still feeding the most-recent tool result back so the agent continues (`open-sse/executors/gitlab.ts`, [#6220](https://github.com/diegosouzapw/OmniRoute/issues/6220)). The unit test covers the bounding logic; the upstream 422→200 clearing is VPS-only (Hard Rule #18). Regression guard: `tests/unit/gitlab-tool-exchange-bounded-6220.test.ts`.
+
+- **fix(i18n):** the provider-detail (`/dashboard/providers/[id]`) connection-status filter labels no longer render as `__MISSING__:All` / `__MISSING__:Active` / `__MISSING__:Error` / `__MISSING__:Banned` / `__MISSING__:CreditsExhausted` in non-English locales (notably pt-BR) ([#6290](https://github.com/diegosouzapw/OmniRoute/issues/6290)). Root cause was **not** the namespace mismatch the issue guessed — the `providers.filter*` keys resolve correctly in `en.json`; the debt lived in the locale mirrors (`src/i18n/messages/*.json`), where these five keys carried the `__MISSING__:` sync sentinel in ~15 locales and were absent entirely in ~26 others, so next-intl found the key and echoed the sentinel verbatim. All 40 non-English/-Chinese mirrors now ship real translations for the five `providers.filter*` labels. Regression guard: `tests/unit/i18n-provider-filter-keys-6290.test.ts`. (thanks @diegosouzapw)
+
+- **fix(providers):** the `copilot-m365-web` streaming executor now emits `debug`-level WebSocket diagnostics ([#6210](https://github.com/diegosouzapw/OmniRoute/issues/6210)) — the outbound WS URL (with the `access_token` **redacted** via `redactWsUrl()`), handshake success/failure, and each received SignalR frame's `type`/`target`. Previously the streaming path logged nothing, so an empty `content:null` response (the M365 Education / Starter tier symptom fixed in #6234) was undiagnosable even at `APP_LOG_LEVEL=debug`. The change is debug-level and side-effect-free — it does not alter streaming behavior or the frame parser, and the token never reaches the logs. Regression guard: `tests/unit/copilot-m365-web-logging-6210.test.ts` (thanks @qpeyba)
+
+- **fix(resilience):** a round-robin combo no longer returns `503 all upstream accounts are unavailable` when a compatibility-rejected target is actually healthy ([#6238](https://github.com/diegosouzapw/OmniRoute/issues/6238)). `filterTargetsByRequestCompatibility` drops request-incompatible targets (tool/vision/structured-output unsupported, or below the required context window) **before** any availability check runs, and its `compatible.length === 0` safety net only fired when _all_ targets were filtered — not when the kept targets later all turned out runtime-unavailable (circuit-open / cooldown / no credentials). So a combo could 503 while a compat-rejected-but-healthy provider sat unused. `handleRoundRobinCombo` now keeps the compat-rejected set and, when every compat-kept target was skipped without a single real attempt, probes those rejected targets as a **last-resort fallback tier** (via the new pure `open-sse/services/combo/comboCompatFallback.ts`) before crystallizing the 503. Regression guard: `tests/unit/combo-roundrobin-compat-fallback-6238.test.ts`. (thanks @ThongAccount)
+
+- **fix(startup):** best-effort self-heal for a corrupted Turbopack dev cache on Windows ([#6289](https://github.com/diegosouzapw/OmniRoute/issues/6289)). On Windows, `pnpm dev` can fail at startup when Turbopack `mmap`s a persistent-cache SST file and the OS refuses the mapping (`os error 1455` — "paging file too small"), which Turbopack surfaces as a misleading `Module not found: Can't resolve '@/shared/utils/machine'`. This is a **known upstream Turbopack cache-corruption bug — not our code**. The dev launcher (`scripts/dev/run-next.mjs`) now wraps `nextApp.prepare()` and, when it rejects with that signature (`isTurbopackCacheCorruption` in the new `scripts/dev/turbopackCacheHeal.mjs`), purges `.build/next/**/cache/turbopack` and retries **once** with a clear log. **Caveat — best-effort only:** the corruption often surfaces as a runtime overlay rather than a `prepare()` rejection, so this cannot always intercept it; the reliable remedy remains manually deleting the Turbopack cache dir. Regression guard: `tests/unit/turbopack-cache-heal-6289.test.ts`. (thanks @chirag127)
+
+- **fix(providers):** qodercli PAT auth no longer fails with `spawn qodercli ENOENT` on Windows ([#6263](https://github.com/diegosouzapw/OmniRoute/issues/6263)) — `spawnQoderCli` spawned the bare `qodercli` name with `shell:false` and an unenriched env, so the npm `.cmd` wrapper under `%APPDATA%\npm` (a user-PATH directory) was never resolved. It now resolves the absolute `.cmd`/`.exe` path through the existing `getCliRuntimeStatus("qoder")` resolver in `src/shared/services/cliRuntime.ts` (memoized), spawns with `shell` when the target is a `.cmd`/`.bat`, and uses the cliRuntime-enriched env (PATH + PATHEXT + APPDATA); the ENOENT error now lists the searched paths plus the `CLI_QODER_BIN` override. End-to-end spawn on a real Windows host is host-only (Hard Rule #18); the path-resolution logic is unit-tested. Regression guard: `tests/unit/qodercli-windows-resolve-6263.test.ts`. (thanks @chirag127)
+
+- **fix(sse):** the reasoning-token buffer no longer inflates **probe-sized `max_tokens`** ([#6274](https://github.com/diegosouzapw/OmniRoute/issues/6274)) — Claude Code's `/model` capability check sends `max_tokens: 1`, but for a thinking-capable model with a large output cap (e.g. `glm-5.2`) the #3587 headroom heuristic (`max(current + 1000, ceil(current * 1.5))`) rewrote it to `1001` and forwarded that upstream, wasting tokens on a request that was never a genuine reasoning budget. `resolveReasoningBufferedMaxTokens()` (`open-sse/services/reasoningTokenBuffer.ts`) now short-circuits and returns the caller's value verbatim when it is below the new `REASONING_BUFFER_MIN_TRIGGER` (256) threshold — a tiny explicit limit is a probe, not a reasoning request. Real budgets still receive the #3587 headroom unchanged, and the guard runs after the existing capability checks so unknown / non-reasoning models keep returning `null`. Regression guard: `tests/unit/reasoning-token-buffer-6274.test.ts`. (thanks @brightfiscalband)
+
+- **fix(cli):** `omniroute reset-password` now works as a real subcommand, and password resets over piped (non-TTY) stdin actually apply ([#6261](https://github.com/diegosouzapw/OmniRoute/issues/6261), [#6258](https://github.com/diegosouzapw/OmniRoute/issues/6258)). Two coupled defects: (1) **#6261** — `bin/omniroute.mjs` routed everything through Commander with only two pre-Commander bypasses (`--mcp`, `reset-encrypted-columns`), so `omniroute reset-password` was rejected as an unknown command; only the separate `omniroute-reset-password` bin worked, while the docs falsely advertised the subcommand (incl. a bogus "legacy alias still works"). A pre-Commander bypass mirroring `reset-encrypted-columns` now dynamically imports `bin/reset-password.mjs` (which self-executes) before Commander parses; the three doc lines were corrected. (2) **#6258** — `bin/reset-password.mjs` issued two sequential `rl.question` prompts; under piped stdin the second read never settled at EOF, so `main()` never reached `resetManagementPassword` and the reset was a silent no-op (both prompts printed, no success, password unchanged). The CLI now detects non-TTY stdin and reads it once (first line = password, second line = confirm if present, else reused), adds a `--password-stdin` flag (entire stdin is the password, no confirmation), and exits `0` explicitly so the success line always flushes; interactive TTY behavior is unchanged. Regression guard: `tests/unit/reset-password-cli-6261-6258.test.ts` (3). (thanks @chirag127)
+
+- **fix(db):** the mass-migration **safety abort** now tells the operator how to bypass it and stops flooding the log ([#6260](https://github.com/diegosouzapw/OmniRoute/issues/6260)) — after restoring a backup that wiped the migration tracking table, `runMigrations()` threw the abort on every downstream `ensureDbInitialized()`, re-logging the full banner 11+ times, and the message never mentioned the existing `OMNIROUTE_MAX_PENDING_MIGRATIONS` escape hatch. The abort text now appends a bypass hint (set `OMNIROUTE_MAX_PENDING_MIGRATIONS=0` in `server.env` / `DATA_DIR/.env`), and a new `MigrationSafetyAbortError` is memoized so repeated calls in the same process throw the same instance and emit a single concise line instead of the full cascade. Regression guard: `tests/unit/migration-safety-abort-6260.test.ts`. (thanks @chirag127)
+
+- **fix(auth):** importing a **distinct** Codex/ChatGPT OAuth `auth.json` is no longer falsely rejected as "already exists" when it belongs to a different user in the same workspace ([#6301](https://github.com/diegosouzapw/OmniRoute/issues/6301)). `findExistingCodexConnection` (in `src/lib/oauth/utils/codexAuthImport.ts`) deduped **only** on `providerSpecificData.workspaceId === accountId`, where `accountId` is the shared `chatgpt_account_id`/`tokens.account_id` — so two members of the same ChatGPT Team collapsed onto a single connection (409 `duplicate_account`). The id_token's `https://api.openai.com/auth` claim carries a per-user `chatgpt_user_id` alongside the workspace id (the device-flow path already persisted it as `chatgptUserId`, but the import path did not). Now `parseAndValidateCodexAuth` extracts `userId` (`chatgpt_user_id` → `user_id` → JWT `sub`) into `ParsedCodexAuth`, the create/update paths persist `chatgptUserId` in `providerSpecificData` (mirroring `codex.ts`), and dedup keys on `workspaceId` **AND** `chatgptUserId` — with a backward-compat fallback to legacy accountId-only matching when no stored connection for that workspace records a `chatgptUserId`, so genuinely-same accounts still dedup. Regression guard: `tests/unit/codex-auth-import-userid-dedup-6301.test.ts` (4). (thanks @anungma)
+
+- **fix(providers):** importing models for the **venice-web** provider no longer fails with a red "Provider venice-web does not support models listing" ([#6269](https://github.com/diegosouzapw/OmniRoute/issues/6269)). `venice-web` is a web-cookie provider with an executor but no upstream `/v1/models` endpoint and no registry `models`, so the models route fell through to the tail `400`. Mirroring the `jules`/`linkup-search`/`ollama-search` fix (#5569), it now ships a static local catalog entry in `src/lib/providers/staticModels.ts` — seeding the current Venice lineup (`venice-uncensored`, `llama-3.3-70b`, `qwen3-235b`, `qwen3-4b`, `deepseek-r1-671b`; Venice rotates its catalog, see docs.venice.ai/models/overview) — so the route returns `200` with `source:"local_catalog"`, `intentional:true`. Regression guard: `tests/unit/static-models-venice-web-6269.test.ts`. (thanks @chirag127)
+
+- **fix(api):** the specialty model catalogs (`/v1/embeddings`, `/v1/images`, `/v1/music`, `/v1/videos` model lists) are now derived from the **unified catalog filtered by a predicate** (`getSpecialtyModelsResponse`) instead of ad-hoc per-route logic, so they consistently respect active-credential visibility and stay in sync with the main catalog ([#6303](https://github.com/diegosouzapw/OmniRoute/pull/6303)). Regression guard: `tests/unit/specialty-model-catalog-routes.test.ts`. (thanks @makcimbx)
+- **fix(api):** the agent-bridge server route now resolves the MITM manager via a **dynamic `import("@/mitm/manager.runtime")`** so Turbopack does not statically pull the stub (or over-bundle the manager), and the agent-skills generator anchors its output base path with `path.join(process.cwd(), …)` so Turbopack's static analyzer stops tracing the whole project root ([#6329](https://github.com/diegosouzapw/OmniRoute/issues/6329), [#6366](https://github.com/diegosouzapw/OmniRoute/pull/6366)). Regression guard: `tests/unit/agent-bridge-server-route-dynamic-import.test.ts`. (thanks @Iammilansoni)
+- **fix(api):** internal probes (combo-test, cloud-sync verify) now pick a **management-scoped / allow-all API key** instead of naively grabbing `getApiKeys()[0]` — a restricted `self:usage` first row made the probe fail with "Model X is not allowed for this API key" even when the combo path was healthy (`pickApiKeyForInternalUse` in `src/lib/db/apiKeys.ts`). The API-manager model editor also falls back to `/api/models?all=true` when `/v1/models` is catalog-protected ([#6372](https://github.com/diegosouzapw/OmniRoute/pull/6372)). Regression guard: `tests/unit/pick-internal-api-key-6372.test.ts`. (thanks @jmengit)
+- **fix(live-ws):** the Live Dashboard WebSocket server now **rejects on bind failure** (e.g. `EADDRINUSE` when the API bridge already holds the port) instead of letting the error surface as an unhandled `error` event that crash-loops the process — the `error` listener is attached to `wss` (not `server`) and releases the EventBus subscription on a failed start ([#6324](https://github.com/diegosouzapw/OmniRoute/issues/6324)). Regression guard: `tests/unit/live-ws-eaddrinuse-6324.test.ts`. (thanks @vinayakkulkarni)
+- **fix(dashboard):** the Home provider-topology widget now trusts the live provider-metrics snapshot — it uses `topology.errorProvider` and live `activeRequests` directly instead of re-deriving state from a stale `lastErrorAt` or applying a frontend timeout filter, so the topology reflects real-time provider health ([#6322](https://github.com/diegosouzapw/OmniRoute/pull/6322)). Regression guard: `tests/unit/home-provider-topology-live-state.test.ts`. (thanks @xz-dev)
+- **fix(sse):** strip zero-width markers from streamed **tool-call arguments** — a follow-up to [#5857](https://github.com/diegosouzapw/OmniRoute/pull/5857). That PR removed injected zero-width joiners (U+200D) from streamed assistant text/reasoning but deliberately left tool-call argument JSON byte-exact. The request-side obfuscation (`open-sse/services/claudeCodeObfuscation.ts`) injects ZWJ into agent words — including the temp path inside the Bash tool description — and Claude models copy that verbatim into generated commands, which are delivered as tool-call arguments rather than assistant text. As a result the ZWJ survived and corrupted code blocks (e.g. a temp path rendered with an invisible joiner). Now `open-sse/handlers/responseSanitizer.ts` strips zero-width code points from tool-call argument strings at every emit site (OpenAI non-stream/stream chat `tool_calls` + legacy `function_call`, native Responses `function_call` items, the OpenAI→Responses conversion, and the native Responses streaming `response.function_call_arguments.delta/.done` events). Only zero-width code points are removed; JSON structure and all other bytes stay identical (no parse/restringify), so normal arguments remain byte-exact. Regression guard: 6 new cases in `tests/unit/response-sanitizer.test.ts` (suite 50/50).
+
+- **fix(nodejs):** the default app log path now resolves under `DATA_DIR` (`~/.omniroute/logs/application/app.log`) instead of `process.cwd()` ([#6197](https://github.com/diegosouzapw/OmniRoute/issues/6197)) — the globally-installed CLI runs from an arbitrary working directory, so anchoring the default to cwd made file logging silently write to (or no-op under) an unrelated directory, contradicting the documented `.env.example` default. `getAppLogFilePath()` now computes the default lazily via the pure `resolveDataDir()` resolver (honours a per-process `DATA_DIR`, no directory-creation side effect); an explicit `APP_LOG_FILE_PATH` still wins. Regression guard: `tests/unit/logenv-datadir-path-6197.test.ts` (3).
+
+- **fix(docker):** AgentBridge/`startMitm` no longer aborts in containers/headless when the Antigravity-default DNS step can't write `/etc/hosts` ([#6127](https://github.com/diegosouzapw/OmniRoute/issues/6127)), and the privileged command's stderr now reaches `app.log` instead of only a bare exit code hitting the toast ([#6198](https://github.com/diegosouzapw/OmniRoute/issues/6198)). The default DNS step (`addDNSEntry`) was called unguarded while cert install and the two sibling DNS steps were each best-effort — in the runtime Docker image (`USER node`, no `sudo`, read-only `/etc/hosts`) it threw `Command failed with code 1` out of `startMitmInternal` and killed the whole start, discarding the stderr. The three DNS steps are extracted into a best-effort `provisionDnsEntries()` where each failure is logged with the full `err` (stderr included, folded in by `systemCommands.ts`) and never aborts the start. Regression guard: `tests/unit/mitm-dns-graceful-degrade-6127.test.ts` (4).
+
+- **fix(providers):** copilot-m365-web now supports the M365 Education "Starter / OfficeWebIncludedCopilot" tier and no longer returns an empty `content:null` stream ([#6210](https://github.com/diegosouzapw/OmniRoute/issues/6210)). Two gaps: (1) `buildWsUrl()` hardcoded the individual-consumer scenario (`OfficeWebPaidConsumerCopilot`, `isEdu=false`) — the EDU tier is now opt-in via `providerSpecificData.tier="edu"`, emitting `scenario=OfficeWebIncludedCopilot`/`isEdu=true` (the individual path is unchanged); (2) the EDU/GPT-5.5 path streams deltas via `arguments[0].writeAtCursor` (incremental) instead of only `messages[].text` (accumulated snapshots), which the parser dropped — a new `accumulateBotContent()` folds both formats, with `type:2 item.result.message` as a last-resort fallback. Regression guard: `tests/unit/copilot-m365-edu-writeatcursor-6210.test.ts` (10). (thanks @qpeyba)
+
+- **fix(providers):** GitLab Duo executor now feeds tool results back into the prompt instead of looping ([#6220](https://github.com/diegosouzapw/OmniRoute/issues/6220)) — `buildPrompt()` branched only on `system`/`user` and took `userParts.at(-1)`, silently dropping the `assistant{tool_calls}` + `tool{result}` turns the client appended, so the reconstructed prompt was byte-identical to turn 1 and the model re-emitted the same `<tool>` call forever. When a tool exchange is present the full conversation is now serialized, folding each tool result back keyed by its `tool_call_id`; simple conversations keep the legacy shape. Complements the tool_call emission from [#6051](https://github.com/diegosouzapw/OmniRoute/issues/6051) (the `kilo-duplicate` label was a false positive — different, sequential defect). Regression guard: `tests/unit/gitlab-tool-result-feedback-6220.test.ts` (4).
+
+- **fix(providers):** opencode-go/opencode-zen can now synthesize the OpenCode CLI identity headers Cloudflare requires on VPS egress ([#5997](https://github.com/diegosouzapw/OmniRoute/issues/5997)) — on a datacenter VPS, `opencode.ai/zen/go/v1/chat/completions` 403s (HTML challenge) requests lacking CLI identity, while the reporter's control curl proved that `User-Agent: opencode-cli/1.0.0` + `x-opencode-client: cli` + `x-opencode-project: default` + fresh request/session UUIDs succeed. Opt-in via `OPENCODE_SYNTHESIZE_CLI_HEADERS=true` (values overridable via `OPENCODE_GO_USER_AGENT`/`OPENCODE_USER_AGENT`/`OPENCODE_CLIENT`/`OPENCODE_PROJECT`); it fills only headers the client did not already send. Kept **off by default** — the forward-only path is deliberate (fabricating a wrong value risks upstream rejection; a prior dedup regressed with `opencode/local`), so this replaces the fragile local header-injection shim without changing default behavior. Regression guard: `tests/unit/opencode-cli-headers-synthesis-5997.test.ts` (6). (thanks @aleksesipenko)
+- fix(resilience): sticky session affinity now evicts and fails over to another account when the pinned account is exhausted/unavailable (#6219)
+
+- fix(sse): Responses API passthrough now drops internal commentary-phase output before forwarding to clients (gated by RESPONSES_PASSTHROUGH_DROP_COMMENTARY, default on) (#6199)
+
+- **fix(sse):** tool-call function schemas with a root `type: null` are now coerced to `type: "object"` before dispatch ([#6359](https://github.com/diegosouzapw/OmniRoute/issues/6359)) — clients like the Codex app emit `parameters: { type: null, ... }` for some tools, which OpenAI-compatible upstreams reject with `400 Invalid schema for function '...': schema must be a JSON Schema of 'type: "object"', got 'type: null'`, failing the whole request. `toolSchemaSanitizer` already stripped the null; it now re-adds the mandatory root `"object"` type (and empty `properties`/open `additionalProperties` when absent). Combinator roots (`anyOf`/`oneOf`/`allOf`) and explicit root types are left untouched. Regression guard: 5 new cases in `tests/unit/tool-schema-sanitizer.test.mjs`.
+
+- **fix(docker):** AgentBridge no longer fails to start on npm/Electron/VPS installs with "MITM manager stub reached at runtime" ([#6344](https://github.com/diegosouzapw/OmniRoute/issues/6344)) — v3.8.45 flipped the production bundler default to Turbopack, but `next.config.mjs` aliased `@/mitm/manager` to its Docker-only degraded stub **unconditionally**. That was harmless while Docker (which sets the alias intentionally for #3390 graceful degradation) was the sole Turbopack consumer, but once every artifact built with Turbopack the stub shipped to all non-Docker users and `startMitm` threw on the first Agent-Bridge start. The alias is now opt-in via `OMNIROUTE_MITM_STUB=1` (set only by the Dockerfile) through the shared `scripts/build/mitm-stub-flag.mjs` helper; default builds bundle the real manager. Regression guard: `tests/unit/mitm-stub-alias-6344.test.mjs` (4).
+
+- **fix(proxy):** stop the v3.8.44 proxy regression that leaked the real IP and disabled healthy proxies ([#6246](https://github.com/diegosouzapw/OmniRoute/issues/6246)). Two coupled defects from the new health scheduler: (1) **IP leak** — when a proxy assigned to a connection was marked `inactive`, resolution fell through to a **direct** egress instead of blocking, exposing the operator's real IP; (2) **over-deactivation** — the sweep flipped a proxy to `inactive` on the **first** failed probe and counted our own 5s timeout / a probe-target `5xx` as the proxy's fault, so healthy paid proxies vanished from egress selection ("my proxies are not being used anymore"). Fix: the sweep decision is extracted into a pure, network-free `decideProxyHealthAction` (`src/lib/proxyHealth/decision.ts`) — by default the health check now **only counts/logs and never downgrades status** (a proxy is downgraded/removed only with `PROXY_AUTO_REMOVE=true`, after `PROXY_AUTO_REMOVE_AFTER` **consecutive** conclusive failures); probes are classified tri-state so an inconclusive result (our timeout, or a `5xx` from the probe target) never penalizes the proxy, and the probe timeout is raised 5s→15s. Separately, `safeResolveProxy` now **fails closed** via the existing policy: a connection whose assigned proxy is dead is blocked instead of leaking direct (`hasBlockingProxyAssignment`), honoring the explicit `proxy off` toggles and the `PROXY_FAIL_OPEN=true` opt-out. Existing proxies stuck `inactive` by the old behavior need a one-time manual re-activate (the operator owns proxy status). Regression guards: `tests/unit/proxy-health-decide-action-6246.test.ts`, `tests/unit/proxy-assigned-unavailable-6246.test.ts`.
+
+- **fix(proxy):** make "Test All" read-only and add bulk enable/disable ([#6246](https://github.com/diegosouzapw/OmniRoute/issues/6246)). Complements the core fail-closed / scheduler fix (#6296) with the two remaining reporter asks. (1) The **"Test All" button** (`POST /api/settings/proxies/auto-test`) used to flip a proxy to `inactive` on a failed reachability probe; since the egress selector excludes `inactive` proxies, a flaky probe (an unreachable `httpbin.org`, a proxy that blocks `HEAD`, or a slow paid proxy) silently disabled every proxy that failed — "Test All" is now **read-only by default** (only the operator sets a proxy active/inactive; opt back into the legacy test-and-set with `PROXY_HEALTH_AUTO_DEACTIVATE=true`). (2) Adds a **bulk enable/disable** proxies endpoint + toolbar action (`POST /api/settings/proxies/batch-activate`) so an operator can re-activate proxies in one click. Regression guard: `tests/unit/proxy-health-6246.test.ts`. (thanks @tenshiak)
+
+- **chatcore (tools): stop the default 128-tool cap from silently dropping opencode's `task`/MCP tools.** opencode (used as an MCP/agent host) sends a large tool list; when it exceeds the speculative `MAX_TOOLS_LIMIT` (128) default, `truncateToolList` did a blind `tools.slice(0, 128)`, dropping every tool past index 128 — including opencode's built-in `task` tool (subagent launch) and many MCP tools, so models routed through OmniRoute could no longer spawn subagents or reach part of their tools. The cap exists to avoid upstream `400`s for providers with real hard limits (e.g. grok-cli 200), so it is kept for those: detection of the opencode client (`isOpencodeClient` — any `x-opencode-*` header, or `opencode` in the user-agent) now only bypasses the **speculative 128 default**, never a known provider ceiling. Precedence is explicit — a proactive/detected provider limit always truncates (even for opencode); otherwise opencode forwards its full tool list; otherwise the unchanged 128 default applies to every other client. Refactors `getEffectiveToolLimit` into `getKnownToolLimit(provider) ?? DEFAULT_LIMIT` (byte-identical for existing callers) and fixes a cosmetic debug-log that reported the truncated count instead of the original. Regression guard: `tests/unit/tool-limit-detector.test.ts`.
+
+- **fix(mitm):** the macOS MITM-cert install check now matches the system keychain again. `security find-certificate -a -Z` prints the SHA-1 as a colon-less hex string, but the installed-check compared it against `getCertFingerprint()`'s colon-separated form, so the substring match never hit — the cert was reported as not-installed and re-prompted for the sudo install on every run. Fingerprints are now normalized (colons stripped, upper-cased) on both sides via the extracted `macCertOutputHasFingerprint` helper. Regression guard: `tests/unit/mitm-cert-mac-fingerprint.test.ts`. ([#6204](https://github.com/diegosouzapw/OmniRoute/pull/6204), closes [#6134](https://github.com/diegosouzapw/OmniRoute/issues/6134) — thanks @rianonehub)
+
+- **fix(api):** `/v1/messages/count_tokens` now counts `tool_use`, `tool_result` and `thinking` content blocks (and array-form `system` prompts) in the local-estimation path, instead of only `text`. Real agentic conversations keep ~95% of their tokens inside tool results; the previous estimate returned near-zero for them, which silently broke Claude Code's auto-compaction (context grew past the window with no compaction until the upstream API rejected the request). The real provider-side count path is unchanged. Regression guard: `tests/unit/messages-count-tokens-route.test.ts`. ([#6221](https://github.com/diegosouzapw/OmniRoute/pull/6221) — thanks @luweiCN)
+
+- **fix(antigravity):** strip a trailing assistant prefill turn for Vertex Claude models to avoid upstream 400s ([#6114](https://github.com/diegosouzapw/OmniRoute/pull/6114)). Regression guard: `tests/unit/antigravity-claude-prefill-strip.test.ts`. (thanks @anki1kr)
+
+- **fix(security):** the mutable cloud-agent routes (`/api/cloud/credentials/update`, `/api/cloud/models/alias`) now require management auth instead of being treated as public. They were classified as public API routes, so a request without management credentials could update stored cloud-agent credentials and model aliases. They are removed from the public-route set, classified as management routes in the authz pipeline, and gated by `requireManagementAuth`; cloud **read**/auth routes stay public. Regression guards: `tests/unit/cloud-write-auth.test.ts`, `tests/unit/authz/classify.test.ts`, `tests/unit/public-api-routes.test.ts`. ([#6233](https://github.com/diegosouzapw/OmniRoute/pull/6233) — thanks @vittoroliveira-dev)
+
+- **refactor(dashboard):** extract the onboarding-wizard "Open provider details" link target into a pure, unit-tested `buildProviderDetailsHref(connection)` helper. The wizard already routes by `connection.id` (the node UUID) rather than the provider category slug (#6144/#6145); this hardens that behavior behind a tested helper that guards a missing id/connection. Regression guard: `tests/unit/provider-onboarding-href.test.ts`. ([#6166](https://github.com/diegosouzapw/OmniRoute/pull/6166) — thanks @KooshaPari)
+
+- **fix(security):** the doubao synthetic device-id generator now derives its digits via an unbiased crypto-random draw (rejection sampling over `crypto.randomBytes()`) instead of a `% 10` reduction, closing a CodeQL `js/biased-cryptographic-random` finding.
+- **fix(agentSkills):** the GitHub-skills generator now resolves `outputDir` to an absolute path before writing, fixing a regression introduced by #6366 (relative-to-cwd base path) that could write generated skill files to the wrong directory.
+- **fix(security):** `/api/keys/{id}/devices` now checks the HTTP method before auth/validation, returning a `405` for non-GET/DELETE verbs instead of a misleading `401`/`500` (closes a `dast-smoke` QUERY-method finding).
+- **fix(quality):** clear the last 2 heavy quality-gate reds on the release tip (cycle pre-flight).
+- **fix(mitm):** the test suite and CI can never mutate the OS trust store — `OMNIROUTE_SKIP_SYSTEM_TRUST=1` is set globally for tests/CI so `installCert`/`uninstallCert`/`installTproxyCa` skip the privileged OS dispatch ([#6310](https://github.com/diegosouzapw/OmniRoute/pull/6310); full detail is under the [3.8.45] section below — this branch received it via the parallel-cycle sync-back).
+- **fix(api):** `POST /api/github-skills` now Zod-validates its request body; documented the new quality-gate env vars and pinned the merge-integrity GitHub Actions to a commit SHA.
+- **fix(skills):** generate the missing `omni-github-skills` registry entry and align the agent-skills catalog-count tests (follow-up to #6186).
+- **fix(quality):** clear the cycle's 11 net-new ESLint errors and make `validate-release-green` suppressions-aware.
+
+### 📝 Maintenance
+
+- **i18n(it):** add 118 missing Italian (`it`) translations (net-additive — no existing keys dropped, valid JSON), improving Italian UI coverage. ([#6212](https://github.com/diegosouzapw/OmniRoute/pull/6212) — thanks @serverless83)
+- **chore(providers):** remove deprecated **MiMo V2** model entries from the catalogs (xiaomi-mimo, opencode-go, zenmux-free, audio TTS) — the upstream V2 line is superseded by MiMo V2.5; drops `mimo-v2-tts`, `mimo-v2-pro`, `mimo-v2-omni`, `mimo-v2-flash`, `mimo-v2-flash-free` and realigns the provider-catalog tests. ([#6248](https://github.com/diegosouzapw/OmniRoute/pull/6248) — thanks @backryun)
+- **chore(release):** ~50 commits on this branch are v3.8.45 pre-flight/hardening fixes and CI-perf work that landed here via the parallel-cycle sync-back (`sync-next-cycle.mjs`, Hard Rule #21) **after** the `v3.8.45` git tag was cut, and are already fully documented under the **[3.8.45]** section below — listed here only so the per-cycle commit-coverage check (`npm run release:uncovered`) doesn't flag them as gaps. Provider/catalog/UX/backend fixes: #6041, #6078, #6108, #6135, #6148, #6149, #6154, #6158, #6161, #6162, #6163, #6164, #6165, #6170, #6177, #6178, #6181, #6186, #6187, #6191, #6193, #6194, #6195, #6200, #6205, #6208, #6209, #6211, #6213, #6223, #6224, #6225, #6226, #6227, #6228, #6229, #6230, #6235, #6291, #6292. CI/release-pipeline work: #6167, #6203, #6214, #6215, #6218, #6273, #6275, #6283, #6284, #6285, #6300, #6305.
+- **chore(release):** additional zero-ref release-cycle plumbing on this branch, kept out of `release:uncovered` on purpose (no `#N` in the commit subject to cite): opening the v3.8.46 cycle, opening/closing the v3.8.45 cycle, the finalized [3.8.45] CHANGELOG i18n sync-back to 42 mirrors, the v3.8.45 cognitive/cyclomatic and file-size drift rebaselines, ESLint stale-suppression pruning (4,273 → 4,233), and clearing test-masking/docs-all pre-flight reds for v3.8.45.
+
+### ⚡ Performance & Infrastructure
+
+- **perf(release-green):** the pre-flight validator (`scripts/quality/validate-release-green.mjs`) now runs its 4 slow suites (unit / vitest / integration / pack-artifact) **concurrently** via `Promise.all` — pre-flight wall time drops from ~the sum of the suites to ~the slowest one (~30min saved per round; Phase 0 was the nº1 bottleneck of the v3.8.45 release benchmark, 2h54 of 6h34 e2e). Guard: `tests/unit/validate-release-green.test.ts` ("runs the slow suites CONCURRENTLY"). ([#6319](https://github.com/diegosouzapw/OmniRoute/pull/6319))
+- **fix(ci):** `scripts/release/sync-next-cycle.mjs` — two defects found live in its first production run (v3.8.45 Phase 5): (1) the `git()` helper's default 1 MiB `maxBuffer` crashed with `ENOBUFS` on `git show origin/main:CHANGELOG.md` (the CHANGELOG alone is >1 MiB) — widened to 64 MiB; (2) the i18n resync only propagated the `[NEXT]` (TBD) section, leaving the just-shipped finalized section as "— TBD" in all 42 mirrors — it now also syncs `[prevVersion]` bounded by the heading below it (new exported pure helper `versionAfter`). Guards: +5 tests in `tests/unit/sync-next-cycle.test.ts` (8/8). ([#6327](https://github.com/diegosouzapw/OmniRoute/pull/6327))
+- **test(ci):** concurrency-sensitive flaky tests are **quarantined into a serial pass** (`tests/unit/serial/`, `--test-concurrency=1`, appended to every unit runner incl. sharded variants — the serial pass is sharded too so concurrent shard jobs never self-collide). Initial set: `glm-coding-plan-monthly-3580`, `quota-division-blocks`, `provider-health-autopilot`, `combo-health-autopilot` — the class behind the ~28min CI wedges/re-runs (two live 1h+ wedges cancelled during this PR's own validation). Discovery + TIA gates track the new glob; systemic root cause (async logger writing after teardown) tracked in [#6360](https://github.com/diegosouzapw/OmniRoute/issues/6360). Guard: `tests/unit/test-serial-quarantine.test.ts` (4). ([#6347](https://github.com/diegosouzapw/OmniRoute/pull/6347))
+
+### 🙌 Contributors
+
+Thanks to everyone whose work landed in v3.8.46:
+
+| Contributor                                                    | PRs / Issues           |
+| -------------------------------------------------------------- | ---------------------- |
+| [@2220258345](https://github.com/2220258345)                   | direct commit / report |
+| [@abdofallah](https://github.com/abdofallah)                   | direct commit / report |
+| [@adentdk](https://github.com/adentdk)                         | direct commit / report |
+| [@aleksesipenko](https://github.com/aleksesipenko)             | direct commit / report |
+| [@anki1kr](https://github.com/anki1kr)                         | direct commit / report |
+| [@anungma](https://github.com/anungma)                         | direct commit / report |
+| [@arpicato](https://github.com/arpicato)                       | direct commit / report |
+| [@backryun](https://github.com/backryun)                       | #6248                  |
+| [@binsarjr](https://github.com/binsarjr)                       | direct commit / report |
+| [@brightfiscalband](https://github.com/brightfiscalband)       | direct commit / report |
+| [@chirag127](https://github.com/chirag127)                     | #6501, #6506           |
+| [@developerjillur](https://github.com/developerjillur)         | direct commit / report |
+| [@dilneiss](https://github.com/dilneiss)                       | #6499                  |
+| [@dtybnrj](https://github.com/dtybnrj)                         | direct commit / report |
+| [@Forcerecon](https://github.com/Forcerecon)                   | direct commit / report |
+| [@hao3039032](https://github.com/hao3039032)                   | direct commit / report |
+| [@Iammilansoni](https://github.com/Iammilansoni)               | #6150, #6245           |
+| [@jmengit](https://github.com/jmengit)                         | direct commit / report |
+| [@jordansilly77-stack](https://github.com/jordansilly77-stack) | direct commit / report |
+| [@JxnLexn](https://github.com/JxnLexn)                         | direct commit / report |
+| [@KooshaPari](https://github.com/KooshaPari)                   | #6166                  |
+| [@loopyd](https://github.com/loopyd)                           | direct commit / report |
+| [@luweiCN](https://github.com/luweiCN)                         | #6221                  |
+| [@makcimbx](https://github.com/makcimbx)                       | direct commit / report |
+| [@muflifadla38](https://github.com/muflifadla38)               | direct commit / report |
+| [@newnol](https://github.com/newnol)                           | direct commit / report |
+| [@ofekbetzalel](https://github.com/ofekbetzalel)               | direct commit / report |
+| [@ohahe52-dot](https://github.com/ohahe52-dot)                 | direct commit / report |
+| [@phidinhmanh](https://github.com/phidinhmanh)                 | direct commit / report |
+| [@powellnorma](https://github.com/powellnorma)                 | direct commit / report |
+| [@qpeyba](https://github.com/qpeyba)                           | direct commit / report |
+| [@RaviTharuma](https://github.com/RaviTharuma)                 | direct commit / report |
+| [@RCrushMe](https://github.com/RCrushMe)                       | direct commit / report |
+| [@rianonehub](https://github.com/rianonehub)                   | #6134, #6204           |
+| [@serverless83](https://github.com/serverless83)               | #6212                  |
+| [@swingtempo](https://github.com/swingtempo)                   | direct commit / report |
+| [@tenshiak](https://github.com/tenshiak)                       | direct commit / report |
+| [@ThongAccount](https://github.com/ThongAccount)               | direct commit / report |
+| [@UnrealAryan](https://github.com/UnrealAryan)                 | direct commit / report |
+| [@vinayakkulkarni](https://github.com/vinayakkulkarni)         | direct commit / report |
+| [@vittoroliveira-dev](https://github.com/vittoroliveira-dev)   | #6233                  |
+| [@warelik](https://github.com/warelik)                         | direct commit / report |
+| [@xxy9468615](https://github.com/xxy9468615)                   | direct commit / report |
+| [@xz-dev](https://github.com/xz-dev)                           | direct commit / report |
+| [@yanpaing007](https://github.com/yanpaing007)                 | direct commit / report |
+| [@diegosouzapw](https://github.com/diegosouzapw)               | maintainer             |
+
+---
+
 ## [3.8.45] — 2026-07-06
 
 ### ✨ New Features

docs/i18n/fr/CHANGELOG.md

@@ -6,6 +6,213 @@
 
 ## [3.8.31] — 2026-06-20
 
+## [3.8.46] — TBD
+
+### ✨ New Features
+
+- **feat(sse):** **hide paid-only models from `auto/*` routing** when `hidePaidModels` is on ([#6512](https://github.com/diegosouzapw/OmniRoute/issues/6512)) — follow-up to #6328/#6495. PR #6495 hid paid-only models from the `GET /v1/models` listing, but `auto/*` combos (`auto/best-coding`, `auto/glm`, …) could still pick a paid-only backend into their candidate pool → a 402/403 at request time. `createVirtualAutoCombo` now filters the candidate pool through the new pure `open-sse/services/autoCombo/paidModelFilter.ts` (`filterPaidOnlyCandidates`), applying the same free-model predicate #6495 uses in `catalog.ts` (`providerHasFreeModels(provider) && isFreeModel(provider, {id})`) whenever `settings.hidePaidModels === true`. Applied before the category/tier/family narrowing, so it covers every `auto/*` combo; an all-paid pool degrades to the existing graceful empty-pool path. **Opt-in — default OFF leaves the pool unchanged** (identity). Regression guard: `tests/unit/autoCombo/paid-model-filter-6512.test.ts` (4, incl. the default-off identity guard).
+- **feat(sse):** **provider-family auto combos** — `auto/glm`, `auto/minimax`, `auto/mimo`, `auto/zai`, `auto/gemma`, `auto/llama`, `auto/gemini` ([#6453](https://github.com/diegosouzapw/OmniRoute/issues/6453)) — new routable ids that materialize an on-demand virtual combo spanning whatever installed backends currently expose that model family, degrading gracefully as backends rotate. A new pure `open-sse/services/autoCombo/modelFamily.ts` (`detectModelFamily`) classifies by model-id prefix for six families; `zai` is instead resolved by provider id (z.ai's hosted API serves the same `glm-*` model ids as every other GLM backend, so `auto/zai` means "route to my z.ai backend specifically" vs `auto/glm`'s "any connected GLM backend"). Reuses the existing `createVirtualAutoCombo` on-demand materialization path (no DB writes) and the `/v1/models` catalog advertising loop. Regression guard: `tests/unit/autoCombo/provider-family-combos.test.ts` (11).
+- **feat(proxy):** native **proxy-pool round-robin / egress IP rotation** ([#6365](https://github.com/diegosouzapw/OmniRoute/issues/6365)) — a scope (global / provider / account) can now hold **multiple** proxies as a pool with a rotation strategy, so outbound requests cycle their egress IP instead of pinning one proxy per scope. Migration `117_proxy_pool_rotation.sql` lifts the `UNIQUE(scope, scope_id)` constraint (rebuild via the canonical rename/copy/drop; existing single assignments become 1-element pools) and adds a `proxy_scope_rotation` companion table holding the per-scope strategy + a persisted monotonic round-robin cursor. Strategies: `round-robin` (default, monotonic cursor — never `Math.random`), `random`, and `sticky-per-N-min`. Resolution (`resolveProxyForScopeFromRegistry` / `resolveProxyForConnectionFromRegistry`) now fetches the alive, position-ordered candidate set (unchanged `PROXY_ALIVE_PREDICATE`) and applies the strategy; an empty / all-dead pool still returns `null` — the #6246 fail-closed guard is **untouched** (never falls through to direct egress). Backend + DB only; dashboard pool-builder UI is a follow-up. Regression guard: `tests/unit/proxy-pool-rotation-6365.test.ts` (8, incl. fail-closed + backward-compat).
+- **feat(providers):** end-to-end **tool/function calling on the native Gemini `/v1beta` endpoint** ([#6222](https://github.com/diegosouzapw/OmniRoute/issues/6222)) — both directions of the Gemini↔OpenAI conversion now preserve tool data (previously silently dropped). Request side: `convertGeminiToInternal` (extracted to its own testable module) maps `tools[].functionDeclarations` → OpenAI `tools`, prior `functionCall` parts → assistant `tool_calls`, and `functionResponse` parts → `tool`-role messages. Response side: `convertOpenAIResponseToGemini` emits `parts[].functionCall {name,args}` from `message.tool_calls`, and the streaming `openAIChunkToGeminiChunk` accumulates fragmented `tool_calls` deltas by index into complete `functionCall` parts. The non-Gemini client paths (Claude, OpenAI-Responses) already preserved tool calls — this closes the gap specific to the native Gemini surface. Regression guard: `tests/unit/v1beta-gemini-tool-calling-6222.test.ts` (6, incl. a streaming SSE round-trip).
+- **feat(providers):** copilot-m365-web **enterprise / work tier** support ([#6334](https://github.com/diegosouzapw/OmniRoute/issues/6334)) — mirrors the EDU-tier pattern (#6210): `M365ConnectionParams` gains an `agent` field, a new opt-in `M365_ENTERPRISE_OVERRIDES` preset (`agent=work`, `scenario=officeweb`, `licenseType=Premium`) applies via `providerSpecificData.tier="enterprise"` (alias `"work"`), and `agent` is also overridable directly via `providerSpecificData.agent`. `buildWsUrl` was hardcoding `agent="web"` (the one enterprise-distinguishing param with no override path), so a Premium work account handshook then returned an empty stream. The individual and EDU paths are untouched. Kilo's dup flag vs #6210 (EDU tier) was a false positive — different tier. Regression guard: `tests/unit/copilot-m365-enterprise-6334.test.ts` (7). End-to-end confirmation on a real Premium work account is a live-VPS validation follow-up (Hard Rule #18). (thanks @Forcerecon)
+- **feat(api):** standardized, provider-agnostic **`effort` + `thinking` request params** ([#6241](https://github.com/diegosouzapw/OmniRoute/issues/6241)) — a thin standardization layer over the existing mature per-provider reasoning plumbing (no provider mapper touched). `providerChatCompletionSchema` gains a canonical `effort` (reusing the shared `none/low/medium/high/xhigh` vocabulary — the UI tiers `extra`/`max` collapse onto `xhigh`) and a boolean `thinking`. A pure `normalizeReasoningRequest` (wired once in `src/sse/handlers/chat.ts`, before any reasoning field is read) folds them onto the fields the translators already consume (`reasoning_effort` / `reasoning.effort` / `thinking`), so they fan out to Anthropic / Gemini / xAI / Responses — an explicit client `reasoning_effort` / object-shaped `thinking` always wins (backward-compatible). `/models` additively exposes `supportsThinking` + `effort_tiers` so the frontend can render the toggles (UI component is a follow-up). Regression guard: `tests/unit/effort-thinking-standardization-6241.test.ts` (12). (thanks @Iammilansoni, @shabeer)
+- **feat(combo):** new **`pipeline` (sequential) combo strategy** ([#6297](https://github.com/diegosouzapw/OmniRoute/issues/6297)) — the 18th routing strategy runs targets **in order**, threading each step's output into the next step's input, with an optional per-step `prompt` (system instruction); only the final step's response is returned. Distinct from `fusion` (parallel fan-out + judge). Implemented as a self-contained `open-sse/services/pipeline.ts` (sibling to `fusion.ts`), dispatched from `combo.ts`; the step list reuses `combo.models` order and reads an optional `prompt` off each target (backward-compatible — ignored by every other strategy). Intermediate steps run non-streaming with tools stripped (complete prose to thread forward); the final step keeps the client's `stream` flag + tools. A failing/empty/unparseable intermediate step fails the whole pipeline explicitly via a sanitized error (never silently swallowed). Kilo's dup flag vs #563 was a false positive (that's model→chain selection; this is a sequential chain). Regression guard: `tests/unit/combo-pipeline-strategy.test.ts` (5). (thanks @ofekbetzalel)
+- **feat(ci):** `check:test-masking` now flags **inline-reimplemented prod conditions** ([#6348](https://github.com/diegosouzapw/OmniRoute/issues/6348)) — a new **report-only** subcheck (v2, 6A.10 family) catches the wrong-shape contract test: a test that recomputes the condition under test inline instead of importing/exercising the real function (the #6216 class, where `=== 500` → `>= 500` stayed green because the test re-implemented the branch). For each added/modified test file it warns when the file textually duplicates a ≥3-token conditional from a production file touched in the same PR **and** does not import the symbol/module owning it, via a pure, fixture-tested `findReimplementedConditions()` with an allowlist mirroring `assertReductionAllowlist`. Report-only for now (does not fail the gate) — to be promoted to blocking after a triage cycle. Regression guard: `tests/unit/check-test-masking.test.ts` (45).
+- **feat(sse):** per-connection routing override (native vs CLIProxyAPI) ([#6339](https://github.com/diegosouzapw/OmniRoute/issues/6339)) — the previously-dead `isCliproxyapiDeepModeEnabled` helper is now wired into `resolveExecutorWithProxy`: a single connection can opt itself into the CLIProxyAPI passthrough executor via `providerSpecificData.cliproxyapiMode="claude-native"`, with precedence **connection override > provider `upstream_proxy_config` mode > default**. `resolveExecutorWithProxy` now receives the resolved connection's `providerSpecificData` (threaded from `chatCore.ts`), so one connection can deep-route while the provider's other connections stay native — no DB schema change (the toggle rides in `providerSpecificData`). Also resolves the same-provider mixing ask in #6340. Regression guard: `tests/unit/chatcore-executor-proxy.test.ts` (9). (thanks @RaviTharuma)
+- **feat(dashboard):** "Add session cookie" modal now shows a prominent **"Open ‹host› →"** link to the provider's own site ([#6268](https://github.com/diegosouzapw/OmniRoute/issues/6268)) — every `-web` cookie-session provider (chatgpt-web, claude-web, gemini-web, kimi-web, lmarena, qwen-web, m365-copilot-web, …) renders a one-click external link (opening the provider's login/home page in a new tab) so operators no longer tab away to retype the URL mid-setup. The host resolves from a pure, unit-tested `resolveWebProviderHost()` (prefers `WEB_COOKIE_PROVIDERS[id].website`, falls back to the registry `baseUrl` origin); non-web providers render exactly as before. Kilo's dup flag vs #6265 (modal-too-small-on-1080p) was a false positive — distinct concern. Regression guard: `tests/unit/resolve-web-provider-host.test.ts` (5). (thanks @chirag127)
+- **feat(providers):** add **DigitalOcean** AI (serverless inference) as an OpenAI-compatible API-key provider ([#6373](https://github.com/diegosouzapw/OmniRoute/pull/6373)) — base `https://inference.do-ai.run/v1`, wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/digitalocean/`, `src/shared/constants/providers/apikey/inference-hosts.ts`). Regression guard: `tests/unit/digitalocean-provider.test.ts`. (thanks @newnol)
+- **feat(providers):** add **Huancheng Public API** (`hcnsec`) as an OpenAI-compatible regional provider ([#6410](https://github.com/diegosouzapw/OmniRoute/pull/6410)) — Xinjiang Huancheng Cybersecurity's public LLM platform (base `https://api.hcnsec.cn/v1`, free credits via daily check-ins), wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/hcnsec/`, `src/shared/constants/providers/apikey/regional.ts`). Regression guard: `tests/unit/hcnsec-provider.test.ts`. (thanks @UnrealAryan)
+- **feat(dashboard):** the web-session credential guide now shows an **"Open {host}" link** ([#6316](https://github.com/diegosouzapw/OmniRoute/pull/6316)) to the provider's sign-in site (derived from the provider `website` via `getProviderWebsiteHost`), so you can jump straight to the page where the cookie/session must be captured. Regression guard: `tests/unit/web-session-provider-link-6316.test.ts`. (thanks @jordansilly77-stack)
+- **feat(cerebras):** add the **Gemma 4 31B** model (`gemma-4-31b`) to the Cerebras registry + pricing table ([#6331](https://github.com/diegosouzapw/OmniRoute/pull/6331)). Regression guard extends `tests/unit/t28-model-catalog-updates.test.ts`. (thanks @backryun)
+- **feat(providers):** add **Yuanbao (web)** as a cookie-session provider ([#6196](https://github.com/diegosouzapw/OmniRoute/issues/6196)) — `yuanbao-web` (Tencent Yuanbao, `yuanbao.tencent.com`) with cookie-only auth (`hy_user`/`hy_token` + public agent id), SSE→OpenAI translation incl. `reasoning_content`, exposing DeepSeek V3/R1 + Hunyuan / Hunyuan-T1. Regression guard: `tests/unit/providers-yuanbao-web.test.ts`. `together-web` was **deferred** (no verifiable web-session endpoint — needs a captured request) and `huggingchat-web` **dropped** (the existing `huggingchat` already is a web-cookie provider). (thanks @chirag127)
+- **feat(providers):** route the built-in **agentrouter** through the dynamic Claude-Code wire image ([#6056](https://github.com/diegosouzapw/OmniRoute/issues/6056)) — a small static allow-set (`CC_WIRE_IMAGE_BUILTINS` in `open-sse/services/ccWireImageBuiltins.ts`), consulted by `isClaudeCodeCompatible` / `isClaudeCodeCompatibleProvider` / `applyFingerprint`, makes agentrouter adopt the CC wire-image headers + fingerprint **while guarding the CC baseUrl/auth branches** so it keeps its own registry `baseUrl` and `x-api-key` auth. Regression guard: `tests/unit/agentrouter-cc-wire-image.test.ts` (asserts the wire image is applied AND agentrouter's baseUrl/auth are preserved). Live WAF-acceptance against agentrouter.org is a VPS validation follow-up (Hard Rule #18).
+- **feat(providers):** **bulk-add API keys for Cloudflare Workers AI** ([#6174](https://github.com/diegosouzapw/OmniRoute/issues/6174)) — `cloudflare-ai` is removed from the bulk-add exclusion list and the bulk parser gains a 3-field `name|accountId|apiKey` mode; the bulk route now builds a **per-entry** `providerSpecificData` so each key carries its own `accountId` (fixing the previous shared-object reuse), and both the create + key-validation paths receive it. Regression guard: `tests/unit/bulk-api-key-parser-cloudflare.test.ts`. (thanks @muflifadla38)
+- **feat(dashboard):** routing/settings UX clarity ([#6147](https://github.com/diegosouzapw/OmniRoute/issues/6147)) — (1) weighted combos show the **effective routing share %** next to each weight when weights don't sum to 100 (`WeightTotalBar.tsx`); (2) the status widget's user-facing **"Cloud Sync" label is renamed** to "Remote Settings Sync" (`CloudSyncStatus.tsx`; internal ids/state untouched); (3) built-in providers gain an **opt-in advanced base-URL override** (`isBaseUrlOverrideEligibleProvider`, hidden behind an "Advanced" toggle, reusing the existing `providerSpecificData.baseUrl` persistence — not globally widened). Regression guard: `tests/unit/routing-settings-ux-6147.test.ts`.
+- **feat(combo):** add an option to **disable session stickiness**, per-combo or globally — round-robin / random combos can rotate to a different connection on every request instead of pinning a whole conversation to one connection by its first-message hash. Resolution precedence per-combo `config.disableSessionStickiness` → global `settings.disableSessionStickiness` → default `false` (preserves the #3825 prompt-cache/504 fix); gates **both** stickiness call sites in `open-sse/services/combo.ts`. Exposed as a global toggle (Combo Defaults) and a per-combo Inherit/on/off control. ([#6168](https://github.com/diegosouzapw/OmniRoute/issues/6168)) Regression guard: `tests/unit/combo-disable-session-stickiness.test.ts`. (thanks @RCrushMe)
+- **feat(docker):** add the `OMNIROUTE_NO_SUDO` env flag for root-less / user-namespaced deployments — the MITM cert-trust command path (`resolveSudoSpawn` in `src/mitm/systemCommands.ts`) now strips the leading `sudo` when the flag is truthy, in addition to the existing root / sudo-missing cases, so the Proxy Agent runs without `sudo` (the operator trusts the CA manually, e.g. via `NODE_EXTRA_CA_CERTS`). Argv-array `spawn` preserved — no shell interpolation (Hard Rule #13). ([#6122](https://github.com/diegosouzapw/OmniRoute/issues/6122)) Regression guard: `tests/unit/mitm-systemCommands-no-sudo.test.ts`. (thanks @powellnorma)
+- **feat(providers):** add **Requesty** as an OpenAI-compatible gateway provider (BYOK, base `https://router.requesty.ai/v1`, ~200 free requests/day) — wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/requesty/`, `src/shared/constants/providers/apikey/gateways.ts`). ([#6120](https://github.com/diegosouzapw/OmniRoute/issues/6120)) Regression guard: `tests/unit/requesty-provider.test.ts`. (thanks @chirag127)
+- **feat(dashboard):** add **configured-only / available-only filters** to the Free Provider Rankings page ([#6150](https://github.com/diegosouzapw/OmniRoute/issues/6150)) — hide providers you haven't configured, or whose connections are all rate-limited / out of quota, via server-side query params (`?configuredOnly` / `?availableOnly` on `GET /api/free-provider-rankings`) backed by a testable lib helper reusing the in-process connection state (no Redis). Both filters default off, so the default view is unchanged; this supersedes the earlier client-side "Configured Only" toggle (#6245) with an available-only dimension and unit-tested logic. Regression guard: `tests/unit/freeProviderRankings-filters.test.ts`.
+
+### 🔧 Bug Fixes
+
+- **fix(dashboard):** adding a second API key connection for the same provider no longer silently overwrites the first — the Add-API-key modal now derives a unique default connection name (`main`, then `main-2`, `main-3`, …) so the backend name-based upsert can't collide (#6499 — thanks @dilneiss).
+- **fix(compression):** the session-dedup engine now also deduplicates a large multi-line block repeated **within a single message** (intra-message dedup), not just across turns; the compression-preview API surfaces a `fallbackReason`, and the fusion panel reports how many models were rate-limited vs failed on total-panel failure (#6501 — thanks @chirag127).
+- **fix(compression):** a stacked-pipeline step naming an unregistered engine now surfaces a `validationErrors` entry instead of silently no-op'ing, so misconfigured pipelines are visible in the preview API (#6506 — thanks @chirag127).
+- **feat(usage):** add a **Codex reset-credit redemption** flow to the Provider Limits UI ([#6361](https://github.com/diegosouzapw/OmniRoute/pull/6361)) — a `useCodexResetCreditRedemption` hook + `/api/usage/codex-reset-credit` route + `codexResetCredits` lib let you redeem banked Codex reset credits from the quota card. Regression guard: `tests/unit/codex-reset-credits.test.ts`. (thanks @JxnLexn)
+- **feat(glm):** add **team-plan quota settings** for `glm-cn` connections ([#6351](https://github.com/diegosouzapw/OmniRoute/pull/6351)) — a dedicated `GlmTeamQuotaFields` form section (team quota id / limits) threaded through the Add/Edit connection modals, persisted via `providerSpecificData`, with the GLM usage service reading the team quota. Regression guards: `tests/unit/glm-team-quota.test.ts`, `provider-specific-data-schema.test.ts`. (thanks @hao3039032)
+- **feat(providers):** add **TinyFish** web-fetch/search support ([#6349](https://github.com/diegosouzapw/OmniRoute/pull/6349)) — a `tinyfish-fetch` executor + `/v1/web/fetch` route + MCP web-fetch tool, registered as a specialty-media provider with request-validation and a search-provider catalog entry. Regression guards: `tests/unit/executor-tinyfish-fetch.test.ts`, `web-fetch-handler.test.ts`, `mcp-web-fetch-tool.test.ts`, `provider-validation-tinyfish.test.ts`. (thanks @dtybnrj)
+- **fix(cli):** `omniroute launch-codex` now spawns `codex.cmd` through a shell on Windows (the npm `.cmd` shim is unresolvable by bare `spawn` → ENOENT), mirroring the qodercli Windows fix (#6263) ([#6312](https://github.com/diegosouzapw/OmniRoute/pull/6312)). Regression guard: `tests/unit/launch-codex-windows-spawn-6312.test.ts`. (thanks @swingtempo)
+- **fix(codex):** isolate the **Spark** quota from the shared Codex quota and stabilize the quota UI ordering / hydration so per-scope limits render consistently ([#6336](https://github.com/diegosouzapw/OmniRoute/pull/6336)). Regression guards: `tests/unit/codex-quota-selection-hydration.test.ts`, `provider-limits-ui.test.ts` + 3 more. (thanks @xz-dev)
+- **feat(api):** add a `hidePaidModels` setting that filters paid-only models out of the `/v1/models` catalog. Regression guard: `tests/unit/models-catalog-hide-paid.test.ts`. (thanks @chirag127)
+- **fix(api-manager):** the fallback model picker now preserves combos instead of dropping them when a primary model is unavailable ([#6443](https://github.com/diegosouzapw/OmniRoute/pull/6443)). Regression guard: `tests/unit/api-manager-page-static.test.ts`. (thanks @jmengit)
+- **fix(providers):** recoverable Antigravity / Cloud-Code (Gemini Code Assist) `403` responses ([#6452](https://github.com/diegosouzapw/OmniRoute/pull/6452)) are now classified as a retryable project-config error instead of a terminal account ban, so a fixable project/API-disabled 403 no longer forces a ~1-year cooldown / full OAuth reconnect. Regression guard: `tests/unit/errorclassifier-antigravity-403.test.ts`. (thanks @developerjillur)
+- **fix(mitm):** `sanitizeHeaders` now redacts `Set-Cookie` response headers so upstream session cookies never leak into logs / diagnostics ([#6451](https://github.com/diegosouzapw/OmniRoute/pull/6451)). Regression guard: `tests/unit/mitm-sanitize-headers.test.ts`. (thanks @developerjillur)
+- **fix(api):** `/api/compression/preview` now accepts `mode: "caveman"` and correctly handles stacked / zero-compression previews ([#6425](https://github.com/diegosouzapw/OmniRoute/issues/6425)). Regression guard: `tests/unit/api/compression-preview-caveman-and-stacked-6425.test.ts`. (thanks @chirag127)
+- **feat(providers):** add **Zed** hosted LLM aggregator as a native-app provider ([#6118](https://github.com/diegosouzapw/OmniRoute/pull/6118)) — OAuth sign-in via the Zed hosted flow, registered through the shared provider registry + executor. Regression guards: `tests/unit/zed-oauth-provider.test.ts`, `zed-import-utils.test.ts`, `zed-docker-detect.test.ts`, `mitm-handler-zed.test.ts`. VPS-validated via live operator login (Hard Rule #18).
+- **fix(oauth):** the Kiro SSO-cache auto-import now **preserves the IDC region** — cross-region Amazon Q / Kiro profiles imported from the SSO cache are no longer collapsed to the default region ([#6113](https://github.com/diegosouzapw/OmniRoute/pull/6113)). Regression guard: `tests/unit/kiro-auto-import-idc-2059.test.ts`. VPS-validated via live operator login (Hard Rule #18).
+- **fix(dashboard):** passthrough model aliases no longer collide when two namespaced model ids share a last segment (port from 9router#1850, [#6431](https://github.com/diegosouzapw/OmniRoute/pull/6431)). `enx/gpt-5.5` and `enx/codebuddy/gpt-5.5` both auto-generated the alias `gpt-5.5`, so the second model could never be added (the UI just alerted "alias already exists"). Aliases are now disambiguated deterministically — bare last segment when free, then parent-qualified (`codebuddy-gpt-5.5`), then a numeric suffix — while re-adding the exact same model id is still blocked. Regression guard: `tests/unit/passthrough-alias-1850.test.ts`. (thanks @arpicato)
+- **fix(translator):** preserve a Gemini `functionResponse` co-located with other parts (another `functionCall`, or trailing `text`) in the same content when translating **Gemini → OpenAI** ([#6376](https://github.com/diegosouzapw/OmniRoute/pull/6376)). `convertGeminiContent()` early-returned the tool message on the first `functionResponse` part, dropping any co-located parts; such contents are now pre-split (one tool message per `functionResponse`, emitted first, plus one message for the remaining parts). Regression guard: `tests/unit/gemini-to-openai-function-response.test.ts`. (thanks @warelik)
+- **fix(headroom):** detect a python interpreter managed by **mise / pyenv / asdf / conda** (port from 9router#2353, [#6382](https://github.com/diegosouzapw/OmniRoute/pull/6382)). Headroom's python probe (`src/lib/headroom/detect.ts`) searched a hardcoded `PATH`, but version managers expose their interpreters via shim dirs that only join `PATH` through interactive-shell activation — which the non-interactive server never runs, so a managed python (≥3.10) was invisible and Headroom reported it missing. The search path now prepends the well-known shim/bin dirs (`~/.local/share/mise/shims`, `~/.pyenv/shims`, `~/.asdf/shims`, `$CONDA_PREFIX/bin`, `~/.local/bin`, respecting `MISE_DATA_DIR`/`PYENV_ROOT`/`ASDF_DATA_DIR` when set), and a new `HEADROOM_PYTHON` env override lets operators point straight at their interpreter (mirroring `HEADROOM_URL`). Still shell-free (`execFileSync`). Regression guard: `tests/unit/headroom-detect.test.ts` (5). (thanks @loopyd)
+- **fix(executors):** strip the OpenAI-Codex/Claude-CLI `client_metadata` passthrough field for **NVIDIA** requests (port from 9router#1887, [#6411](https://github.com/diegosouzapw/OmniRoute/pull/6411)). NVIDIA's OpenAI-compatible wrapper rejects it with `400 Unsupported parameter`, the same class already handled for `cerebras`/`mistral`; `nvidia` (executor `default`) was missing from the strip allowlist so Codex/Claude-Code passthrough requests 400'd. Regression guard: `tests/unit/executor-default-strip-client-metadata.test.ts` (+nvidia case). (thanks @phidinhmanh)
+- **fix(translator):** strip the Claude-style `thinking` field for **NVIDIA `z-ai/glm-5.2`** (port from 9router#2023, [#6413](https://github.com/diegosouzapw/OmniRoute/pull/6413)). NVIDIA's OpenAI-compatible wrapper 400s on `thinking` (a Claude-format client routed here leaves a `thinking:{type:"adaptive"}`); the existing strip rule only dropped `reasoning`. Same class already handled for `minimax-m2.7`. Regression guard: `tests/unit/nvidia-minimax-thinking-strip.test.ts` (+glm-5.2 case). (thanks @phidinhmanh)
+- **fix(translator):** suppress the streamed `</think>` close marker for the **Antigravity IDE** client (port from 9router#1061, [#6415](https://github.com/diegosouzapw/OmniRoute/pull/6415)). On thinking-only turns Antigravity rendered a bare `</think>` as the sole visible content, tripping its loop-detection and wasting requests. Antigravity's UA (`vscode/<v> (Antigravity/<v>)`) is added to the marker-suppress allowlist (alongside OpenCode); Claude Code / Cursor still get the marker, and `x-omniroute-thinking-marker: on` force-restores it. Regression guard: `tests/unit/think-close-marker-suppress-5245.test.ts`. (thanks @abdofallah)
+- **fix(executors):** strip nested `reasoning_content` from messages for **Mistral** (port from 9router#1649, [#6417](https://github.com/diegosouzapw/OmniRoute/pull/6417)). Mistral's API returns `422 extra_forbidden` when an assistant message carries `reasoning_content` (replayed thinking from a prior turn, e.g. via the Codex `/responses` path); the generic top-level 400 field-downgrade retry never covered the nested per-message field. `DefaultExecutor` now strips it for provider `mistral` only, so DeepSeek (which requires replayed `reasoning_content`) is unaffected. Regression guard: `tests/unit/mistral-strip-reasoning-content-1649.test.ts`. (thanks @xxy9468615)
+- **fix(executors):** strip the `client_metadata` passthrough field on the **OpenCode** path (port from 9router#1442, [#6418](https://github.com/diegosouzapw/OmniRoute/pull/6418)). OpenCode upstreams (e.g. `kimi-k2.6` via opencode-go) reject it with `400 "Extra inputs are not permitted, field: 'client_metadata'"`; the DefaultExecutor strip only covered cerebras/mistral and `OpencodeExecutor` extends `BaseExecutor` directly, so nothing removed it there. Regression guard: `tests/unit/opencode-strip-client-metadata-1442.test.ts`. (thanks @yanpaing007)
+- **fix(executors):** inject the `reasoning_content` echo for the native **Moonshot Kimi** provider (port from 9router#1480, [#6419](https://github.com/diegosouzapw/OmniRoute/pull/6419)). Kimi (executor `default`) is a thinking-mode upstream that 400s with "reasoning_content must be passed back" when a prior assistant turn lacks it; the placeholder injection was only wired into the OpenCode meta-provider, so direct multi-turn Kimi conversations failed. Scoped to `kimi` (gateway-served models matching the thinking-model name pattern are unaffected). Regression guard: `tests/unit/kimi-native-reasoning-injected-1480.test.ts`. (thanks @2220258345)
+- **fix(executors):** recover from a strict gateway's `context_management: Extra inputs are not permitted` 400 (port from 9router#1468, [#6420](https://github.com/diegosouzapw/OmniRoute/pull/6420)). **Claude Code** always sends a top-level `context_management` field; strict anthropic-compatible gateways reject it. The dedicated context-editing 400-fallback only fired when OmniRoute's own `contextEditing` feature was enabled (default off), so a client-sent field passed through untouched and 400'd. `context_management` is now in the generic reactive field-strip list, so it's stripped-and-retried once regardless of the feature flag (with correct request re-signing for claude-compatible relays). Regression guard: `tests/unit/provider-field-strips.test.ts`. (thanks @ohahe52-dot)
+- **fix(network):** enable **RFC 8305 Happy Eyeballs** (`autoSelectFamily`) on the direct-egress undici dispatcher (port from 9router#1237, [#6423](https://github.com/diegosouzapw/OmniRoute/pull/6423)). When DNS returns both IPv6 (AAAA) and IPv4 (A) and the IPv6 route is broken (e.g. a NAT64 `64:ff9b::` prefix without routing), undici tried IPv6 first and hung until `ETIMEDOUT` (then a 502 + account lockout), even though `curl` reached the same host. The direct dispatcher now races both families and uses whichever connects first. Proxy paths pin family via `proxyTls` and are unaffected. Regression guard: `tests/unit/direct-dispatcher-pipelining-4580.test.ts`. (thanks @adentdk)
+- **fix(combo):** round-robin now advances the rotation pointer past the model that **actually served**, not the eagerly-scheduled one (port from 9router#948, [#6428](https://github.com/diegosouzapw/OmniRoute/pull/6428)). With `stickyLimit: 1` (true round-robin), when the scheduled model failed and a _different_ model served via fallback, the counter had already advanced +1 from the scheduled index — so the next request reused the fallback-served model, degrading round-robin into hot-spotting on whichever model was healthy. The pointer now advances to the served index + 1 (mirroring the sticky-limit>1 path). Session-stickiness (#3825) and distribution are preserved. Regression guard: `tests/unit/combo-rr-fallback-advance-948.test.ts`. (thanks @binsarjr)
+- **fix(sse):** a non-string `model` field is now rejected with a `400` before the resolver, instead of crashing downstream `.toLowerCase()`/`.split()` calls into an empty-body `500` that escapes the error sanitizer ([#6407](https://github.com/diegosouzapw/OmniRoute/issues/6407)). Regression guard: `tests/unit/chat-non-string-model-6407.test.ts`. (thanks @chirag127)
+- **fix(api):** unknown `/api/*` routes now return a JSON `404` (instead of the dashboard HTML shell) and scalar chat params (`model`/`temperature`/etc.) are validated **before** the provider lookup so malformed requests fail fast with a clear `400` ([#6424](https://github.com/diegosouzapw/OmniRoute/issues/6424), [#6412](https://github.com/diegosouzapw/OmniRoute/issues/6412)). Regression guards: `tests/unit/api/api-catchall-json-404.test.ts`, `tests/unit/chat-early-schema-validation-6412.test.ts`. (thanks @chirag127)
+- **fix(api):** `/v1/chat/completions` now rejects a non-JSON `Content-Type` with a `400` before parsing the body ([#6414](https://github.com/diegosouzapw/OmniRoute/issues/6414)). Regression guard: `tests/unit/v1-chat-completions-content-type-6414.test.ts`. (thanks @chirag127)
+- **fix(api):** the `X-OmniRoute-Compression` response header is now echoed on `/v1/chat/completions` and `/v1/completions` ([#6422](https://github.com/diegosouzapw/OmniRoute/issues/6422)). Regression guard: `tests/unit/compression-header-echo-6422.test.ts`. (thanks @chirag127)
+- **fix(api):** concurrent `GET /v1/models` requests are coalesced into a single catalog build ([#6408](https://github.com/diegosouzapw/OmniRoute/issues/6408)). Regression guard: `tests/unit/v1-models-concurrent-6408.test.ts`. (thanks @chirag127)
+- **fix(api):** `/v1/completions` now echoes the requested `body.model` in its JSON + streamed responses ([#6429](https://github.com/diegosouzapw/OmniRoute/pull/6429)). Regression guard: `tests/unit/completions-body-model-echo.test.ts`. (thanks @chirag127)
+- **fix(api):** env-var master keys now see the full `/v1/models` catalog ([#6406](https://github.com/diegosouzapw/OmniRoute/issues/6406)). Regression guard: `tests/unit/models-catalog-envkey-6406.test.ts`. (thanks @chirag127)
+- **fix(api):** non-streaming `/v1/completions` responses now echo `body.model` aligned with the `X-OmniRoute-Model` header ([#6426](https://github.com/diegosouzapw/OmniRoute/issues/6426)). Regression guard: `tests/unit/v1-completions-model-header-match-6426.test.ts`. (thanks @chirag127)
+- **fix(api):** unknown `/v1/*` routes now return a JSON `404 not_found` instead of the Next.js dashboard HTML shell ([#6405](https://github.com/diegosouzapw/OmniRoute/issues/6405)). Regression guard: `tests/unit/api/v1-catchall-json-404.test.ts`. (thanks @chirag127)
+- **fix(api):** the per-connection provider models route now degrades to the shipped catalog when a provider's `/models` endpoint answers with a redirect ([#6267](https://github.com/diegosouzapw/OmniRoute/issues/6267)) — a `qwen-web` import failed with a raw `Redirect blocked … (307)` 503. `safeOutboundFetch` throws `REDIRECT_BLOCKED` on the 307, `getSafeOutboundFetchErrorStatus` maps it to 503, and `buildDiscoveryErrorFallbackResponse` treated every 503 as a hard error — so the non-empty `getModelsByProviderId("qwen-web")` catalog was never surfaced. A models-endpoint redirect is not a fixable-config error (unlike `URL_GUARD_BLOCKED`/`INVALID_URL`, which stay hard errors), so it now falls back to the local/cached catalog before the 503 short-circuit. General fix — covers any config-driven provider that 307s. Regression guard: `tests/unit/provider-models-qwen-web-redirect-6267.test.ts`. (thanks @chirag127)
+- **fix(api):** the per-connection provider models route (MCP `list_models_catalog` + the dashboard import view) now merges USER-ADDED custom models into its response ([#6247](https://github.com/diegosouzapw/OmniRoute/issues/6247)) — custom models live in the `key_value` namespace `customModels`, which the live REST `/api/v1/models` already merges, but `src/app/api/providers/[id]/models/route.ts` never read `getCustomModels`, so custom models were dropped on both the discovery-success and local_catalog paths. They are now folded into the returned model list (deduped by id, stamped `owned_by: provider`), fixing MCP + the dashboard import view in one place. Regression guard: `tests/unit/provider-models-custom-merge-6247.test.ts`. (thanks @RCrushMe)
+
+- **fix(providers):** GitLab Duo tool-calling follow-up turns no longer fail upstream with `422 {"detail":"Validation error"}` (tokens 0/0, rejected pre-inference). The #6234 tool-result-feedback fix serialized the **entire** multi-turn conversation into GitLab's single-file `code_suggestions` (`small_file`) generation endpoint — folded history that turn-N sent as an oversized `current_file.content_above_cursor` **and** duplicated verbatim into `user_instruction`, tripping the AI-Gateway's `small_file` validation guard. The executor now **bounds** that prompt: it keeps system + latest user message + the most-recent tool round (dropping older turns), caps oversized tool results, and stops duplicating the full prompt into `user_instruction` (which now carries only the short latest user message) — while still feeding the most-recent tool result back so the agent continues (`open-sse/executors/gitlab.ts`, [#6220](https://github.com/diegosouzapw/OmniRoute/issues/6220)). The unit test covers the bounding logic; the upstream 422→200 clearing is VPS-only (Hard Rule #18). Regression guard: `tests/unit/gitlab-tool-exchange-bounded-6220.test.ts`.
+
+- **fix(i18n):** the provider-detail (`/dashboard/providers/[id]`) connection-status filter labels no longer render as `__MISSING__:All` / `__MISSING__:Active` / `__MISSING__:Error` / `__MISSING__:Banned` / `__MISSING__:CreditsExhausted` in non-English locales (notably pt-BR) ([#6290](https://github.com/diegosouzapw/OmniRoute/issues/6290)). Root cause was **not** the namespace mismatch the issue guessed — the `providers.filter*` keys resolve correctly in `en.json`; the debt lived in the locale mirrors (`src/i18n/messages/*.json`), where these five keys carried the `__MISSING__:` sync sentinel in ~15 locales and were absent entirely in ~26 others, so next-intl found the key and echoed the sentinel verbatim. All 40 non-English/-Chinese mirrors now ship real translations for the five `providers.filter*` labels. Regression guard: `tests/unit/i18n-provider-filter-keys-6290.test.ts`. (thanks @diegosouzapw)
+
+- **fix(providers):** the `copilot-m365-web` streaming executor now emits `debug`-level WebSocket diagnostics ([#6210](https://github.com/diegosouzapw/OmniRoute/issues/6210)) — the outbound WS URL (with the `access_token` **redacted** via `redactWsUrl()`), handshake success/failure, and each received SignalR frame's `type`/`target`. Previously the streaming path logged nothing, so an empty `content:null` response (the M365 Education / Starter tier symptom fixed in #6234) was undiagnosable even at `APP_LOG_LEVEL=debug`. The change is debug-level and side-effect-free — it does not alter streaming behavior or the frame parser, and the token never reaches the logs. Regression guard: `tests/unit/copilot-m365-web-logging-6210.test.ts` (thanks @qpeyba)
+
+- **fix(resilience):** a round-robin combo no longer returns `503 all upstream accounts are unavailable` when a compatibility-rejected target is actually healthy ([#6238](https://github.com/diegosouzapw/OmniRoute/issues/6238)). `filterTargetsByRequestCompatibility` drops request-incompatible targets (tool/vision/structured-output unsupported, or below the required context window) **before** any availability check runs, and its `compatible.length === 0` safety net only fired when _all_ targets were filtered — not when the kept targets later all turned out runtime-unavailable (circuit-open / cooldown / no credentials). So a combo could 503 while a compat-rejected-but-healthy provider sat unused. `handleRoundRobinCombo` now keeps the compat-rejected set and, when every compat-kept target was skipped without a single real attempt, probes those rejected targets as a **last-resort fallback tier** (via the new pure `open-sse/services/combo/comboCompatFallback.ts`) before crystallizing the 503. Regression guard: `tests/unit/combo-roundrobin-compat-fallback-6238.test.ts`. (thanks @ThongAccount)
+
+- **fix(startup):** best-effort self-heal for a corrupted Turbopack dev cache on Windows ([#6289](https://github.com/diegosouzapw/OmniRoute/issues/6289)). On Windows, `pnpm dev` can fail at startup when Turbopack `mmap`s a persistent-cache SST file and the OS refuses the mapping (`os error 1455` — "paging file too small"), which Turbopack surfaces as a misleading `Module not found: Can't resolve '@/shared/utils/machine'`. This is a **known upstream Turbopack cache-corruption bug — not our code**. The dev launcher (`scripts/dev/run-next.mjs`) now wraps `nextApp.prepare()` and, when it rejects with that signature (`isTurbopackCacheCorruption` in the new `scripts/dev/turbopackCacheHeal.mjs`), purges `.build/next/**/cache/turbopack` and retries **once** with a clear log. **Caveat — best-effort only:** the corruption often surfaces as a runtime overlay rather than a `prepare()` rejection, so this cannot always intercept it; the reliable remedy remains manually deleting the Turbopack cache dir. Regression guard: `tests/unit/turbopack-cache-heal-6289.test.ts`. (thanks @chirag127)
+
+- **fix(providers):** qodercli PAT auth no longer fails with `spawn qodercli ENOENT` on Windows ([#6263](https://github.com/diegosouzapw/OmniRoute/issues/6263)) — `spawnQoderCli` spawned the bare `qodercli` name with `shell:false` and an unenriched env, so the npm `.cmd` wrapper under `%APPDATA%\npm` (a user-PATH directory) was never resolved. It now resolves the absolute `.cmd`/`.exe` path through the existing `getCliRuntimeStatus("qoder")` resolver in `src/shared/services/cliRuntime.ts` (memoized), spawns with `shell` when the target is a `.cmd`/`.bat`, and uses the cliRuntime-enriched env (PATH + PATHEXT + APPDATA); the ENOENT error now lists the searched paths plus the `CLI_QODER_BIN` override. End-to-end spawn on a real Windows host is host-only (Hard Rule #18); the path-resolution logic is unit-tested. Regression guard: `tests/unit/qodercli-windows-resolve-6263.test.ts`. (thanks @chirag127)
+
+- **fix(sse):** the reasoning-token buffer no longer inflates **probe-sized `max_tokens`** ([#6274](https://github.com/diegosouzapw/OmniRoute/issues/6274)) — Claude Code's `/model` capability check sends `max_tokens: 1`, but for a thinking-capable model with a large output cap (e.g. `glm-5.2`) the #3587 headroom heuristic (`max(current + 1000, ceil(current * 1.5))`) rewrote it to `1001` and forwarded that upstream, wasting tokens on a request that was never a genuine reasoning budget. `resolveReasoningBufferedMaxTokens()` (`open-sse/services/reasoningTokenBuffer.ts`) now short-circuits and returns the caller's value verbatim when it is below the new `REASONING_BUFFER_MIN_TRIGGER` (256) threshold — a tiny explicit limit is a probe, not a reasoning request. Real budgets still receive the #3587 headroom unchanged, and the guard runs after the existing capability checks so unknown / non-reasoning models keep returning `null`. Regression guard: `tests/unit/reasoning-token-buffer-6274.test.ts`. (thanks @brightfiscalband)
+
+- **fix(cli):** `omniroute reset-password` now works as a real subcommand, and password resets over piped (non-TTY) stdin actually apply ([#6261](https://github.com/diegosouzapw/OmniRoute/issues/6261), [#6258](https://github.com/diegosouzapw/OmniRoute/issues/6258)). Two coupled defects: (1) **#6261** — `bin/omniroute.mjs` routed everything through Commander with only two pre-Commander bypasses (`--mcp`, `reset-encrypted-columns`), so `omniroute reset-password` was rejected as an unknown command; only the separate `omniroute-reset-password` bin worked, while the docs falsely advertised the subcommand (incl. a bogus "legacy alias still works"). A pre-Commander bypass mirroring `reset-encrypted-columns` now dynamically imports `bin/reset-password.mjs` (which self-executes) before Commander parses; the three doc lines were corrected. (2) **#6258** — `bin/reset-password.mjs` issued two sequential `rl.question` prompts; under piped stdin the second read never settled at EOF, so `main()` never reached `resetManagementPassword` and the reset was a silent no-op (both prompts printed, no success, password unchanged). The CLI now detects non-TTY stdin and reads it once (first line = password, second line = confirm if present, else reused), adds a `--password-stdin` flag (entire stdin is the password, no confirmation), and exits `0` explicitly so the success line always flushes; interactive TTY behavior is unchanged. Regression guard: `tests/unit/reset-password-cli-6261-6258.test.ts` (3). (thanks @chirag127)
+
+- **fix(db):** the mass-migration **safety abort** now tells the operator how to bypass it and stops flooding the log ([#6260](https://github.com/diegosouzapw/OmniRoute/issues/6260)) — after restoring a backup that wiped the migration tracking table, `runMigrations()` threw the abort on every downstream `ensureDbInitialized()`, re-logging the full banner 11+ times, and the message never mentioned the existing `OMNIROUTE_MAX_PENDING_MIGRATIONS` escape hatch. The abort text now appends a bypass hint (set `OMNIROUTE_MAX_PENDING_MIGRATIONS=0` in `server.env` / `DATA_DIR/.env`), and a new `MigrationSafetyAbortError` is memoized so repeated calls in the same process throw the same instance and emit a single concise line instead of the full cascade. Regression guard: `tests/unit/migration-safety-abort-6260.test.ts`. (thanks @chirag127)
+
+- **fix(auth):** importing a **distinct** Codex/ChatGPT OAuth `auth.json` is no longer falsely rejected as "already exists" when it belongs to a different user in the same workspace ([#6301](https://github.com/diegosouzapw/OmniRoute/issues/6301)). `findExistingCodexConnection` (in `src/lib/oauth/utils/codexAuthImport.ts`) deduped **only** on `providerSpecificData.workspaceId === accountId`, where `accountId` is the shared `chatgpt_account_id`/`tokens.account_id` — so two members of the same ChatGPT Team collapsed onto a single connection (409 `duplicate_account`). The id_token's `https://api.openai.com/auth` claim carries a per-user `chatgpt_user_id` alongside the workspace id (the device-flow path already persisted it as `chatgptUserId`, but the import path did not). Now `parseAndValidateCodexAuth` extracts `userId` (`chatgpt_user_id` → `user_id` → JWT `sub`) into `ParsedCodexAuth`, the create/update paths persist `chatgptUserId` in `providerSpecificData` (mirroring `codex.ts`), and dedup keys on `workspaceId` **AND** `chatgptUserId` — with a backward-compat fallback to legacy accountId-only matching when no stored connection for that workspace records a `chatgptUserId`, so genuinely-same accounts still dedup. Regression guard: `tests/unit/codex-auth-import-userid-dedup-6301.test.ts` (4). (thanks @anungma)
+
+- **fix(providers):** importing models for the **venice-web** provider no longer fails with a red "Provider venice-web does not support models listing" ([#6269](https://github.com/diegosouzapw/OmniRoute/issues/6269)). `venice-web` is a web-cookie provider with an executor but no upstream `/v1/models` endpoint and no registry `models`, so the models route fell through to the tail `400`. Mirroring the `jules`/`linkup-search`/`ollama-search` fix (#5569), it now ships a static local catalog entry in `src/lib/providers/staticModels.ts` — seeding the current Venice lineup (`venice-uncensored`, `llama-3.3-70b`, `qwen3-235b`, `qwen3-4b`, `deepseek-r1-671b`; Venice rotates its catalog, see docs.venice.ai/models/overview) — so the route returns `200` with `source:"local_catalog"`, `intentional:true`. Regression guard: `tests/unit/static-models-venice-web-6269.test.ts`. (thanks @chirag127)
+
+- **fix(api):** the specialty model catalogs (`/v1/embeddings`, `/v1/images`, `/v1/music`, `/v1/videos` model lists) are now derived from the **unified catalog filtered by a predicate** (`getSpecialtyModelsResponse`) instead of ad-hoc per-route logic, so they consistently respect active-credential visibility and stay in sync with the main catalog ([#6303](https://github.com/diegosouzapw/OmniRoute/pull/6303)). Regression guard: `tests/unit/specialty-model-catalog-routes.test.ts`. (thanks @makcimbx)
+- **fix(api):** the agent-bridge server route now resolves the MITM manager via a **dynamic `import("@/mitm/manager.runtime")`** so Turbopack does not statically pull the stub (or over-bundle the manager), and the agent-skills generator anchors its output base path with `path.join(process.cwd(), …)` so Turbopack's static analyzer stops tracing the whole project root ([#6329](https://github.com/diegosouzapw/OmniRoute/issues/6329), [#6366](https://github.com/diegosouzapw/OmniRoute/pull/6366)). Regression guard: `tests/unit/agent-bridge-server-route-dynamic-import.test.ts`. (thanks @Iammilansoni)
+- **fix(api):** internal probes (combo-test, cloud-sync verify) now pick a **management-scoped / allow-all API key** instead of naively grabbing `getApiKeys()[0]` — a restricted `self:usage` first row made the probe fail with "Model X is not allowed for this API key" even when the combo path was healthy (`pickApiKeyForInternalUse` in `src/lib/db/apiKeys.ts`). The API-manager model editor also falls back to `/api/models?all=true` when `/v1/models` is catalog-protected ([#6372](https://github.com/diegosouzapw/OmniRoute/pull/6372)). Regression guard: `tests/unit/pick-internal-api-key-6372.test.ts`. (thanks @jmengit)
+- **fix(live-ws):** the Live Dashboard WebSocket server now **rejects on bind failure** (e.g. `EADDRINUSE` when the API bridge already holds the port) instead of letting the error surface as an unhandled `error` event that crash-loops the process — the `error` listener is attached to `wss` (not `server`) and releases the EventBus subscription on a failed start ([#6324](https://github.com/diegosouzapw/OmniRoute/issues/6324)). Regression guard: `tests/unit/live-ws-eaddrinuse-6324.test.ts`. (thanks @vinayakkulkarni)
+- **fix(dashboard):** the Home provider-topology widget now trusts the live provider-metrics snapshot — it uses `topology.errorProvider` and live `activeRequests` directly instead of re-deriving state from a stale `lastErrorAt` or applying a frontend timeout filter, so the topology reflects real-time provider health ([#6322](https://github.com/diegosouzapw/OmniRoute/pull/6322)). Regression guard: `tests/unit/home-provider-topology-live-state.test.ts`. (thanks @xz-dev)
+- **fix(sse):** strip zero-width markers from streamed **tool-call arguments** — a follow-up to [#5857](https://github.com/diegosouzapw/OmniRoute/pull/5857). That PR removed injected zero-width joiners (U+200D) from streamed assistant text/reasoning but deliberately left tool-call argument JSON byte-exact. The request-side obfuscation (`open-sse/services/claudeCodeObfuscation.ts`) injects ZWJ into agent words — including the temp path inside the Bash tool description — and Claude models copy that verbatim into generated commands, which are delivered as tool-call arguments rather than assistant text. As a result the ZWJ survived and corrupted code blocks (e.g. a temp path rendered with an invisible joiner). Now `open-sse/handlers/responseSanitizer.ts` strips zero-width code points from tool-call argument strings at every emit site (OpenAI non-stream/stream chat `tool_calls` + legacy `function_call`, native Responses `function_call` items, the OpenAI→Responses conversion, and the native Responses streaming `response.function_call_arguments.delta/.done` events). Only zero-width code points are removed; JSON structure and all other bytes stay identical (no parse/restringify), so normal arguments remain byte-exact. Regression guard: 6 new cases in `tests/unit/response-sanitizer.test.ts` (suite 50/50).
+
+- **fix(nodejs):** the default app log path now resolves under `DATA_DIR` (`~/.omniroute/logs/application/app.log`) instead of `process.cwd()` ([#6197](https://github.com/diegosouzapw/OmniRoute/issues/6197)) — the globally-installed CLI runs from an arbitrary working directory, so anchoring the default to cwd made file logging silently write to (or no-op under) an unrelated directory, contradicting the documented `.env.example` default. `getAppLogFilePath()` now computes the default lazily via the pure `resolveDataDir()` resolver (honours a per-process `DATA_DIR`, no directory-creation side effect); an explicit `APP_LOG_FILE_PATH` still wins. Regression guard: `tests/unit/logenv-datadir-path-6197.test.ts` (3).
+
+- **fix(docker):** AgentBridge/`startMitm` no longer aborts in containers/headless when the Antigravity-default DNS step can't write `/etc/hosts` ([#6127](https://github.com/diegosouzapw/OmniRoute/issues/6127)), and the privileged command's stderr now reaches `app.log` instead of only a bare exit code hitting the toast ([#6198](https://github.com/diegosouzapw/OmniRoute/issues/6198)). The default DNS step (`addDNSEntry`) was called unguarded while cert install and the two sibling DNS steps were each best-effort — in the runtime Docker image (`USER node`, no `sudo`, read-only `/etc/hosts`) it threw `Command failed with code 1` out of `startMitmInternal` and killed the whole start, discarding the stderr. The three DNS steps are extracted into a best-effort `provisionDnsEntries()` where each failure is logged with the full `err` (stderr included, folded in by `systemCommands.ts`) and never aborts the start. Regression guard: `tests/unit/mitm-dns-graceful-degrade-6127.test.ts` (4).
+
+- **fix(providers):** copilot-m365-web now supports the M365 Education "Starter / OfficeWebIncludedCopilot" tier and no longer returns an empty `content:null` stream ([#6210](https://github.com/diegosouzapw/OmniRoute/issues/6210)). Two gaps: (1) `buildWsUrl()` hardcoded the individual-consumer scenario (`OfficeWebPaidConsumerCopilot`, `isEdu=false`) — the EDU tier is now opt-in via `providerSpecificData.tier="edu"`, emitting `scenario=OfficeWebIncludedCopilot`/`isEdu=true` (the individual path is unchanged); (2) the EDU/GPT-5.5 path streams deltas via `arguments[0].writeAtCursor` (incremental) instead of only `messages[].text` (accumulated snapshots), which the parser dropped — a new `accumulateBotContent()` folds both formats, with `type:2 item.result.message` as a last-resort fallback. Regression guard: `tests/unit/copilot-m365-edu-writeatcursor-6210.test.ts` (10). (thanks @qpeyba)
+
+- **fix(providers):** GitLab Duo executor now feeds tool results back into the prompt instead of looping ([#6220](https://github.com/diegosouzapw/OmniRoute/issues/6220)) — `buildPrompt()` branched only on `system`/`user` and took `userParts.at(-1)`, silently dropping the `assistant{tool_calls}` + `tool{result}` turns the client appended, so the reconstructed prompt was byte-identical to turn 1 and the model re-emitted the same `<tool>` call forever. When a tool exchange is present the full conversation is now serialized, folding each tool result back keyed by its `tool_call_id`; simple conversations keep the legacy shape. Complements the tool_call emission from [#6051](https://github.com/diegosouzapw/OmniRoute/issues/6051) (the `kilo-duplicate` label was a false positive — different, sequential defect). Regression guard: `tests/unit/gitlab-tool-result-feedback-6220.test.ts` (4).
+
+- **fix(providers):** opencode-go/opencode-zen can now synthesize the OpenCode CLI identity headers Cloudflare requires on VPS egress ([#5997](https://github.com/diegosouzapw/OmniRoute/issues/5997)) — on a datacenter VPS, `opencode.ai/zen/go/v1/chat/completions` 403s (HTML challenge) requests lacking CLI identity, while the reporter's control curl proved that `User-Agent: opencode-cli/1.0.0` + `x-opencode-client: cli` + `x-opencode-project: default` + fresh request/session UUIDs succeed. Opt-in via `OPENCODE_SYNTHESIZE_CLI_HEADERS=true` (values overridable via `OPENCODE_GO_USER_AGENT`/`OPENCODE_USER_AGENT`/`OPENCODE_CLIENT`/`OPENCODE_PROJECT`); it fills only headers the client did not already send. Kept **off by default** — the forward-only path is deliberate (fabricating a wrong value risks upstream rejection; a prior dedup regressed with `opencode/local`), so this replaces the fragile local header-injection shim without changing default behavior. Regression guard: `tests/unit/opencode-cli-headers-synthesis-5997.test.ts` (6). (thanks @aleksesipenko)
+- fix(resilience): sticky session affinity now evicts and fails over to another account when the pinned account is exhausted/unavailable (#6219)
+
+- fix(sse): Responses API passthrough now drops internal commentary-phase output before forwarding to clients (gated by RESPONSES_PASSTHROUGH_DROP_COMMENTARY, default on) (#6199)
+
+- **fix(sse):** tool-call function schemas with a root `type: null` are now coerced to `type: "object"` before dispatch ([#6359](https://github.com/diegosouzapw/OmniRoute/issues/6359)) — clients like the Codex app emit `parameters: { type: null, ... }` for some tools, which OpenAI-compatible upstreams reject with `400 Invalid schema for function '...': schema must be a JSON Schema of 'type: "object"', got 'type: null'`, failing the whole request. `toolSchemaSanitizer` already stripped the null; it now re-adds the mandatory root `"object"` type (and empty `properties`/open `additionalProperties` when absent). Combinator roots (`anyOf`/`oneOf`/`allOf`) and explicit root types are left untouched. Regression guard: 5 new cases in `tests/unit/tool-schema-sanitizer.test.mjs`.
+
+- **fix(docker):** AgentBridge no longer fails to start on npm/Electron/VPS installs with "MITM manager stub reached at runtime" ([#6344](https://github.com/diegosouzapw/OmniRoute/issues/6344)) — v3.8.45 flipped the production bundler default to Turbopack, but `next.config.mjs` aliased `@/mitm/manager` to its Docker-only degraded stub **unconditionally**. That was harmless while Docker (which sets the alias intentionally for #3390 graceful degradation) was the sole Turbopack consumer, but once every artifact built with Turbopack the stub shipped to all non-Docker users and `startMitm` threw on the first Agent-Bridge start. The alias is now opt-in via `OMNIROUTE_MITM_STUB=1` (set only by the Dockerfile) through the shared `scripts/build/mitm-stub-flag.mjs` helper; default builds bundle the real manager. Regression guard: `tests/unit/mitm-stub-alias-6344.test.mjs` (4).
+
+- **fix(proxy):** stop the v3.8.44 proxy regression that leaked the real IP and disabled healthy proxies ([#6246](https://github.com/diegosouzapw/OmniRoute/issues/6246)). Two coupled defects from the new health scheduler: (1) **IP leak** — when a proxy assigned to a connection was marked `inactive`, resolution fell through to a **direct** egress instead of blocking, exposing the operator's real IP; (2) **over-deactivation** — the sweep flipped a proxy to `inactive` on the **first** failed probe and counted our own 5s timeout / a probe-target `5xx` as the proxy's fault, so healthy paid proxies vanished from egress selection ("my proxies are not being used anymore"). Fix: the sweep decision is extracted into a pure, network-free `decideProxyHealthAction` (`src/lib/proxyHealth/decision.ts`) — by default the health check now **only counts/logs and never downgrades status** (a proxy is downgraded/removed only with `PROXY_AUTO_REMOVE=true`, after `PROXY_AUTO_REMOVE_AFTER` **consecutive** conclusive failures); probes are classified tri-state so an inconclusive result (our timeout, or a `5xx` from the probe target) never penalizes the proxy, and the probe timeout is raised 5s→15s. Separately, `safeResolveProxy` now **fails closed** via the existing policy: a connection whose assigned proxy is dead is blocked instead of leaking direct (`hasBlockingProxyAssignment`), honoring the explicit `proxy off` toggles and the `PROXY_FAIL_OPEN=true` opt-out. Existing proxies stuck `inactive` by the old behavior need a one-time manual re-activate (the operator owns proxy status). Regression guards: `tests/unit/proxy-health-decide-action-6246.test.ts`, `tests/unit/proxy-assigned-unavailable-6246.test.ts`.
+
+- **fix(proxy):** make "Test All" read-only and add bulk enable/disable ([#6246](https://github.com/diegosouzapw/OmniRoute/issues/6246)). Complements the core fail-closed / scheduler fix (#6296) with the two remaining reporter asks. (1) The **"Test All" button** (`POST /api/settings/proxies/auto-test`) used to flip a proxy to `inactive` on a failed reachability probe; since the egress selector excludes `inactive` proxies, a flaky probe (an unreachable `httpbin.org`, a proxy that blocks `HEAD`, or a slow paid proxy) silently disabled every proxy that failed — "Test All" is now **read-only by default** (only the operator sets a proxy active/inactive; opt back into the legacy test-and-set with `PROXY_HEALTH_AUTO_DEACTIVATE=true`). (2) Adds a **bulk enable/disable** proxies endpoint + toolbar action (`POST /api/settings/proxies/batch-activate`) so an operator can re-activate proxies in one click. Regression guard: `tests/unit/proxy-health-6246.test.ts`. (thanks @tenshiak)
+
+- **chatcore (tools): stop the default 128-tool cap from silently dropping opencode's `task`/MCP tools.** opencode (used as an MCP/agent host) sends a large tool list; when it exceeds the speculative `MAX_TOOLS_LIMIT` (128) default, `truncateToolList` did a blind `tools.slice(0, 128)`, dropping every tool past index 128 — including opencode's built-in `task` tool (subagent launch) and many MCP tools, so models routed through OmniRoute could no longer spawn subagents or reach part of their tools. The cap exists to avoid upstream `400`s for providers with real hard limits (e.g. grok-cli 200), so it is kept for those: detection of the opencode client (`isOpencodeClient` — any `x-opencode-*` header, or `opencode` in the user-agent) now only bypasses the **speculative 128 default**, never a known provider ceiling. Precedence is explicit — a proactive/detected provider limit always truncates (even for opencode); otherwise opencode forwards its full tool list; otherwise the unchanged 128 default applies to every other client. Refactors `getEffectiveToolLimit` into `getKnownToolLimit(provider) ?? DEFAULT_LIMIT` (byte-identical for existing callers) and fixes a cosmetic debug-log that reported the truncated count instead of the original. Regression guard: `tests/unit/tool-limit-detector.test.ts`.
+
+- **fix(mitm):** the macOS MITM-cert install check now matches the system keychain again. `security find-certificate -a -Z` prints the SHA-1 as a colon-less hex string, but the installed-check compared it against `getCertFingerprint()`'s colon-separated form, so the substring match never hit — the cert was reported as not-installed and re-prompted for the sudo install on every run. Fingerprints are now normalized (colons stripped, upper-cased) on both sides via the extracted `macCertOutputHasFingerprint` helper. Regression guard: `tests/unit/mitm-cert-mac-fingerprint.test.ts`. ([#6204](https://github.com/diegosouzapw/OmniRoute/pull/6204), closes [#6134](https://github.com/diegosouzapw/OmniRoute/issues/6134) — thanks @rianonehub)
+
+- **fix(api):** `/v1/messages/count_tokens` now counts `tool_use`, `tool_result` and `thinking` content blocks (and array-form `system` prompts) in the local-estimation path, instead of only `text`. Real agentic conversations keep ~95% of their tokens inside tool results; the previous estimate returned near-zero for them, which silently broke Claude Code's auto-compaction (context grew past the window with no compaction until the upstream API rejected the request). The real provider-side count path is unchanged. Regression guard: `tests/unit/messages-count-tokens-route.test.ts`. ([#6221](https://github.com/diegosouzapw/OmniRoute/pull/6221) — thanks @luweiCN)
+
+- **fix(antigravity):** strip a trailing assistant prefill turn for Vertex Claude models to avoid upstream 400s ([#6114](https://github.com/diegosouzapw/OmniRoute/pull/6114)). Regression guard: `tests/unit/antigravity-claude-prefill-strip.test.ts`. (thanks @anki1kr)
+
+- **fix(security):** the mutable cloud-agent routes (`/api/cloud/credentials/update`, `/api/cloud/models/alias`) now require management auth instead of being treated as public. They were classified as public API routes, so a request without management credentials could update stored cloud-agent credentials and model aliases. They are removed from the public-route set, classified as management routes in the authz pipeline, and gated by `requireManagementAuth`; cloud **read**/auth routes stay public. Regression guards: `tests/unit/cloud-write-auth.test.ts`, `tests/unit/authz/classify.test.ts`, `tests/unit/public-api-routes.test.ts`. ([#6233](https://github.com/diegosouzapw/OmniRoute/pull/6233) — thanks @vittoroliveira-dev)
+
+- **refactor(dashboard):** extract the onboarding-wizard "Open provider details" link target into a pure, unit-tested `buildProviderDetailsHref(connection)` helper. The wizard already routes by `connection.id` (the node UUID) rather than the provider category slug (#6144/#6145); this hardens that behavior behind a tested helper that guards a missing id/connection. Regression guard: `tests/unit/provider-onboarding-href.test.ts`. ([#6166](https://github.com/diegosouzapw/OmniRoute/pull/6166) — thanks @KooshaPari)
+
+- **fix(security):** the doubao synthetic device-id generator now derives its digits via an unbiased crypto-random draw (rejection sampling over `crypto.randomBytes()`) instead of a `% 10` reduction, closing a CodeQL `js/biased-cryptographic-random` finding.
+- **fix(agentSkills):** the GitHub-skills generator now resolves `outputDir` to an absolute path before writing, fixing a regression introduced by #6366 (relative-to-cwd base path) that could write generated skill files to the wrong directory.
+- **fix(security):** `/api/keys/{id}/devices` now checks the HTTP method before auth/validation, returning a `405` for non-GET/DELETE verbs instead of a misleading `401`/`500` (closes a `dast-smoke` QUERY-method finding).
+- **fix(quality):** clear the last 2 heavy quality-gate reds on the release tip (cycle pre-flight).
+- **fix(mitm):** the test suite and CI can never mutate the OS trust store — `OMNIROUTE_SKIP_SYSTEM_TRUST=1` is set globally for tests/CI so `installCert`/`uninstallCert`/`installTproxyCa` skip the privileged OS dispatch ([#6310](https://github.com/diegosouzapw/OmniRoute/pull/6310); full detail is under the [3.8.45] section below — this branch received it via the parallel-cycle sync-back).
+- **fix(api):** `POST /api/github-skills` now Zod-validates its request body; documented the new quality-gate env vars and pinned the merge-integrity GitHub Actions to a commit SHA.
+- **fix(skills):** generate the missing `omni-github-skills` registry entry and align the agent-skills catalog-count tests (follow-up to #6186).
+- **fix(quality):** clear the cycle's 11 net-new ESLint errors and make `validate-release-green` suppressions-aware.
+
+### 📝 Maintenance
+
+- **i18n(it):** add 118 missing Italian (`it`) translations (net-additive — no existing keys dropped, valid JSON), improving Italian UI coverage. ([#6212](https://github.com/diegosouzapw/OmniRoute/pull/6212) — thanks @serverless83)
+- **chore(providers):** remove deprecated **MiMo V2** model entries from the catalogs (xiaomi-mimo, opencode-go, zenmux-free, audio TTS) — the upstream V2 line is superseded by MiMo V2.5; drops `mimo-v2-tts`, `mimo-v2-pro`, `mimo-v2-omni`, `mimo-v2-flash`, `mimo-v2-flash-free` and realigns the provider-catalog tests. ([#6248](https://github.com/diegosouzapw/OmniRoute/pull/6248) — thanks @backryun)
+- **chore(release):** ~50 commits on this branch are v3.8.45 pre-flight/hardening fixes and CI-perf work that landed here via the parallel-cycle sync-back (`sync-next-cycle.mjs`, Hard Rule #21) **after** the `v3.8.45` git tag was cut, and are already fully documented under the **[3.8.45]** section below — listed here only so the per-cycle commit-coverage check (`npm run release:uncovered`) doesn't flag them as gaps. Provider/catalog/UX/backend fixes: #6041, #6078, #6108, #6135, #6148, #6149, #6154, #6158, #6161, #6162, #6163, #6164, #6165, #6170, #6177, #6178, #6181, #6186, #6187, #6191, #6193, #6194, #6195, #6200, #6205, #6208, #6209, #6211, #6213, #6223, #6224, #6225, #6226, #6227, #6228, #6229, #6230, #6235, #6291, #6292. CI/release-pipeline work: #6167, #6203, #6214, #6215, #6218, #6273, #6275, #6283, #6284, #6285, #6300, #6305.
+- **chore(release):** additional zero-ref release-cycle plumbing on this branch, kept out of `release:uncovered` on purpose (no `#N` in the commit subject to cite): opening the v3.8.46 cycle, opening/closing the v3.8.45 cycle, the finalized [3.8.45] CHANGELOG i18n sync-back to 42 mirrors, the v3.8.45 cognitive/cyclomatic and file-size drift rebaselines, ESLint stale-suppression pruning (4,273 → 4,233), and clearing test-masking/docs-all pre-flight reds for v3.8.45.
+
+### ⚡ Performance & Infrastructure
+
+- **perf(release-green):** the pre-flight validator (`scripts/quality/validate-release-green.mjs`) now runs its 4 slow suites (unit / vitest / integration / pack-artifact) **concurrently** via `Promise.all` — pre-flight wall time drops from ~the sum of the suites to ~the slowest one (~30min saved per round; Phase 0 was the nº1 bottleneck of the v3.8.45 release benchmark, 2h54 of 6h34 e2e). Guard: `tests/unit/validate-release-green.test.ts` ("runs the slow suites CONCURRENTLY"). ([#6319](https://github.com/diegosouzapw/OmniRoute/pull/6319))
+- **fix(ci):** `scripts/release/sync-next-cycle.mjs` — two defects found live in its first production run (v3.8.45 Phase 5): (1) the `git()` helper's default 1 MiB `maxBuffer` crashed with `ENOBUFS` on `git show origin/main:CHANGELOG.md` (the CHANGELOG alone is >1 MiB) — widened to 64 MiB; (2) the i18n resync only propagated the `[NEXT]` (TBD) section, leaving the just-shipped finalized section as "— TBD" in all 42 mirrors — it now also syncs `[prevVersion]` bounded by the heading below it (new exported pure helper `versionAfter`). Guards: +5 tests in `tests/unit/sync-next-cycle.test.ts` (8/8). ([#6327](https://github.com/diegosouzapw/OmniRoute/pull/6327))
+- **test(ci):** concurrency-sensitive flaky tests are **quarantined into a serial pass** (`tests/unit/serial/`, `--test-concurrency=1`, appended to every unit runner incl. sharded variants — the serial pass is sharded too so concurrent shard jobs never self-collide). Initial set: `glm-coding-plan-monthly-3580`, `quota-division-blocks`, `provider-health-autopilot`, `combo-health-autopilot` — the class behind the ~28min CI wedges/re-runs (two live 1h+ wedges cancelled during this PR's own validation). Discovery + TIA gates track the new glob; systemic root cause (async logger writing after teardown) tracked in [#6360](https://github.com/diegosouzapw/OmniRoute/issues/6360). Guard: `tests/unit/test-serial-quarantine.test.ts` (4). ([#6347](https://github.com/diegosouzapw/OmniRoute/pull/6347))
+
+### 🙌 Contributors
+
+Thanks to everyone whose work landed in v3.8.46:
+
+| Contributor                                                    | PRs / Issues           |
+| -------------------------------------------------------------- | ---------------------- |
+| [@2220258345](https://github.com/2220258345)                   | direct commit / report |
+| [@abdofallah](https://github.com/abdofallah)                   | direct commit / report |
+| [@adentdk](https://github.com/adentdk)                         | direct commit / report |
+| [@aleksesipenko](https://github.com/aleksesipenko)             | direct commit / report |
+| [@anki1kr](https://github.com/anki1kr)                         | direct commit / report |
+| [@anungma](https://github.com/anungma)                         | direct commit / report |
+| [@arpicato](https://github.com/arpicato)                       | direct commit / report |
+| [@backryun](https://github.com/backryun)                       | #6248                  |
+| [@binsarjr](https://github.com/binsarjr)                       | direct commit / report |
+| [@brightfiscalband](https://github.com/brightfiscalband)       | direct commit / report |
+| [@chirag127](https://github.com/chirag127)                     | #6501, #6506           |
+| [@developerjillur](https://github.com/developerjillur)         | direct commit / report |
+| [@dilneiss](https://github.com/dilneiss)                       | #6499                  |
+| [@dtybnrj](https://github.com/dtybnrj)                         | direct commit / report |
+| [@Forcerecon](https://github.com/Forcerecon)                   | direct commit / report |
+| [@hao3039032](https://github.com/hao3039032)                   | direct commit / report |
+| [@Iammilansoni](https://github.com/Iammilansoni)               | #6150, #6245           |
+| [@jmengit](https://github.com/jmengit)                         | direct commit / report |
+| [@jordansilly77-stack](https://github.com/jordansilly77-stack) | direct commit / report |
+| [@JxnLexn](https://github.com/JxnLexn)                         | direct commit / report |
+| [@KooshaPari](https://github.com/KooshaPari)                   | #6166                  |
+| [@loopyd](https://github.com/loopyd)                           | direct commit / report |
+| [@luweiCN](https://github.com/luweiCN)                         | #6221                  |
+| [@makcimbx](https://github.com/makcimbx)                       | direct commit / report |
+| [@muflifadla38](https://github.com/muflifadla38)               | direct commit / report |
+| [@newnol](https://github.com/newnol)                           | direct commit / report |
+| [@ofekbetzalel](https://github.com/ofekbetzalel)               | direct commit / report |
+| [@ohahe52-dot](https://github.com/ohahe52-dot)                 | direct commit / report |
+| [@phidinhmanh](https://github.com/phidinhmanh)                 | direct commit / report |
+| [@powellnorma](https://github.com/powellnorma)                 | direct commit / report |
+| [@qpeyba](https://github.com/qpeyba)                           | direct commit / report |
+| [@RaviTharuma](https://github.com/RaviTharuma)                 | direct commit / report |
+| [@RCrushMe](https://github.com/RCrushMe)                       | direct commit / report |
+| [@rianonehub](https://github.com/rianonehub)                   | #6134, #6204           |
+| [@serverless83](https://github.com/serverless83)               | #6212                  |
+| [@swingtempo](https://github.com/swingtempo)                   | direct commit / report |
+| [@tenshiak](https://github.com/tenshiak)                       | direct commit / report |
+| [@ThongAccount](https://github.com/ThongAccount)               | direct commit / report |
+| [@UnrealAryan](https://github.com/UnrealAryan)                 | direct commit / report |
+| [@vinayakkulkarni](https://github.com/vinayakkulkarni)         | direct commit / report |
+| [@vittoroliveira-dev](https://github.com/vittoroliveira-dev)   | #6233                  |
+| [@warelik](https://github.com/warelik)                         | direct commit / report |
+| [@xxy9468615](https://github.com/xxy9468615)                   | direct commit / report |
+| [@xz-dev](https://github.com/xz-dev)                           | direct commit / report |
+| [@yanpaing007](https://github.com/yanpaing007)                 | direct commit / report |
+| [@diegosouzapw](https://github.com/diegosouzapw)               | maintainer             |
+
+---
+
 ## [3.8.45] — 2026-07-06
 
 ### ✨ New Features

docs/i18n/gu/CHANGELOG.md

@@ -6,6 +6,213 @@
 
 ## [3.8.31] — 2026-06-20
 
+## [3.8.46] — TBD
+
+### ✨ New Features
+
+- **feat(sse):** **hide paid-only models from `auto/*` routing** when `hidePaidModels` is on ([#6512](https://github.com/diegosouzapw/OmniRoute/issues/6512)) — follow-up to #6328/#6495. PR #6495 hid paid-only models from the `GET /v1/models` listing, but `auto/*` combos (`auto/best-coding`, `auto/glm`, …) could still pick a paid-only backend into their candidate pool → a 402/403 at request time. `createVirtualAutoCombo` now filters the candidate pool through the new pure `open-sse/services/autoCombo/paidModelFilter.ts` (`filterPaidOnlyCandidates`), applying the same free-model predicate #6495 uses in `catalog.ts` (`providerHasFreeModels(provider) && isFreeModel(provider, {id})`) whenever `settings.hidePaidModels === true`. Applied before the category/tier/family narrowing, so it covers every `auto/*` combo; an all-paid pool degrades to the existing graceful empty-pool path. **Opt-in — default OFF leaves the pool unchanged** (identity). Regression guard: `tests/unit/autoCombo/paid-model-filter-6512.test.ts` (4, incl. the default-off identity guard).
+- **feat(sse):** **provider-family auto combos** — `auto/glm`, `auto/minimax`, `auto/mimo`, `auto/zai`, `auto/gemma`, `auto/llama`, `auto/gemini` ([#6453](https://github.com/diegosouzapw/OmniRoute/issues/6453)) — new routable ids that materialize an on-demand virtual combo spanning whatever installed backends currently expose that model family, degrading gracefully as backends rotate. A new pure `open-sse/services/autoCombo/modelFamily.ts` (`detectModelFamily`) classifies by model-id prefix for six families; `zai` is instead resolved by provider id (z.ai's hosted API serves the same `glm-*` model ids as every other GLM backend, so `auto/zai` means "route to my z.ai backend specifically" vs `auto/glm`'s "any connected GLM backend"). Reuses the existing `createVirtualAutoCombo` on-demand materialization path (no DB writes) and the `/v1/models` catalog advertising loop. Regression guard: `tests/unit/autoCombo/provider-family-combos.test.ts` (11).
+- **feat(proxy):** native **proxy-pool round-robin / egress IP rotation** ([#6365](https://github.com/diegosouzapw/OmniRoute/issues/6365)) — a scope (global / provider / account) can now hold **multiple** proxies as a pool with a rotation strategy, so outbound requests cycle their egress IP instead of pinning one proxy per scope. Migration `117_proxy_pool_rotation.sql` lifts the `UNIQUE(scope, scope_id)` constraint (rebuild via the canonical rename/copy/drop; existing single assignments become 1-element pools) and adds a `proxy_scope_rotation` companion table holding the per-scope strategy + a persisted monotonic round-robin cursor. Strategies: `round-robin` (default, monotonic cursor — never `Math.random`), `random`, and `sticky-per-N-min`. Resolution (`resolveProxyForScopeFromRegistry` / `resolveProxyForConnectionFromRegistry`) now fetches the alive, position-ordered candidate set (unchanged `PROXY_ALIVE_PREDICATE`) and applies the strategy; an empty / all-dead pool still returns `null` — the #6246 fail-closed guard is **untouched** (never falls through to direct egress). Backend + DB only; dashboard pool-builder UI is a follow-up. Regression guard: `tests/unit/proxy-pool-rotation-6365.test.ts` (8, incl. fail-closed + backward-compat).
+- **feat(providers):** end-to-end **tool/function calling on the native Gemini `/v1beta` endpoint** ([#6222](https://github.com/diegosouzapw/OmniRoute/issues/6222)) — both directions of the Gemini↔OpenAI conversion now preserve tool data (previously silently dropped). Request side: `convertGeminiToInternal` (extracted to its own testable module) maps `tools[].functionDeclarations` → OpenAI `tools`, prior `functionCall` parts → assistant `tool_calls`, and `functionResponse` parts → `tool`-role messages. Response side: `convertOpenAIResponseToGemini` emits `parts[].functionCall {name,args}` from `message.tool_calls`, and the streaming `openAIChunkToGeminiChunk` accumulates fragmented `tool_calls` deltas by index into complete `functionCall` parts. The non-Gemini client paths (Claude, OpenAI-Responses) already preserved tool calls — this closes the gap specific to the native Gemini surface. Regression guard: `tests/unit/v1beta-gemini-tool-calling-6222.test.ts` (6, incl. a streaming SSE round-trip).
+- **feat(providers):** copilot-m365-web **enterprise / work tier** support ([#6334](https://github.com/diegosouzapw/OmniRoute/issues/6334)) — mirrors the EDU-tier pattern (#6210): `M365ConnectionParams` gains an `agent` field, a new opt-in `M365_ENTERPRISE_OVERRIDES` preset (`agent=work`, `scenario=officeweb`, `licenseType=Premium`) applies via `providerSpecificData.tier="enterprise"` (alias `"work"`), and `agent` is also overridable directly via `providerSpecificData.agent`. `buildWsUrl` was hardcoding `agent="web"` (the one enterprise-distinguishing param with no override path), so a Premium work account handshook then returned an empty stream. The individual and EDU paths are untouched. Kilo's dup flag vs #6210 (EDU tier) was a false positive — different tier. Regression guard: `tests/unit/copilot-m365-enterprise-6334.test.ts` (7). End-to-end confirmation on a real Premium work account is a live-VPS validation follow-up (Hard Rule #18). (thanks @Forcerecon)
+- **feat(api):** standardized, provider-agnostic **`effort` + `thinking` request params** ([#6241](https://github.com/diegosouzapw/OmniRoute/issues/6241)) — a thin standardization layer over the existing mature per-provider reasoning plumbing (no provider mapper touched). `providerChatCompletionSchema` gains a canonical `effort` (reusing the shared `none/low/medium/high/xhigh` vocabulary — the UI tiers `extra`/`max` collapse onto `xhigh`) and a boolean `thinking`. A pure `normalizeReasoningRequest` (wired once in `src/sse/handlers/chat.ts`, before any reasoning field is read) folds them onto the fields the translators already consume (`reasoning_effort` / `reasoning.effort` / `thinking`), so they fan out to Anthropic / Gemini / xAI / Responses — an explicit client `reasoning_effort` / object-shaped `thinking` always wins (backward-compatible). `/models` additively exposes `supportsThinking` + `effort_tiers` so the frontend can render the toggles (UI component is a follow-up). Regression guard: `tests/unit/effort-thinking-standardization-6241.test.ts` (12). (thanks @Iammilansoni, @shabeer)
+- **feat(combo):** new **`pipeline` (sequential) combo strategy** ([#6297](https://github.com/diegosouzapw/OmniRoute/issues/6297)) — the 18th routing strategy runs targets **in order**, threading each step's output into the next step's input, with an optional per-step `prompt` (system instruction); only the final step's response is returned. Distinct from `fusion` (parallel fan-out + judge). Implemented as a self-contained `open-sse/services/pipeline.ts` (sibling to `fusion.ts`), dispatched from `combo.ts`; the step list reuses `combo.models` order and reads an optional `prompt` off each target (backward-compatible — ignored by every other strategy). Intermediate steps run non-streaming with tools stripped (complete prose to thread forward); the final step keeps the client's `stream` flag + tools. A failing/empty/unparseable intermediate step fails the whole pipeline explicitly via a sanitized error (never silently swallowed). Kilo's dup flag vs #563 was a false positive (that's model→chain selection; this is a sequential chain). Regression guard: `tests/unit/combo-pipeline-strategy.test.ts` (5). (thanks @ofekbetzalel)
+- **feat(ci):** `check:test-masking` now flags **inline-reimplemented prod conditions** ([#6348](https://github.com/diegosouzapw/OmniRoute/issues/6348)) — a new **report-only** subcheck (v2, 6A.10 family) catches the wrong-shape contract test: a test that recomputes the condition under test inline instead of importing/exercising the real function (the #6216 class, where `=== 500` → `>= 500` stayed green because the test re-implemented the branch). For each added/modified test file it warns when the file textually duplicates a ≥3-token conditional from a production file touched in the same PR **and** does not import the symbol/module owning it, via a pure, fixture-tested `findReimplementedConditions()` with an allowlist mirroring `assertReductionAllowlist`. Report-only for now (does not fail the gate) — to be promoted to blocking after a triage cycle. Regression guard: `tests/unit/check-test-masking.test.ts` (45).
+- **feat(sse):** per-connection routing override (native vs CLIProxyAPI) ([#6339](https://github.com/diegosouzapw/OmniRoute/issues/6339)) — the previously-dead `isCliproxyapiDeepModeEnabled` helper is now wired into `resolveExecutorWithProxy`: a single connection can opt itself into the CLIProxyAPI passthrough executor via `providerSpecificData.cliproxyapiMode="claude-native"`, with precedence **connection override > provider `upstream_proxy_config` mode > default**. `resolveExecutorWithProxy` now receives the resolved connection's `providerSpecificData` (threaded from `chatCore.ts`), so one connection can deep-route while the provider's other connections stay native — no DB schema change (the toggle rides in `providerSpecificData`). Also resolves the same-provider mixing ask in #6340. Regression guard: `tests/unit/chatcore-executor-proxy.test.ts` (9). (thanks @RaviTharuma)
+- **feat(dashboard):** "Add session cookie" modal now shows a prominent **"Open ‹host› →"** link to the provider's own site ([#6268](https://github.com/diegosouzapw/OmniRoute/issues/6268)) — every `-web` cookie-session provider (chatgpt-web, claude-web, gemini-web, kimi-web, lmarena, qwen-web, m365-copilot-web, …) renders a one-click external link (opening the provider's login/home page in a new tab) so operators no longer tab away to retype the URL mid-setup. The host resolves from a pure, unit-tested `resolveWebProviderHost()` (prefers `WEB_COOKIE_PROVIDERS[id].website`, falls back to the registry `baseUrl` origin); non-web providers render exactly as before. Kilo's dup flag vs #6265 (modal-too-small-on-1080p) was a false positive — distinct concern. Regression guard: `tests/unit/resolve-web-provider-host.test.ts` (5). (thanks @chirag127)
+- **feat(providers):** add **DigitalOcean** AI (serverless inference) as an OpenAI-compatible API-key provider ([#6373](https://github.com/diegosouzapw/OmniRoute/pull/6373)) — base `https://inference.do-ai.run/v1`, wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/digitalocean/`, `src/shared/constants/providers/apikey/inference-hosts.ts`). Regression guard: `tests/unit/digitalocean-provider.test.ts`. (thanks @newnol)
+- **feat(providers):** add **Huancheng Public API** (`hcnsec`) as an OpenAI-compatible regional provider ([#6410](https://github.com/diegosouzapw/OmniRoute/pull/6410)) — Xinjiang Huancheng Cybersecurity's public LLM platform (base `https://api.hcnsec.cn/v1`, free credits via daily check-ins), wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/hcnsec/`, `src/shared/constants/providers/apikey/regional.ts`). Regression guard: `tests/unit/hcnsec-provider.test.ts`. (thanks @UnrealAryan)
+- **feat(dashboard):** the web-session credential guide now shows an **"Open {host}" link** ([#6316](https://github.com/diegosouzapw/OmniRoute/pull/6316)) to the provider's sign-in site (derived from the provider `website` via `getProviderWebsiteHost`), so you can jump straight to the page where the cookie/session must be captured. Regression guard: `tests/unit/web-session-provider-link-6316.test.ts`. (thanks @jordansilly77-stack)
+- **feat(cerebras):** add the **Gemma 4 31B** model (`gemma-4-31b`) to the Cerebras registry + pricing table ([#6331](https://github.com/diegosouzapw/OmniRoute/pull/6331)). Regression guard extends `tests/unit/t28-model-catalog-updates.test.ts`. (thanks @backryun)
+- **feat(providers):** add **Yuanbao (web)** as a cookie-session provider ([#6196](https://github.com/diegosouzapw/OmniRoute/issues/6196)) — `yuanbao-web` (Tencent Yuanbao, `yuanbao.tencent.com`) with cookie-only auth (`hy_user`/`hy_token` + public agent id), SSE→OpenAI translation incl. `reasoning_content`, exposing DeepSeek V3/R1 + Hunyuan / Hunyuan-T1. Regression guard: `tests/unit/providers-yuanbao-web.test.ts`. `together-web` was **deferred** (no verifiable web-session endpoint — needs a captured request) and `huggingchat-web` **dropped** (the existing `huggingchat` already is a web-cookie provider). (thanks @chirag127)
+- **feat(providers):** route the built-in **agentrouter** through the dynamic Claude-Code wire image ([#6056](https://github.com/diegosouzapw/OmniRoute/issues/6056)) — a small static allow-set (`CC_WIRE_IMAGE_BUILTINS` in `open-sse/services/ccWireImageBuiltins.ts`), consulted by `isClaudeCodeCompatible` / `isClaudeCodeCompatibleProvider` / `applyFingerprint`, makes agentrouter adopt the CC wire-image headers + fingerprint **while guarding the CC baseUrl/auth branches** so it keeps its own registry `baseUrl` and `x-api-key` auth. Regression guard: `tests/unit/agentrouter-cc-wire-image.test.ts` (asserts the wire image is applied AND agentrouter's baseUrl/auth are preserved). Live WAF-acceptance against agentrouter.org is a VPS validation follow-up (Hard Rule #18).
+- **feat(providers):** **bulk-add API keys for Cloudflare Workers AI** ([#6174](https://github.com/diegosouzapw/OmniRoute/issues/6174)) — `cloudflare-ai` is removed from the bulk-add exclusion list and the bulk parser gains a 3-field `name|accountId|apiKey` mode; the bulk route now builds a **per-entry** `providerSpecificData` so each key carries its own `accountId` (fixing the previous shared-object reuse), and both the create + key-validation paths receive it. Regression guard: `tests/unit/bulk-api-key-parser-cloudflare.test.ts`. (thanks @muflifadla38)
+- **feat(dashboard):** routing/settings UX clarity ([#6147](https://github.com/diegosouzapw/OmniRoute/issues/6147)) — (1) weighted combos show the **effective routing share %** next to each weight when weights don't sum to 100 (`WeightTotalBar.tsx`); (2) the status widget's user-facing **"Cloud Sync" label is renamed** to "Remote Settings Sync" (`CloudSyncStatus.tsx`; internal ids/state untouched); (3) built-in providers gain an **opt-in advanced base-URL override** (`isBaseUrlOverrideEligibleProvider`, hidden behind an "Advanced" toggle, reusing the existing `providerSpecificData.baseUrl` persistence — not globally widened). Regression guard: `tests/unit/routing-settings-ux-6147.test.ts`.
+- **feat(combo):** add an option to **disable session stickiness**, per-combo or globally — round-robin / random combos can rotate to a different connection on every request instead of pinning a whole conversation to one connection by its first-message hash. Resolution precedence per-combo `config.disableSessionStickiness` → global `settings.disableSessionStickiness` → default `false` (preserves the #3825 prompt-cache/504 fix); gates **both** stickiness call sites in `open-sse/services/combo.ts`. Exposed as a global toggle (Combo Defaults) and a per-combo Inherit/on/off control. ([#6168](https://github.com/diegosouzapw/OmniRoute/issues/6168)) Regression guard: `tests/unit/combo-disable-session-stickiness.test.ts`. (thanks @RCrushMe)
+- **feat(docker):** add the `OMNIROUTE_NO_SUDO` env flag for root-less / user-namespaced deployments — the MITM cert-trust command path (`resolveSudoSpawn` in `src/mitm/systemCommands.ts`) now strips the leading `sudo` when the flag is truthy, in addition to the existing root / sudo-missing cases, so the Proxy Agent runs without `sudo` (the operator trusts the CA manually, e.g. via `NODE_EXTRA_CA_CERTS`). Argv-array `spawn` preserved — no shell interpolation (Hard Rule #13). ([#6122](https://github.com/diegosouzapw/OmniRoute/issues/6122)) Regression guard: `tests/unit/mitm-systemCommands-no-sudo.test.ts`. (thanks @powellnorma)
+- **feat(providers):** add **Requesty** as an OpenAI-compatible gateway provider (BYOK, base `https://router.requesty.ai/v1`, ~200 free requests/day) — wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/requesty/`, `src/shared/constants/providers/apikey/gateways.ts`). ([#6120](https://github.com/diegosouzapw/OmniRoute/issues/6120)) Regression guard: `tests/unit/requesty-provider.test.ts`. (thanks @chirag127)
+- **feat(dashboard):** add **configured-only / available-only filters** to the Free Provider Rankings page ([#6150](https://github.com/diegosouzapw/OmniRoute/issues/6150)) — hide providers you haven't configured, or whose connections are all rate-limited / out of quota, via server-side query params (`?configuredOnly` / `?availableOnly` on `GET /api/free-provider-rankings`) backed by a testable lib helper reusing the in-process connection state (no Redis). Both filters default off, so the default view is unchanged; this supersedes the earlier client-side "Configured Only" toggle (#6245) with an available-only dimension and unit-tested logic. Regression guard: `tests/unit/freeProviderRankings-filters.test.ts`.
+
+### 🔧 Bug Fixes
+
+- **fix(dashboard):** adding a second API key connection for the same provider no longer silently overwrites the first — the Add-API-key modal now derives a unique default connection name (`main`, then `main-2`, `main-3`, …) so the backend name-based upsert can't collide (#6499 — thanks @dilneiss).
+- **fix(compression):** the session-dedup engine now also deduplicates a large multi-line block repeated **within a single message** (intra-message dedup), not just across turns; the compression-preview API surfaces a `fallbackReason`, and the fusion panel reports how many models were rate-limited vs failed on total-panel failure (#6501 — thanks @chirag127).
+- **fix(compression):** a stacked-pipeline step naming an unregistered engine now surfaces a `validationErrors` entry instead of silently no-op'ing, so misconfigured pipelines are visible in the preview API (#6506 — thanks @chirag127).
+- **feat(usage):** add a **Codex reset-credit redemption** flow to the Provider Limits UI ([#6361](https://github.com/diegosouzapw/OmniRoute/pull/6361)) — a `useCodexResetCreditRedemption` hook + `/api/usage/codex-reset-credit` route + `codexResetCredits` lib let you redeem banked Codex reset credits from the quota card. Regression guard: `tests/unit/codex-reset-credits.test.ts`. (thanks @JxnLexn)
+- **feat(glm):** add **team-plan quota settings** for `glm-cn` connections ([#6351](https://github.com/diegosouzapw/OmniRoute/pull/6351)) — a dedicated `GlmTeamQuotaFields` form section (team quota id / limits) threaded through the Add/Edit connection modals, persisted via `providerSpecificData`, with the GLM usage service reading the team quota. Regression guards: `tests/unit/glm-team-quota.test.ts`, `provider-specific-data-schema.test.ts`. (thanks @hao3039032)
+- **feat(providers):** add **TinyFish** web-fetch/search support ([#6349](https://github.com/diegosouzapw/OmniRoute/pull/6349)) — a `tinyfish-fetch` executor + `/v1/web/fetch` route + MCP web-fetch tool, registered as a specialty-media provider with request-validation and a search-provider catalog entry. Regression guards: `tests/unit/executor-tinyfish-fetch.test.ts`, `web-fetch-handler.test.ts`, `mcp-web-fetch-tool.test.ts`, `provider-validation-tinyfish.test.ts`. (thanks @dtybnrj)
+- **fix(cli):** `omniroute launch-codex` now spawns `codex.cmd` through a shell on Windows (the npm `.cmd` shim is unresolvable by bare `spawn` → ENOENT), mirroring the qodercli Windows fix (#6263) ([#6312](https://github.com/diegosouzapw/OmniRoute/pull/6312)). Regression guard: `tests/unit/launch-codex-windows-spawn-6312.test.ts`. (thanks @swingtempo)
+- **fix(codex):** isolate the **Spark** quota from the shared Codex quota and stabilize the quota UI ordering / hydration so per-scope limits render consistently ([#6336](https://github.com/diegosouzapw/OmniRoute/pull/6336)). Regression guards: `tests/unit/codex-quota-selection-hydration.test.ts`, `provider-limits-ui.test.ts` + 3 more. (thanks @xz-dev)
+- **feat(api):** add a `hidePaidModels` setting that filters paid-only models out of the `/v1/models` catalog. Regression guard: `tests/unit/models-catalog-hide-paid.test.ts`. (thanks @chirag127)
+- **fix(api-manager):** the fallback model picker now preserves combos instead of dropping them when a primary model is unavailable ([#6443](https://github.com/diegosouzapw/OmniRoute/pull/6443)). Regression guard: `tests/unit/api-manager-page-static.test.ts`. (thanks @jmengit)
+- **fix(providers):** recoverable Antigravity / Cloud-Code (Gemini Code Assist) `403` responses ([#6452](https://github.com/diegosouzapw/OmniRoute/pull/6452)) are now classified as a retryable project-config error instead of a terminal account ban, so a fixable project/API-disabled 403 no longer forces a ~1-year cooldown / full OAuth reconnect. Regression guard: `tests/unit/errorclassifier-antigravity-403.test.ts`. (thanks @developerjillur)
+- **fix(mitm):** `sanitizeHeaders` now redacts `Set-Cookie` response headers so upstream session cookies never leak into logs / diagnostics ([#6451](https://github.com/diegosouzapw/OmniRoute/pull/6451)). Regression guard: `tests/unit/mitm-sanitize-headers.test.ts`. (thanks @developerjillur)
+- **fix(api):** `/api/compression/preview` now accepts `mode: "caveman"` and correctly handles stacked / zero-compression previews ([#6425](https://github.com/diegosouzapw/OmniRoute/issues/6425)). Regression guard: `tests/unit/api/compression-preview-caveman-and-stacked-6425.test.ts`. (thanks @chirag127)
+- **feat(providers):** add **Zed** hosted LLM aggregator as a native-app provider ([#6118](https://github.com/diegosouzapw/OmniRoute/pull/6118)) — OAuth sign-in via the Zed hosted flow, registered through the shared provider registry + executor. Regression guards: `tests/unit/zed-oauth-provider.test.ts`, `zed-import-utils.test.ts`, `zed-docker-detect.test.ts`, `mitm-handler-zed.test.ts`. VPS-validated via live operator login (Hard Rule #18).
+- **fix(oauth):** the Kiro SSO-cache auto-import now **preserves the IDC region** — cross-region Amazon Q / Kiro profiles imported from the SSO cache are no longer collapsed to the default region ([#6113](https://github.com/diegosouzapw/OmniRoute/pull/6113)). Regression guard: `tests/unit/kiro-auto-import-idc-2059.test.ts`. VPS-validated via live operator login (Hard Rule #18).
+- **fix(dashboard):** passthrough model aliases no longer collide when two namespaced model ids share a last segment (port from 9router#1850, [#6431](https://github.com/diegosouzapw/OmniRoute/pull/6431)). `enx/gpt-5.5` and `enx/codebuddy/gpt-5.5` both auto-generated the alias `gpt-5.5`, so the second model could never be added (the UI just alerted "alias already exists"). Aliases are now disambiguated deterministically — bare last segment when free, then parent-qualified (`codebuddy-gpt-5.5`), then a numeric suffix — while re-adding the exact same model id is still blocked. Regression guard: `tests/unit/passthrough-alias-1850.test.ts`. (thanks @arpicato)
+- **fix(translator):** preserve a Gemini `functionResponse` co-located with other parts (another `functionCall`, or trailing `text`) in the same content when translating **Gemini → OpenAI** ([#6376](https://github.com/diegosouzapw/OmniRoute/pull/6376)). `convertGeminiContent()` early-returned the tool message on the first `functionResponse` part, dropping any co-located parts; such contents are now pre-split (one tool message per `functionResponse`, emitted first, plus one message for the remaining parts). Regression guard: `tests/unit/gemini-to-openai-function-response.test.ts`. (thanks @warelik)
+- **fix(headroom):** detect a python interpreter managed by **mise / pyenv / asdf / conda** (port from 9router#2353, [#6382](https://github.com/diegosouzapw/OmniRoute/pull/6382)). Headroom's python probe (`src/lib/headroom/detect.ts`) searched a hardcoded `PATH`, but version managers expose their interpreters via shim dirs that only join `PATH` through interactive-shell activation — which the non-interactive server never runs, so a managed python (≥3.10) was invisible and Headroom reported it missing. The search path now prepends the well-known shim/bin dirs (`~/.local/share/mise/shims`, `~/.pyenv/shims`, `~/.asdf/shims`, `$CONDA_PREFIX/bin`, `~/.local/bin`, respecting `MISE_DATA_DIR`/`PYENV_ROOT`/`ASDF_DATA_DIR` when set), and a new `HEADROOM_PYTHON` env override lets operators point straight at their interpreter (mirroring `HEADROOM_URL`). Still shell-free (`execFileSync`). Regression guard: `tests/unit/headroom-detect.test.ts` (5). (thanks @loopyd)
+- **fix(executors):** strip the OpenAI-Codex/Claude-CLI `client_metadata` passthrough field for **NVIDIA** requests (port from 9router#1887, [#6411](https://github.com/diegosouzapw/OmniRoute/pull/6411)). NVIDIA's OpenAI-compatible wrapper rejects it with `400 Unsupported parameter`, the same class already handled for `cerebras`/`mistral`; `nvidia` (executor `default`) was missing from the strip allowlist so Codex/Claude-Code passthrough requests 400'd. Regression guard: `tests/unit/executor-default-strip-client-metadata.test.ts` (+nvidia case). (thanks @phidinhmanh)
+- **fix(translator):** strip the Claude-style `thinking` field for **NVIDIA `z-ai/glm-5.2`** (port from 9router#2023, [#6413](https://github.com/diegosouzapw/OmniRoute/pull/6413)). NVIDIA's OpenAI-compatible wrapper 400s on `thinking` (a Claude-format client routed here leaves a `thinking:{type:"adaptive"}`); the existing strip rule only dropped `reasoning`. Same class already handled for `minimax-m2.7`. Regression guard: `tests/unit/nvidia-minimax-thinking-strip.test.ts` (+glm-5.2 case). (thanks @phidinhmanh)
+- **fix(translator):** suppress the streamed `</think>` close marker for the **Antigravity IDE** client (port from 9router#1061, [#6415](https://github.com/diegosouzapw/OmniRoute/pull/6415)). On thinking-only turns Antigravity rendered a bare `</think>` as the sole visible content, tripping its loop-detection and wasting requests. Antigravity's UA (`vscode/<v> (Antigravity/<v>)`) is added to the marker-suppress allowlist (alongside OpenCode); Claude Code / Cursor still get the marker, and `x-omniroute-thinking-marker: on` force-restores it. Regression guard: `tests/unit/think-close-marker-suppress-5245.test.ts`. (thanks @abdofallah)
+- **fix(executors):** strip nested `reasoning_content` from messages for **Mistral** (port from 9router#1649, [#6417](https://github.com/diegosouzapw/OmniRoute/pull/6417)). Mistral's API returns `422 extra_forbidden` when an assistant message carries `reasoning_content` (replayed thinking from a prior turn, e.g. via the Codex `/responses` path); the generic top-level 400 field-downgrade retry never covered the nested per-message field. `DefaultExecutor` now strips it for provider `mistral` only, so DeepSeek (which requires replayed `reasoning_content`) is unaffected. Regression guard: `tests/unit/mistral-strip-reasoning-content-1649.test.ts`. (thanks @xxy9468615)
+- **fix(executors):** strip the `client_metadata` passthrough field on the **OpenCode** path (port from 9router#1442, [#6418](https://github.com/diegosouzapw/OmniRoute/pull/6418)). OpenCode upstreams (e.g. `kimi-k2.6` via opencode-go) reject it with `400 "Extra inputs are not permitted, field: 'client_metadata'"`; the DefaultExecutor strip only covered cerebras/mistral and `OpencodeExecutor` extends `BaseExecutor` directly, so nothing removed it there. Regression guard: `tests/unit/opencode-strip-client-metadata-1442.test.ts`. (thanks @yanpaing007)
+- **fix(executors):** inject the `reasoning_content` echo for the native **Moonshot Kimi** provider (port from 9router#1480, [#6419](https://github.com/diegosouzapw/OmniRoute/pull/6419)). Kimi (executor `default`) is a thinking-mode upstream that 400s with "reasoning_content must be passed back" when a prior assistant turn lacks it; the placeholder injection was only wired into the OpenCode meta-provider, so direct multi-turn Kimi conversations failed. Scoped to `kimi` (gateway-served models matching the thinking-model name pattern are unaffected). Regression guard: `tests/unit/kimi-native-reasoning-injected-1480.test.ts`. (thanks @2220258345)
+- **fix(executors):** recover from a strict gateway's `context_management: Extra inputs are not permitted` 400 (port from 9router#1468, [#6420](https://github.com/diegosouzapw/OmniRoute/pull/6420)). **Claude Code** always sends a top-level `context_management` field; strict anthropic-compatible gateways reject it. The dedicated context-editing 400-fallback only fired when OmniRoute's own `contextEditing` feature was enabled (default off), so a client-sent field passed through untouched and 400'd. `context_management` is now in the generic reactive field-strip list, so it's stripped-and-retried once regardless of the feature flag (with correct request re-signing for claude-compatible relays). Regression guard: `tests/unit/provider-field-strips.test.ts`. (thanks @ohahe52-dot)
+- **fix(network):** enable **RFC 8305 Happy Eyeballs** (`autoSelectFamily`) on the direct-egress undici dispatcher (port from 9router#1237, [#6423](https://github.com/diegosouzapw/OmniRoute/pull/6423)). When DNS returns both IPv6 (AAAA) and IPv4 (A) and the IPv6 route is broken (e.g. a NAT64 `64:ff9b::` prefix without routing), undici tried IPv6 first and hung until `ETIMEDOUT` (then a 502 + account lockout), even though `curl` reached the same host. The direct dispatcher now races both families and uses whichever connects first. Proxy paths pin family via `proxyTls` and are unaffected. Regression guard: `tests/unit/direct-dispatcher-pipelining-4580.test.ts`. (thanks @adentdk)
+- **fix(combo):** round-robin now advances the rotation pointer past the model that **actually served**, not the eagerly-scheduled one (port from 9router#948, [#6428](https://github.com/diegosouzapw/OmniRoute/pull/6428)). With `stickyLimit: 1` (true round-robin), when the scheduled model failed and a _different_ model served via fallback, the counter had already advanced +1 from the scheduled index — so the next request reused the fallback-served model, degrading round-robin into hot-spotting on whichever model was healthy. The pointer now advances to the served index + 1 (mirroring the sticky-limit>1 path). Session-stickiness (#3825) and distribution are preserved. Regression guard: `tests/unit/combo-rr-fallback-advance-948.test.ts`. (thanks @binsarjr)
+- **fix(sse):** a non-string `model` field is now rejected with a `400` before the resolver, instead of crashing downstream `.toLowerCase()`/`.split()` calls into an empty-body `500` that escapes the error sanitizer ([#6407](https://github.com/diegosouzapw/OmniRoute/issues/6407)). Regression guard: `tests/unit/chat-non-string-model-6407.test.ts`. (thanks @chirag127)
+- **fix(api):** unknown `/api/*` routes now return a JSON `404` (instead of the dashboard HTML shell) and scalar chat params (`model`/`temperature`/etc.) are validated **before** the provider lookup so malformed requests fail fast with a clear `400` ([#6424](https://github.com/diegosouzapw/OmniRoute/issues/6424), [#6412](https://github.com/diegosouzapw/OmniRoute/issues/6412)). Regression guards: `tests/unit/api/api-catchall-json-404.test.ts`, `tests/unit/chat-early-schema-validation-6412.test.ts`. (thanks @chirag127)
+- **fix(api):** `/v1/chat/completions` now rejects a non-JSON `Content-Type` with a `400` before parsing the body ([#6414](https://github.com/diegosouzapw/OmniRoute/issues/6414)). Regression guard: `tests/unit/v1-chat-completions-content-type-6414.test.ts`. (thanks @chirag127)
+- **fix(api):** the `X-OmniRoute-Compression` response header is now echoed on `/v1/chat/completions` and `/v1/completions` ([#6422](https://github.com/diegosouzapw/OmniRoute/issues/6422)). Regression guard: `tests/unit/compression-header-echo-6422.test.ts`. (thanks @chirag127)
+- **fix(api):** concurrent `GET /v1/models` requests are coalesced into a single catalog build ([#6408](https://github.com/diegosouzapw/OmniRoute/issues/6408)). Regression guard: `tests/unit/v1-models-concurrent-6408.test.ts`. (thanks @chirag127)
+- **fix(api):** `/v1/completions` now echoes the requested `body.model` in its JSON + streamed responses ([#6429](https://github.com/diegosouzapw/OmniRoute/pull/6429)). Regression guard: `tests/unit/completions-body-model-echo.test.ts`. (thanks @chirag127)
+- **fix(api):** env-var master keys now see the full `/v1/models` catalog ([#6406](https://github.com/diegosouzapw/OmniRoute/issues/6406)). Regression guard: `tests/unit/models-catalog-envkey-6406.test.ts`. (thanks @chirag127)
+- **fix(api):** non-streaming `/v1/completions` responses now echo `body.model` aligned with the `X-OmniRoute-Model` header ([#6426](https://github.com/diegosouzapw/OmniRoute/issues/6426)). Regression guard: `tests/unit/v1-completions-model-header-match-6426.test.ts`. (thanks @chirag127)
+- **fix(api):** unknown `/v1/*` routes now return a JSON `404 not_found` instead of the Next.js dashboard HTML shell ([#6405](https://github.com/diegosouzapw/OmniRoute/issues/6405)). Regression guard: `tests/unit/api/v1-catchall-json-404.test.ts`. (thanks @chirag127)
+- **fix(api):** the per-connection provider models route now degrades to the shipped catalog when a provider's `/models` endpoint answers with a redirect ([#6267](https://github.com/diegosouzapw/OmniRoute/issues/6267)) — a `qwen-web` import failed with a raw `Redirect blocked … (307)` 503. `safeOutboundFetch` throws `REDIRECT_BLOCKED` on the 307, `getSafeOutboundFetchErrorStatus` maps it to 503, and `buildDiscoveryErrorFallbackResponse` treated every 503 as a hard error — so the non-empty `getModelsByProviderId("qwen-web")` catalog was never surfaced. A models-endpoint redirect is not a fixable-config error (unlike `URL_GUARD_BLOCKED`/`INVALID_URL`, which stay hard errors), so it now falls back to the local/cached catalog before the 503 short-circuit. General fix — covers any config-driven provider that 307s. Regression guard: `tests/unit/provider-models-qwen-web-redirect-6267.test.ts`. (thanks @chirag127)
+- **fix(api):** the per-connection provider models route (MCP `list_models_catalog` + the dashboard import view) now merges USER-ADDED custom models into its response ([#6247](https://github.com/diegosouzapw/OmniRoute/issues/6247)) — custom models live in the `key_value` namespace `customModels`, which the live REST `/api/v1/models` already merges, but `src/app/api/providers/[id]/models/route.ts` never read `getCustomModels`, so custom models were dropped on both the discovery-success and local_catalog paths. They are now folded into the returned model list (deduped by id, stamped `owned_by: provider`), fixing MCP + the dashboard import view in one place. Regression guard: `tests/unit/provider-models-custom-merge-6247.test.ts`. (thanks @RCrushMe)
+
+- **fix(providers):** GitLab Duo tool-calling follow-up turns no longer fail upstream with `422 {"detail":"Validation error"}` (tokens 0/0, rejected pre-inference). The #6234 tool-result-feedback fix serialized the **entire** multi-turn conversation into GitLab's single-file `code_suggestions` (`small_file`) generation endpoint — folded history that turn-N sent as an oversized `current_file.content_above_cursor` **and** duplicated verbatim into `user_instruction`, tripping the AI-Gateway's `small_file` validation guard. The executor now **bounds** that prompt: it keeps system + latest user message + the most-recent tool round (dropping older turns), caps oversized tool results, and stops duplicating the full prompt into `user_instruction` (which now carries only the short latest user message) — while still feeding the most-recent tool result back so the agent continues (`open-sse/executors/gitlab.ts`, [#6220](https://github.com/diegosouzapw/OmniRoute/issues/6220)). The unit test covers the bounding logic; the upstream 422→200 clearing is VPS-only (Hard Rule #18). Regression guard: `tests/unit/gitlab-tool-exchange-bounded-6220.test.ts`.
+
+- **fix(i18n):** the provider-detail (`/dashboard/providers/[id]`) connection-status filter labels no longer render as `__MISSING__:All` / `__MISSING__:Active` / `__MISSING__:Error` / `__MISSING__:Banned` / `__MISSING__:CreditsExhausted` in non-English locales (notably pt-BR) ([#6290](https://github.com/diegosouzapw/OmniRoute/issues/6290)). Root cause was **not** the namespace mismatch the issue guessed — the `providers.filter*` keys resolve correctly in `en.json`; the debt lived in the locale mirrors (`src/i18n/messages/*.json`), where these five keys carried the `__MISSING__:` sync sentinel in ~15 locales and were absent entirely in ~26 others, so next-intl found the key and echoed the sentinel verbatim. All 40 non-English/-Chinese mirrors now ship real translations for the five `providers.filter*` labels. Regression guard: `tests/unit/i18n-provider-filter-keys-6290.test.ts`. (thanks @diegosouzapw)
+
+- **fix(providers):** the `copilot-m365-web` streaming executor now emits `debug`-level WebSocket diagnostics ([#6210](https://github.com/diegosouzapw/OmniRoute/issues/6210)) — the outbound WS URL (with the `access_token` **redacted** via `redactWsUrl()`), handshake success/failure, and each received SignalR frame's `type`/`target`. Previously the streaming path logged nothing, so an empty `content:null` response (the M365 Education / Starter tier symptom fixed in #6234) was undiagnosable even at `APP_LOG_LEVEL=debug`. The change is debug-level and side-effect-free — it does not alter streaming behavior or the frame parser, and the token never reaches the logs. Regression guard: `tests/unit/copilot-m365-web-logging-6210.test.ts` (thanks @qpeyba)
+
+- **fix(resilience):** a round-robin combo no longer returns `503 all upstream accounts are unavailable` when a compatibility-rejected target is actually healthy ([#6238](https://github.com/diegosouzapw/OmniRoute/issues/6238)). `filterTargetsByRequestCompatibility` drops request-incompatible targets (tool/vision/structured-output unsupported, or below the required context window) **before** any availability check runs, and its `compatible.length === 0` safety net only fired when _all_ targets were filtered — not when the kept targets later all turned out runtime-unavailable (circuit-open / cooldown / no credentials). So a combo could 503 while a compat-rejected-but-healthy provider sat unused. `handleRoundRobinCombo` now keeps the compat-rejected set and, when every compat-kept target was skipped without a single real attempt, probes those rejected targets as a **last-resort fallback tier** (via the new pure `open-sse/services/combo/comboCompatFallback.ts`) before crystallizing the 503. Regression guard: `tests/unit/combo-roundrobin-compat-fallback-6238.test.ts`. (thanks @ThongAccount)
+
+- **fix(startup):** best-effort self-heal for a corrupted Turbopack dev cache on Windows ([#6289](https://github.com/diegosouzapw/OmniRoute/issues/6289)). On Windows, `pnpm dev` can fail at startup when Turbopack `mmap`s a persistent-cache SST file and the OS refuses the mapping (`os error 1455` — "paging file too small"), which Turbopack surfaces as a misleading `Module not found: Can't resolve '@/shared/utils/machine'`. This is a **known upstream Turbopack cache-corruption bug — not our code**. The dev launcher (`scripts/dev/run-next.mjs`) now wraps `nextApp.prepare()` and, when it rejects with that signature (`isTurbopackCacheCorruption` in the new `scripts/dev/turbopackCacheHeal.mjs`), purges `.build/next/**/cache/turbopack` and retries **once** with a clear log. **Caveat — best-effort only:** the corruption often surfaces as a runtime overlay rather than a `prepare()` rejection, so this cannot always intercept it; the reliable remedy remains manually deleting the Turbopack cache dir. Regression guard: `tests/unit/turbopack-cache-heal-6289.test.ts`. (thanks @chirag127)
+
+- **fix(providers):** qodercli PAT auth no longer fails with `spawn qodercli ENOENT` on Windows ([#6263](https://github.com/diegosouzapw/OmniRoute/issues/6263)) — `spawnQoderCli` spawned the bare `qodercli` name with `shell:false` and an unenriched env, so the npm `.cmd` wrapper under `%APPDATA%\npm` (a user-PATH directory) was never resolved. It now resolves the absolute `.cmd`/`.exe` path through the existing `getCliRuntimeStatus("qoder")` resolver in `src/shared/services/cliRuntime.ts` (memoized), spawns with `shell` when the target is a `.cmd`/`.bat`, and uses the cliRuntime-enriched env (PATH + PATHEXT + APPDATA); the ENOENT error now lists the searched paths plus the `CLI_QODER_BIN` override. End-to-end spawn on a real Windows host is host-only (Hard Rule #18); the path-resolution logic is unit-tested. Regression guard: `tests/unit/qodercli-windows-resolve-6263.test.ts`. (thanks @chirag127)
+
+- **fix(sse):** the reasoning-token buffer no longer inflates **probe-sized `max_tokens`** ([#6274](https://github.com/diegosouzapw/OmniRoute/issues/6274)) — Claude Code's `/model` capability check sends `max_tokens: 1`, but for a thinking-capable model with a large output cap (e.g. `glm-5.2`) the #3587 headroom heuristic (`max(current + 1000, ceil(current * 1.5))`) rewrote it to `1001` and forwarded that upstream, wasting tokens on a request that was never a genuine reasoning budget. `resolveReasoningBufferedMaxTokens()` (`open-sse/services/reasoningTokenBuffer.ts`) now short-circuits and returns the caller's value verbatim when it is below the new `REASONING_BUFFER_MIN_TRIGGER` (256) threshold — a tiny explicit limit is a probe, not a reasoning request. Real budgets still receive the #3587 headroom unchanged, and the guard runs after the existing capability checks so unknown / non-reasoning models keep returning `null`. Regression guard: `tests/unit/reasoning-token-buffer-6274.test.ts`. (thanks @brightfiscalband)
+
+- **fix(cli):** `omniroute reset-password` now works as a real subcommand, and password resets over piped (non-TTY) stdin actually apply ([#6261](https://github.com/diegosouzapw/OmniRoute/issues/6261), [#6258](https://github.com/diegosouzapw/OmniRoute/issues/6258)). Two coupled defects: (1) **#6261** — `bin/omniroute.mjs` routed everything through Commander with only two pre-Commander bypasses (`--mcp`, `reset-encrypted-columns`), so `omniroute reset-password` was rejected as an unknown command; only the separate `omniroute-reset-password` bin worked, while the docs falsely advertised the subcommand (incl. a bogus "legacy alias still works"). A pre-Commander bypass mirroring `reset-encrypted-columns` now dynamically imports `bin/reset-password.mjs` (which self-executes) before Commander parses; the three doc lines were corrected. (2) **#6258** — `bin/reset-password.mjs` issued two sequential `rl.question` prompts; under piped stdin the second read never settled at EOF, so `main()` never reached `resetManagementPassword` and the reset was a silent no-op (both prompts printed, no success, password unchanged). The CLI now detects non-TTY stdin and reads it once (first line = password, second line = confirm if present, else reused), adds a `--password-stdin` flag (entire stdin is the password, no confirmation), and exits `0` explicitly so the success line always flushes; interactive TTY behavior is unchanged. Regression guard: `tests/unit/reset-password-cli-6261-6258.test.ts` (3). (thanks @chirag127)
+
+- **fix(db):** the mass-migration **safety abort** now tells the operator how to bypass it and stops flooding the log ([#6260](https://github.com/diegosouzapw/OmniRoute/issues/6260)) — after restoring a backup that wiped the migration tracking table, `runMigrations()` threw the abort on every downstream `ensureDbInitialized()`, re-logging the full banner 11+ times, and the message never mentioned the existing `OMNIROUTE_MAX_PENDING_MIGRATIONS` escape hatch. The abort text now appends a bypass hint (set `OMNIROUTE_MAX_PENDING_MIGRATIONS=0` in `server.env` / `DATA_DIR/.env`), and a new `MigrationSafetyAbortError` is memoized so repeated calls in the same process throw the same instance and emit a single concise line instead of the full cascade. Regression guard: `tests/unit/migration-safety-abort-6260.test.ts`. (thanks @chirag127)
+
+- **fix(auth):** importing a **distinct** Codex/ChatGPT OAuth `auth.json` is no longer falsely rejected as "already exists" when it belongs to a different user in the same workspace ([#6301](https://github.com/diegosouzapw/OmniRoute/issues/6301)). `findExistingCodexConnection` (in `src/lib/oauth/utils/codexAuthImport.ts`) deduped **only** on `providerSpecificData.workspaceId === accountId`, where `accountId` is the shared `chatgpt_account_id`/`tokens.account_id` — so two members of the same ChatGPT Team collapsed onto a single connection (409 `duplicate_account`). The id_token's `https://api.openai.com/auth` claim carries a per-user `chatgpt_user_id` alongside the workspace id (the device-flow path already persisted it as `chatgptUserId`, but the import path did not). Now `parseAndValidateCodexAuth` extracts `userId` (`chatgpt_user_id` → `user_id` → JWT `sub`) into `ParsedCodexAuth`, the create/update paths persist `chatgptUserId` in `providerSpecificData` (mirroring `codex.ts`), and dedup keys on `workspaceId` **AND** `chatgptUserId` — with a backward-compat fallback to legacy accountId-only matching when no stored connection for that workspace records a `chatgptUserId`, so genuinely-same accounts still dedup. Regression guard: `tests/unit/codex-auth-import-userid-dedup-6301.test.ts` (4). (thanks @anungma)
+
+- **fix(providers):** importing models for the **venice-web** provider no longer fails with a red "Provider venice-web does not support models listing" ([#6269](https://github.com/diegosouzapw/OmniRoute/issues/6269)). `venice-web` is a web-cookie provider with an executor but no upstream `/v1/models` endpoint and no registry `models`, so the models route fell through to the tail `400`. Mirroring the `jules`/`linkup-search`/`ollama-search` fix (#5569), it now ships a static local catalog entry in `src/lib/providers/staticModels.ts` — seeding the current Venice lineup (`venice-uncensored`, `llama-3.3-70b`, `qwen3-235b`, `qwen3-4b`, `deepseek-r1-671b`; Venice rotates its catalog, see docs.venice.ai/models/overview) — so the route returns `200` with `source:"local_catalog"`, `intentional:true`. Regression guard: `tests/unit/static-models-venice-web-6269.test.ts`. (thanks @chirag127)
+
+- **fix(api):** the specialty model catalogs (`/v1/embeddings`, `/v1/images`, `/v1/music`, `/v1/videos` model lists) are now derived from the **unified catalog filtered by a predicate** (`getSpecialtyModelsResponse`) instead of ad-hoc per-route logic, so they consistently respect active-credential visibility and stay in sync with the main catalog ([#6303](https://github.com/diegosouzapw/OmniRoute/pull/6303)). Regression guard: `tests/unit/specialty-model-catalog-routes.test.ts`. (thanks @makcimbx)
+- **fix(api):** the agent-bridge server route now resolves the MITM manager via a **dynamic `import("@/mitm/manager.runtime")`** so Turbopack does not statically pull the stub (or over-bundle the manager), and the agent-skills generator anchors its output base path with `path.join(process.cwd(), …)` so Turbopack's static analyzer stops tracing the whole project root ([#6329](https://github.com/diegosouzapw/OmniRoute/issues/6329), [#6366](https://github.com/diegosouzapw/OmniRoute/pull/6366)). Regression guard: `tests/unit/agent-bridge-server-route-dynamic-import.test.ts`. (thanks @Iammilansoni)
+- **fix(api):** internal probes (combo-test, cloud-sync verify) now pick a **management-scoped / allow-all API key** instead of naively grabbing `getApiKeys()[0]` — a restricted `self:usage` first row made the probe fail with "Model X is not allowed for this API key" even when the combo path was healthy (`pickApiKeyForInternalUse` in `src/lib/db/apiKeys.ts`). The API-manager model editor also falls back to `/api/models?all=true` when `/v1/models` is catalog-protected ([#6372](https://github.com/diegosouzapw/OmniRoute/pull/6372)). Regression guard: `tests/unit/pick-internal-api-key-6372.test.ts`. (thanks @jmengit)
+- **fix(live-ws):** the Live Dashboard WebSocket server now **rejects on bind failure** (e.g. `EADDRINUSE` when the API bridge already holds the port) instead of letting the error surface as an unhandled `error` event that crash-loops the process — the `error` listener is attached to `wss` (not `server`) and releases the EventBus subscription on a failed start ([#6324](https://github.com/diegosouzapw/OmniRoute/issues/6324)). Regression guard: `tests/unit/live-ws-eaddrinuse-6324.test.ts`. (thanks @vinayakkulkarni)
+- **fix(dashboard):** the Home provider-topology widget now trusts the live provider-metrics snapshot — it uses `topology.errorProvider` and live `activeRequests` directly instead of re-deriving state from a stale `lastErrorAt` or applying a frontend timeout filter, so the topology reflects real-time provider health ([#6322](https://github.com/diegosouzapw/OmniRoute/pull/6322)). Regression guard: `tests/unit/home-provider-topology-live-state.test.ts`. (thanks @xz-dev)
+- **fix(sse):** strip zero-width markers from streamed **tool-call arguments** — a follow-up to [#5857](https://github.com/diegosouzapw/OmniRoute/pull/5857). That PR removed injected zero-width joiners (U+200D) from streamed assistant text/reasoning but deliberately left tool-call argument JSON byte-exact. The request-side obfuscation (`open-sse/services/claudeCodeObfuscation.ts`) injects ZWJ into agent words — including the temp path inside the Bash tool description — and Claude models copy that verbatim into generated commands, which are delivered as tool-call arguments rather than assistant text. As a result the ZWJ survived and corrupted code blocks (e.g. a temp path rendered with an invisible joiner). Now `open-sse/handlers/responseSanitizer.ts` strips zero-width code points from tool-call argument strings at every emit site (OpenAI non-stream/stream chat `tool_calls` + legacy `function_call`, native Responses `function_call` items, the OpenAI→Responses conversion, and the native Responses streaming `response.function_call_arguments.delta/.done` events). Only zero-width code points are removed; JSON structure and all other bytes stay identical (no parse/restringify), so normal arguments remain byte-exact. Regression guard: 6 new cases in `tests/unit/response-sanitizer.test.ts` (suite 50/50).
+
+- **fix(nodejs):** the default app log path now resolves under `DATA_DIR` (`~/.omniroute/logs/application/app.log`) instead of `process.cwd()` ([#6197](https://github.com/diegosouzapw/OmniRoute/issues/6197)) — the globally-installed CLI runs from an arbitrary working directory, so anchoring the default to cwd made file logging silently write to (or no-op under) an unrelated directory, contradicting the documented `.env.example` default. `getAppLogFilePath()` now computes the default lazily via the pure `resolveDataDir()` resolver (honours a per-process `DATA_DIR`, no directory-creation side effect); an explicit `APP_LOG_FILE_PATH` still wins. Regression guard: `tests/unit/logenv-datadir-path-6197.test.ts` (3).
+
+- **fix(docker):** AgentBridge/`startMitm` no longer aborts in containers/headless when the Antigravity-default DNS step can't write `/etc/hosts` ([#6127](https://github.com/diegosouzapw/OmniRoute/issues/6127)), and the privileged command's stderr now reaches `app.log` instead of only a bare exit code hitting the toast ([#6198](https://github.com/diegosouzapw/OmniRoute/issues/6198)). The default DNS step (`addDNSEntry`) was called unguarded while cert install and the two sibling DNS steps were each best-effort — in the runtime Docker image (`USER node`, no `sudo`, read-only `/etc/hosts`) it threw `Command failed with code 1` out of `startMitmInternal` and killed the whole start, discarding the stderr. The three DNS steps are extracted into a best-effort `provisionDnsEntries()` where each failure is logged with the full `err` (stderr included, folded in by `systemCommands.ts`) and never aborts the start. Regression guard: `tests/unit/mitm-dns-graceful-degrade-6127.test.ts` (4).
+
+- **fix(providers):** copilot-m365-web now supports the M365 Education "Starter / OfficeWebIncludedCopilot" tier and no longer returns an empty `content:null` stream ([#6210](https://github.com/diegosouzapw/OmniRoute/issues/6210)). Two gaps: (1) `buildWsUrl()` hardcoded the individual-consumer scenario (`OfficeWebPaidConsumerCopilot`, `isEdu=false`) — the EDU tier is now opt-in via `providerSpecificData.tier="edu"`, emitting `scenario=OfficeWebIncludedCopilot`/`isEdu=true` (the individual path is unchanged); (2) the EDU/GPT-5.5 path streams deltas via `arguments[0].writeAtCursor` (incremental) instead of only `messages[].text` (accumulated snapshots), which the parser dropped — a new `accumulateBotContent()` folds both formats, with `type:2 item.result.message` as a last-resort fallback. Regression guard: `tests/unit/copilot-m365-edu-writeatcursor-6210.test.ts` (10). (thanks @qpeyba)
+
+- **fix(providers):** GitLab Duo executor now feeds tool results back into the prompt instead of looping ([#6220](https://github.com/diegosouzapw/OmniRoute/issues/6220)) — `buildPrompt()` branched only on `system`/`user` and took `userParts.at(-1)`, silently dropping the `assistant{tool_calls}` + `tool{result}` turns the client appended, so the reconstructed prompt was byte-identical to turn 1 and the model re-emitted the same `<tool>` call forever. When a tool exchange is present the full conversation is now serialized, folding each tool result back keyed by its `tool_call_id`; simple conversations keep the legacy shape. Complements the tool_call emission from [#6051](https://github.com/diegosouzapw/OmniRoute/issues/6051) (the `kilo-duplicate` label was a false positive — different, sequential defect). Regression guard: `tests/unit/gitlab-tool-result-feedback-6220.test.ts` (4).
+
+- **fix(providers):** opencode-go/opencode-zen can now synthesize the OpenCode CLI identity headers Cloudflare requires on VPS egress ([#5997](https://github.com/diegosouzapw/OmniRoute/issues/5997)) — on a datacenter VPS, `opencode.ai/zen/go/v1/chat/completions` 403s (HTML challenge) requests lacking CLI identity, while the reporter's control curl proved that `User-Agent: opencode-cli/1.0.0` + `x-opencode-client: cli` + `x-opencode-project: default` + fresh request/session UUIDs succeed. Opt-in via `OPENCODE_SYNTHESIZE_CLI_HEADERS=true` (values overridable via `OPENCODE_GO_USER_AGENT`/`OPENCODE_USER_AGENT`/`OPENCODE_CLIENT`/`OPENCODE_PROJECT`); it fills only headers the client did not already send. Kept **off by default** — the forward-only path is deliberate (fabricating a wrong value risks upstream rejection; a prior dedup regressed with `opencode/local`), so this replaces the fragile local header-injection shim without changing default behavior. Regression guard: `tests/unit/opencode-cli-headers-synthesis-5997.test.ts` (6). (thanks @aleksesipenko)
+- fix(resilience): sticky session affinity now evicts and fails over to another account when the pinned account is exhausted/unavailable (#6219)
+
+- fix(sse): Responses API passthrough now drops internal commentary-phase output before forwarding to clients (gated by RESPONSES_PASSTHROUGH_DROP_COMMENTARY, default on) (#6199)
+
+- **fix(sse):** tool-call function schemas with a root `type: null` are now coerced to `type: "object"` before dispatch ([#6359](https://github.com/diegosouzapw/OmniRoute/issues/6359)) — clients like the Codex app emit `parameters: { type: null, ... }` for some tools, which OpenAI-compatible upstreams reject with `400 Invalid schema for function '...': schema must be a JSON Schema of 'type: "object"', got 'type: null'`, failing the whole request. `toolSchemaSanitizer` already stripped the null; it now re-adds the mandatory root `"object"` type (and empty `properties`/open `additionalProperties` when absent). Combinator roots (`anyOf`/`oneOf`/`allOf`) and explicit root types are left untouched. Regression guard: 5 new cases in `tests/unit/tool-schema-sanitizer.test.mjs`.
+
+- **fix(docker):** AgentBridge no longer fails to start on npm/Electron/VPS installs with "MITM manager stub reached at runtime" ([#6344](https://github.com/diegosouzapw/OmniRoute/issues/6344)) — v3.8.45 flipped the production bundler default to Turbopack, but `next.config.mjs` aliased `@/mitm/manager` to its Docker-only degraded stub **unconditionally**. That was harmless while Docker (which sets the alias intentionally for #3390 graceful degradation) was the sole Turbopack consumer, but once every artifact built with Turbopack the stub shipped to all non-Docker users and `startMitm` threw on the first Agent-Bridge start. The alias is now opt-in via `OMNIROUTE_MITM_STUB=1` (set only by the Dockerfile) through the shared `scripts/build/mitm-stub-flag.mjs` helper; default builds bundle the real manager. Regression guard: `tests/unit/mitm-stub-alias-6344.test.mjs` (4).
+
+- **fix(proxy):** stop the v3.8.44 proxy regression that leaked the real IP and disabled healthy proxies ([#6246](https://github.com/diegosouzapw/OmniRoute/issues/6246)). Two coupled defects from the new health scheduler: (1) **IP leak** — when a proxy assigned to a connection was marked `inactive`, resolution fell through to a **direct** egress instead of blocking, exposing the operator's real IP; (2) **over-deactivation** — the sweep flipped a proxy to `inactive` on the **first** failed probe and counted our own 5s timeout / a probe-target `5xx` as the proxy's fault, so healthy paid proxies vanished from egress selection ("my proxies are not being used anymore"). Fix: the sweep decision is extracted into a pure, network-free `decideProxyHealthAction` (`src/lib/proxyHealth/decision.ts`) — by default the health check now **only counts/logs and never downgrades status** (a proxy is downgraded/removed only with `PROXY_AUTO_REMOVE=true`, after `PROXY_AUTO_REMOVE_AFTER` **consecutive** conclusive failures); probes are classified tri-state so an inconclusive result (our timeout, or a `5xx` from the probe target) never penalizes the proxy, and the probe timeout is raised 5s→15s. Separately, `safeResolveProxy` now **fails closed** via the existing policy: a connection whose assigned proxy is dead is blocked instead of leaking direct (`hasBlockingProxyAssignment`), honoring the explicit `proxy off` toggles and the `PROXY_FAIL_OPEN=true` opt-out. Existing proxies stuck `inactive` by the old behavior need a one-time manual re-activate (the operator owns proxy status). Regression guards: `tests/unit/proxy-health-decide-action-6246.test.ts`, `tests/unit/proxy-assigned-unavailable-6246.test.ts`.
+
+- **fix(proxy):** make "Test All" read-only and add bulk enable/disable ([#6246](https://github.com/diegosouzapw/OmniRoute/issues/6246)). Complements the core fail-closed / scheduler fix (#6296) with the two remaining reporter asks. (1) The **"Test All" button** (`POST /api/settings/proxies/auto-test`) used to flip a proxy to `inactive` on a failed reachability probe; since the egress selector excludes `inactive` proxies, a flaky probe (an unreachable `httpbin.org`, a proxy that blocks `HEAD`, or a slow paid proxy) silently disabled every proxy that failed — "Test All" is now **read-only by default** (only the operator sets a proxy active/inactive; opt back into the legacy test-and-set with `PROXY_HEALTH_AUTO_DEACTIVATE=true`). (2) Adds a **bulk enable/disable** proxies endpoint + toolbar action (`POST /api/settings/proxies/batch-activate`) so an operator can re-activate proxies in one click. Regression guard: `tests/unit/proxy-health-6246.test.ts`. (thanks @tenshiak)
+
+- **chatcore (tools): stop the default 128-tool cap from silently dropping opencode's `task`/MCP tools.** opencode (used as an MCP/agent host) sends a large tool list; when it exceeds the speculative `MAX_TOOLS_LIMIT` (128) default, `truncateToolList` did a blind `tools.slice(0, 128)`, dropping every tool past index 128 — including opencode's built-in `task` tool (subagent launch) and many MCP tools, so models routed through OmniRoute could no longer spawn subagents or reach part of their tools. The cap exists to avoid upstream `400`s for providers with real hard limits (e.g. grok-cli 200), so it is kept for those: detection of the opencode client (`isOpencodeClient` — any `x-opencode-*` header, or `opencode` in the user-agent) now only bypasses the **speculative 128 default**, never a known provider ceiling. Precedence is explicit — a proactive/detected provider limit always truncates (even for opencode); otherwise opencode forwards its full tool list; otherwise the unchanged 128 default applies to every other client. Refactors `getEffectiveToolLimit` into `getKnownToolLimit(provider) ?? DEFAULT_LIMIT` (byte-identical for existing callers) and fixes a cosmetic debug-log that reported the truncated count instead of the original. Regression guard: `tests/unit/tool-limit-detector.test.ts`.
+
+- **fix(mitm):** the macOS MITM-cert install check now matches the system keychain again. `security find-certificate -a -Z` prints the SHA-1 as a colon-less hex string, but the installed-check compared it against `getCertFingerprint()`'s colon-separated form, so the substring match never hit — the cert was reported as not-installed and re-prompted for the sudo install on every run. Fingerprints are now normalized (colons stripped, upper-cased) on both sides via the extracted `macCertOutputHasFingerprint` helper. Regression guard: `tests/unit/mitm-cert-mac-fingerprint.test.ts`. ([#6204](https://github.com/diegosouzapw/OmniRoute/pull/6204), closes [#6134](https://github.com/diegosouzapw/OmniRoute/issues/6134) — thanks @rianonehub)
+
+- **fix(api):** `/v1/messages/count_tokens` now counts `tool_use`, `tool_result` and `thinking` content blocks (and array-form `system` prompts) in the local-estimation path, instead of only `text`. Real agentic conversations keep ~95% of their tokens inside tool results; the previous estimate returned near-zero for them, which silently broke Claude Code's auto-compaction (context grew past the window with no compaction until the upstream API rejected the request). The real provider-side count path is unchanged. Regression guard: `tests/unit/messages-count-tokens-route.test.ts`. ([#6221](https://github.com/diegosouzapw/OmniRoute/pull/6221) — thanks @luweiCN)
+
+- **fix(antigravity):** strip a trailing assistant prefill turn for Vertex Claude models to avoid upstream 400s ([#6114](https://github.com/diegosouzapw/OmniRoute/pull/6114)). Regression guard: `tests/unit/antigravity-claude-prefill-strip.test.ts`. (thanks @anki1kr)
+
+- **fix(security):** the mutable cloud-agent routes (`/api/cloud/credentials/update`, `/api/cloud/models/alias`) now require management auth instead of being treated as public. They were classified as public API routes, so a request without management credentials could update stored cloud-agent credentials and model aliases. They are removed from the public-route set, classified as management routes in the authz pipeline, and gated by `requireManagementAuth`; cloud **read**/auth routes stay public. Regression guards: `tests/unit/cloud-write-auth.test.ts`, `tests/unit/authz/classify.test.ts`, `tests/unit/public-api-routes.test.ts`. ([#6233](https://github.com/diegosouzapw/OmniRoute/pull/6233) — thanks @vittoroliveira-dev)
+
+- **refactor(dashboard):** extract the onboarding-wizard "Open provider details" link target into a pure, unit-tested `buildProviderDetailsHref(connection)` helper. The wizard already routes by `connection.id` (the node UUID) rather than the provider category slug (#6144/#6145); this hardens that behavior behind a tested helper that guards a missing id/connection. Regression guard: `tests/unit/provider-onboarding-href.test.ts`. ([#6166](https://github.com/diegosouzapw/OmniRoute/pull/6166) — thanks @KooshaPari)
+
+- **fix(security):** the doubao synthetic device-id generator now derives its digits via an unbiased crypto-random draw (rejection sampling over `crypto.randomBytes()`) instead of a `% 10` reduction, closing a CodeQL `js/biased-cryptographic-random` finding.
+- **fix(agentSkills):** the GitHub-skills generator now resolves `outputDir` to an absolute path before writing, fixing a regression introduced by #6366 (relative-to-cwd base path) that could write generated skill files to the wrong directory.
+- **fix(security):** `/api/keys/{id}/devices` now checks the HTTP method before auth/validation, returning a `405` for non-GET/DELETE verbs instead of a misleading `401`/`500` (closes a `dast-smoke` QUERY-method finding).
+- **fix(quality):** clear the last 2 heavy quality-gate reds on the release tip (cycle pre-flight).
+- **fix(mitm):** the test suite and CI can never mutate the OS trust store — `OMNIROUTE_SKIP_SYSTEM_TRUST=1` is set globally for tests/CI so `installCert`/`uninstallCert`/`installTproxyCa` skip the privileged OS dispatch ([#6310](https://github.com/diegosouzapw/OmniRoute/pull/6310); full detail is under the [3.8.45] section below — this branch received it via the parallel-cycle sync-back).
+- **fix(api):** `POST /api/github-skills` now Zod-validates its request body; documented the new quality-gate env vars and pinned the merge-integrity GitHub Actions to a commit SHA.
+- **fix(skills):** generate the missing `omni-github-skills` registry entry and align the agent-skills catalog-count tests (follow-up to #6186).
+- **fix(quality):** clear the cycle's 11 net-new ESLint errors and make `validate-release-green` suppressions-aware.
+
+### 📝 Maintenance
+
+- **i18n(it):** add 118 missing Italian (`it`) translations (net-additive — no existing keys dropped, valid JSON), improving Italian UI coverage. ([#6212](https://github.com/diegosouzapw/OmniRoute/pull/6212) — thanks @serverless83)
+- **chore(providers):** remove deprecated **MiMo V2** model entries from the catalogs (xiaomi-mimo, opencode-go, zenmux-free, audio TTS) — the upstream V2 line is superseded by MiMo V2.5; drops `mimo-v2-tts`, `mimo-v2-pro`, `mimo-v2-omni`, `mimo-v2-flash`, `mimo-v2-flash-free` and realigns the provider-catalog tests. ([#6248](https://github.com/diegosouzapw/OmniRoute/pull/6248) — thanks @backryun)
+- **chore(release):** ~50 commits on this branch are v3.8.45 pre-flight/hardening fixes and CI-perf work that landed here via the parallel-cycle sync-back (`sync-next-cycle.mjs`, Hard Rule #21) **after** the `v3.8.45` git tag was cut, and are already fully documented under the **[3.8.45]** section below — listed here only so the per-cycle commit-coverage check (`npm run release:uncovered`) doesn't flag them as gaps. Provider/catalog/UX/backend fixes: #6041, #6078, #6108, #6135, #6148, #6149, #6154, #6158, #6161, #6162, #6163, #6164, #6165, #6170, #6177, #6178, #6181, #6186, #6187, #6191, #6193, #6194, #6195, #6200, #6205, #6208, #6209, #6211, #6213, #6223, #6224, #6225, #6226, #6227, #6228, #6229, #6230, #6235, #6291, #6292. CI/release-pipeline work: #6167, #6203, #6214, #6215, #6218, #6273, #6275, #6283, #6284, #6285, #6300, #6305.
+- **chore(release):** additional zero-ref release-cycle plumbing on this branch, kept out of `release:uncovered` on purpose (no `#N` in the commit subject to cite): opening the v3.8.46 cycle, opening/closing the v3.8.45 cycle, the finalized [3.8.45] CHANGELOG i18n sync-back to 42 mirrors, the v3.8.45 cognitive/cyclomatic and file-size drift rebaselines, ESLint stale-suppression pruning (4,273 → 4,233), and clearing test-masking/docs-all pre-flight reds for v3.8.45.
+
+### ⚡ Performance & Infrastructure
+
+- **perf(release-green):** the pre-flight validator (`scripts/quality/validate-release-green.mjs`) now runs its 4 slow suites (unit / vitest / integration / pack-artifact) **concurrently** via `Promise.all` — pre-flight wall time drops from ~the sum of the suites to ~the slowest one (~30min saved per round; Phase 0 was the nº1 bottleneck of the v3.8.45 release benchmark, 2h54 of 6h34 e2e). Guard: `tests/unit/validate-release-green.test.ts` ("runs the slow suites CONCURRENTLY"). ([#6319](https://github.com/diegosouzapw/OmniRoute/pull/6319))
+- **fix(ci):** `scripts/release/sync-next-cycle.mjs` — two defects found live in its first production run (v3.8.45 Phase 5): (1) the `git()` helper's default 1 MiB `maxBuffer` crashed with `ENOBUFS` on `git show origin/main:CHANGELOG.md` (the CHANGELOG alone is >1 MiB) — widened to 64 MiB; (2) the i18n resync only propagated the `[NEXT]` (TBD) section, leaving the just-shipped finalized section as "— TBD" in all 42 mirrors — it now also syncs `[prevVersion]` bounded by the heading below it (new exported pure helper `versionAfter`). Guards: +5 tests in `tests/unit/sync-next-cycle.test.ts` (8/8). ([#6327](https://github.com/diegosouzapw/OmniRoute/pull/6327))
+- **test(ci):** concurrency-sensitive flaky tests are **quarantined into a serial pass** (`tests/unit/serial/`, `--test-concurrency=1`, appended to every unit runner incl. sharded variants — the serial pass is sharded too so concurrent shard jobs never self-collide). Initial set: `glm-coding-plan-monthly-3580`, `quota-division-blocks`, `provider-health-autopilot`, `combo-health-autopilot` — the class behind the ~28min CI wedges/re-runs (two live 1h+ wedges cancelled during this PR's own validation). Discovery + TIA gates track the new glob; systemic root cause (async logger writing after teardown) tracked in [#6360](https://github.com/diegosouzapw/OmniRoute/issues/6360). Guard: `tests/unit/test-serial-quarantine.test.ts` (4). ([#6347](https://github.com/diegosouzapw/OmniRoute/pull/6347))
+
+### 🙌 Contributors
+
+Thanks to everyone whose work landed in v3.8.46:
+
+| Contributor                                                    | PRs / Issues           |
+| -------------------------------------------------------------- | ---------------------- |
+| [@2220258345](https://github.com/2220258345)                   | direct commit / report |
+| [@abdofallah](https://github.com/abdofallah)                   | direct commit / report |
+| [@adentdk](https://github.com/adentdk)                         | direct commit / report |
+| [@aleksesipenko](https://github.com/aleksesipenko)             | direct commit / report |
+| [@anki1kr](https://github.com/anki1kr)                         | direct commit / report |
+| [@anungma](https://github.com/anungma)                         | direct commit / report |
+| [@arpicato](https://github.com/arpicato)                       | direct commit / report |
+| [@backryun](https://github.com/backryun)                       | #6248                  |
+| [@binsarjr](https://github.com/binsarjr)                       | direct commit / report |
+| [@brightfiscalband](https://github.com/brightfiscalband)       | direct commit / report |
+| [@chirag127](https://github.com/chirag127)                     | #6501, #6506           |
+| [@developerjillur](https://github.com/developerjillur)         | direct commit / report |
+| [@dilneiss](https://github.com/dilneiss)                       | #6499                  |
+| [@dtybnrj](https://github.com/dtybnrj)                         | direct commit / report |
+| [@Forcerecon](https://github.com/Forcerecon)                   | direct commit / report |
+| [@hao3039032](https://github.com/hao3039032)                   | direct commit / report |
+| [@Iammilansoni](https://github.com/Iammilansoni)               | #6150, #6245           |
+| [@jmengit](https://github.com/jmengit)                         | direct commit / report |
+| [@jordansilly77-stack](https://github.com/jordansilly77-stack) | direct commit / report |
+| [@JxnLexn](https://github.com/JxnLexn)                         | direct commit / report |
+| [@KooshaPari](https://github.com/KooshaPari)                   | #6166                  |
+| [@loopyd](https://github.com/loopyd)                           | direct commit / report |
+| [@luweiCN](https://github.com/luweiCN)                         | #6221                  |
+| [@makcimbx](https://github.com/makcimbx)                       | direct commit / report |
+| [@muflifadla38](https://github.com/muflifadla38)               | direct commit / report |
+| [@newnol](https://github.com/newnol)                           | direct commit / report |
+| [@ofekbetzalel](https://github.com/ofekbetzalel)               | direct commit / report |
+| [@ohahe52-dot](https://github.com/ohahe52-dot)                 | direct commit / report |
+| [@phidinhmanh](https://github.com/phidinhmanh)                 | direct commit / report |
+| [@powellnorma](https://github.com/powellnorma)                 | direct commit / report |
+| [@qpeyba](https://github.com/qpeyba)                           | direct commit / report |
+| [@RaviTharuma](https://github.com/RaviTharuma)                 | direct commit / report |
+| [@RCrushMe](https://github.com/RCrushMe)                       | direct commit / report |
+| [@rianonehub](https://github.com/rianonehub)                   | #6134, #6204           |
+| [@serverless83](https://github.com/serverless83)               | #6212                  |
+| [@swingtempo](https://github.com/swingtempo)                   | direct commit / report |
+| [@tenshiak](https://github.com/tenshiak)                       | direct commit / report |
+| [@ThongAccount](https://github.com/ThongAccount)               | direct commit / report |
+| [@UnrealAryan](https://github.com/UnrealAryan)                 | direct commit / report |
+| [@vinayakkulkarni](https://github.com/vinayakkulkarni)         | direct commit / report |
+| [@vittoroliveira-dev](https://github.com/vittoroliveira-dev)   | #6233                  |
+| [@warelik](https://github.com/warelik)                         | direct commit / report |
+| [@xxy9468615](https://github.com/xxy9468615)                   | direct commit / report |
+| [@xz-dev](https://github.com/xz-dev)                           | direct commit / report |
+| [@yanpaing007](https://github.com/yanpaing007)                 | direct commit / report |
+| [@diegosouzapw](https://github.com/diegosouzapw)               | maintainer             |
+
+---
+
 ## [3.8.45] — 2026-07-06
 
 ### ✨ New Features

docs/i18n/he/CHANGELOG.md

@@ -6,6 +6,213 @@
 
 ## [3.8.31] — 2026-06-20
 
+## [3.8.46] — TBD
+
+### ✨ New Features
+
+- **feat(sse):** **hide paid-only models from `auto/*` routing** when `hidePaidModels` is on ([#6512](https://github.com/diegosouzapw/OmniRoute/issues/6512)) — follow-up to #6328/#6495. PR #6495 hid paid-only models from the `GET /v1/models` listing, but `auto/*` combos (`auto/best-coding`, `auto/glm`, …) could still pick a paid-only backend into their candidate pool → a 402/403 at request time. `createVirtualAutoCombo` now filters the candidate pool through the new pure `open-sse/services/autoCombo/paidModelFilter.ts` (`filterPaidOnlyCandidates`), applying the same free-model predicate #6495 uses in `catalog.ts` (`providerHasFreeModels(provider) && isFreeModel(provider, {id})`) whenever `settings.hidePaidModels === true`. Applied before the category/tier/family narrowing, so it covers every `auto/*` combo; an all-paid pool degrades to the existing graceful empty-pool path. **Opt-in — default OFF leaves the pool unchanged** (identity). Regression guard: `tests/unit/autoCombo/paid-model-filter-6512.test.ts` (4, incl. the default-off identity guard).
+- **feat(sse):** **provider-family auto combos** — `auto/glm`, `auto/minimax`, `auto/mimo`, `auto/zai`, `auto/gemma`, `auto/llama`, `auto/gemini` ([#6453](https://github.com/diegosouzapw/OmniRoute/issues/6453)) — new routable ids that materialize an on-demand virtual combo spanning whatever installed backends currently expose that model family, degrading gracefully as backends rotate. A new pure `open-sse/services/autoCombo/modelFamily.ts` (`detectModelFamily`) classifies by model-id prefix for six families; `zai` is instead resolved by provider id (z.ai's hosted API serves the same `glm-*` model ids as every other GLM backend, so `auto/zai` means "route to my z.ai backend specifically" vs `auto/glm`'s "any connected GLM backend"). Reuses the existing `createVirtualAutoCombo` on-demand materialization path (no DB writes) and the `/v1/models` catalog advertising loop. Regression guard: `tests/unit/autoCombo/provider-family-combos.test.ts` (11).
+- **feat(proxy):** native **proxy-pool round-robin / egress IP rotation** ([#6365](https://github.com/diegosouzapw/OmniRoute/issues/6365)) — a scope (global / provider / account) can now hold **multiple** proxies as a pool with a rotation strategy, so outbound requests cycle their egress IP instead of pinning one proxy per scope. Migration `117_proxy_pool_rotation.sql` lifts the `UNIQUE(scope, scope_id)` constraint (rebuild via the canonical rename/copy/drop; existing single assignments become 1-element pools) and adds a `proxy_scope_rotation` companion table holding the per-scope strategy + a persisted monotonic round-robin cursor. Strategies: `round-robin` (default, monotonic cursor — never `Math.random`), `random`, and `sticky-per-N-min`. Resolution (`resolveProxyForScopeFromRegistry` / `resolveProxyForConnectionFromRegistry`) now fetches the alive, position-ordered candidate set (unchanged `PROXY_ALIVE_PREDICATE`) and applies the strategy; an empty / all-dead pool still returns `null` — the #6246 fail-closed guard is **untouched** (never falls through to direct egress). Backend + DB only; dashboard pool-builder UI is a follow-up. Regression guard: `tests/unit/proxy-pool-rotation-6365.test.ts` (8, incl. fail-closed + backward-compat).
+- **feat(providers):** end-to-end **tool/function calling on the native Gemini `/v1beta` endpoint** ([#6222](https://github.com/diegosouzapw/OmniRoute/issues/6222)) — both directions of the Gemini↔OpenAI conversion now preserve tool data (previously silently dropped). Request side: `convertGeminiToInternal` (extracted to its own testable module) maps `tools[].functionDeclarations` → OpenAI `tools`, prior `functionCall` parts → assistant `tool_calls`, and `functionResponse` parts → `tool`-role messages. Response side: `convertOpenAIResponseToGemini` emits `parts[].functionCall {name,args}` from `message.tool_calls`, and the streaming `openAIChunkToGeminiChunk` accumulates fragmented `tool_calls` deltas by index into complete `functionCall` parts. The non-Gemini client paths (Claude, OpenAI-Responses) already preserved tool calls — this closes the gap specific to the native Gemini surface. Regression guard: `tests/unit/v1beta-gemini-tool-calling-6222.test.ts` (6, incl. a streaming SSE round-trip).
+- **feat(providers):** copilot-m365-web **enterprise / work tier** support ([#6334](https://github.com/diegosouzapw/OmniRoute/issues/6334)) — mirrors the EDU-tier pattern (#6210): `M365ConnectionParams` gains an `agent` field, a new opt-in `M365_ENTERPRISE_OVERRIDES` preset (`agent=work`, `scenario=officeweb`, `licenseType=Premium`) applies via `providerSpecificData.tier="enterprise"` (alias `"work"`), and `agent` is also overridable directly via `providerSpecificData.agent`. `buildWsUrl` was hardcoding `agent="web"` (the one enterprise-distinguishing param with no override path), so a Premium work account handshook then returned an empty stream. The individual and EDU paths are untouched. Kilo's dup flag vs #6210 (EDU tier) was a false positive — different tier. Regression guard: `tests/unit/copilot-m365-enterprise-6334.test.ts` (7). End-to-end confirmation on a real Premium work account is a live-VPS validation follow-up (Hard Rule #18). (thanks @Forcerecon)
+- **feat(api):** standardized, provider-agnostic **`effort` + `thinking` request params** ([#6241](https://github.com/diegosouzapw/OmniRoute/issues/6241)) — a thin standardization layer over the existing mature per-provider reasoning plumbing (no provider mapper touched). `providerChatCompletionSchema` gains a canonical `effort` (reusing the shared `none/low/medium/high/xhigh` vocabulary — the UI tiers `extra`/`max` collapse onto `xhigh`) and a boolean `thinking`. A pure `normalizeReasoningRequest` (wired once in `src/sse/handlers/chat.ts`, before any reasoning field is read) folds them onto the fields the translators already consume (`reasoning_effort` / `reasoning.effort` / `thinking`), so they fan out to Anthropic / Gemini / xAI / Responses — an explicit client `reasoning_effort` / object-shaped `thinking` always wins (backward-compatible). `/models` additively exposes `supportsThinking` + `effort_tiers` so the frontend can render the toggles (UI component is a follow-up). Regression guard: `tests/unit/effort-thinking-standardization-6241.test.ts` (12). (thanks @Iammilansoni, @shabeer)
+- **feat(combo):** new **`pipeline` (sequential) combo strategy** ([#6297](https://github.com/diegosouzapw/OmniRoute/issues/6297)) — the 18th routing strategy runs targets **in order**, threading each step's output into the next step's input, with an optional per-step `prompt` (system instruction); only the final step's response is returned. Distinct from `fusion` (parallel fan-out + judge). Implemented as a self-contained `open-sse/services/pipeline.ts` (sibling to `fusion.ts`), dispatched from `combo.ts`; the step list reuses `combo.models` order and reads an optional `prompt` off each target (backward-compatible — ignored by every other strategy). Intermediate steps run non-streaming with tools stripped (complete prose to thread forward); the final step keeps the client's `stream` flag + tools. A failing/empty/unparseable intermediate step fails the whole pipeline explicitly via a sanitized error (never silently swallowed). Kilo's dup flag vs #563 was a false positive (that's model→chain selection; this is a sequential chain). Regression guard: `tests/unit/combo-pipeline-strategy.test.ts` (5). (thanks @ofekbetzalel)
+- **feat(ci):** `check:test-masking` now flags **inline-reimplemented prod conditions** ([#6348](https://github.com/diegosouzapw/OmniRoute/issues/6348)) — a new **report-only** subcheck (v2, 6A.10 family) catches the wrong-shape contract test: a test that recomputes the condition under test inline instead of importing/exercising the real function (the #6216 class, where `=== 500` → `>= 500` stayed green because the test re-implemented the branch). For each added/modified test file it warns when the file textually duplicates a ≥3-token conditional from a production file touched in the same PR **and** does not import the symbol/module owning it, via a pure, fixture-tested `findReimplementedConditions()` with an allowlist mirroring `assertReductionAllowlist`. Report-only for now (does not fail the gate) — to be promoted to blocking after a triage cycle. Regression guard: `tests/unit/check-test-masking.test.ts` (45).
+- **feat(sse):** per-connection routing override (native vs CLIProxyAPI) ([#6339](https://github.com/diegosouzapw/OmniRoute/issues/6339)) — the previously-dead `isCliproxyapiDeepModeEnabled` helper is now wired into `resolveExecutorWithProxy`: a single connection can opt itself into the CLIProxyAPI passthrough executor via `providerSpecificData.cliproxyapiMode="claude-native"`, with precedence **connection override > provider `upstream_proxy_config` mode > default**. `resolveExecutorWithProxy` now receives the resolved connection's `providerSpecificData` (threaded from `chatCore.ts`), so one connection can deep-route while the provider's other connections stay native — no DB schema change (the toggle rides in `providerSpecificData`). Also resolves the same-provider mixing ask in #6340. Regression guard: `tests/unit/chatcore-executor-proxy.test.ts` (9). (thanks @RaviTharuma)
+- **feat(dashboard):** "Add session cookie" modal now shows a prominent **"Open ‹host› →"** link to the provider's own site ([#6268](https://github.com/diegosouzapw/OmniRoute/issues/6268)) — every `-web` cookie-session provider (chatgpt-web, claude-web, gemini-web, kimi-web, lmarena, qwen-web, m365-copilot-web, …) renders a one-click external link (opening the provider's login/home page in a new tab) so operators no longer tab away to retype the URL mid-setup. The host resolves from a pure, unit-tested `resolveWebProviderHost()` (prefers `WEB_COOKIE_PROVIDERS[id].website`, falls back to the registry `baseUrl` origin); non-web providers render exactly as before. Kilo's dup flag vs #6265 (modal-too-small-on-1080p) was a false positive — distinct concern. Regression guard: `tests/unit/resolve-web-provider-host.test.ts` (5). (thanks @chirag127)
+- **feat(providers):** add **DigitalOcean** AI (serverless inference) as an OpenAI-compatible API-key provider ([#6373](https://github.com/diegosouzapw/OmniRoute/pull/6373)) — base `https://inference.do-ai.run/v1`, wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/digitalocean/`, `src/shared/constants/providers/apikey/inference-hosts.ts`). Regression guard: `tests/unit/digitalocean-provider.test.ts`. (thanks @newnol)
+- **feat(providers):** add **Huancheng Public API** (`hcnsec`) as an OpenAI-compatible regional provider ([#6410](https://github.com/diegosouzapw/OmniRoute/pull/6410)) — Xinjiang Huancheng Cybersecurity's public LLM platform (base `https://api.hcnsec.cn/v1`, free credits via daily check-ins), wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/hcnsec/`, `src/shared/constants/providers/apikey/regional.ts`). Regression guard: `tests/unit/hcnsec-provider.test.ts`. (thanks @UnrealAryan)
+- **feat(dashboard):** the web-session credential guide now shows an **"Open {host}" link** ([#6316](https://github.com/diegosouzapw/OmniRoute/pull/6316)) to the provider's sign-in site (derived from the provider `website` via `getProviderWebsiteHost`), so you can jump straight to the page where the cookie/session must be captured. Regression guard: `tests/unit/web-session-provider-link-6316.test.ts`. (thanks @jordansilly77-stack)
+- **feat(cerebras):** add the **Gemma 4 31B** model (`gemma-4-31b`) to the Cerebras registry + pricing table ([#6331](https://github.com/diegosouzapw/OmniRoute/pull/6331)). Regression guard extends `tests/unit/t28-model-catalog-updates.test.ts`. (thanks @backryun)
+- **feat(providers):** add **Yuanbao (web)** as a cookie-session provider ([#6196](https://github.com/diegosouzapw/OmniRoute/issues/6196)) — `yuanbao-web` (Tencent Yuanbao, `yuanbao.tencent.com`) with cookie-only auth (`hy_user`/`hy_token` + public agent id), SSE→OpenAI translation incl. `reasoning_content`, exposing DeepSeek V3/R1 + Hunyuan / Hunyuan-T1. Regression guard: `tests/unit/providers-yuanbao-web.test.ts`. `together-web` was **deferred** (no verifiable web-session endpoint — needs a captured request) and `huggingchat-web` **dropped** (the existing `huggingchat` already is a web-cookie provider). (thanks @chirag127)
+- **feat(providers):** route the built-in **agentrouter** through the dynamic Claude-Code wire image ([#6056](https://github.com/diegosouzapw/OmniRoute/issues/6056)) — a small static allow-set (`CC_WIRE_IMAGE_BUILTINS` in `open-sse/services/ccWireImageBuiltins.ts`), consulted by `isClaudeCodeCompatible` / `isClaudeCodeCompatibleProvider` / `applyFingerprint`, makes agentrouter adopt the CC wire-image headers + fingerprint **while guarding the CC baseUrl/auth branches** so it keeps its own registry `baseUrl` and `x-api-key` auth. Regression guard: `tests/unit/agentrouter-cc-wire-image.test.ts` (asserts the wire image is applied AND agentrouter's baseUrl/auth are preserved). Live WAF-acceptance against agentrouter.org is a VPS validation follow-up (Hard Rule #18).
+- **feat(providers):** **bulk-add API keys for Cloudflare Workers AI** ([#6174](https://github.com/diegosouzapw/OmniRoute/issues/6174)) — `cloudflare-ai` is removed from the bulk-add exclusion list and the bulk parser gains a 3-field `name|accountId|apiKey` mode; the bulk route now builds a **per-entry** `providerSpecificData` so each key carries its own `accountId` (fixing the previous shared-object reuse), and both the create + key-validation paths receive it. Regression guard: `tests/unit/bulk-api-key-parser-cloudflare.test.ts`. (thanks @muflifadla38)
+- **feat(dashboard):** routing/settings UX clarity ([#6147](https://github.com/diegosouzapw/OmniRoute/issues/6147)) — (1) weighted combos show the **effective routing share %** next to each weight when weights don't sum to 100 (`WeightTotalBar.tsx`); (2) the status widget's user-facing **"Cloud Sync" label is renamed** to "Remote Settings Sync" (`CloudSyncStatus.tsx`; internal ids/state untouched); (3) built-in providers gain an **opt-in advanced base-URL override** (`isBaseUrlOverrideEligibleProvider`, hidden behind an "Advanced" toggle, reusing the existing `providerSpecificData.baseUrl` persistence — not globally widened). Regression guard: `tests/unit/routing-settings-ux-6147.test.ts`.
+- **feat(combo):** add an option to **disable session stickiness**, per-combo or globally — round-robin / random combos can rotate to a different connection on every request instead of pinning a whole conversation to one connection by its first-message hash. Resolution precedence per-combo `config.disableSessionStickiness` → global `settings.disableSessionStickiness` → default `false` (preserves the #3825 prompt-cache/504 fix); gates **both** stickiness call sites in `open-sse/services/combo.ts`. Exposed as a global toggle (Combo Defaults) and a per-combo Inherit/on/off control. ([#6168](https://github.com/diegosouzapw/OmniRoute/issues/6168)) Regression guard: `tests/unit/combo-disable-session-stickiness.test.ts`. (thanks @RCrushMe)
+- **feat(docker):** add the `OMNIROUTE_NO_SUDO` env flag for root-less / user-namespaced deployments — the MITM cert-trust command path (`resolveSudoSpawn` in `src/mitm/systemCommands.ts`) now strips the leading `sudo` when the flag is truthy, in addition to the existing root / sudo-missing cases, so the Proxy Agent runs without `sudo` (the operator trusts the CA manually, e.g. via `NODE_EXTRA_CA_CERTS`). Argv-array `spawn` preserved — no shell interpolation (Hard Rule #13). ([#6122](https://github.com/diegosouzapw/OmniRoute/issues/6122)) Regression guard: `tests/unit/mitm-systemCommands-no-sudo.test.ts`. (thanks @powellnorma)
+- **feat(providers):** add **Requesty** as an OpenAI-compatible gateway provider (BYOK, base `https://router.requesty.ai/v1`, ~200 free requests/day) — wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/requesty/`, `src/shared/constants/providers/apikey/gateways.ts`). ([#6120](https://github.com/diegosouzapw/OmniRoute/issues/6120)) Regression guard: `tests/unit/requesty-provider.test.ts`. (thanks @chirag127)
+- **feat(dashboard):** add **configured-only / available-only filters** to the Free Provider Rankings page ([#6150](https://github.com/diegosouzapw/OmniRoute/issues/6150)) — hide providers you haven't configured, or whose connections are all rate-limited / out of quota, via server-side query params (`?configuredOnly` / `?availableOnly` on `GET /api/free-provider-rankings`) backed by a testable lib helper reusing the in-process connection state (no Redis). Both filters default off, so the default view is unchanged; this supersedes the earlier client-side "Configured Only" toggle (#6245) with an available-only dimension and unit-tested logic. Regression guard: `tests/unit/freeProviderRankings-filters.test.ts`.
+
+### 🔧 Bug Fixes
+
+- **fix(dashboard):** adding a second API key connection for the same provider no longer silently overwrites the first — the Add-API-key modal now derives a unique default connection name (`main`, then `main-2`, `main-3`, …) so the backend name-based upsert can't collide (#6499 — thanks @dilneiss).
+- **fix(compression):** the session-dedup engine now also deduplicates a large multi-line block repeated **within a single message** (intra-message dedup), not just across turns; the compression-preview API surfaces a `fallbackReason`, and the fusion panel reports how many models were rate-limited vs failed on total-panel failure (#6501 — thanks @chirag127).
+- **fix(compression):** a stacked-pipeline step naming an unregistered engine now surfaces a `validationErrors` entry instead of silently no-op'ing, so misconfigured pipelines are visible in the preview API (#6506 — thanks @chirag127).
+- **feat(usage):** add a **Codex reset-credit redemption** flow to the Provider Limits UI ([#6361](https://github.com/diegosouzapw/OmniRoute/pull/6361)) — a `useCodexResetCreditRedemption` hook + `/api/usage/codex-reset-credit` route + `codexResetCredits` lib let you redeem banked Codex reset credits from the quota card. Regression guard: `tests/unit/codex-reset-credits.test.ts`. (thanks @JxnLexn)
+- **feat(glm):** add **team-plan quota settings** for `glm-cn` connections ([#6351](https://github.com/diegosouzapw/OmniRoute/pull/6351)) — a dedicated `GlmTeamQuotaFields` form section (team quota id / limits) threaded through the Add/Edit connection modals, persisted via `providerSpecificData`, with the GLM usage service reading the team quota. Regression guards: `tests/unit/glm-team-quota.test.ts`, `provider-specific-data-schema.test.ts`. (thanks @hao3039032)
+- **feat(providers):** add **TinyFish** web-fetch/search support ([#6349](https://github.com/diegosouzapw/OmniRoute/pull/6349)) — a `tinyfish-fetch` executor + `/v1/web/fetch` route + MCP web-fetch tool, registered as a specialty-media provider with request-validation and a search-provider catalog entry. Regression guards: `tests/unit/executor-tinyfish-fetch.test.ts`, `web-fetch-handler.test.ts`, `mcp-web-fetch-tool.test.ts`, `provider-validation-tinyfish.test.ts`. (thanks @dtybnrj)
+- **fix(cli):** `omniroute launch-codex` now spawns `codex.cmd` through a shell on Windows (the npm `.cmd` shim is unresolvable by bare `spawn` → ENOENT), mirroring the qodercli Windows fix (#6263) ([#6312](https://github.com/diegosouzapw/OmniRoute/pull/6312)). Regression guard: `tests/unit/launch-codex-windows-spawn-6312.test.ts`. (thanks @swingtempo)
+- **fix(codex):** isolate the **Spark** quota from the shared Codex quota and stabilize the quota UI ordering / hydration so per-scope limits render consistently ([#6336](https://github.com/diegosouzapw/OmniRoute/pull/6336)). Regression guards: `tests/unit/codex-quota-selection-hydration.test.ts`, `provider-limits-ui.test.ts` + 3 more. (thanks @xz-dev)
+- **feat(api):** add a `hidePaidModels` setting that filters paid-only models out of the `/v1/models` catalog. Regression guard: `tests/unit/models-catalog-hide-paid.test.ts`. (thanks @chirag127)
+- **fix(api-manager):** the fallback model picker now preserves combos instead of dropping them when a primary model is unavailable ([#6443](https://github.com/diegosouzapw/OmniRoute/pull/6443)). Regression guard: `tests/unit/api-manager-page-static.test.ts`. (thanks @jmengit)
+- **fix(providers):** recoverable Antigravity / Cloud-Code (Gemini Code Assist) `403` responses ([#6452](https://github.com/diegosouzapw/OmniRoute/pull/6452)) are now classified as a retryable project-config error instead of a terminal account ban, so a fixable project/API-disabled 403 no longer forces a ~1-year cooldown / full OAuth reconnect. Regression guard: `tests/unit/errorclassifier-antigravity-403.test.ts`. (thanks @developerjillur)
+- **fix(mitm):** `sanitizeHeaders` now redacts `Set-Cookie` response headers so upstream session cookies never leak into logs / diagnostics ([#6451](https://github.com/diegosouzapw/OmniRoute/pull/6451)). Regression guard: `tests/unit/mitm-sanitize-headers.test.ts`. (thanks @developerjillur)
+- **fix(api):** `/api/compression/preview` now accepts `mode: "caveman"` and correctly handles stacked / zero-compression previews ([#6425](https://github.com/diegosouzapw/OmniRoute/issues/6425)). Regression guard: `tests/unit/api/compression-preview-caveman-and-stacked-6425.test.ts`. (thanks @chirag127)
+- **feat(providers):** add **Zed** hosted LLM aggregator as a native-app provider ([#6118](https://github.com/diegosouzapw/OmniRoute/pull/6118)) — OAuth sign-in via the Zed hosted flow, registered through the shared provider registry + executor. Regression guards: `tests/unit/zed-oauth-provider.test.ts`, `zed-import-utils.test.ts`, `zed-docker-detect.test.ts`, `mitm-handler-zed.test.ts`. VPS-validated via live operator login (Hard Rule #18).
+- **fix(oauth):** the Kiro SSO-cache auto-import now **preserves the IDC region** — cross-region Amazon Q / Kiro profiles imported from the SSO cache are no longer collapsed to the default region ([#6113](https://github.com/diegosouzapw/OmniRoute/pull/6113)). Regression guard: `tests/unit/kiro-auto-import-idc-2059.test.ts`. VPS-validated via live operator login (Hard Rule #18).
+- **fix(dashboard):** passthrough model aliases no longer collide when two namespaced model ids share a last segment (port from 9router#1850, [#6431](https://github.com/diegosouzapw/OmniRoute/pull/6431)). `enx/gpt-5.5` and `enx/codebuddy/gpt-5.5` both auto-generated the alias `gpt-5.5`, so the second model could never be added (the UI just alerted "alias already exists"). Aliases are now disambiguated deterministically — bare last segment when free, then parent-qualified (`codebuddy-gpt-5.5`), then a numeric suffix — while re-adding the exact same model id is still blocked. Regression guard: `tests/unit/passthrough-alias-1850.test.ts`. (thanks @arpicato)
+- **fix(translator):** preserve a Gemini `functionResponse` co-located with other parts (another `functionCall`, or trailing `text`) in the same content when translating **Gemini → OpenAI** ([#6376](https://github.com/diegosouzapw/OmniRoute/pull/6376)). `convertGeminiContent()` early-returned the tool message on the first `functionResponse` part, dropping any co-located parts; such contents are now pre-split (one tool message per `functionResponse`, emitted first, plus one message for the remaining parts). Regression guard: `tests/unit/gemini-to-openai-function-response.test.ts`. (thanks @warelik)
+- **fix(headroom):** detect a python interpreter managed by **mise / pyenv / asdf / conda** (port from 9router#2353, [#6382](https://github.com/diegosouzapw/OmniRoute/pull/6382)). Headroom's python probe (`src/lib/headroom/detect.ts`) searched a hardcoded `PATH`, but version managers expose their interpreters via shim dirs that only join `PATH` through interactive-shell activation — which the non-interactive server never runs, so a managed python (≥3.10) was invisible and Headroom reported it missing. The search path now prepends the well-known shim/bin dirs (`~/.local/share/mise/shims`, `~/.pyenv/shims`, `~/.asdf/shims`, `$CONDA_PREFIX/bin`, `~/.local/bin`, respecting `MISE_DATA_DIR`/`PYENV_ROOT`/`ASDF_DATA_DIR` when set), and a new `HEADROOM_PYTHON` env override lets operators point straight at their interpreter (mirroring `HEADROOM_URL`). Still shell-free (`execFileSync`). Regression guard: `tests/unit/headroom-detect.test.ts` (5). (thanks @loopyd)
+- **fix(executors):** strip the OpenAI-Codex/Claude-CLI `client_metadata` passthrough field for **NVIDIA** requests (port from 9router#1887, [#6411](https://github.com/diegosouzapw/OmniRoute/pull/6411)). NVIDIA's OpenAI-compatible wrapper rejects it with `400 Unsupported parameter`, the same class already handled for `cerebras`/`mistral`; `nvidia` (executor `default`) was missing from the strip allowlist so Codex/Claude-Code passthrough requests 400'd. Regression guard: `tests/unit/executor-default-strip-client-metadata.test.ts` (+nvidia case). (thanks @phidinhmanh)
+- **fix(translator):** strip the Claude-style `thinking` field for **NVIDIA `z-ai/glm-5.2`** (port from 9router#2023, [#6413](https://github.com/diegosouzapw/OmniRoute/pull/6413)). NVIDIA's OpenAI-compatible wrapper 400s on `thinking` (a Claude-format client routed here leaves a `thinking:{type:"adaptive"}`); the existing strip rule only dropped `reasoning`. Same class already handled for `minimax-m2.7`. Regression guard: `tests/unit/nvidia-minimax-thinking-strip.test.ts` (+glm-5.2 case). (thanks @phidinhmanh)
+- **fix(translator):** suppress the streamed `</think>` close marker for the **Antigravity IDE** client (port from 9router#1061, [#6415](https://github.com/diegosouzapw/OmniRoute/pull/6415)). On thinking-only turns Antigravity rendered a bare `</think>` as the sole visible content, tripping its loop-detection and wasting requests. Antigravity's UA (`vscode/<v> (Antigravity/<v>)`) is added to the marker-suppress allowlist (alongside OpenCode); Claude Code / Cursor still get the marker, and `x-omniroute-thinking-marker: on` force-restores it. Regression guard: `tests/unit/think-close-marker-suppress-5245.test.ts`. (thanks @abdofallah)
+- **fix(executors):** strip nested `reasoning_content` from messages for **Mistral** (port from 9router#1649, [#6417](https://github.com/diegosouzapw/OmniRoute/pull/6417)). Mistral's API returns `422 extra_forbidden` when an assistant message carries `reasoning_content` (replayed thinking from a prior turn, e.g. via the Codex `/responses` path); the generic top-level 400 field-downgrade retry never covered the nested per-message field. `DefaultExecutor` now strips it for provider `mistral` only, so DeepSeek (which requires replayed `reasoning_content`) is unaffected. Regression guard: `tests/unit/mistral-strip-reasoning-content-1649.test.ts`. (thanks @xxy9468615)
+- **fix(executors):** strip the `client_metadata` passthrough field on the **OpenCode** path (port from 9router#1442, [#6418](https://github.com/diegosouzapw/OmniRoute/pull/6418)). OpenCode upstreams (e.g. `kimi-k2.6` via opencode-go) reject it with `400 "Extra inputs are not permitted, field: 'client_metadata'"`; the DefaultExecutor strip only covered cerebras/mistral and `OpencodeExecutor` extends `BaseExecutor` directly, so nothing removed it there. Regression guard: `tests/unit/opencode-strip-client-metadata-1442.test.ts`. (thanks @yanpaing007)
+- **fix(executors):** inject the `reasoning_content` echo for the native **Moonshot Kimi** provider (port from 9router#1480, [#6419](https://github.com/diegosouzapw/OmniRoute/pull/6419)). Kimi (executor `default`) is a thinking-mode upstream that 400s with "reasoning_content must be passed back" when a prior assistant turn lacks it; the placeholder injection was only wired into the OpenCode meta-provider, so direct multi-turn Kimi conversations failed. Scoped to `kimi` (gateway-served models matching the thinking-model name pattern are unaffected). Regression guard: `tests/unit/kimi-native-reasoning-injected-1480.test.ts`. (thanks @2220258345)
+- **fix(executors):** recover from a strict gateway's `context_management: Extra inputs are not permitted` 400 (port from 9router#1468, [#6420](https://github.com/diegosouzapw/OmniRoute/pull/6420)). **Claude Code** always sends a top-level `context_management` field; strict anthropic-compatible gateways reject it. The dedicated context-editing 400-fallback only fired when OmniRoute's own `contextEditing` feature was enabled (default off), so a client-sent field passed through untouched and 400'd. `context_management` is now in the generic reactive field-strip list, so it's stripped-and-retried once regardless of the feature flag (with correct request re-signing for claude-compatible relays). Regression guard: `tests/unit/provider-field-strips.test.ts`. (thanks @ohahe52-dot)
+- **fix(network):** enable **RFC 8305 Happy Eyeballs** (`autoSelectFamily`) on the direct-egress undici dispatcher (port from 9router#1237, [#6423](https://github.com/diegosouzapw/OmniRoute/pull/6423)). When DNS returns both IPv6 (AAAA) and IPv4 (A) and the IPv6 route is broken (e.g. a NAT64 `64:ff9b::` prefix without routing), undici tried IPv6 first and hung until `ETIMEDOUT` (then a 502 + account lockout), even though `curl` reached the same host. The direct dispatcher now races both families and uses whichever connects first. Proxy paths pin family via `proxyTls` and are unaffected. Regression guard: `tests/unit/direct-dispatcher-pipelining-4580.test.ts`. (thanks @adentdk)
+- **fix(combo):** round-robin now advances the rotation pointer past the model that **actually served**, not the eagerly-scheduled one (port from 9router#948, [#6428](https://github.com/diegosouzapw/OmniRoute/pull/6428)). With `stickyLimit: 1` (true round-robin), when the scheduled model failed and a _different_ model served via fallback, the counter had already advanced +1 from the scheduled index — so the next request reused the fallback-served model, degrading round-robin into hot-spotting on whichever model was healthy. The pointer now advances to the served index + 1 (mirroring the sticky-limit>1 path). Session-stickiness (#3825) and distribution are preserved. Regression guard: `tests/unit/combo-rr-fallback-advance-948.test.ts`. (thanks @binsarjr)
+- **fix(sse):** a non-string `model` field is now rejected with a `400` before the resolver, instead of crashing downstream `.toLowerCase()`/`.split()` calls into an empty-body `500` that escapes the error sanitizer ([#6407](https://github.com/diegosouzapw/OmniRoute/issues/6407)). Regression guard: `tests/unit/chat-non-string-model-6407.test.ts`. (thanks @chirag127)
+- **fix(api):** unknown `/api/*` routes now return a JSON `404` (instead of the dashboard HTML shell) and scalar chat params (`model`/`temperature`/etc.) are validated **before** the provider lookup so malformed requests fail fast with a clear `400` ([#6424](https://github.com/diegosouzapw/OmniRoute/issues/6424), [#6412](https://github.com/diegosouzapw/OmniRoute/issues/6412)). Regression guards: `tests/unit/api/api-catchall-json-404.test.ts`, `tests/unit/chat-early-schema-validation-6412.test.ts`. (thanks @chirag127)
+- **fix(api):** `/v1/chat/completions` now rejects a non-JSON `Content-Type` with a `400` before parsing the body ([#6414](https://github.com/diegosouzapw/OmniRoute/issues/6414)). Regression guard: `tests/unit/v1-chat-completions-content-type-6414.test.ts`. (thanks @chirag127)
+- **fix(api):** the `X-OmniRoute-Compression` response header is now echoed on `/v1/chat/completions` and `/v1/completions` ([#6422](https://github.com/diegosouzapw/OmniRoute/issues/6422)). Regression guard: `tests/unit/compression-header-echo-6422.test.ts`. (thanks @chirag127)
+- **fix(api):** concurrent `GET /v1/models` requests are coalesced into a single catalog build ([#6408](https://github.com/diegosouzapw/OmniRoute/issues/6408)). Regression guard: `tests/unit/v1-models-concurrent-6408.test.ts`. (thanks @chirag127)
+- **fix(api):** `/v1/completions` now echoes the requested `body.model` in its JSON + streamed responses ([#6429](https://github.com/diegosouzapw/OmniRoute/pull/6429)). Regression guard: `tests/unit/completions-body-model-echo.test.ts`. (thanks @chirag127)
+- **fix(api):** env-var master keys now see the full `/v1/models` catalog ([#6406](https://github.com/diegosouzapw/OmniRoute/issues/6406)). Regression guard: `tests/unit/models-catalog-envkey-6406.test.ts`. (thanks @chirag127)
+- **fix(api):** non-streaming `/v1/completions` responses now echo `body.model` aligned with the `X-OmniRoute-Model` header ([#6426](https://github.com/diegosouzapw/OmniRoute/issues/6426)). Regression guard: `tests/unit/v1-completions-model-header-match-6426.test.ts`. (thanks @chirag127)
+- **fix(api):** unknown `/v1/*` routes now return a JSON `404 not_found` instead of the Next.js dashboard HTML shell ([#6405](https://github.com/diegosouzapw/OmniRoute/issues/6405)). Regression guard: `tests/unit/api/v1-catchall-json-404.test.ts`. (thanks @chirag127)
+- **fix(api):** the per-connection provider models route now degrades to the shipped catalog when a provider's `/models` endpoint answers with a redirect ([#6267](https://github.com/diegosouzapw/OmniRoute/issues/6267)) — a `qwen-web` import failed with a raw `Redirect blocked … (307)` 503. `safeOutboundFetch` throws `REDIRECT_BLOCKED` on the 307, `getSafeOutboundFetchErrorStatus` maps it to 503, and `buildDiscoveryErrorFallbackResponse` treated every 503 as a hard error — so the non-empty `getModelsByProviderId("qwen-web")` catalog was never surfaced. A models-endpoint redirect is not a fixable-config error (unlike `URL_GUARD_BLOCKED`/`INVALID_URL`, which stay hard errors), so it now falls back to the local/cached catalog before the 503 short-circuit. General fix — covers any config-driven provider that 307s. Regression guard: `tests/unit/provider-models-qwen-web-redirect-6267.test.ts`. (thanks @chirag127)
+- **fix(api):** the per-connection provider models route (MCP `list_models_catalog` + the dashboard import view) now merges USER-ADDED custom models into its response ([#6247](https://github.com/diegosouzapw/OmniRoute/issues/6247)) — custom models live in the `key_value` namespace `customModels`, which the live REST `/api/v1/models` already merges, but `src/app/api/providers/[id]/models/route.ts` never read `getCustomModels`, so custom models were dropped on both the discovery-success and local_catalog paths. They are now folded into the returned model list (deduped by id, stamped `owned_by: provider`), fixing MCP + the dashboard import view in one place. Regression guard: `tests/unit/provider-models-custom-merge-6247.test.ts`. (thanks @RCrushMe)
+
+- **fix(providers):** GitLab Duo tool-calling follow-up turns no longer fail upstream with `422 {"detail":"Validation error"}` (tokens 0/0, rejected pre-inference). The #6234 tool-result-feedback fix serialized the **entire** multi-turn conversation into GitLab's single-file `code_suggestions` (`small_file`) generation endpoint — folded history that turn-N sent as an oversized `current_file.content_above_cursor` **and** duplicated verbatim into `user_instruction`, tripping the AI-Gateway's `small_file` validation guard. The executor now **bounds** that prompt: it keeps system + latest user message + the most-recent tool round (dropping older turns), caps oversized tool results, and stops duplicating the full prompt into `user_instruction` (which now carries only the short latest user message) — while still feeding the most-recent tool result back so the agent continues (`open-sse/executors/gitlab.ts`, [#6220](https://github.com/diegosouzapw/OmniRoute/issues/6220)). The unit test covers the bounding logic; the upstream 422→200 clearing is VPS-only (Hard Rule #18). Regression guard: `tests/unit/gitlab-tool-exchange-bounded-6220.test.ts`.
+
+- **fix(i18n):** the provider-detail (`/dashboard/providers/[id]`) connection-status filter labels no longer render as `__MISSING__:All` / `__MISSING__:Active` / `__MISSING__:Error` / `__MISSING__:Banned` / `__MISSING__:CreditsExhausted` in non-English locales (notably pt-BR) ([#6290](https://github.com/diegosouzapw/OmniRoute/issues/6290)). Root cause was **not** the namespace mismatch the issue guessed — the `providers.filter*` keys resolve correctly in `en.json`; the debt lived in the locale mirrors (`src/i18n/messages/*.json`), where these five keys carried the `__MISSING__:` sync sentinel in ~15 locales and were absent entirely in ~26 others, so next-intl found the key and echoed the sentinel verbatim. All 40 non-English/-Chinese mirrors now ship real translations for the five `providers.filter*` labels. Regression guard: `tests/unit/i18n-provider-filter-keys-6290.test.ts`. (thanks @diegosouzapw)
+
+- **fix(providers):** the `copilot-m365-web` streaming executor now emits `debug`-level WebSocket diagnostics ([#6210](https://github.com/diegosouzapw/OmniRoute/issues/6210)) — the outbound WS URL (with the `access_token` **redacted** via `redactWsUrl()`), handshake success/failure, and each received SignalR frame's `type`/`target`. Previously the streaming path logged nothing, so an empty `content:null` response (the M365 Education / Starter tier symptom fixed in #6234) was undiagnosable even at `APP_LOG_LEVEL=debug`. The change is debug-level and side-effect-free — it does not alter streaming behavior or the frame parser, and the token never reaches the logs. Regression guard: `tests/unit/copilot-m365-web-logging-6210.test.ts` (thanks @qpeyba)
+
+- **fix(resilience):** a round-robin combo no longer returns `503 all upstream accounts are unavailable` when a compatibility-rejected target is actually healthy ([#6238](https://github.com/diegosouzapw/OmniRoute/issues/6238)). `filterTargetsByRequestCompatibility` drops request-incompatible targets (tool/vision/structured-output unsupported, or below the required context window) **before** any availability check runs, and its `compatible.length === 0` safety net only fired when _all_ targets were filtered — not when the kept targets later all turned out runtime-unavailable (circuit-open / cooldown / no credentials). So a combo could 503 while a compat-rejected-but-healthy provider sat unused. `handleRoundRobinCombo` now keeps the compat-rejected set and, when every compat-kept target was skipped without a single real attempt, probes those rejected targets as a **last-resort fallback tier** (via the new pure `open-sse/services/combo/comboCompatFallback.ts`) before crystallizing the 503. Regression guard: `tests/unit/combo-roundrobin-compat-fallback-6238.test.ts`. (thanks @ThongAccount)
+
+- **fix(startup):** best-effort self-heal for a corrupted Turbopack dev cache on Windows ([#6289](https://github.com/diegosouzapw/OmniRoute/issues/6289)). On Windows, `pnpm dev` can fail at startup when Turbopack `mmap`s a persistent-cache SST file and the OS refuses the mapping (`os error 1455` — "paging file too small"), which Turbopack surfaces as a misleading `Module not found: Can't resolve '@/shared/utils/machine'`. This is a **known upstream Turbopack cache-corruption bug — not our code**. The dev launcher (`scripts/dev/run-next.mjs`) now wraps `nextApp.prepare()` and, when it rejects with that signature (`isTurbopackCacheCorruption` in the new `scripts/dev/turbopackCacheHeal.mjs`), purges `.build/next/**/cache/turbopack` and retries **once** with a clear log. **Caveat — best-effort only:** the corruption often surfaces as a runtime overlay rather than a `prepare()` rejection, so this cannot always intercept it; the reliable remedy remains manually deleting the Turbopack cache dir. Regression guard: `tests/unit/turbopack-cache-heal-6289.test.ts`. (thanks @chirag127)
+
+- **fix(providers):** qodercli PAT auth no longer fails with `spawn qodercli ENOENT` on Windows ([#6263](https://github.com/diegosouzapw/OmniRoute/issues/6263)) — `spawnQoderCli` spawned the bare `qodercli` name with `shell:false` and an unenriched env, so the npm `.cmd` wrapper under `%APPDATA%\npm` (a user-PATH directory) was never resolved. It now resolves the absolute `.cmd`/`.exe` path through the existing `getCliRuntimeStatus("qoder")` resolver in `src/shared/services/cliRuntime.ts` (memoized), spawns with `shell` when the target is a `.cmd`/`.bat`, and uses the cliRuntime-enriched env (PATH + PATHEXT + APPDATA); the ENOENT error now lists the searched paths plus the `CLI_QODER_BIN` override. End-to-end spawn on a real Windows host is host-only (Hard Rule #18); the path-resolution logic is unit-tested. Regression guard: `tests/unit/qodercli-windows-resolve-6263.test.ts`. (thanks @chirag127)
+
+- **fix(sse):** the reasoning-token buffer no longer inflates **probe-sized `max_tokens`** ([#6274](https://github.com/diegosouzapw/OmniRoute/issues/6274)) — Claude Code's `/model` capability check sends `max_tokens: 1`, but for a thinking-capable model with a large output cap (e.g. `glm-5.2`) the #3587 headroom heuristic (`max(current + 1000, ceil(current * 1.5))`) rewrote it to `1001` and forwarded that upstream, wasting tokens on a request that was never a genuine reasoning budget. `resolveReasoningBufferedMaxTokens()` (`open-sse/services/reasoningTokenBuffer.ts`) now short-circuits and returns the caller's value verbatim when it is below the new `REASONING_BUFFER_MIN_TRIGGER` (256) threshold — a tiny explicit limit is a probe, not a reasoning request. Real budgets still receive the #3587 headroom unchanged, and the guard runs after the existing capability checks so unknown / non-reasoning models keep returning `null`. Regression guard: `tests/unit/reasoning-token-buffer-6274.test.ts`. (thanks @brightfiscalband)
+
+- **fix(cli):** `omniroute reset-password` now works as a real subcommand, and password resets over piped (non-TTY) stdin actually apply ([#6261](https://github.com/diegosouzapw/OmniRoute/issues/6261), [#6258](https://github.com/diegosouzapw/OmniRoute/issues/6258)). Two coupled defects: (1) **#6261** — `bin/omniroute.mjs` routed everything through Commander with only two pre-Commander bypasses (`--mcp`, `reset-encrypted-columns`), so `omniroute reset-password` was rejected as an unknown command; only the separate `omniroute-reset-password` bin worked, while the docs falsely advertised the subcommand (incl. a bogus "legacy alias still works"). A pre-Commander bypass mirroring `reset-encrypted-columns` now dynamically imports `bin/reset-password.mjs` (which self-executes) before Commander parses; the three doc lines were corrected. (2) **#6258** — `bin/reset-password.mjs` issued two sequential `rl.question` prompts; under piped stdin the second read never settled at EOF, so `main()` never reached `resetManagementPassword` and the reset was a silent no-op (both prompts printed, no success, password unchanged). The CLI now detects non-TTY stdin and reads it once (first line = password, second line = confirm if present, else reused), adds a `--password-stdin` flag (entire stdin is the password, no confirmation), and exits `0` explicitly so the success line always flushes; interactive TTY behavior is unchanged. Regression guard: `tests/unit/reset-password-cli-6261-6258.test.ts` (3). (thanks @chirag127)
+
+- **fix(db):** the mass-migration **safety abort** now tells the operator how to bypass it and stops flooding the log ([#6260](https://github.com/diegosouzapw/OmniRoute/issues/6260)) — after restoring a backup that wiped the migration tracking table, `runMigrations()` threw the abort on every downstream `ensureDbInitialized()`, re-logging the full banner 11+ times, and the message never mentioned the existing `OMNIROUTE_MAX_PENDING_MIGRATIONS` escape hatch. The abort text now appends a bypass hint (set `OMNIROUTE_MAX_PENDING_MIGRATIONS=0` in `server.env` / `DATA_DIR/.env`), and a new `MigrationSafetyAbortError` is memoized so repeated calls in the same process throw the same instance and emit a single concise line instead of the full cascade. Regression guard: `tests/unit/migration-safety-abort-6260.test.ts`. (thanks @chirag127)
+
+- **fix(auth):** importing a **distinct** Codex/ChatGPT OAuth `auth.json` is no longer falsely rejected as "already exists" when it belongs to a different user in the same workspace ([#6301](https://github.com/diegosouzapw/OmniRoute/issues/6301)). `findExistingCodexConnection` (in `src/lib/oauth/utils/codexAuthImport.ts`) deduped **only** on `providerSpecificData.workspaceId === accountId`, where `accountId` is the shared `chatgpt_account_id`/`tokens.account_id` — so two members of the same ChatGPT Team collapsed onto a single connection (409 `duplicate_account`). The id_token's `https://api.openai.com/auth` claim carries a per-user `chatgpt_user_id` alongside the workspace id (the device-flow path already persisted it as `chatgptUserId`, but the import path did not). Now `parseAndValidateCodexAuth` extracts `userId` (`chatgpt_user_id` → `user_id` → JWT `sub`) into `ParsedCodexAuth`, the create/update paths persist `chatgptUserId` in `providerSpecificData` (mirroring `codex.ts`), and dedup keys on `workspaceId` **AND** `chatgptUserId` — with a backward-compat fallback to legacy accountId-only matching when no stored connection for that workspace records a `chatgptUserId`, so genuinely-same accounts still dedup. Regression guard: `tests/unit/codex-auth-import-userid-dedup-6301.test.ts` (4). (thanks @anungma)
+
+- **fix(providers):** importing models for the **venice-web** provider no longer fails with a red "Provider venice-web does not support models listing" ([#6269](https://github.com/diegosouzapw/OmniRoute/issues/6269)). `venice-web` is a web-cookie provider with an executor but no upstream `/v1/models` endpoint and no registry `models`, so the models route fell through to the tail `400`. Mirroring the `jules`/`linkup-search`/`ollama-search` fix (#5569), it now ships a static local catalog entry in `src/lib/providers/staticModels.ts` — seeding the current Venice lineup (`venice-uncensored`, `llama-3.3-70b`, `qwen3-235b`, `qwen3-4b`, `deepseek-r1-671b`; Venice rotates its catalog, see docs.venice.ai/models/overview) — so the route returns `200` with `source:"local_catalog"`, `intentional:true`. Regression guard: `tests/unit/static-models-venice-web-6269.test.ts`. (thanks @chirag127)
+
+- **fix(api):** the specialty model catalogs (`/v1/embeddings`, `/v1/images`, `/v1/music`, `/v1/videos` model lists) are now derived from the **unified catalog filtered by a predicate** (`getSpecialtyModelsResponse`) instead of ad-hoc per-route logic, so they consistently respect active-credential visibility and stay in sync with the main catalog ([#6303](https://github.com/diegosouzapw/OmniRoute/pull/6303)). Regression guard: `tests/unit/specialty-model-catalog-routes.test.ts`. (thanks @makcimbx)
+- **fix(api):** the agent-bridge server route now resolves the MITM manager via a **dynamic `import("@/mitm/manager.runtime")`** so Turbopack does not statically pull the stub (or over-bundle the manager), and the agent-skills generator anchors its output base path with `path.join(process.cwd(), …)` so Turbopack's static analyzer stops tracing the whole project root ([#6329](https://github.com/diegosouzapw/OmniRoute/issues/6329), [#6366](https://github.com/diegosouzapw/OmniRoute/pull/6366)). Regression guard: `tests/unit/agent-bridge-server-route-dynamic-import.test.ts`. (thanks @Iammilansoni)
+- **fix(api):** internal probes (combo-test, cloud-sync verify) now pick a **management-scoped / allow-all API key** instead of naively grabbing `getApiKeys()[0]` — a restricted `self:usage` first row made the probe fail with "Model X is not allowed for this API key" even when the combo path was healthy (`pickApiKeyForInternalUse` in `src/lib/db/apiKeys.ts`). The API-manager model editor also falls back to `/api/models?all=true` when `/v1/models` is catalog-protected ([#6372](https://github.com/diegosouzapw/OmniRoute/pull/6372)). Regression guard: `tests/unit/pick-internal-api-key-6372.test.ts`. (thanks @jmengit)
+- **fix(live-ws):** the Live Dashboard WebSocket server now **rejects on bind failure** (e.g. `EADDRINUSE` when the API bridge already holds the port) instead of letting the error surface as an unhandled `error` event that crash-loops the process — the `error` listener is attached to `wss` (not `server`) and releases the EventBus subscription on a failed start ([#6324](https://github.com/diegosouzapw/OmniRoute/issues/6324)). Regression guard: `tests/unit/live-ws-eaddrinuse-6324.test.ts`. (thanks @vinayakkulkarni)
+- **fix(dashboard):** the Home provider-topology widget now trusts the live provider-metrics snapshot — it uses `topology.errorProvider` and live `activeRequests` directly instead of re-deriving state from a stale `lastErrorAt` or applying a frontend timeout filter, so the topology reflects real-time provider health ([#6322](https://github.com/diegosouzapw/OmniRoute/pull/6322)). Regression guard: `tests/unit/home-provider-topology-live-state.test.ts`. (thanks @xz-dev)
+- **fix(sse):** strip zero-width markers from streamed **tool-call arguments** — a follow-up to [#5857](https://github.com/diegosouzapw/OmniRoute/pull/5857). That PR removed injected zero-width joiners (U+200D) from streamed assistant text/reasoning but deliberately left tool-call argument JSON byte-exact. The request-side obfuscation (`open-sse/services/claudeCodeObfuscation.ts`) injects ZWJ into agent words — including the temp path inside the Bash tool description — and Claude models copy that verbatim into generated commands, which are delivered as tool-call arguments rather than assistant text. As a result the ZWJ survived and corrupted code blocks (e.g. a temp path rendered with an invisible joiner). Now `open-sse/handlers/responseSanitizer.ts` strips zero-width code points from tool-call argument strings at every emit site (OpenAI non-stream/stream chat `tool_calls` + legacy `function_call`, native Responses `function_call` items, the OpenAI→Responses conversion, and the native Responses streaming `response.function_call_arguments.delta/.done` events). Only zero-width code points are removed; JSON structure and all other bytes stay identical (no parse/restringify), so normal arguments remain byte-exact. Regression guard: 6 new cases in `tests/unit/response-sanitizer.test.ts` (suite 50/50).
+
+- **fix(nodejs):** the default app log path now resolves under `DATA_DIR` (`~/.omniroute/logs/application/app.log`) instead of `process.cwd()` ([#6197](https://github.com/diegosouzapw/OmniRoute/issues/6197)) — the globally-installed CLI runs from an arbitrary working directory, so anchoring the default to cwd made file logging silently write to (or no-op under) an unrelated directory, contradicting the documented `.env.example` default. `getAppLogFilePath()` now computes the default lazily via the pure `resolveDataDir()` resolver (honours a per-process `DATA_DIR`, no directory-creation side effect); an explicit `APP_LOG_FILE_PATH` still wins. Regression guard: `tests/unit/logenv-datadir-path-6197.test.ts` (3).
+
+- **fix(docker):** AgentBridge/`startMitm` no longer aborts in containers/headless when the Antigravity-default DNS step can't write `/etc/hosts` ([#6127](https://github.com/diegosouzapw/OmniRoute/issues/6127)), and the privileged command's stderr now reaches `app.log` instead of only a bare exit code hitting the toast ([#6198](https://github.com/diegosouzapw/OmniRoute/issues/6198)). The default DNS step (`addDNSEntry`) was called unguarded while cert install and the two sibling DNS steps were each best-effort — in the runtime Docker image (`USER node`, no `sudo`, read-only `/etc/hosts`) it threw `Command failed with code 1` out of `startMitmInternal` and killed the whole start, discarding the stderr. The three DNS steps are extracted into a best-effort `provisionDnsEntries()` where each failure is logged with the full `err` (stderr included, folded in by `systemCommands.ts`) and never aborts the start. Regression guard: `tests/unit/mitm-dns-graceful-degrade-6127.test.ts` (4).
+
+- **fix(providers):** copilot-m365-web now supports the M365 Education "Starter / OfficeWebIncludedCopilot" tier and no longer returns an empty `content:null` stream ([#6210](https://github.com/diegosouzapw/OmniRoute/issues/6210)). Two gaps: (1) `buildWsUrl()` hardcoded the individual-consumer scenario (`OfficeWebPaidConsumerCopilot`, `isEdu=false`) — the EDU tier is now opt-in via `providerSpecificData.tier="edu"`, emitting `scenario=OfficeWebIncludedCopilot`/`isEdu=true` (the individual path is unchanged); (2) the EDU/GPT-5.5 path streams deltas via `arguments[0].writeAtCursor` (incremental) instead of only `messages[].text` (accumulated snapshots), which the parser dropped — a new `accumulateBotContent()` folds both formats, with `type:2 item.result.message` as a last-resort fallback. Regression guard: `tests/unit/copilot-m365-edu-writeatcursor-6210.test.ts` (10). (thanks @qpeyba)
+
+- **fix(providers):** GitLab Duo executor now feeds tool results back into the prompt instead of looping ([#6220](https://github.com/diegosouzapw/OmniRoute/issues/6220)) — `buildPrompt()` branched only on `system`/`user` and took `userParts.at(-1)`, silently dropping the `assistant{tool_calls}` + `tool{result}` turns the client appended, so the reconstructed prompt was byte-identical to turn 1 and the model re-emitted the same `<tool>` call forever. When a tool exchange is present the full conversation is now serialized, folding each tool result back keyed by its `tool_call_id`; simple conversations keep the legacy shape. Complements the tool_call emission from [#6051](https://github.com/diegosouzapw/OmniRoute/issues/6051) (the `kilo-duplicate` label was a false positive — different, sequential defect). Regression guard: `tests/unit/gitlab-tool-result-feedback-6220.test.ts` (4).
+
+- **fix(providers):** opencode-go/opencode-zen can now synthesize the OpenCode CLI identity headers Cloudflare requires on VPS egress ([#5997](https://github.com/diegosouzapw/OmniRoute/issues/5997)) — on a datacenter VPS, `opencode.ai/zen/go/v1/chat/completions` 403s (HTML challenge) requests lacking CLI identity, while the reporter's control curl proved that `User-Agent: opencode-cli/1.0.0` + `x-opencode-client: cli` + `x-opencode-project: default` + fresh request/session UUIDs succeed. Opt-in via `OPENCODE_SYNTHESIZE_CLI_HEADERS=true` (values overridable via `OPENCODE_GO_USER_AGENT`/`OPENCODE_USER_AGENT`/`OPENCODE_CLIENT`/`OPENCODE_PROJECT`); it fills only headers the client did not already send. Kept **off by default** — the forward-only path is deliberate (fabricating a wrong value risks upstream rejection; a prior dedup regressed with `opencode/local`), so this replaces the fragile local header-injection shim without changing default behavior. Regression guard: `tests/unit/opencode-cli-headers-synthesis-5997.test.ts` (6). (thanks @aleksesipenko)
+- fix(resilience): sticky session affinity now evicts and fails over to another account when the pinned account is exhausted/unavailable (#6219)
+
+- fix(sse): Responses API passthrough now drops internal commentary-phase output before forwarding to clients (gated by RESPONSES_PASSTHROUGH_DROP_COMMENTARY, default on) (#6199)
+
+- **fix(sse):** tool-call function schemas with a root `type: null` are now coerced to `type: "object"` before dispatch ([#6359](https://github.com/diegosouzapw/OmniRoute/issues/6359)) — clients like the Codex app emit `parameters: { type: null, ... }` for some tools, which OpenAI-compatible upstreams reject with `400 Invalid schema for function '...': schema must be a JSON Schema of 'type: "object"', got 'type: null'`, failing the whole request. `toolSchemaSanitizer` already stripped the null; it now re-adds the mandatory root `"object"` type (and empty `properties`/open `additionalProperties` when absent). Combinator roots (`anyOf`/`oneOf`/`allOf`) and explicit root types are left untouched. Regression guard: 5 new cases in `tests/unit/tool-schema-sanitizer.test.mjs`.
+
+- **fix(docker):** AgentBridge no longer fails to start on npm/Electron/VPS installs with "MITM manager stub reached at runtime" ([#6344](https://github.com/diegosouzapw/OmniRoute/issues/6344)) — v3.8.45 flipped the production bundler default to Turbopack, but `next.config.mjs` aliased `@/mitm/manager` to its Docker-only degraded stub **unconditionally**. That was harmless while Docker (which sets the alias intentionally for #3390 graceful degradation) was the sole Turbopack consumer, but once every artifact built with Turbopack the stub shipped to all non-Docker users and `startMitm` threw on the first Agent-Bridge start. The alias is now opt-in via `OMNIROUTE_MITM_STUB=1` (set only by the Dockerfile) through the shared `scripts/build/mitm-stub-flag.mjs` helper; default builds bundle the real manager. Regression guard: `tests/unit/mitm-stub-alias-6344.test.mjs` (4).
+
+- **fix(proxy):** stop the v3.8.44 proxy regression that leaked the real IP and disabled healthy proxies ([#6246](https://github.com/diegosouzapw/OmniRoute/issues/6246)). Two coupled defects from the new health scheduler: (1) **IP leak** — when a proxy assigned to a connection was marked `inactive`, resolution fell through to a **direct** egress instead of blocking, exposing the operator's real IP; (2) **over-deactivation** — the sweep flipped a proxy to `inactive` on the **first** failed probe and counted our own 5s timeout / a probe-target `5xx` as the proxy's fault, so healthy paid proxies vanished from egress selection ("my proxies are not being used anymore"). Fix: the sweep decision is extracted into a pure, network-free `decideProxyHealthAction` (`src/lib/proxyHealth/decision.ts`) — by default the health check now **only counts/logs and never downgrades status** (a proxy is downgraded/removed only with `PROXY_AUTO_REMOVE=true`, after `PROXY_AUTO_REMOVE_AFTER` **consecutive** conclusive failures); probes are classified tri-state so an inconclusive result (our timeout, or a `5xx` from the probe target) never penalizes the proxy, and the probe timeout is raised 5s→15s. Separately, `safeResolveProxy` now **fails closed** via the existing policy: a connection whose assigned proxy is dead is blocked instead of leaking direct (`hasBlockingProxyAssignment`), honoring the explicit `proxy off` toggles and the `PROXY_FAIL_OPEN=true` opt-out. Existing proxies stuck `inactive` by the old behavior need a one-time manual re-activate (the operator owns proxy status). Regression guards: `tests/unit/proxy-health-decide-action-6246.test.ts`, `tests/unit/proxy-assigned-unavailable-6246.test.ts`.
+
+- **fix(proxy):** make "Test All" read-only and add bulk enable/disable ([#6246](https://github.com/diegosouzapw/OmniRoute/issues/6246)). Complements the core fail-closed / scheduler fix (#6296) with the two remaining reporter asks. (1) The **"Test All" button** (`POST /api/settings/proxies/auto-test`) used to flip a proxy to `inactive` on a failed reachability probe; since the egress selector excludes `inactive` proxies, a flaky probe (an unreachable `httpbin.org`, a proxy that blocks `HEAD`, or a slow paid proxy) silently disabled every proxy that failed — "Test All" is now **read-only by default** (only the operator sets a proxy active/inactive; opt back into the legacy test-and-set with `PROXY_HEALTH_AUTO_DEACTIVATE=true`). (2) Adds a **bulk enable/disable** proxies endpoint + toolbar action (`POST /api/settings/proxies/batch-activate`) so an operator can re-activate proxies in one click. Regression guard: `tests/unit/proxy-health-6246.test.ts`. (thanks @tenshiak)
+
+- **chatcore (tools): stop the default 128-tool cap from silently dropping opencode's `task`/MCP tools.** opencode (used as an MCP/agent host) sends a large tool list; when it exceeds the speculative `MAX_TOOLS_LIMIT` (128) default, `truncateToolList` did a blind `tools.slice(0, 128)`, dropping every tool past index 128 — including opencode's built-in `task` tool (subagent launch) and many MCP tools, so models routed through OmniRoute could no longer spawn subagents or reach part of their tools. The cap exists to avoid upstream `400`s for providers with real hard limits (e.g. grok-cli 200), so it is kept for those: detection of the opencode client (`isOpencodeClient` — any `x-opencode-*` header, or `opencode` in the user-agent) now only bypasses the **speculative 128 default**, never a known provider ceiling. Precedence is explicit — a proactive/detected provider limit always truncates (even for opencode); otherwise opencode forwards its full tool list; otherwise the unchanged 128 default applies to every other client. Refactors `getEffectiveToolLimit` into `getKnownToolLimit(provider) ?? DEFAULT_LIMIT` (byte-identical for existing callers) and fixes a cosmetic debug-log that reported the truncated count instead of the original. Regression guard: `tests/unit/tool-limit-detector.test.ts`.
+
+- **fix(mitm):** the macOS MITM-cert install check now matches the system keychain again. `security find-certificate -a -Z` prints the SHA-1 as a colon-less hex string, but the installed-check compared it against `getCertFingerprint()`'s colon-separated form, so the substring match never hit — the cert was reported as not-installed and re-prompted for the sudo install on every run. Fingerprints are now normalized (colons stripped, upper-cased) on both sides via the extracted `macCertOutputHasFingerprint` helper. Regression guard: `tests/unit/mitm-cert-mac-fingerprint.test.ts`. ([#6204](https://github.com/diegosouzapw/OmniRoute/pull/6204), closes [#6134](https://github.com/diegosouzapw/OmniRoute/issues/6134) — thanks @rianonehub)
+
+- **fix(api):** `/v1/messages/count_tokens` now counts `tool_use`, `tool_result` and `thinking` content blocks (and array-form `system` prompts) in the local-estimation path, instead of only `text`. Real agentic conversations keep ~95% of their tokens inside tool results; the previous estimate returned near-zero for them, which silently broke Claude Code's auto-compaction (context grew past the window with no compaction until the upstream API rejected the request). The real provider-side count path is unchanged. Regression guard: `tests/unit/messages-count-tokens-route.test.ts`. ([#6221](https://github.com/diegosouzapw/OmniRoute/pull/6221) — thanks @luweiCN)
+
+- **fix(antigravity):** strip a trailing assistant prefill turn for Vertex Claude models to avoid upstream 400s ([#6114](https://github.com/diegosouzapw/OmniRoute/pull/6114)). Regression guard: `tests/unit/antigravity-claude-prefill-strip.test.ts`. (thanks @anki1kr)
+
+- **fix(security):** the mutable cloud-agent routes (`/api/cloud/credentials/update`, `/api/cloud/models/alias`) now require management auth instead of being treated as public. They were classified as public API routes, so a request without management credentials could update stored cloud-agent credentials and model aliases. They are removed from the public-route set, classified as management routes in the authz pipeline, and gated by `requireManagementAuth`; cloud **read**/auth routes stay public. Regression guards: `tests/unit/cloud-write-auth.test.ts`, `tests/unit/authz/classify.test.ts`, `tests/unit/public-api-routes.test.ts`. ([#6233](https://github.com/diegosouzapw/OmniRoute/pull/6233) — thanks @vittoroliveira-dev)
+
+- **refactor(dashboard):** extract the onboarding-wizard "Open provider details" link target into a pure, unit-tested `buildProviderDetailsHref(connection)` helper. The wizard already routes by `connection.id` (the node UUID) rather than the provider category slug (#6144/#6145); this hardens that behavior behind a tested helper that guards a missing id/connection. Regression guard: `tests/unit/provider-onboarding-href.test.ts`. ([#6166](https://github.com/diegosouzapw/OmniRoute/pull/6166) — thanks @KooshaPari)
+
+- **fix(security):** the doubao synthetic device-id generator now derives its digits via an unbiased crypto-random draw (rejection sampling over `crypto.randomBytes()`) instead of a `% 10` reduction, closing a CodeQL `js/biased-cryptographic-random` finding.
+- **fix(agentSkills):** the GitHub-skills generator now resolves `outputDir` to an absolute path before writing, fixing a regression introduced by #6366 (relative-to-cwd base path) that could write generated skill files to the wrong directory.
+- **fix(security):** `/api/keys/{id}/devices` now checks the HTTP method before auth/validation, returning a `405` for non-GET/DELETE verbs instead of a misleading `401`/`500` (closes a `dast-smoke` QUERY-method finding).
+- **fix(quality):** clear the last 2 heavy quality-gate reds on the release tip (cycle pre-flight).
+- **fix(mitm):** the test suite and CI can never mutate the OS trust store — `OMNIROUTE_SKIP_SYSTEM_TRUST=1` is set globally for tests/CI so `installCert`/`uninstallCert`/`installTproxyCa` skip the privileged OS dispatch ([#6310](https://github.com/diegosouzapw/OmniRoute/pull/6310); full detail is under the [3.8.45] section below — this branch received it via the parallel-cycle sync-back).
+- **fix(api):** `POST /api/github-skills` now Zod-validates its request body; documented the new quality-gate env vars and pinned the merge-integrity GitHub Actions to a commit SHA.
+- **fix(skills):** generate the missing `omni-github-skills` registry entry and align the agent-skills catalog-count tests (follow-up to #6186).
+- **fix(quality):** clear the cycle's 11 net-new ESLint errors and make `validate-release-green` suppressions-aware.
+
+### 📝 Maintenance
+
+- **i18n(it):** add 118 missing Italian (`it`) translations (net-additive — no existing keys dropped, valid JSON), improving Italian UI coverage. ([#6212](https://github.com/diegosouzapw/OmniRoute/pull/6212) — thanks @serverless83)
+- **chore(providers):** remove deprecated **MiMo V2** model entries from the catalogs (xiaomi-mimo, opencode-go, zenmux-free, audio TTS) — the upstream V2 line is superseded by MiMo V2.5; drops `mimo-v2-tts`, `mimo-v2-pro`, `mimo-v2-omni`, `mimo-v2-flash`, `mimo-v2-flash-free` and realigns the provider-catalog tests. ([#6248](https://github.com/diegosouzapw/OmniRoute/pull/6248) — thanks @backryun)
+- **chore(release):** ~50 commits on this branch are v3.8.45 pre-flight/hardening fixes and CI-perf work that landed here via the parallel-cycle sync-back (`sync-next-cycle.mjs`, Hard Rule #21) **after** the `v3.8.45` git tag was cut, and are already fully documented under the **[3.8.45]** section below — listed here only so the per-cycle commit-coverage check (`npm run release:uncovered`) doesn't flag them as gaps. Provider/catalog/UX/backend fixes: #6041, #6078, #6108, #6135, #6148, #6149, #6154, #6158, #6161, #6162, #6163, #6164, #6165, #6170, #6177, #6178, #6181, #6186, #6187, #6191, #6193, #6194, #6195, #6200, #6205, #6208, #6209, #6211, #6213, #6223, #6224, #6225, #6226, #6227, #6228, #6229, #6230, #6235, #6291, #6292. CI/release-pipeline work: #6167, #6203, #6214, #6215, #6218, #6273, #6275, #6283, #6284, #6285, #6300, #6305.
+- **chore(release):** additional zero-ref release-cycle plumbing on this branch, kept out of `release:uncovered` on purpose (no `#N` in the commit subject to cite): opening the v3.8.46 cycle, opening/closing the v3.8.45 cycle, the finalized [3.8.45] CHANGELOG i18n sync-back to 42 mirrors, the v3.8.45 cognitive/cyclomatic and file-size drift rebaselines, ESLint stale-suppression pruning (4,273 → 4,233), and clearing test-masking/docs-all pre-flight reds for v3.8.45.
+
+### ⚡ Performance & Infrastructure
+
+- **perf(release-green):** the pre-flight validator (`scripts/quality/validate-release-green.mjs`) now runs its 4 slow suites (unit / vitest / integration / pack-artifact) **concurrently** via `Promise.all` — pre-flight wall time drops from ~the sum of the suites to ~the slowest one (~30min saved per round; Phase 0 was the nº1 bottleneck of the v3.8.45 release benchmark, 2h54 of 6h34 e2e). Guard: `tests/unit/validate-release-green.test.ts` ("runs the slow suites CONCURRENTLY"). ([#6319](https://github.com/diegosouzapw/OmniRoute/pull/6319))
+- **fix(ci):** `scripts/release/sync-next-cycle.mjs` — two defects found live in its first production run (v3.8.45 Phase 5): (1) the `git()` helper's default 1 MiB `maxBuffer` crashed with `ENOBUFS` on `git show origin/main:CHANGELOG.md` (the CHANGELOG alone is >1 MiB) — widened to 64 MiB; (2) the i18n resync only propagated the `[NEXT]` (TBD) section, leaving the just-shipped finalized section as "— TBD" in all 42 mirrors — it now also syncs `[prevVersion]` bounded by the heading below it (new exported pure helper `versionAfter`). Guards: +5 tests in `tests/unit/sync-next-cycle.test.ts` (8/8). ([#6327](https://github.com/diegosouzapw/OmniRoute/pull/6327))
+- **test(ci):** concurrency-sensitive flaky tests are **quarantined into a serial pass** (`tests/unit/serial/`, `--test-concurrency=1`, appended to every unit runner incl. sharded variants — the serial pass is sharded too so concurrent shard jobs never self-collide). Initial set: `glm-coding-plan-monthly-3580`, `quota-division-blocks`, `provider-health-autopilot`, `combo-health-autopilot` — the class behind the ~28min CI wedges/re-runs (two live 1h+ wedges cancelled during this PR's own validation). Discovery + TIA gates track the new glob; systemic root cause (async logger writing after teardown) tracked in [#6360](https://github.com/diegosouzapw/OmniRoute/issues/6360). Guard: `tests/unit/test-serial-quarantine.test.ts` (4). ([#6347](https://github.com/diegosouzapw/OmniRoute/pull/6347))
+
+### 🙌 Contributors
+
+Thanks to everyone whose work landed in v3.8.46:
+
+| Contributor                                                    | PRs / Issues           |
+| -------------------------------------------------------------- | ---------------------- |
+| [@2220258345](https://github.com/2220258345)                   | direct commit / report |
+| [@abdofallah](https://github.com/abdofallah)                   | direct commit / report |
+| [@adentdk](https://github.com/adentdk)                         | direct commit / report |
+| [@aleksesipenko](https://github.com/aleksesipenko)             | direct commit / report |
+| [@anki1kr](https://github.com/anki1kr)                         | direct commit / report |
+| [@anungma](https://github.com/anungma)                         | direct commit / report |
+| [@arpicato](https://github.com/arpicato)                       | direct commit / report |
+| [@backryun](https://github.com/backryun)                       | #6248                  |
+| [@binsarjr](https://github.com/binsarjr)                       | direct commit / report |
+| [@brightfiscalband](https://github.com/brightfiscalband)       | direct commit / report |
+| [@chirag127](https://github.com/chirag127)                     | #6501, #6506           |
+| [@developerjillur](https://github.com/developerjillur)         | direct commit / report |
+| [@dilneiss](https://github.com/dilneiss)                       | #6499                  |
+| [@dtybnrj](https://github.com/dtybnrj)                         | direct commit / report |
+| [@Forcerecon](https://github.com/Forcerecon)                   | direct commit / report |
+| [@hao3039032](https://github.com/hao3039032)                   | direct commit / report |
+| [@Iammilansoni](https://github.com/Iammilansoni)               | #6150, #6245           |
+| [@jmengit](https://github.com/jmengit)                         | direct commit / report |
+| [@jordansilly77-stack](https://github.com/jordansilly77-stack) | direct commit / report |
+| [@JxnLexn](https://github.com/JxnLexn)                         | direct commit / report |
+| [@KooshaPari](https://github.com/KooshaPari)                   | #6166                  |
+| [@loopyd](https://github.com/loopyd)                           | direct commit / report |
+| [@luweiCN](https://github.com/luweiCN)                         | #6221                  |
+| [@makcimbx](https://github.com/makcimbx)                       | direct commit / report |
+| [@muflifadla38](https://github.com/muflifadla38)               | direct commit / report |
+| [@newnol](https://github.com/newnol)                           | direct commit / report |
+| [@ofekbetzalel](https://github.com/ofekbetzalel)               | direct commit / report |
+| [@ohahe52-dot](https://github.com/ohahe52-dot)                 | direct commit / report |
+| [@phidinhmanh](https://github.com/phidinhmanh)                 | direct commit / report |
+| [@powellnorma](https://github.com/powellnorma)                 | direct commit / report |
+| [@qpeyba](https://github.com/qpeyba)                           | direct commit / report |
+| [@RaviTharuma](https://github.com/RaviTharuma)                 | direct commit / report |
+| [@RCrushMe](https://github.com/RCrushMe)                       | direct commit / report |
+| [@rianonehub](https://github.com/rianonehub)                   | #6134, #6204           |
+| [@serverless83](https://github.com/serverless83)               | #6212                  |
+| [@swingtempo](https://github.com/swingtempo)                   | direct commit / report |
+| [@tenshiak](https://github.com/tenshiak)                       | direct commit / report |
+| [@ThongAccount](https://github.com/ThongAccount)               | direct commit / report |
+| [@UnrealAryan](https://github.com/UnrealAryan)                 | direct commit / report |
+| [@vinayakkulkarni](https://github.com/vinayakkulkarni)         | direct commit / report |
+| [@vittoroliveira-dev](https://github.com/vittoroliveira-dev)   | #6233                  |
+| [@warelik](https://github.com/warelik)                         | direct commit / report |
+| [@xxy9468615](https://github.com/xxy9468615)                   | direct commit / report |
+| [@xz-dev](https://github.com/xz-dev)                           | direct commit / report |
+| [@yanpaing007](https://github.com/yanpaing007)                 | direct commit / report |
+| [@diegosouzapw](https://github.com/diegosouzapw)               | maintainer             |
+
+---
+
 ## [3.8.45] — 2026-07-06
 
 ### ✨ New Features

docs/i18n/hi/CHANGELOG.md

@@ -6,6 +6,213 @@
 
 ## [3.8.31] — 2026-06-20
 
+## [3.8.46] — TBD
+
+### ✨ New Features
+
+- **feat(sse):** **hide paid-only models from `auto/*` routing** when `hidePaidModels` is on ([#6512](https://github.com/diegosouzapw/OmniRoute/issues/6512)) — follow-up to #6328/#6495. PR #6495 hid paid-only models from the `GET /v1/models` listing, but `auto/*` combos (`auto/best-coding`, `auto/glm`, …) could still pick a paid-only backend into their candidate pool → a 402/403 at request time. `createVirtualAutoCombo` now filters the candidate pool through the new pure `open-sse/services/autoCombo/paidModelFilter.ts` (`filterPaidOnlyCandidates`), applying the same free-model predicate #6495 uses in `catalog.ts` (`providerHasFreeModels(provider) && isFreeModel(provider, {id})`) whenever `settings.hidePaidModels === true`. Applied before the category/tier/family narrowing, so it covers every `auto/*` combo; an all-paid pool degrades to the existing graceful empty-pool path. **Opt-in — default OFF leaves the pool unchanged** (identity). Regression guard: `tests/unit/autoCombo/paid-model-filter-6512.test.ts` (4, incl. the default-off identity guard).
+- **feat(sse):** **provider-family auto combos** — `auto/glm`, `auto/minimax`, `auto/mimo`, `auto/zai`, `auto/gemma`, `auto/llama`, `auto/gemini` ([#6453](https://github.com/diegosouzapw/OmniRoute/issues/6453)) — new routable ids that materialize an on-demand virtual combo spanning whatever installed backends currently expose that model family, degrading gracefully as backends rotate. A new pure `open-sse/services/autoCombo/modelFamily.ts` (`detectModelFamily`) classifies by model-id prefix for six families; `zai` is instead resolved by provider id (z.ai's hosted API serves the same `glm-*` model ids as every other GLM backend, so `auto/zai` means "route to my z.ai backend specifically" vs `auto/glm`'s "any connected GLM backend"). Reuses the existing `createVirtualAutoCombo` on-demand materialization path (no DB writes) and the `/v1/models` catalog advertising loop. Regression guard: `tests/unit/autoCombo/provider-family-combos.test.ts` (11).
+- **feat(proxy):** native **proxy-pool round-robin / egress IP rotation** ([#6365](https://github.com/diegosouzapw/OmniRoute/issues/6365)) — a scope (global / provider / account) can now hold **multiple** proxies as a pool with a rotation strategy, so outbound requests cycle their egress IP instead of pinning one proxy per scope. Migration `117_proxy_pool_rotation.sql` lifts the `UNIQUE(scope, scope_id)` constraint (rebuild via the canonical rename/copy/drop; existing single assignments become 1-element pools) and adds a `proxy_scope_rotation` companion table holding the per-scope strategy + a persisted monotonic round-robin cursor. Strategies: `round-robin` (default, monotonic cursor — never `Math.random`), `random`, and `sticky-per-N-min`. Resolution (`resolveProxyForScopeFromRegistry` / `resolveProxyForConnectionFromRegistry`) now fetches the alive, position-ordered candidate set (unchanged `PROXY_ALIVE_PREDICATE`) and applies the strategy; an empty / all-dead pool still returns `null` — the #6246 fail-closed guard is **untouched** (never falls through to direct egress). Backend + DB only; dashboard pool-builder UI is a follow-up. Regression guard: `tests/unit/proxy-pool-rotation-6365.test.ts` (8, incl. fail-closed + backward-compat).
+- **feat(providers):** end-to-end **tool/function calling on the native Gemini `/v1beta` endpoint** ([#6222](https://github.com/diegosouzapw/OmniRoute/issues/6222)) — both directions of the Gemini↔OpenAI conversion now preserve tool data (previously silently dropped). Request side: `convertGeminiToInternal` (extracted to its own testable module) maps `tools[].functionDeclarations` → OpenAI `tools`, prior `functionCall` parts → assistant `tool_calls`, and `functionResponse` parts → `tool`-role messages. Response side: `convertOpenAIResponseToGemini` emits `parts[].functionCall {name,args}` from `message.tool_calls`, and the streaming `openAIChunkToGeminiChunk` accumulates fragmented `tool_calls` deltas by index into complete `functionCall` parts. The non-Gemini client paths (Claude, OpenAI-Responses) already preserved tool calls — this closes the gap specific to the native Gemini surface. Regression guard: `tests/unit/v1beta-gemini-tool-calling-6222.test.ts` (6, incl. a streaming SSE round-trip).
+- **feat(providers):** copilot-m365-web **enterprise / work tier** support ([#6334](https://github.com/diegosouzapw/OmniRoute/issues/6334)) — mirrors the EDU-tier pattern (#6210): `M365ConnectionParams` gains an `agent` field, a new opt-in `M365_ENTERPRISE_OVERRIDES` preset (`agent=work`, `scenario=officeweb`, `licenseType=Premium`) applies via `providerSpecificData.tier="enterprise"` (alias `"work"`), and `agent` is also overridable directly via `providerSpecificData.agent`. `buildWsUrl` was hardcoding `agent="web"` (the one enterprise-distinguishing param with no override path), so a Premium work account handshook then returned an empty stream. The individual and EDU paths are untouched. Kilo's dup flag vs #6210 (EDU tier) was a false positive — different tier. Regression guard: `tests/unit/copilot-m365-enterprise-6334.test.ts` (7). End-to-end confirmation on a real Premium work account is a live-VPS validation follow-up (Hard Rule #18). (thanks @Forcerecon)
+- **feat(api):** standardized, provider-agnostic **`effort` + `thinking` request params** ([#6241](https://github.com/diegosouzapw/OmniRoute/issues/6241)) — a thin standardization layer over the existing mature per-provider reasoning plumbing (no provider mapper touched). `providerChatCompletionSchema` gains a canonical `effort` (reusing the shared `none/low/medium/high/xhigh` vocabulary — the UI tiers `extra`/`max` collapse onto `xhigh`) and a boolean `thinking`. A pure `normalizeReasoningRequest` (wired once in `src/sse/handlers/chat.ts`, before any reasoning field is read) folds them onto the fields the translators already consume (`reasoning_effort` / `reasoning.effort` / `thinking`), so they fan out to Anthropic / Gemini / xAI / Responses — an explicit client `reasoning_effort` / object-shaped `thinking` always wins (backward-compatible). `/models` additively exposes `supportsThinking` + `effort_tiers` so the frontend can render the toggles (UI component is a follow-up). Regression guard: `tests/unit/effort-thinking-standardization-6241.test.ts` (12). (thanks @Iammilansoni, @shabeer)
+- **feat(combo):** new **`pipeline` (sequential) combo strategy** ([#6297](https://github.com/diegosouzapw/OmniRoute/issues/6297)) — the 18th routing strategy runs targets **in order**, threading each step's output into the next step's input, with an optional per-step `prompt` (system instruction); only the final step's response is returned. Distinct from `fusion` (parallel fan-out + judge). Implemented as a self-contained `open-sse/services/pipeline.ts` (sibling to `fusion.ts`), dispatched from `combo.ts`; the step list reuses `combo.models` order and reads an optional `prompt` off each target (backward-compatible — ignored by every other strategy). Intermediate steps run non-streaming with tools stripped (complete prose to thread forward); the final step keeps the client's `stream` flag + tools. A failing/empty/unparseable intermediate step fails the whole pipeline explicitly via a sanitized error (never silently swallowed). Kilo's dup flag vs #563 was a false positive (that's model→chain selection; this is a sequential chain). Regression guard: `tests/unit/combo-pipeline-strategy.test.ts` (5). (thanks @ofekbetzalel)
+- **feat(ci):** `check:test-masking` now flags **inline-reimplemented prod conditions** ([#6348](https://github.com/diegosouzapw/OmniRoute/issues/6348)) — a new **report-only** subcheck (v2, 6A.10 family) catches the wrong-shape contract test: a test that recomputes the condition under test inline instead of importing/exercising the real function (the #6216 class, where `=== 500` → `>= 500` stayed green because the test re-implemented the branch). For each added/modified test file it warns when the file textually duplicates a ≥3-token conditional from a production file touched in the same PR **and** does not import the symbol/module owning it, via a pure, fixture-tested `findReimplementedConditions()` with an allowlist mirroring `assertReductionAllowlist`. Report-only for now (does not fail the gate) — to be promoted to blocking after a triage cycle. Regression guard: `tests/unit/check-test-masking.test.ts` (45).
+- **feat(sse):** per-connection routing override (native vs CLIProxyAPI) ([#6339](https://github.com/diegosouzapw/OmniRoute/issues/6339)) — the previously-dead `isCliproxyapiDeepModeEnabled` helper is now wired into `resolveExecutorWithProxy`: a single connection can opt itself into the CLIProxyAPI passthrough executor via `providerSpecificData.cliproxyapiMode="claude-native"`, with precedence **connection override > provider `upstream_proxy_config` mode > default**. `resolveExecutorWithProxy` now receives the resolved connection's `providerSpecificData` (threaded from `chatCore.ts`), so one connection can deep-route while the provider's other connections stay native — no DB schema change (the toggle rides in `providerSpecificData`). Also resolves the same-provider mixing ask in #6340. Regression guard: `tests/unit/chatcore-executor-proxy.test.ts` (9). (thanks @RaviTharuma)
+- **feat(dashboard):** "Add session cookie" modal now shows a prominent **"Open ‹host› →"** link to the provider's own site ([#6268](https://github.com/diegosouzapw/OmniRoute/issues/6268)) — every `-web` cookie-session provider (chatgpt-web, claude-web, gemini-web, kimi-web, lmarena, qwen-web, m365-copilot-web, …) renders a one-click external link (opening the provider's login/home page in a new tab) so operators no longer tab away to retype the URL mid-setup. The host resolves from a pure, unit-tested `resolveWebProviderHost()` (prefers `WEB_COOKIE_PROVIDERS[id].website`, falls back to the registry `baseUrl` origin); non-web providers render exactly as before. Kilo's dup flag vs #6265 (modal-too-small-on-1080p) was a false positive — distinct concern. Regression guard: `tests/unit/resolve-web-provider-host.test.ts` (5). (thanks @chirag127)
+- **feat(providers):** add **DigitalOcean** AI (serverless inference) as an OpenAI-compatible API-key provider ([#6373](https://github.com/diegosouzapw/OmniRoute/pull/6373)) — base `https://inference.do-ai.run/v1`, wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/digitalocean/`, `src/shared/constants/providers/apikey/inference-hosts.ts`). Regression guard: `tests/unit/digitalocean-provider.test.ts`. (thanks @newnol)
+- **feat(providers):** add **Huancheng Public API** (`hcnsec`) as an OpenAI-compatible regional provider ([#6410](https://github.com/diegosouzapw/OmniRoute/pull/6410)) — Xinjiang Huancheng Cybersecurity's public LLM platform (base `https://api.hcnsec.cn/v1`, free credits via daily check-ins), wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/hcnsec/`, `src/shared/constants/providers/apikey/regional.ts`). Regression guard: `tests/unit/hcnsec-provider.test.ts`. (thanks @UnrealAryan)
+- **feat(dashboard):** the web-session credential guide now shows an **"Open {host}" link** ([#6316](https://github.com/diegosouzapw/OmniRoute/pull/6316)) to the provider's sign-in site (derived from the provider `website` via `getProviderWebsiteHost`), so you can jump straight to the page where the cookie/session must be captured. Regression guard: `tests/unit/web-session-provider-link-6316.test.ts`. (thanks @jordansilly77-stack)
+- **feat(cerebras):** add the **Gemma 4 31B** model (`gemma-4-31b`) to the Cerebras registry + pricing table ([#6331](https://github.com/diegosouzapw/OmniRoute/pull/6331)). Regression guard extends `tests/unit/t28-model-catalog-updates.test.ts`. (thanks @backryun)
+- **feat(providers):** add **Yuanbao (web)** as a cookie-session provider ([#6196](https://github.com/diegosouzapw/OmniRoute/issues/6196)) — `yuanbao-web` (Tencent Yuanbao, `yuanbao.tencent.com`) with cookie-only auth (`hy_user`/`hy_token` + public agent id), SSE→OpenAI translation incl. `reasoning_content`, exposing DeepSeek V3/R1 + Hunyuan / Hunyuan-T1. Regression guard: `tests/unit/providers-yuanbao-web.test.ts`. `together-web` was **deferred** (no verifiable web-session endpoint — needs a captured request) and `huggingchat-web` **dropped** (the existing `huggingchat` already is a web-cookie provider). (thanks @chirag127)
+- **feat(providers):** route the built-in **agentrouter** through the dynamic Claude-Code wire image ([#6056](https://github.com/diegosouzapw/OmniRoute/issues/6056)) — a small static allow-set (`CC_WIRE_IMAGE_BUILTINS` in `open-sse/services/ccWireImageBuiltins.ts`), consulted by `isClaudeCodeCompatible` / `isClaudeCodeCompatibleProvider` / `applyFingerprint`, makes agentrouter adopt the CC wire-image headers + fingerprint **while guarding the CC baseUrl/auth branches** so it keeps its own registry `baseUrl` and `x-api-key` auth. Regression guard: `tests/unit/agentrouter-cc-wire-image.test.ts` (asserts the wire image is applied AND agentrouter's baseUrl/auth are preserved). Live WAF-acceptance against agentrouter.org is a VPS validation follow-up (Hard Rule #18).
+- **feat(providers):** **bulk-add API keys for Cloudflare Workers AI** ([#6174](https://github.com/diegosouzapw/OmniRoute/issues/6174)) — `cloudflare-ai` is removed from the bulk-add exclusion list and the bulk parser gains a 3-field `name|accountId|apiKey` mode; the bulk route now builds a **per-entry** `providerSpecificData` so each key carries its own `accountId` (fixing the previous shared-object reuse), and both the create + key-validation paths receive it. Regression guard: `tests/unit/bulk-api-key-parser-cloudflare.test.ts`. (thanks @muflifadla38)
+- **feat(dashboard):** routing/settings UX clarity ([#6147](https://github.com/diegosouzapw/OmniRoute/issues/6147)) — (1) weighted combos show the **effective routing share %** next to each weight when weights don't sum to 100 (`WeightTotalBar.tsx`); (2) the status widget's user-facing **"Cloud Sync" label is renamed** to "Remote Settings Sync" (`CloudSyncStatus.tsx`; internal ids/state untouched); (3) built-in providers gain an **opt-in advanced base-URL override** (`isBaseUrlOverrideEligibleProvider`, hidden behind an "Advanced" toggle, reusing the existing `providerSpecificData.baseUrl` persistence — not globally widened). Regression guard: `tests/unit/routing-settings-ux-6147.test.ts`.
+- **feat(combo):** add an option to **disable session stickiness**, per-combo or globally — round-robin / random combos can rotate to a different connection on every request instead of pinning a whole conversation to one connection by its first-message hash. Resolution precedence per-combo `config.disableSessionStickiness` → global `settings.disableSessionStickiness` → default `false` (preserves the #3825 prompt-cache/504 fix); gates **both** stickiness call sites in `open-sse/services/combo.ts`. Exposed as a global toggle (Combo Defaults) and a per-combo Inherit/on/off control. ([#6168](https://github.com/diegosouzapw/OmniRoute/issues/6168)) Regression guard: `tests/unit/combo-disable-session-stickiness.test.ts`. (thanks @RCrushMe)
+- **feat(docker):** add the `OMNIROUTE_NO_SUDO` env flag for root-less / user-namespaced deployments — the MITM cert-trust command path (`resolveSudoSpawn` in `src/mitm/systemCommands.ts`) now strips the leading `sudo` when the flag is truthy, in addition to the existing root / sudo-missing cases, so the Proxy Agent runs without `sudo` (the operator trusts the CA manually, e.g. via `NODE_EXTRA_CA_CERTS`). Argv-array `spawn` preserved — no shell interpolation (Hard Rule #13). ([#6122](https://github.com/diegosouzapw/OmniRoute/issues/6122)) Regression guard: `tests/unit/mitm-systemCommands-no-sudo.test.ts`. (thanks @powellnorma)
+- **feat(providers):** add **Requesty** as an OpenAI-compatible gateway provider (BYOK, base `https://router.requesty.ai/v1`, ~200 free requests/day) — wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/requesty/`, `src/shared/constants/providers/apikey/gateways.ts`). ([#6120](https://github.com/diegosouzapw/OmniRoute/issues/6120)) Regression guard: `tests/unit/requesty-provider.test.ts`. (thanks @chirag127)
+- **feat(dashboard):** add **configured-only / available-only filters** to the Free Provider Rankings page ([#6150](https://github.com/diegosouzapw/OmniRoute/issues/6150)) — hide providers you haven't configured, or whose connections are all rate-limited / out of quota, via server-side query params (`?configuredOnly` / `?availableOnly` on `GET /api/free-provider-rankings`) backed by a testable lib helper reusing the in-process connection state (no Redis). Both filters default off, so the default view is unchanged; this supersedes the earlier client-side "Configured Only" toggle (#6245) with an available-only dimension and unit-tested logic. Regression guard: `tests/unit/freeProviderRankings-filters.test.ts`.
+
+### 🔧 Bug Fixes
+
+- **fix(dashboard):** adding a second API key connection for the same provider no longer silently overwrites the first — the Add-API-key modal now derives a unique default connection name (`main`, then `main-2`, `main-3`, …) so the backend name-based upsert can't collide (#6499 — thanks @dilneiss).
+- **fix(compression):** the session-dedup engine now also deduplicates a large multi-line block repeated **within a single message** (intra-message dedup), not just across turns; the compression-preview API surfaces a `fallbackReason`, and the fusion panel reports how many models were rate-limited vs failed on total-panel failure (#6501 — thanks @chirag127).
+- **fix(compression):** a stacked-pipeline step naming an unregistered engine now surfaces a `validationErrors` entry instead of silently no-op'ing, so misconfigured pipelines are visible in the preview API (#6506 — thanks @chirag127).
+- **feat(usage):** add a **Codex reset-credit redemption** flow to the Provider Limits UI ([#6361](https://github.com/diegosouzapw/OmniRoute/pull/6361)) — a `useCodexResetCreditRedemption` hook + `/api/usage/codex-reset-credit` route + `codexResetCredits` lib let you redeem banked Codex reset credits from the quota card. Regression guard: `tests/unit/codex-reset-credits.test.ts`. (thanks @JxnLexn)
+- **feat(glm):** add **team-plan quota settings** for `glm-cn` connections ([#6351](https://github.com/diegosouzapw/OmniRoute/pull/6351)) — a dedicated `GlmTeamQuotaFields` form section (team quota id / limits) threaded through the Add/Edit connection modals, persisted via `providerSpecificData`, with the GLM usage service reading the team quota. Regression guards: `tests/unit/glm-team-quota.test.ts`, `provider-specific-data-schema.test.ts`. (thanks @hao3039032)
+- **feat(providers):** add **TinyFish** web-fetch/search support ([#6349](https://github.com/diegosouzapw/OmniRoute/pull/6349)) — a `tinyfish-fetch` executor + `/v1/web/fetch` route + MCP web-fetch tool, registered as a specialty-media provider with request-validation and a search-provider catalog entry. Regression guards: `tests/unit/executor-tinyfish-fetch.test.ts`, `web-fetch-handler.test.ts`, `mcp-web-fetch-tool.test.ts`, `provider-validation-tinyfish.test.ts`. (thanks @dtybnrj)
+- **fix(cli):** `omniroute launch-codex` now spawns `codex.cmd` through a shell on Windows (the npm `.cmd` shim is unresolvable by bare `spawn` → ENOENT), mirroring the qodercli Windows fix (#6263) ([#6312](https://github.com/diegosouzapw/OmniRoute/pull/6312)). Regression guard: `tests/unit/launch-codex-windows-spawn-6312.test.ts`. (thanks @swingtempo)
+- **fix(codex):** isolate the **Spark** quota from the shared Codex quota and stabilize the quota UI ordering / hydration so per-scope limits render consistently ([#6336](https://github.com/diegosouzapw/OmniRoute/pull/6336)). Regression guards: `tests/unit/codex-quota-selection-hydration.test.ts`, `provider-limits-ui.test.ts` + 3 more. (thanks @xz-dev)
+- **feat(api):** add a `hidePaidModels` setting that filters paid-only models out of the `/v1/models` catalog. Regression guard: `tests/unit/models-catalog-hide-paid.test.ts`. (thanks @chirag127)
+- **fix(api-manager):** the fallback model picker now preserves combos instead of dropping them when a primary model is unavailable ([#6443](https://github.com/diegosouzapw/OmniRoute/pull/6443)). Regression guard: `tests/unit/api-manager-page-static.test.ts`. (thanks @jmengit)
+- **fix(providers):** recoverable Antigravity / Cloud-Code (Gemini Code Assist) `403` responses ([#6452](https://github.com/diegosouzapw/OmniRoute/pull/6452)) are now classified as a retryable project-config error instead of a terminal account ban, so a fixable project/API-disabled 403 no longer forces a ~1-year cooldown / full OAuth reconnect. Regression guard: `tests/unit/errorclassifier-antigravity-403.test.ts`. (thanks @developerjillur)
+- **fix(mitm):** `sanitizeHeaders` now redacts `Set-Cookie` response headers so upstream session cookies never leak into logs / diagnostics ([#6451](https://github.com/diegosouzapw/OmniRoute/pull/6451)). Regression guard: `tests/unit/mitm-sanitize-headers.test.ts`. (thanks @developerjillur)
+- **fix(api):** `/api/compression/preview` now accepts `mode: "caveman"` and correctly handles stacked / zero-compression previews ([#6425](https://github.com/diegosouzapw/OmniRoute/issues/6425)). Regression guard: `tests/unit/api/compression-preview-caveman-and-stacked-6425.test.ts`. (thanks @chirag127)
+- **feat(providers):** add **Zed** hosted LLM aggregator as a native-app provider ([#6118](https://github.com/diegosouzapw/OmniRoute/pull/6118)) — OAuth sign-in via the Zed hosted flow, registered through the shared provider registry + executor. Regression guards: `tests/unit/zed-oauth-provider.test.ts`, `zed-import-utils.test.ts`, `zed-docker-detect.test.ts`, `mitm-handler-zed.test.ts`. VPS-validated via live operator login (Hard Rule #18).
+- **fix(oauth):** the Kiro SSO-cache auto-import now **preserves the IDC region** — cross-region Amazon Q / Kiro profiles imported from the SSO cache are no longer collapsed to the default region ([#6113](https://github.com/diegosouzapw/OmniRoute/pull/6113)). Regression guard: `tests/unit/kiro-auto-import-idc-2059.test.ts`. VPS-validated via live operator login (Hard Rule #18).
+- **fix(dashboard):** passthrough model aliases no longer collide when two namespaced model ids share a last segment (port from 9router#1850, [#6431](https://github.com/diegosouzapw/OmniRoute/pull/6431)). `enx/gpt-5.5` and `enx/codebuddy/gpt-5.5` both auto-generated the alias `gpt-5.5`, so the second model could never be added (the UI just alerted "alias already exists"). Aliases are now disambiguated deterministically — bare last segment when free, then parent-qualified (`codebuddy-gpt-5.5`), then a numeric suffix — while re-adding the exact same model id is still blocked. Regression guard: `tests/unit/passthrough-alias-1850.test.ts`. (thanks @arpicato)
+- **fix(translator):** preserve a Gemini `functionResponse` co-located with other parts (another `functionCall`, or trailing `text`) in the same content when translating **Gemini → OpenAI** ([#6376](https://github.com/diegosouzapw/OmniRoute/pull/6376)). `convertGeminiContent()` early-returned the tool message on the first `functionResponse` part, dropping any co-located parts; such contents are now pre-split (one tool message per `functionResponse`, emitted first, plus one message for the remaining parts). Regression guard: `tests/unit/gemini-to-openai-function-response.test.ts`. (thanks @warelik)
+- **fix(headroom):** detect a python interpreter managed by **mise / pyenv / asdf / conda** (port from 9router#2353, [#6382](https://github.com/diegosouzapw/OmniRoute/pull/6382)). Headroom's python probe (`src/lib/headroom/detect.ts`) searched a hardcoded `PATH`, but version managers expose their interpreters via shim dirs that only join `PATH` through interactive-shell activation — which the non-interactive server never runs, so a managed python (≥3.10) was invisible and Headroom reported it missing. The search path now prepends the well-known shim/bin dirs (`~/.local/share/mise/shims`, `~/.pyenv/shims`, `~/.asdf/shims`, `$CONDA_PREFIX/bin`, `~/.local/bin`, respecting `MISE_DATA_DIR`/`PYENV_ROOT`/`ASDF_DATA_DIR` when set), and a new `HEADROOM_PYTHON` env override lets operators point straight at their interpreter (mirroring `HEADROOM_URL`). Still shell-free (`execFileSync`). Regression guard: `tests/unit/headroom-detect.test.ts` (5). (thanks @loopyd)
+- **fix(executors):** strip the OpenAI-Codex/Claude-CLI `client_metadata` passthrough field for **NVIDIA** requests (port from 9router#1887, [#6411](https://github.com/diegosouzapw/OmniRoute/pull/6411)). NVIDIA's OpenAI-compatible wrapper rejects it with `400 Unsupported parameter`, the same class already handled for `cerebras`/`mistral`; `nvidia` (executor `default`) was missing from the strip allowlist so Codex/Claude-Code passthrough requests 400'd. Regression guard: `tests/unit/executor-default-strip-client-metadata.test.ts` (+nvidia case). (thanks @phidinhmanh)
+- **fix(translator):** strip the Claude-style `thinking` field for **NVIDIA `z-ai/glm-5.2`** (port from 9router#2023, [#6413](https://github.com/diegosouzapw/OmniRoute/pull/6413)). NVIDIA's OpenAI-compatible wrapper 400s on `thinking` (a Claude-format client routed here leaves a `thinking:{type:"adaptive"}`); the existing strip rule only dropped `reasoning`. Same class already handled for `minimax-m2.7`. Regression guard: `tests/unit/nvidia-minimax-thinking-strip.test.ts` (+glm-5.2 case). (thanks @phidinhmanh)
+- **fix(translator):** suppress the streamed `</think>` close marker for the **Antigravity IDE** client (port from 9router#1061, [#6415](https://github.com/diegosouzapw/OmniRoute/pull/6415)). On thinking-only turns Antigravity rendered a bare `</think>` as the sole visible content, tripping its loop-detection and wasting requests. Antigravity's UA (`vscode/<v> (Antigravity/<v>)`) is added to the marker-suppress allowlist (alongside OpenCode); Claude Code / Cursor still get the marker, and `x-omniroute-thinking-marker: on` force-restores it. Regression guard: `tests/unit/think-close-marker-suppress-5245.test.ts`. (thanks @abdofallah)
+- **fix(executors):** strip nested `reasoning_content` from messages for **Mistral** (port from 9router#1649, [#6417](https://github.com/diegosouzapw/OmniRoute/pull/6417)). Mistral's API returns `422 extra_forbidden` when an assistant message carries `reasoning_content` (replayed thinking from a prior turn, e.g. via the Codex `/responses` path); the generic top-level 400 field-downgrade retry never covered the nested per-message field. `DefaultExecutor` now strips it for provider `mistral` only, so DeepSeek (which requires replayed `reasoning_content`) is unaffected. Regression guard: `tests/unit/mistral-strip-reasoning-content-1649.test.ts`. (thanks @xxy9468615)
+- **fix(executors):** strip the `client_metadata` passthrough field on the **OpenCode** path (port from 9router#1442, [#6418](https://github.com/diegosouzapw/OmniRoute/pull/6418)). OpenCode upstreams (e.g. `kimi-k2.6` via opencode-go) reject it with `400 "Extra inputs are not permitted, field: 'client_metadata'"`; the DefaultExecutor strip only covered cerebras/mistral and `OpencodeExecutor` extends `BaseExecutor` directly, so nothing removed it there. Regression guard: `tests/unit/opencode-strip-client-metadata-1442.test.ts`. (thanks @yanpaing007)
+- **fix(executors):** inject the `reasoning_content` echo for the native **Moonshot Kimi** provider (port from 9router#1480, [#6419](https://github.com/diegosouzapw/OmniRoute/pull/6419)). Kimi (executor `default`) is a thinking-mode upstream that 400s with "reasoning_content must be passed back" when a prior assistant turn lacks it; the placeholder injection was only wired into the OpenCode meta-provider, so direct multi-turn Kimi conversations failed. Scoped to `kimi` (gateway-served models matching the thinking-model name pattern are unaffected). Regression guard: `tests/unit/kimi-native-reasoning-injected-1480.test.ts`. (thanks @2220258345)
+- **fix(executors):** recover from a strict gateway's `context_management: Extra inputs are not permitted` 400 (port from 9router#1468, [#6420](https://github.com/diegosouzapw/OmniRoute/pull/6420)). **Claude Code** always sends a top-level `context_management` field; strict anthropic-compatible gateways reject it. The dedicated context-editing 400-fallback only fired when OmniRoute's own `contextEditing` feature was enabled (default off), so a client-sent field passed through untouched and 400'd. `context_management` is now in the generic reactive field-strip list, so it's stripped-and-retried once regardless of the feature flag (with correct request re-signing for claude-compatible relays). Regression guard: `tests/unit/provider-field-strips.test.ts`. (thanks @ohahe52-dot)
+- **fix(network):** enable **RFC 8305 Happy Eyeballs** (`autoSelectFamily`) on the direct-egress undici dispatcher (port from 9router#1237, [#6423](https://github.com/diegosouzapw/OmniRoute/pull/6423)). When DNS returns both IPv6 (AAAA) and IPv4 (A) and the IPv6 route is broken (e.g. a NAT64 `64:ff9b::` prefix without routing), undici tried IPv6 first and hung until `ETIMEDOUT` (then a 502 + account lockout), even though `curl` reached the same host. The direct dispatcher now races both families and uses whichever connects first. Proxy paths pin family via `proxyTls` and are unaffected. Regression guard: `tests/unit/direct-dispatcher-pipelining-4580.test.ts`. (thanks @adentdk)
+- **fix(combo):** round-robin now advances the rotation pointer past the model that **actually served**, not the eagerly-scheduled one (port from 9router#948, [#6428](https://github.com/diegosouzapw/OmniRoute/pull/6428)). With `stickyLimit: 1` (true round-robin), when the scheduled model failed and a _different_ model served via fallback, the counter had already advanced +1 from the scheduled index — so the next request reused the fallback-served model, degrading round-robin into hot-spotting on whichever model was healthy. The pointer now advances to the served index + 1 (mirroring the sticky-limit>1 path). Session-stickiness (#3825) and distribution are preserved. Regression guard: `tests/unit/combo-rr-fallback-advance-948.test.ts`. (thanks @binsarjr)
+- **fix(sse):** a non-string `model` field is now rejected with a `400` before the resolver, instead of crashing downstream `.toLowerCase()`/`.split()` calls into an empty-body `500` that escapes the error sanitizer ([#6407](https://github.com/diegosouzapw/OmniRoute/issues/6407)). Regression guard: `tests/unit/chat-non-string-model-6407.test.ts`. (thanks @chirag127)
+- **fix(api):** unknown `/api/*` routes now return a JSON `404` (instead of the dashboard HTML shell) and scalar chat params (`model`/`temperature`/etc.) are validated **before** the provider lookup so malformed requests fail fast with a clear `400` ([#6424](https://github.com/diegosouzapw/OmniRoute/issues/6424), [#6412](https://github.com/diegosouzapw/OmniRoute/issues/6412)). Regression guards: `tests/unit/api/api-catchall-json-404.test.ts`, `tests/unit/chat-early-schema-validation-6412.test.ts`. (thanks @chirag127)
+- **fix(api):** `/v1/chat/completions` now rejects a non-JSON `Content-Type` with a `400` before parsing the body ([#6414](https://github.com/diegosouzapw/OmniRoute/issues/6414)). Regression guard: `tests/unit/v1-chat-completions-content-type-6414.test.ts`. (thanks @chirag127)
+- **fix(api):** the `X-OmniRoute-Compression` response header is now echoed on `/v1/chat/completions` and `/v1/completions` ([#6422](https://github.com/diegosouzapw/OmniRoute/issues/6422)). Regression guard: `tests/unit/compression-header-echo-6422.test.ts`. (thanks @chirag127)
+- **fix(api):** concurrent `GET /v1/models` requests are coalesced into a single catalog build ([#6408](https://github.com/diegosouzapw/OmniRoute/issues/6408)). Regression guard: `tests/unit/v1-models-concurrent-6408.test.ts`. (thanks @chirag127)
+- **fix(api):** `/v1/completions` now echoes the requested `body.model` in its JSON + streamed responses ([#6429](https://github.com/diegosouzapw/OmniRoute/pull/6429)). Regression guard: `tests/unit/completions-body-model-echo.test.ts`. (thanks @chirag127)
+- **fix(api):** env-var master keys now see the full `/v1/models` catalog ([#6406](https://github.com/diegosouzapw/OmniRoute/issues/6406)). Regression guard: `tests/unit/models-catalog-envkey-6406.test.ts`. (thanks @chirag127)
+- **fix(api):** non-streaming `/v1/completions` responses now echo `body.model` aligned with the `X-OmniRoute-Model` header ([#6426](https://github.com/diegosouzapw/OmniRoute/issues/6426)). Regression guard: `tests/unit/v1-completions-model-header-match-6426.test.ts`. (thanks @chirag127)
+- **fix(api):** unknown `/v1/*` routes now return a JSON `404 not_found` instead of the Next.js dashboard HTML shell ([#6405](https://github.com/diegosouzapw/OmniRoute/issues/6405)). Regression guard: `tests/unit/api/v1-catchall-json-404.test.ts`. (thanks @chirag127)
+- **fix(api):** the per-connection provider models route now degrades to the shipped catalog when a provider's `/models` endpoint answers with a redirect ([#6267](https://github.com/diegosouzapw/OmniRoute/issues/6267)) — a `qwen-web` import failed with a raw `Redirect blocked … (307)` 503. `safeOutboundFetch` throws `REDIRECT_BLOCKED` on the 307, `getSafeOutboundFetchErrorStatus` maps it to 503, and `buildDiscoveryErrorFallbackResponse` treated every 503 as a hard error — so the non-empty `getModelsByProviderId("qwen-web")` catalog was never surfaced. A models-endpoint redirect is not a fixable-config error (unlike `URL_GUARD_BLOCKED`/`INVALID_URL`, which stay hard errors), so it now falls back to the local/cached catalog before the 503 short-circuit. General fix — covers any config-driven provider that 307s. Regression guard: `tests/unit/provider-models-qwen-web-redirect-6267.test.ts`. (thanks @chirag127)
+- **fix(api):** the per-connection provider models route (MCP `list_models_catalog` + the dashboard import view) now merges USER-ADDED custom models into its response ([#6247](https://github.com/diegosouzapw/OmniRoute/issues/6247)) — custom models live in the `key_value` namespace `customModels`, which the live REST `/api/v1/models` already merges, but `src/app/api/providers/[id]/models/route.ts` never read `getCustomModels`, so custom models were dropped on both the discovery-success and local_catalog paths. They are now folded into the returned model list (deduped by id, stamped `owned_by: provider`), fixing MCP + the dashboard import view in one place. Regression guard: `tests/unit/provider-models-custom-merge-6247.test.ts`. (thanks @RCrushMe)
+
+- **fix(providers):** GitLab Duo tool-calling follow-up turns no longer fail upstream with `422 {"detail":"Validation error"}` (tokens 0/0, rejected pre-inference). The #6234 tool-result-feedback fix serialized the **entire** multi-turn conversation into GitLab's single-file `code_suggestions` (`small_file`) generation endpoint — folded history that turn-N sent as an oversized `current_file.content_above_cursor` **and** duplicated verbatim into `user_instruction`, tripping the AI-Gateway's `small_file` validation guard. The executor now **bounds** that prompt: it keeps system + latest user message + the most-recent tool round (dropping older turns), caps oversized tool results, and stops duplicating the full prompt into `user_instruction` (which now carries only the short latest user message) — while still feeding the most-recent tool result back so the agent continues (`open-sse/executors/gitlab.ts`, [#6220](https://github.com/diegosouzapw/OmniRoute/issues/6220)). The unit test covers the bounding logic; the upstream 422→200 clearing is VPS-only (Hard Rule #18). Regression guard: `tests/unit/gitlab-tool-exchange-bounded-6220.test.ts`.
+
+- **fix(i18n):** the provider-detail (`/dashboard/providers/[id]`) connection-status filter labels no longer render as `__MISSING__:All` / `__MISSING__:Active` / `__MISSING__:Error` / `__MISSING__:Banned` / `__MISSING__:CreditsExhausted` in non-English locales (notably pt-BR) ([#6290](https://github.com/diegosouzapw/OmniRoute/issues/6290)). Root cause was **not** the namespace mismatch the issue guessed — the `providers.filter*` keys resolve correctly in `en.json`; the debt lived in the locale mirrors (`src/i18n/messages/*.json`), where these five keys carried the `__MISSING__:` sync sentinel in ~15 locales and were absent entirely in ~26 others, so next-intl found the key and echoed the sentinel verbatim. All 40 non-English/-Chinese mirrors now ship real translations for the five `providers.filter*` labels. Regression guard: `tests/unit/i18n-provider-filter-keys-6290.test.ts`. (thanks @diegosouzapw)
+
+- **fix(providers):** the `copilot-m365-web` streaming executor now emits `debug`-level WebSocket diagnostics ([#6210](https://github.com/diegosouzapw/OmniRoute/issues/6210)) — the outbound WS URL (with the `access_token` **redacted** via `redactWsUrl()`), handshake success/failure, and each received SignalR frame's `type`/`target`. Previously the streaming path logged nothing, so an empty `content:null` response (the M365 Education / Starter tier symptom fixed in #6234) was undiagnosable even at `APP_LOG_LEVEL=debug`. The change is debug-level and side-effect-free — it does not alter streaming behavior or the frame parser, and the token never reaches the logs. Regression guard: `tests/unit/copilot-m365-web-logging-6210.test.ts` (thanks @qpeyba)
+
+- **fix(resilience):** a round-robin combo no longer returns `503 all upstream accounts are unavailable` when a compatibility-rejected target is actually healthy ([#6238](https://github.com/diegosouzapw/OmniRoute/issues/6238)). `filterTargetsByRequestCompatibility` drops request-incompatible targets (tool/vision/structured-output unsupported, or below the required context window) **before** any availability check runs, and its `compatible.length === 0` safety net only fired when _all_ targets were filtered — not when the kept targets later all turned out runtime-unavailable (circuit-open / cooldown / no credentials). So a combo could 503 while a compat-rejected-but-healthy provider sat unused. `handleRoundRobinCombo` now keeps the compat-rejected set and, when every compat-kept target was skipped without a single real attempt, probes those rejected targets as a **last-resort fallback tier** (via the new pure `open-sse/services/combo/comboCompatFallback.ts`) before crystallizing the 503. Regression guard: `tests/unit/combo-roundrobin-compat-fallback-6238.test.ts`. (thanks @ThongAccount)
+
+- **fix(startup):** best-effort self-heal for a corrupted Turbopack dev cache on Windows ([#6289](https://github.com/diegosouzapw/OmniRoute/issues/6289)). On Windows, `pnpm dev` can fail at startup when Turbopack `mmap`s a persistent-cache SST file and the OS refuses the mapping (`os error 1455` — "paging file too small"), which Turbopack surfaces as a misleading `Module not found: Can't resolve '@/shared/utils/machine'`. This is a **known upstream Turbopack cache-corruption bug — not our code**. The dev launcher (`scripts/dev/run-next.mjs`) now wraps `nextApp.prepare()` and, when it rejects with that signature (`isTurbopackCacheCorruption` in the new `scripts/dev/turbopackCacheHeal.mjs`), purges `.build/next/**/cache/turbopack` and retries **once** with a clear log. **Caveat — best-effort only:** the corruption often surfaces as a runtime overlay rather than a `prepare()` rejection, so this cannot always intercept it; the reliable remedy remains manually deleting the Turbopack cache dir. Regression guard: `tests/unit/turbopack-cache-heal-6289.test.ts`. (thanks @chirag127)
+
+- **fix(providers):** qodercli PAT auth no longer fails with `spawn qodercli ENOENT` on Windows ([#6263](https://github.com/diegosouzapw/OmniRoute/issues/6263)) — `spawnQoderCli` spawned the bare `qodercli` name with `shell:false` and an unenriched env, so the npm `.cmd` wrapper under `%APPDATA%\npm` (a user-PATH directory) was never resolved. It now resolves the absolute `.cmd`/`.exe` path through the existing `getCliRuntimeStatus("qoder")` resolver in `src/shared/services/cliRuntime.ts` (memoized), spawns with `shell` when the target is a `.cmd`/`.bat`, and uses the cliRuntime-enriched env (PATH + PATHEXT + APPDATA); the ENOENT error now lists the searched paths plus the `CLI_QODER_BIN` override. End-to-end spawn on a real Windows host is host-only (Hard Rule #18); the path-resolution logic is unit-tested. Regression guard: `tests/unit/qodercli-windows-resolve-6263.test.ts`. (thanks @chirag127)
+
+- **fix(sse):** the reasoning-token buffer no longer inflates **probe-sized `max_tokens`** ([#6274](https://github.com/diegosouzapw/OmniRoute/issues/6274)) — Claude Code's `/model` capability check sends `max_tokens: 1`, but for a thinking-capable model with a large output cap (e.g. `glm-5.2`) the #3587 headroom heuristic (`max(current + 1000, ceil(current * 1.5))`) rewrote it to `1001` and forwarded that upstream, wasting tokens on a request that was never a genuine reasoning budget. `resolveReasoningBufferedMaxTokens()` (`open-sse/services/reasoningTokenBuffer.ts`) now short-circuits and returns the caller's value verbatim when it is below the new `REASONING_BUFFER_MIN_TRIGGER` (256) threshold — a tiny explicit limit is a probe, not a reasoning request. Real budgets still receive the #3587 headroom unchanged, and the guard runs after the existing capability checks so unknown / non-reasoning models keep returning `null`. Regression guard: `tests/unit/reasoning-token-buffer-6274.test.ts`. (thanks @brightfiscalband)
+
+- **fix(cli):** `omniroute reset-password` now works as a real subcommand, and password resets over piped (non-TTY) stdin actually apply ([#6261](https://github.com/diegosouzapw/OmniRoute/issues/6261), [#6258](https://github.com/diegosouzapw/OmniRoute/issues/6258)). Two coupled defects: (1) **#6261** — `bin/omniroute.mjs` routed everything through Commander with only two pre-Commander bypasses (`--mcp`, `reset-encrypted-columns`), so `omniroute reset-password` was rejected as an unknown command; only the separate `omniroute-reset-password` bin worked, while the docs falsely advertised the subcommand (incl. a bogus "legacy alias still works"). A pre-Commander bypass mirroring `reset-encrypted-columns` now dynamically imports `bin/reset-password.mjs` (which self-executes) before Commander parses; the three doc lines were corrected. (2) **#6258** — `bin/reset-password.mjs` issued two sequential `rl.question` prompts; under piped stdin the second read never settled at EOF, so `main()` never reached `resetManagementPassword` and the reset was a silent no-op (both prompts printed, no success, password unchanged). The CLI now detects non-TTY stdin and reads it once (first line = password, second line = confirm if present, else reused), adds a `--password-stdin` flag (entire stdin is the password, no confirmation), and exits `0` explicitly so the success line always flushes; interactive TTY behavior is unchanged. Regression guard: `tests/unit/reset-password-cli-6261-6258.test.ts` (3). (thanks @chirag127)
+
+- **fix(db):** the mass-migration **safety abort** now tells the operator how to bypass it and stops flooding the log ([#6260](https://github.com/diegosouzapw/OmniRoute/issues/6260)) — after restoring a backup that wiped the migration tracking table, `runMigrations()` threw the abort on every downstream `ensureDbInitialized()`, re-logging the full banner 11+ times, and the message never mentioned the existing `OMNIROUTE_MAX_PENDING_MIGRATIONS` escape hatch. The abort text now appends a bypass hint (set `OMNIROUTE_MAX_PENDING_MIGRATIONS=0` in `server.env` / `DATA_DIR/.env`), and a new `MigrationSafetyAbortError` is memoized so repeated calls in the same process throw the same instance and emit a single concise line instead of the full cascade. Regression guard: `tests/unit/migration-safety-abort-6260.test.ts`. (thanks @chirag127)
+
+- **fix(auth):** importing a **distinct** Codex/ChatGPT OAuth `auth.json` is no longer falsely rejected as "already exists" when it belongs to a different user in the same workspace ([#6301](https://github.com/diegosouzapw/OmniRoute/issues/6301)). `findExistingCodexConnection` (in `src/lib/oauth/utils/codexAuthImport.ts`) deduped **only** on `providerSpecificData.workspaceId === accountId`, where `accountId` is the shared `chatgpt_account_id`/`tokens.account_id` — so two members of the same ChatGPT Team collapsed onto a single connection (409 `duplicate_account`). The id_token's `https://api.openai.com/auth` claim carries a per-user `chatgpt_user_id` alongside the workspace id (the device-flow path already persisted it as `chatgptUserId`, but the import path did not). Now `parseAndValidateCodexAuth` extracts `userId` (`chatgpt_user_id` → `user_id` → JWT `sub`) into `ParsedCodexAuth`, the create/update paths persist `chatgptUserId` in `providerSpecificData` (mirroring `codex.ts`), and dedup keys on `workspaceId` **AND** `chatgptUserId` — with a backward-compat fallback to legacy accountId-only matching when no stored connection for that workspace records a `chatgptUserId`, so genuinely-same accounts still dedup. Regression guard: `tests/unit/codex-auth-import-userid-dedup-6301.test.ts` (4). (thanks @anungma)
+
+- **fix(providers):** importing models for the **venice-web** provider no longer fails with a red "Provider venice-web does not support models listing" ([#6269](https://github.com/diegosouzapw/OmniRoute/issues/6269)). `venice-web` is a web-cookie provider with an executor but no upstream `/v1/models` endpoint and no registry `models`, so the models route fell through to the tail `400`. Mirroring the `jules`/`linkup-search`/`ollama-search` fix (#5569), it now ships a static local catalog entry in `src/lib/providers/staticModels.ts` — seeding the current Venice lineup (`venice-uncensored`, `llama-3.3-70b`, `qwen3-235b`, `qwen3-4b`, `deepseek-r1-671b`; Venice rotates its catalog, see docs.venice.ai/models/overview) — so the route returns `200` with `source:"local_catalog"`, `intentional:true`. Regression guard: `tests/unit/static-models-venice-web-6269.test.ts`. (thanks @chirag127)
+
+- **fix(api):** the specialty model catalogs (`/v1/embeddings`, `/v1/images`, `/v1/music`, `/v1/videos` model lists) are now derived from the **unified catalog filtered by a predicate** (`getSpecialtyModelsResponse`) instead of ad-hoc per-route logic, so they consistently respect active-credential visibility and stay in sync with the main catalog ([#6303](https://github.com/diegosouzapw/OmniRoute/pull/6303)). Regression guard: `tests/unit/specialty-model-catalog-routes.test.ts`. (thanks @makcimbx)
+- **fix(api):** the agent-bridge server route now resolves the MITM manager via a **dynamic `import("@/mitm/manager.runtime")`** so Turbopack does not statically pull the stub (or over-bundle the manager), and the agent-skills generator anchors its output base path with `path.join(process.cwd(), …)` so Turbopack's static analyzer stops tracing the whole project root ([#6329](https://github.com/diegosouzapw/OmniRoute/issues/6329), [#6366](https://github.com/diegosouzapw/OmniRoute/pull/6366)). Regression guard: `tests/unit/agent-bridge-server-route-dynamic-import.test.ts`. (thanks @Iammilansoni)
+- **fix(api):** internal probes (combo-test, cloud-sync verify) now pick a **management-scoped / allow-all API key** instead of naively grabbing `getApiKeys()[0]` — a restricted `self:usage` first row made the probe fail with "Model X is not allowed for this API key" even when the combo path was healthy (`pickApiKeyForInternalUse` in `src/lib/db/apiKeys.ts`). The API-manager model editor also falls back to `/api/models?all=true` when `/v1/models` is catalog-protected ([#6372](https://github.com/diegosouzapw/OmniRoute/pull/6372)). Regression guard: `tests/unit/pick-internal-api-key-6372.test.ts`. (thanks @jmengit)
+- **fix(live-ws):** the Live Dashboard WebSocket server now **rejects on bind failure** (e.g. `EADDRINUSE` when the API bridge already holds the port) instead of letting the error surface as an unhandled `error` event that crash-loops the process — the `error` listener is attached to `wss` (not `server`) and releases the EventBus subscription on a failed start ([#6324](https://github.com/diegosouzapw/OmniRoute/issues/6324)). Regression guard: `tests/unit/live-ws-eaddrinuse-6324.test.ts`. (thanks @vinayakkulkarni)
+- **fix(dashboard):** the Home provider-topology widget now trusts the live provider-metrics snapshot — it uses `topology.errorProvider` and live `activeRequests` directly instead of re-deriving state from a stale `lastErrorAt` or applying a frontend timeout filter, so the topology reflects real-time provider health ([#6322](https://github.com/diegosouzapw/OmniRoute/pull/6322)). Regression guard: `tests/unit/home-provider-topology-live-state.test.ts`. (thanks @xz-dev)
+- **fix(sse):** strip zero-width markers from streamed **tool-call arguments** — a follow-up to [#5857](https://github.com/diegosouzapw/OmniRoute/pull/5857). That PR removed injected zero-width joiners (U+200D) from streamed assistant text/reasoning but deliberately left tool-call argument JSON byte-exact. The request-side obfuscation (`open-sse/services/claudeCodeObfuscation.ts`) injects ZWJ into agent words — including the temp path inside the Bash tool description — and Claude models copy that verbatim into generated commands, which are delivered as tool-call arguments rather than assistant text. As a result the ZWJ survived and corrupted code blocks (e.g. a temp path rendered with an invisible joiner). Now `open-sse/handlers/responseSanitizer.ts` strips zero-width code points from tool-call argument strings at every emit site (OpenAI non-stream/stream chat `tool_calls` + legacy `function_call`, native Responses `function_call` items, the OpenAI→Responses conversion, and the native Responses streaming `response.function_call_arguments.delta/.done` events). Only zero-width code points are removed; JSON structure and all other bytes stay identical (no parse/restringify), so normal arguments remain byte-exact. Regression guard: 6 new cases in `tests/unit/response-sanitizer.test.ts` (suite 50/50).
+
+- **fix(nodejs):** the default app log path now resolves under `DATA_DIR` (`~/.omniroute/logs/application/app.log`) instead of `process.cwd()` ([#6197](https://github.com/diegosouzapw/OmniRoute/issues/6197)) — the globally-installed CLI runs from an arbitrary working directory, so anchoring the default to cwd made file logging silently write to (or no-op under) an unrelated directory, contradicting the documented `.env.example` default. `getAppLogFilePath()` now computes the default lazily via the pure `resolveDataDir()` resolver (honours a per-process `DATA_DIR`, no directory-creation side effect); an explicit `APP_LOG_FILE_PATH` still wins. Regression guard: `tests/unit/logenv-datadir-path-6197.test.ts` (3).
+
+- **fix(docker):** AgentBridge/`startMitm` no longer aborts in containers/headless when the Antigravity-default DNS step can't write `/etc/hosts` ([#6127](https://github.com/diegosouzapw/OmniRoute/issues/6127)), and the privileged command's stderr now reaches `app.log` instead of only a bare exit code hitting the toast ([#6198](https://github.com/diegosouzapw/OmniRoute/issues/6198)). The default DNS step (`addDNSEntry`) was called unguarded while cert install and the two sibling DNS steps were each best-effort — in the runtime Docker image (`USER node`, no `sudo`, read-only `/etc/hosts`) it threw `Command failed with code 1` out of `startMitmInternal` and killed the whole start, discarding the stderr. The three DNS steps are extracted into a best-effort `provisionDnsEntries()` where each failure is logged with the full `err` (stderr included, folded in by `systemCommands.ts`) and never aborts the start. Regression guard: `tests/unit/mitm-dns-graceful-degrade-6127.test.ts` (4).
+
+- **fix(providers):** copilot-m365-web now supports the M365 Education "Starter / OfficeWebIncludedCopilot" tier and no longer returns an empty `content:null` stream ([#6210](https://github.com/diegosouzapw/OmniRoute/issues/6210)). Two gaps: (1) `buildWsUrl()` hardcoded the individual-consumer scenario (`OfficeWebPaidConsumerCopilot`, `isEdu=false`) — the EDU tier is now opt-in via `providerSpecificData.tier="edu"`, emitting `scenario=OfficeWebIncludedCopilot`/`isEdu=true` (the individual path is unchanged); (2) the EDU/GPT-5.5 path streams deltas via `arguments[0].writeAtCursor` (incremental) instead of only `messages[].text` (accumulated snapshots), which the parser dropped — a new `accumulateBotContent()` folds both formats, with `type:2 item.result.message` as a last-resort fallback. Regression guard: `tests/unit/copilot-m365-edu-writeatcursor-6210.test.ts` (10). (thanks @qpeyba)
+
+- **fix(providers):** GitLab Duo executor now feeds tool results back into the prompt instead of looping ([#6220](https://github.com/diegosouzapw/OmniRoute/issues/6220)) — `buildPrompt()` branched only on `system`/`user` and took `userParts.at(-1)`, silently dropping the `assistant{tool_calls}` + `tool{result}` turns the client appended, so the reconstructed prompt was byte-identical to turn 1 and the model re-emitted the same `<tool>` call forever. When a tool exchange is present the full conversation is now serialized, folding each tool result back keyed by its `tool_call_id`; simple conversations keep the legacy shape. Complements the tool_call emission from [#6051](https://github.com/diegosouzapw/OmniRoute/issues/6051) (the `kilo-duplicate` label was a false positive — different, sequential defect). Regression guard: `tests/unit/gitlab-tool-result-feedback-6220.test.ts` (4).
+
+- **fix(providers):** opencode-go/opencode-zen can now synthesize the OpenCode CLI identity headers Cloudflare requires on VPS egress ([#5997](https://github.com/diegosouzapw/OmniRoute/issues/5997)) — on a datacenter VPS, `opencode.ai/zen/go/v1/chat/completions` 403s (HTML challenge) requests lacking CLI identity, while the reporter's control curl proved that `User-Agent: opencode-cli/1.0.0` + `x-opencode-client: cli` + `x-opencode-project: default` + fresh request/session UUIDs succeed. Opt-in via `OPENCODE_SYNTHESIZE_CLI_HEADERS=true` (values overridable via `OPENCODE_GO_USER_AGENT`/`OPENCODE_USER_AGENT`/`OPENCODE_CLIENT`/`OPENCODE_PROJECT`); it fills only headers the client did not already send. Kept **off by default** — the forward-only path is deliberate (fabricating a wrong value risks upstream rejection; a prior dedup regressed with `opencode/local`), so this replaces the fragile local header-injection shim without changing default behavior. Regression guard: `tests/unit/opencode-cli-headers-synthesis-5997.test.ts` (6). (thanks @aleksesipenko)
+- fix(resilience): sticky session affinity now evicts and fails over to another account when the pinned account is exhausted/unavailable (#6219)
+
+- fix(sse): Responses API passthrough now drops internal commentary-phase output before forwarding to clients (gated by RESPONSES_PASSTHROUGH_DROP_COMMENTARY, default on) (#6199)
+
+- **fix(sse):** tool-call function schemas with a root `type: null` are now coerced to `type: "object"` before dispatch ([#6359](https://github.com/diegosouzapw/OmniRoute/issues/6359)) — clients like the Codex app emit `parameters: { type: null, ... }` for some tools, which OpenAI-compatible upstreams reject with `400 Invalid schema for function '...': schema must be a JSON Schema of 'type: "object"', got 'type: null'`, failing the whole request. `toolSchemaSanitizer` already stripped the null; it now re-adds the mandatory root `"object"` type (and empty `properties`/open `additionalProperties` when absent). Combinator roots (`anyOf`/`oneOf`/`allOf`) and explicit root types are left untouched. Regression guard: 5 new cases in `tests/unit/tool-schema-sanitizer.test.mjs`.
+
+- **fix(docker):** AgentBridge no longer fails to start on npm/Electron/VPS installs with "MITM manager stub reached at runtime" ([#6344](https://github.com/diegosouzapw/OmniRoute/issues/6344)) — v3.8.45 flipped the production bundler default to Turbopack, but `next.config.mjs` aliased `@/mitm/manager` to its Docker-only degraded stub **unconditionally**. That was harmless while Docker (which sets the alias intentionally for #3390 graceful degradation) was the sole Turbopack consumer, but once every artifact built with Turbopack the stub shipped to all non-Docker users and `startMitm` threw on the first Agent-Bridge start. The alias is now opt-in via `OMNIROUTE_MITM_STUB=1` (set only by the Dockerfile) through the shared `scripts/build/mitm-stub-flag.mjs` helper; default builds bundle the real manager. Regression guard: `tests/unit/mitm-stub-alias-6344.test.mjs` (4).
+
+- **fix(proxy):** stop the v3.8.44 proxy regression that leaked the real IP and disabled healthy proxies ([#6246](https://github.com/diegosouzapw/OmniRoute/issues/6246)). Two coupled defects from the new health scheduler: (1) **IP leak** — when a proxy assigned to a connection was marked `inactive`, resolution fell through to a **direct** egress instead of blocking, exposing the operator's real IP; (2) **over-deactivation** — the sweep flipped a proxy to `inactive` on the **first** failed probe and counted our own 5s timeout / a probe-target `5xx` as the proxy's fault, so healthy paid proxies vanished from egress selection ("my proxies are not being used anymore"). Fix: the sweep decision is extracted into a pure, network-free `decideProxyHealthAction` (`src/lib/proxyHealth/decision.ts`) — by default the health check now **only counts/logs and never downgrades status** (a proxy is downgraded/removed only with `PROXY_AUTO_REMOVE=true`, after `PROXY_AUTO_REMOVE_AFTER` **consecutive** conclusive failures); probes are classified tri-state so an inconclusive result (our timeout, or a `5xx` from the probe target) never penalizes the proxy, and the probe timeout is raised 5s→15s. Separately, `safeResolveProxy` now **fails closed** via the existing policy: a connection whose assigned proxy is dead is blocked instead of leaking direct (`hasBlockingProxyAssignment`), honoring the explicit `proxy off` toggles and the `PROXY_FAIL_OPEN=true` opt-out. Existing proxies stuck `inactive` by the old behavior need a one-time manual re-activate (the operator owns proxy status). Regression guards: `tests/unit/proxy-health-decide-action-6246.test.ts`, `tests/unit/proxy-assigned-unavailable-6246.test.ts`.
+
+- **fix(proxy):** make "Test All" read-only and add bulk enable/disable ([#6246](https://github.com/diegosouzapw/OmniRoute/issues/6246)). Complements the core fail-closed / scheduler fix (#6296) with the two remaining reporter asks. (1) The **"Test All" button** (`POST /api/settings/proxies/auto-test`) used to flip a proxy to `inactive` on a failed reachability probe; since the egress selector excludes `inactive` proxies, a flaky probe (an unreachable `httpbin.org`, a proxy that blocks `HEAD`, or a slow paid proxy) silently disabled every proxy that failed — "Test All" is now **read-only by default** (only the operator sets a proxy active/inactive; opt back into the legacy test-and-set with `PROXY_HEALTH_AUTO_DEACTIVATE=true`). (2) Adds a **bulk enable/disable** proxies endpoint + toolbar action (`POST /api/settings/proxies/batch-activate`) so an operator can re-activate proxies in one click. Regression guard: `tests/unit/proxy-health-6246.test.ts`. (thanks @tenshiak)
+
+- **chatcore (tools): stop the default 128-tool cap from silently dropping opencode's `task`/MCP tools.** opencode (used as an MCP/agent host) sends a large tool list; when it exceeds the speculative `MAX_TOOLS_LIMIT` (128) default, `truncateToolList` did a blind `tools.slice(0, 128)`, dropping every tool past index 128 — including opencode's built-in `task` tool (subagent launch) and many MCP tools, so models routed through OmniRoute could no longer spawn subagents or reach part of their tools. The cap exists to avoid upstream `400`s for providers with real hard limits (e.g. grok-cli 200), so it is kept for those: detection of the opencode client (`isOpencodeClient` — any `x-opencode-*` header, or `opencode` in the user-agent) now only bypasses the **speculative 128 default**, never a known provider ceiling. Precedence is explicit — a proactive/detected provider limit always truncates (even for opencode); otherwise opencode forwards its full tool list; otherwise the unchanged 128 default applies to every other client. Refactors `getEffectiveToolLimit` into `getKnownToolLimit(provider) ?? DEFAULT_LIMIT` (byte-identical for existing callers) and fixes a cosmetic debug-log that reported the truncated count instead of the original. Regression guard: `tests/unit/tool-limit-detector.test.ts`.
+
+- **fix(mitm):** the macOS MITM-cert install check now matches the system keychain again. `security find-certificate -a -Z` prints the SHA-1 as a colon-less hex string, but the installed-check compared it against `getCertFingerprint()`'s colon-separated form, so the substring match never hit — the cert was reported as not-installed and re-prompted for the sudo install on every run. Fingerprints are now normalized (colons stripped, upper-cased) on both sides via the extracted `macCertOutputHasFingerprint` helper. Regression guard: `tests/unit/mitm-cert-mac-fingerprint.test.ts`. ([#6204](https://github.com/diegosouzapw/OmniRoute/pull/6204), closes [#6134](https://github.com/diegosouzapw/OmniRoute/issues/6134) — thanks @rianonehub)
+
+- **fix(api):** `/v1/messages/count_tokens` now counts `tool_use`, `tool_result` and `thinking` content blocks (and array-form `system` prompts) in the local-estimation path, instead of only `text`. Real agentic conversations keep ~95% of their tokens inside tool results; the previous estimate returned near-zero for them, which silently broke Claude Code's auto-compaction (context grew past the window with no compaction until the upstream API rejected the request). The real provider-side count path is unchanged. Regression guard: `tests/unit/messages-count-tokens-route.test.ts`. ([#6221](https://github.com/diegosouzapw/OmniRoute/pull/6221) — thanks @luweiCN)
+
+- **fix(antigravity):** strip a trailing assistant prefill turn for Vertex Claude models to avoid upstream 400s ([#6114](https://github.com/diegosouzapw/OmniRoute/pull/6114)). Regression guard: `tests/unit/antigravity-claude-prefill-strip.test.ts`. (thanks @anki1kr)
+
+- **fix(security):** the mutable cloud-agent routes (`/api/cloud/credentials/update`, `/api/cloud/models/alias`) now require management auth instead of being treated as public. They were classified as public API routes, so a request without management credentials could update stored cloud-agent credentials and model aliases. They are removed from the public-route set, classified as management routes in the authz pipeline, and gated by `requireManagementAuth`; cloud **read**/auth routes stay public. Regression guards: `tests/unit/cloud-write-auth.test.ts`, `tests/unit/authz/classify.test.ts`, `tests/unit/public-api-routes.test.ts`. ([#6233](https://github.com/diegosouzapw/OmniRoute/pull/6233) — thanks @vittoroliveira-dev)
+
+- **refactor(dashboard):** extract the onboarding-wizard "Open provider details" link target into a pure, unit-tested `buildProviderDetailsHref(connection)` helper. The wizard already routes by `connection.id` (the node UUID) rather than the provider category slug (#6144/#6145); this hardens that behavior behind a tested helper that guards a missing id/connection. Regression guard: `tests/unit/provider-onboarding-href.test.ts`. ([#6166](https://github.com/diegosouzapw/OmniRoute/pull/6166) — thanks @KooshaPari)
+
+- **fix(security):** the doubao synthetic device-id generator now derives its digits via an unbiased crypto-random draw (rejection sampling over `crypto.randomBytes()`) instead of a `% 10` reduction, closing a CodeQL `js/biased-cryptographic-random` finding.
+- **fix(agentSkills):** the GitHub-skills generator now resolves `outputDir` to an absolute path before writing, fixing a regression introduced by #6366 (relative-to-cwd base path) that could write generated skill files to the wrong directory.
+- **fix(security):** `/api/keys/{id}/devices` now checks the HTTP method before auth/validation, returning a `405` for non-GET/DELETE verbs instead of a misleading `401`/`500` (closes a `dast-smoke` QUERY-method finding).
+- **fix(quality):** clear the last 2 heavy quality-gate reds on the release tip (cycle pre-flight).
+- **fix(mitm):** the test suite and CI can never mutate the OS trust store — `OMNIROUTE_SKIP_SYSTEM_TRUST=1` is set globally for tests/CI so `installCert`/`uninstallCert`/`installTproxyCa` skip the privileged OS dispatch ([#6310](https://github.com/diegosouzapw/OmniRoute/pull/6310); full detail is under the [3.8.45] section below — this branch received it via the parallel-cycle sync-back).
+- **fix(api):** `POST /api/github-skills` now Zod-validates its request body; documented the new quality-gate env vars and pinned the merge-integrity GitHub Actions to a commit SHA.
+- **fix(skills):** generate the missing `omni-github-skills` registry entry and align the agent-skills catalog-count tests (follow-up to #6186).
+- **fix(quality):** clear the cycle's 11 net-new ESLint errors and make `validate-release-green` suppressions-aware.
+
+### 📝 Maintenance
+
+- **i18n(it):** add 118 missing Italian (`it`) translations (net-additive — no existing keys dropped, valid JSON), improving Italian UI coverage. ([#6212](https://github.com/diegosouzapw/OmniRoute/pull/6212) — thanks @serverless83)
+- **chore(providers):** remove deprecated **MiMo V2** model entries from the catalogs (xiaomi-mimo, opencode-go, zenmux-free, audio TTS) — the upstream V2 line is superseded by MiMo V2.5; drops `mimo-v2-tts`, `mimo-v2-pro`, `mimo-v2-omni`, `mimo-v2-flash`, `mimo-v2-flash-free` and realigns the provider-catalog tests. ([#6248](https://github.com/diegosouzapw/OmniRoute/pull/6248) — thanks @backryun)
+- **chore(release):** ~50 commits on this branch are v3.8.45 pre-flight/hardening fixes and CI-perf work that landed here via the parallel-cycle sync-back (`sync-next-cycle.mjs`, Hard Rule #21) **after** the `v3.8.45` git tag was cut, and are already fully documented under the **[3.8.45]** section below — listed here only so the per-cycle commit-coverage check (`npm run release:uncovered`) doesn't flag them as gaps. Provider/catalog/UX/backend fixes: #6041, #6078, #6108, #6135, #6148, #6149, #6154, #6158, #6161, #6162, #6163, #6164, #6165, #6170, #6177, #6178, #6181, #6186, #6187, #6191, #6193, #6194, #6195, #6200, #6205, #6208, #6209, #6211, #6213, #6223, #6224, #6225, #6226, #6227, #6228, #6229, #6230, #6235, #6291, #6292. CI/release-pipeline work: #6167, #6203, #6214, #6215, #6218, #6273, #6275, #6283, #6284, #6285, #6300, #6305.
+- **chore(release):** additional zero-ref release-cycle plumbing on this branch, kept out of `release:uncovered` on purpose (no `#N` in the commit subject to cite): opening the v3.8.46 cycle, opening/closing the v3.8.45 cycle, the finalized [3.8.45] CHANGELOG i18n sync-back to 42 mirrors, the v3.8.45 cognitive/cyclomatic and file-size drift rebaselines, ESLint stale-suppression pruning (4,273 → 4,233), and clearing test-masking/docs-all pre-flight reds for v3.8.45.
+
+### ⚡ Performance & Infrastructure
+
+- **perf(release-green):** the pre-flight validator (`scripts/quality/validate-release-green.mjs`) now runs its 4 slow suites (unit / vitest / integration / pack-artifact) **concurrently** via `Promise.all` — pre-flight wall time drops from ~the sum of the suites to ~the slowest one (~30min saved per round; Phase 0 was the nº1 bottleneck of the v3.8.45 release benchmark, 2h54 of 6h34 e2e). Guard: `tests/unit/validate-release-green.test.ts` ("runs the slow suites CONCURRENTLY"). ([#6319](https://github.com/diegosouzapw/OmniRoute/pull/6319))
+- **fix(ci):** `scripts/release/sync-next-cycle.mjs` — two defects found live in its first production run (v3.8.45 Phase 5): (1) the `git()` helper's default 1 MiB `maxBuffer` crashed with `ENOBUFS` on `git show origin/main:CHANGELOG.md` (the CHANGELOG alone is >1 MiB) — widened to 64 MiB; (2) the i18n resync only propagated the `[NEXT]` (TBD) section, leaving the just-shipped finalized section as "— TBD" in all 42 mirrors — it now also syncs `[prevVersion]` bounded by the heading below it (new exported pure helper `versionAfter`). Guards: +5 tests in `tests/unit/sync-next-cycle.test.ts` (8/8). ([#6327](https://github.com/diegosouzapw/OmniRoute/pull/6327))
+- **test(ci):** concurrency-sensitive flaky tests are **quarantined into a serial pass** (`tests/unit/serial/`, `--test-concurrency=1`, appended to every unit runner incl. sharded variants — the serial pass is sharded too so concurrent shard jobs never self-collide). Initial set: `glm-coding-plan-monthly-3580`, `quota-division-blocks`, `provider-health-autopilot`, `combo-health-autopilot` — the class behind the ~28min CI wedges/re-runs (two live 1h+ wedges cancelled during this PR's own validation). Discovery + TIA gates track the new glob; systemic root cause (async logger writing after teardown) tracked in [#6360](https://github.com/diegosouzapw/OmniRoute/issues/6360). Guard: `tests/unit/test-serial-quarantine.test.ts` (4). ([#6347](https://github.com/diegosouzapw/OmniRoute/pull/6347))
+
+### 🙌 Contributors
+
+Thanks to everyone whose work landed in v3.8.46:
+
+| Contributor                                                    | PRs / Issues           |
+| -------------------------------------------------------------- | ---------------------- |
+| [@2220258345](https://github.com/2220258345)                   | direct commit / report |
+| [@abdofallah](https://github.com/abdofallah)                   | direct commit / report |
+| [@adentdk](https://github.com/adentdk)                         | direct commit / report |
+| [@aleksesipenko](https://github.com/aleksesipenko)             | direct commit / report |
+| [@anki1kr](https://github.com/anki1kr)                         | direct commit / report |
+| [@anungma](https://github.com/anungma)                         | direct commit / report |
+| [@arpicato](https://github.com/arpicato)                       | direct commit / report |
+| [@backryun](https://github.com/backryun)                       | #6248                  |
+| [@binsarjr](https://github.com/binsarjr)                       | direct commit / report |
+| [@brightfiscalband](https://github.com/brightfiscalband)       | direct commit / report |
+| [@chirag127](https://github.com/chirag127)                     | #6501, #6506           |
+| [@developerjillur](https://github.com/developerjillur)         | direct commit / report |
+| [@dilneiss](https://github.com/dilneiss)                       | #6499                  |
+| [@dtybnrj](https://github.com/dtybnrj)                         | direct commit / report |
+| [@Forcerecon](https://github.com/Forcerecon)                   | direct commit / report |
+| [@hao3039032](https://github.com/hao3039032)                   | direct commit / report |
+| [@Iammilansoni](https://github.com/Iammilansoni)               | #6150, #6245           |
+| [@jmengit](https://github.com/jmengit)                         | direct commit / report |
+| [@jordansilly77-stack](https://github.com/jordansilly77-stack) | direct commit / report |
+| [@JxnLexn](https://github.com/JxnLexn)                         | direct commit / report |
+| [@KooshaPari](https://github.com/KooshaPari)                   | #6166                  |
+| [@loopyd](https://github.com/loopyd)                           | direct commit / report |
+| [@luweiCN](https://github.com/luweiCN)                         | #6221                  |
+| [@makcimbx](https://github.com/makcimbx)                       | direct commit / report |
+| [@muflifadla38](https://github.com/muflifadla38)               | direct commit / report |
+| [@newnol](https://github.com/newnol)                           | direct commit / report |
+| [@ofekbetzalel](https://github.com/ofekbetzalel)               | direct commit / report |
+| [@ohahe52-dot](https://github.com/ohahe52-dot)                 | direct commit / report |
+| [@phidinhmanh](https://github.com/phidinhmanh)                 | direct commit / report |
+| [@powellnorma](https://github.com/powellnorma)                 | direct commit / report |
+| [@qpeyba](https://github.com/qpeyba)                           | direct commit / report |
+| [@RaviTharuma](https://github.com/RaviTharuma)                 | direct commit / report |
+| [@RCrushMe](https://github.com/RCrushMe)                       | direct commit / report |
+| [@rianonehub](https://github.com/rianonehub)                   | #6134, #6204           |
+| [@serverless83](https://github.com/serverless83)               | #6212                  |
+| [@swingtempo](https://github.com/swingtempo)                   | direct commit / report |
+| [@tenshiak](https://github.com/tenshiak)                       | direct commit / report |
+| [@ThongAccount](https://github.com/ThongAccount)               | direct commit / report |
+| [@UnrealAryan](https://github.com/UnrealAryan)                 | direct commit / report |
+| [@vinayakkulkarni](https://github.com/vinayakkulkarni)         | direct commit / report |
+| [@vittoroliveira-dev](https://github.com/vittoroliveira-dev)   | #6233                  |
+| [@warelik](https://github.com/warelik)                         | direct commit / report |
+| [@xxy9468615](https://github.com/xxy9468615)                   | direct commit / report |
+| [@xz-dev](https://github.com/xz-dev)                           | direct commit / report |
+| [@yanpaing007](https://github.com/yanpaing007)                 | direct commit / report |
+| [@diegosouzapw](https://github.com/diegosouzapw)               | maintainer             |
+
+---
+
 ## [3.8.45] — 2026-07-06
 
 ### ✨ New Features

docs/i18n/hu/CHANGELOG.md

@@ -6,6 +6,213 @@
 
 ## [3.8.31] — 2026-06-20
 
+## [3.8.46] — TBD
+
+### ✨ New Features
+
+- **feat(sse):** **hide paid-only models from `auto/*` routing** when `hidePaidModels` is on ([#6512](https://github.com/diegosouzapw/OmniRoute/issues/6512)) — follow-up to #6328/#6495. PR #6495 hid paid-only models from the `GET /v1/models` listing, but `auto/*` combos (`auto/best-coding`, `auto/glm`, …) could still pick a paid-only backend into their candidate pool → a 402/403 at request time. `createVirtualAutoCombo` now filters the candidate pool through the new pure `open-sse/services/autoCombo/paidModelFilter.ts` (`filterPaidOnlyCandidates`), applying the same free-model predicate #6495 uses in `catalog.ts` (`providerHasFreeModels(provider) && isFreeModel(provider, {id})`) whenever `settings.hidePaidModels === true`. Applied before the category/tier/family narrowing, so it covers every `auto/*` combo; an all-paid pool degrades to the existing graceful empty-pool path. **Opt-in — default OFF leaves the pool unchanged** (identity). Regression guard: `tests/unit/autoCombo/paid-model-filter-6512.test.ts` (4, incl. the default-off identity guard).
+- **feat(sse):** **provider-family auto combos** — `auto/glm`, `auto/minimax`, `auto/mimo`, `auto/zai`, `auto/gemma`, `auto/llama`, `auto/gemini` ([#6453](https://github.com/diegosouzapw/OmniRoute/issues/6453)) — new routable ids that materialize an on-demand virtual combo spanning whatever installed backends currently expose that model family, degrading gracefully as backends rotate. A new pure `open-sse/services/autoCombo/modelFamily.ts` (`detectModelFamily`) classifies by model-id prefix for six families; `zai` is instead resolved by provider id (z.ai's hosted API serves the same `glm-*` model ids as every other GLM backend, so `auto/zai` means "route to my z.ai backend specifically" vs `auto/glm`'s "any connected GLM backend"). Reuses the existing `createVirtualAutoCombo` on-demand materialization path (no DB writes) and the `/v1/models` catalog advertising loop. Regression guard: `tests/unit/autoCombo/provider-family-combos.test.ts` (11).
+- **feat(proxy):** native **proxy-pool round-robin / egress IP rotation** ([#6365](https://github.com/diegosouzapw/OmniRoute/issues/6365)) — a scope (global / provider / account) can now hold **multiple** proxies as a pool with a rotation strategy, so outbound requests cycle their egress IP instead of pinning one proxy per scope. Migration `117_proxy_pool_rotation.sql` lifts the `UNIQUE(scope, scope_id)` constraint (rebuild via the canonical rename/copy/drop; existing single assignments become 1-element pools) and adds a `proxy_scope_rotation` companion table holding the per-scope strategy + a persisted monotonic round-robin cursor. Strategies: `round-robin` (default, monotonic cursor — never `Math.random`), `random`, and `sticky-per-N-min`. Resolution (`resolveProxyForScopeFromRegistry` / `resolveProxyForConnectionFromRegistry`) now fetches the alive, position-ordered candidate set (unchanged `PROXY_ALIVE_PREDICATE`) and applies the strategy; an empty / all-dead pool still returns `null` — the #6246 fail-closed guard is **untouched** (never falls through to direct egress). Backend + DB only; dashboard pool-builder UI is a follow-up. Regression guard: `tests/unit/proxy-pool-rotation-6365.test.ts` (8, incl. fail-closed + backward-compat).
+- **feat(providers):** end-to-end **tool/function calling on the native Gemini `/v1beta` endpoint** ([#6222](https://github.com/diegosouzapw/OmniRoute/issues/6222)) — both directions of the Gemini↔OpenAI conversion now preserve tool data (previously silently dropped). Request side: `convertGeminiToInternal` (extracted to its own testable module) maps `tools[].functionDeclarations` → OpenAI `tools`, prior `functionCall` parts → assistant `tool_calls`, and `functionResponse` parts → `tool`-role messages. Response side: `convertOpenAIResponseToGemini` emits `parts[].functionCall {name,args}` from `message.tool_calls`, and the streaming `openAIChunkToGeminiChunk` accumulates fragmented `tool_calls` deltas by index into complete `functionCall` parts. The non-Gemini client paths (Claude, OpenAI-Responses) already preserved tool calls — this closes the gap specific to the native Gemini surface. Regression guard: `tests/unit/v1beta-gemini-tool-calling-6222.test.ts` (6, incl. a streaming SSE round-trip).
+- **feat(providers):** copilot-m365-web **enterprise / work tier** support ([#6334](https://github.com/diegosouzapw/OmniRoute/issues/6334)) — mirrors the EDU-tier pattern (#6210): `M365ConnectionParams` gains an `agent` field, a new opt-in `M365_ENTERPRISE_OVERRIDES` preset (`agent=work`, `scenario=officeweb`, `licenseType=Premium`) applies via `providerSpecificData.tier="enterprise"` (alias `"work"`), and `agent` is also overridable directly via `providerSpecificData.agent`. `buildWsUrl` was hardcoding `agent="web"` (the one enterprise-distinguishing param with no override path), so a Premium work account handshook then returned an empty stream. The individual and EDU paths are untouched. Kilo's dup flag vs #6210 (EDU tier) was a false positive — different tier. Regression guard: `tests/unit/copilot-m365-enterprise-6334.test.ts` (7). End-to-end confirmation on a real Premium work account is a live-VPS validation follow-up (Hard Rule #18). (thanks @Forcerecon)
+- **feat(api):** standardized, provider-agnostic **`effort` + `thinking` request params** ([#6241](https://github.com/diegosouzapw/OmniRoute/issues/6241)) — a thin standardization layer over the existing mature per-provider reasoning plumbing (no provider mapper touched). `providerChatCompletionSchema` gains a canonical `effort` (reusing the shared `none/low/medium/high/xhigh` vocabulary — the UI tiers `extra`/`max` collapse onto `xhigh`) and a boolean `thinking`. A pure `normalizeReasoningRequest` (wired once in `src/sse/handlers/chat.ts`, before any reasoning field is read) folds them onto the fields the translators already consume (`reasoning_effort` / `reasoning.effort` / `thinking`), so they fan out to Anthropic / Gemini / xAI / Responses — an explicit client `reasoning_effort` / object-shaped `thinking` always wins (backward-compatible). `/models` additively exposes `supportsThinking` + `effort_tiers` so the frontend can render the toggles (UI component is a follow-up). Regression guard: `tests/unit/effort-thinking-standardization-6241.test.ts` (12). (thanks @Iammilansoni, @shabeer)
+- **feat(combo):** new **`pipeline` (sequential) combo strategy** ([#6297](https://github.com/diegosouzapw/OmniRoute/issues/6297)) — the 18th routing strategy runs targets **in order**, threading each step's output into the next step's input, with an optional per-step `prompt` (system instruction); only the final step's response is returned. Distinct from `fusion` (parallel fan-out + judge). Implemented as a self-contained `open-sse/services/pipeline.ts` (sibling to `fusion.ts`), dispatched from `combo.ts`; the step list reuses `combo.models` order and reads an optional `prompt` off each target (backward-compatible — ignored by every other strategy). Intermediate steps run non-streaming with tools stripped (complete prose to thread forward); the final step keeps the client's `stream` flag + tools. A failing/empty/unparseable intermediate step fails the whole pipeline explicitly via a sanitized error (never silently swallowed). Kilo's dup flag vs #563 was a false positive (that's model→chain selection; this is a sequential chain). Regression guard: `tests/unit/combo-pipeline-strategy.test.ts` (5). (thanks @ofekbetzalel)
+- **feat(ci):** `check:test-masking` now flags **inline-reimplemented prod conditions** ([#6348](https://github.com/diegosouzapw/OmniRoute/issues/6348)) — a new **report-only** subcheck (v2, 6A.10 family) catches the wrong-shape contract test: a test that recomputes the condition under test inline instead of importing/exercising the real function (the #6216 class, where `=== 500` → `>= 500` stayed green because the test re-implemented the branch). For each added/modified test file it warns when the file textually duplicates a ≥3-token conditional from a production file touched in the same PR **and** does not import the symbol/module owning it, via a pure, fixture-tested `findReimplementedConditions()` with an allowlist mirroring `assertReductionAllowlist`. Report-only for now (does not fail the gate) — to be promoted to blocking after a triage cycle. Regression guard: `tests/unit/check-test-masking.test.ts` (45).
+- **feat(sse):** per-connection routing override (native vs CLIProxyAPI) ([#6339](https://github.com/diegosouzapw/OmniRoute/issues/6339)) — the previously-dead `isCliproxyapiDeepModeEnabled` helper is now wired into `resolveExecutorWithProxy`: a single connection can opt itself into the CLIProxyAPI passthrough executor via `providerSpecificData.cliproxyapiMode="claude-native"`, with precedence **connection override > provider `upstream_proxy_config` mode > default**. `resolveExecutorWithProxy` now receives the resolved connection's `providerSpecificData` (threaded from `chatCore.ts`), so one connection can deep-route while the provider's other connections stay native — no DB schema change (the toggle rides in `providerSpecificData`). Also resolves the same-provider mixing ask in #6340. Regression guard: `tests/unit/chatcore-executor-proxy.test.ts` (9). (thanks @RaviTharuma)
+- **feat(dashboard):** "Add session cookie" modal now shows a prominent **"Open ‹host› →"** link to the provider's own site ([#6268](https://github.com/diegosouzapw/OmniRoute/issues/6268)) — every `-web` cookie-session provider (chatgpt-web, claude-web, gemini-web, kimi-web, lmarena, qwen-web, m365-copilot-web, …) renders a one-click external link (opening the provider's login/home page in a new tab) so operators no longer tab away to retype the URL mid-setup. The host resolves from a pure, unit-tested `resolveWebProviderHost()` (prefers `WEB_COOKIE_PROVIDERS[id].website`, falls back to the registry `baseUrl` origin); non-web providers render exactly as before. Kilo's dup flag vs #6265 (modal-too-small-on-1080p) was a false positive — distinct concern. Regression guard: `tests/unit/resolve-web-provider-host.test.ts` (5). (thanks @chirag127)
+- **feat(providers):** add **DigitalOcean** AI (serverless inference) as an OpenAI-compatible API-key provider ([#6373](https://github.com/diegosouzapw/OmniRoute/pull/6373)) — base `https://inference.do-ai.run/v1`, wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/digitalocean/`, `src/shared/constants/providers/apikey/inference-hosts.ts`). Regression guard: `tests/unit/digitalocean-provider.test.ts`. (thanks @newnol)
+- **feat(providers):** add **Huancheng Public API** (`hcnsec`) as an OpenAI-compatible regional provider ([#6410](https://github.com/diegosouzapw/OmniRoute/pull/6410)) — Xinjiang Huancheng Cybersecurity's public LLM platform (base `https://api.hcnsec.cn/v1`, free credits via daily check-ins), wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/hcnsec/`, `src/shared/constants/providers/apikey/regional.ts`). Regression guard: `tests/unit/hcnsec-provider.test.ts`. (thanks @UnrealAryan)
+- **feat(dashboard):** the web-session credential guide now shows an **"Open {host}" link** ([#6316](https://github.com/diegosouzapw/OmniRoute/pull/6316)) to the provider's sign-in site (derived from the provider `website` via `getProviderWebsiteHost`), so you can jump straight to the page where the cookie/session must be captured. Regression guard: `tests/unit/web-session-provider-link-6316.test.ts`. (thanks @jordansilly77-stack)
+- **feat(cerebras):** add the **Gemma 4 31B** model (`gemma-4-31b`) to the Cerebras registry + pricing table ([#6331](https://github.com/diegosouzapw/OmniRoute/pull/6331)). Regression guard extends `tests/unit/t28-model-catalog-updates.test.ts`. (thanks @backryun)
+- **feat(providers):** add **Yuanbao (web)** as a cookie-session provider ([#6196](https://github.com/diegosouzapw/OmniRoute/issues/6196)) — `yuanbao-web` (Tencent Yuanbao, `yuanbao.tencent.com`) with cookie-only auth (`hy_user`/`hy_token` + public agent id), SSE→OpenAI translation incl. `reasoning_content`, exposing DeepSeek V3/R1 + Hunyuan / Hunyuan-T1. Regression guard: `tests/unit/providers-yuanbao-web.test.ts`. `together-web` was **deferred** (no verifiable web-session endpoint — needs a captured request) and `huggingchat-web` **dropped** (the existing `huggingchat` already is a web-cookie provider). (thanks @chirag127)
+- **feat(providers):** route the built-in **agentrouter** through the dynamic Claude-Code wire image ([#6056](https://github.com/diegosouzapw/OmniRoute/issues/6056)) — a small static allow-set (`CC_WIRE_IMAGE_BUILTINS` in `open-sse/services/ccWireImageBuiltins.ts`), consulted by `isClaudeCodeCompatible` / `isClaudeCodeCompatibleProvider` / `applyFingerprint`, makes agentrouter adopt the CC wire-image headers + fingerprint **while guarding the CC baseUrl/auth branches** so it keeps its own registry `baseUrl` and `x-api-key` auth. Regression guard: `tests/unit/agentrouter-cc-wire-image.test.ts` (asserts the wire image is applied AND agentrouter's baseUrl/auth are preserved). Live WAF-acceptance against agentrouter.org is a VPS validation follow-up (Hard Rule #18).
+- **feat(providers):** **bulk-add API keys for Cloudflare Workers AI** ([#6174](https://github.com/diegosouzapw/OmniRoute/issues/6174)) — `cloudflare-ai` is removed from the bulk-add exclusion list and the bulk parser gains a 3-field `name|accountId|apiKey` mode; the bulk route now builds a **per-entry** `providerSpecificData` so each key carries its own `accountId` (fixing the previous shared-object reuse), and both the create + key-validation paths receive it. Regression guard: `tests/unit/bulk-api-key-parser-cloudflare.test.ts`. (thanks @muflifadla38)
+- **feat(dashboard):** routing/settings UX clarity ([#6147](https://github.com/diegosouzapw/OmniRoute/issues/6147)) — (1) weighted combos show the **effective routing share %** next to each weight when weights don't sum to 100 (`WeightTotalBar.tsx`); (2) the status widget's user-facing **"Cloud Sync" label is renamed** to "Remote Settings Sync" (`CloudSyncStatus.tsx`; internal ids/state untouched); (3) built-in providers gain an **opt-in advanced base-URL override** (`isBaseUrlOverrideEligibleProvider`, hidden behind an "Advanced" toggle, reusing the existing `providerSpecificData.baseUrl` persistence — not globally widened). Regression guard: `tests/unit/routing-settings-ux-6147.test.ts`.
+- **feat(combo):** add an option to **disable session stickiness**, per-combo or globally — round-robin / random combos can rotate to a different connection on every request instead of pinning a whole conversation to one connection by its first-message hash. Resolution precedence per-combo `config.disableSessionStickiness` → global `settings.disableSessionStickiness` → default `false` (preserves the #3825 prompt-cache/504 fix); gates **both** stickiness call sites in `open-sse/services/combo.ts`. Exposed as a global toggle (Combo Defaults) and a per-combo Inherit/on/off control. ([#6168](https://github.com/diegosouzapw/OmniRoute/issues/6168)) Regression guard: `tests/unit/combo-disable-session-stickiness.test.ts`. (thanks @RCrushMe)
+- **feat(docker):** add the `OMNIROUTE_NO_SUDO` env flag for root-less / user-namespaced deployments — the MITM cert-trust command path (`resolveSudoSpawn` in `src/mitm/systemCommands.ts`) now strips the leading `sudo` when the flag is truthy, in addition to the existing root / sudo-missing cases, so the Proxy Agent runs without `sudo` (the operator trusts the CA manually, e.g. via `NODE_EXTRA_CA_CERTS`). Argv-array `spawn` preserved — no shell interpolation (Hard Rule #13). ([#6122](https://github.com/diegosouzapw/OmniRoute/issues/6122)) Regression guard: `tests/unit/mitm-systemCommands-no-sudo.test.ts`. (thanks @powellnorma)
+- **feat(providers):** add **Requesty** as an OpenAI-compatible gateway provider (BYOK, base `https://router.requesty.ai/v1`, ~200 free requests/day) — wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/requesty/`, `src/shared/constants/providers/apikey/gateways.ts`). ([#6120](https://github.com/diegosouzapw/OmniRoute/issues/6120)) Regression guard: `tests/unit/requesty-provider.test.ts`. (thanks @chirag127)
+- **feat(dashboard):** add **configured-only / available-only filters** to the Free Provider Rankings page ([#6150](https://github.com/diegosouzapw/OmniRoute/issues/6150)) — hide providers you haven't configured, or whose connections are all rate-limited / out of quota, via server-side query params (`?configuredOnly` / `?availableOnly` on `GET /api/free-provider-rankings`) backed by a testable lib helper reusing the in-process connection state (no Redis). Both filters default off, so the default view is unchanged; this supersedes the earlier client-side "Configured Only" toggle (#6245) with an available-only dimension and unit-tested logic. Regression guard: `tests/unit/freeProviderRankings-filters.test.ts`.
+
+### 🔧 Bug Fixes
+
+- **fix(dashboard):** adding a second API key connection for the same provider no longer silently overwrites the first — the Add-API-key modal now derives a unique default connection name (`main`, then `main-2`, `main-3`, …) so the backend name-based upsert can't collide (#6499 — thanks @dilneiss).
+- **fix(compression):** the session-dedup engine now also deduplicates a large multi-line block repeated **within a single message** (intra-message dedup), not just across turns; the compression-preview API surfaces a `fallbackReason`, and the fusion panel reports how many models were rate-limited vs failed on total-panel failure (#6501 — thanks @chirag127).
+- **fix(compression):** a stacked-pipeline step naming an unregistered engine now surfaces a `validationErrors` entry instead of silently no-op'ing, so misconfigured pipelines are visible in the preview API (#6506 — thanks @chirag127).
+- **feat(usage):** add a **Codex reset-credit redemption** flow to the Provider Limits UI ([#6361](https://github.com/diegosouzapw/OmniRoute/pull/6361)) — a `useCodexResetCreditRedemption` hook + `/api/usage/codex-reset-credit` route + `codexResetCredits` lib let you redeem banked Codex reset credits from the quota card. Regression guard: `tests/unit/codex-reset-credits.test.ts`. (thanks @JxnLexn)
+- **feat(glm):** add **team-plan quota settings** for `glm-cn` connections ([#6351](https://github.com/diegosouzapw/OmniRoute/pull/6351)) — a dedicated `GlmTeamQuotaFields` form section (team quota id / limits) threaded through the Add/Edit connection modals, persisted via `providerSpecificData`, with the GLM usage service reading the team quota. Regression guards: `tests/unit/glm-team-quota.test.ts`, `provider-specific-data-schema.test.ts`. (thanks @hao3039032)
+- **feat(providers):** add **TinyFish** web-fetch/search support ([#6349](https://github.com/diegosouzapw/OmniRoute/pull/6349)) — a `tinyfish-fetch` executor + `/v1/web/fetch` route + MCP web-fetch tool, registered as a specialty-media provider with request-validation and a search-provider catalog entry. Regression guards: `tests/unit/executor-tinyfish-fetch.test.ts`, `web-fetch-handler.test.ts`, `mcp-web-fetch-tool.test.ts`, `provider-validation-tinyfish.test.ts`. (thanks @dtybnrj)
+- **fix(cli):** `omniroute launch-codex` now spawns `codex.cmd` through a shell on Windows (the npm `.cmd` shim is unresolvable by bare `spawn` → ENOENT), mirroring the qodercli Windows fix (#6263) ([#6312](https://github.com/diegosouzapw/OmniRoute/pull/6312)). Regression guard: `tests/unit/launch-codex-windows-spawn-6312.test.ts`. (thanks @swingtempo)
+- **fix(codex):** isolate the **Spark** quota from the shared Codex quota and stabilize the quota UI ordering / hydration so per-scope limits render consistently ([#6336](https://github.com/diegosouzapw/OmniRoute/pull/6336)). Regression guards: `tests/unit/codex-quota-selection-hydration.test.ts`, `provider-limits-ui.test.ts` + 3 more. (thanks @xz-dev)
+- **feat(api):** add a `hidePaidModels` setting that filters paid-only models out of the `/v1/models` catalog. Regression guard: `tests/unit/models-catalog-hide-paid.test.ts`. (thanks @chirag127)
+- **fix(api-manager):** the fallback model picker now preserves combos instead of dropping them when a primary model is unavailable ([#6443](https://github.com/diegosouzapw/OmniRoute/pull/6443)). Regression guard: `tests/unit/api-manager-page-static.test.ts`. (thanks @jmengit)
+- **fix(providers):** recoverable Antigravity / Cloud-Code (Gemini Code Assist) `403` responses ([#6452](https://github.com/diegosouzapw/OmniRoute/pull/6452)) are now classified as a retryable project-config error instead of a terminal account ban, so a fixable project/API-disabled 403 no longer forces a ~1-year cooldown / full OAuth reconnect. Regression guard: `tests/unit/errorclassifier-antigravity-403.test.ts`. (thanks @developerjillur)
+- **fix(mitm):** `sanitizeHeaders` now redacts `Set-Cookie` response headers so upstream session cookies never leak into logs / diagnostics ([#6451](https://github.com/diegosouzapw/OmniRoute/pull/6451)). Regression guard: `tests/unit/mitm-sanitize-headers.test.ts`. (thanks @developerjillur)
+- **fix(api):** `/api/compression/preview` now accepts `mode: "caveman"` and correctly handles stacked / zero-compression previews ([#6425](https://github.com/diegosouzapw/OmniRoute/issues/6425)). Regression guard: `tests/unit/api/compression-preview-caveman-and-stacked-6425.test.ts`. (thanks @chirag127)
+- **feat(providers):** add **Zed** hosted LLM aggregator as a native-app provider ([#6118](https://github.com/diegosouzapw/OmniRoute/pull/6118)) — OAuth sign-in via the Zed hosted flow, registered through the shared provider registry + executor. Regression guards: `tests/unit/zed-oauth-provider.test.ts`, `zed-import-utils.test.ts`, `zed-docker-detect.test.ts`, `mitm-handler-zed.test.ts`. VPS-validated via live operator login (Hard Rule #18).
+- **fix(oauth):** the Kiro SSO-cache auto-import now **preserves the IDC region** — cross-region Amazon Q / Kiro profiles imported from the SSO cache are no longer collapsed to the default region ([#6113](https://github.com/diegosouzapw/OmniRoute/pull/6113)). Regression guard: `tests/unit/kiro-auto-import-idc-2059.test.ts`. VPS-validated via live operator login (Hard Rule #18).
+- **fix(dashboard):** passthrough model aliases no longer collide when two namespaced model ids share a last segment (port from 9router#1850, [#6431](https://github.com/diegosouzapw/OmniRoute/pull/6431)). `enx/gpt-5.5` and `enx/codebuddy/gpt-5.5` both auto-generated the alias `gpt-5.5`, so the second model could never be added (the UI just alerted "alias already exists"). Aliases are now disambiguated deterministically — bare last segment when free, then parent-qualified (`codebuddy-gpt-5.5`), then a numeric suffix — while re-adding the exact same model id is still blocked. Regression guard: `tests/unit/passthrough-alias-1850.test.ts`. (thanks @arpicato)
+- **fix(translator):** preserve a Gemini `functionResponse` co-located with other parts (another `functionCall`, or trailing `text`) in the same content when translating **Gemini → OpenAI** ([#6376](https://github.com/diegosouzapw/OmniRoute/pull/6376)). `convertGeminiContent()` early-returned the tool message on the first `functionResponse` part, dropping any co-located parts; such contents are now pre-split (one tool message per `functionResponse`, emitted first, plus one message for the remaining parts). Regression guard: `tests/unit/gemini-to-openai-function-response.test.ts`. (thanks @warelik)
+- **fix(headroom):** detect a python interpreter managed by **mise / pyenv / asdf / conda** (port from 9router#2353, [#6382](https://github.com/diegosouzapw/OmniRoute/pull/6382)). Headroom's python probe (`src/lib/headroom/detect.ts`) searched a hardcoded `PATH`, but version managers expose their interpreters via shim dirs that only join `PATH` through interactive-shell activation — which the non-interactive server never runs, so a managed python (≥3.10) was invisible and Headroom reported it missing. The search path now prepends the well-known shim/bin dirs (`~/.local/share/mise/shims`, `~/.pyenv/shims`, `~/.asdf/shims`, `$CONDA_PREFIX/bin`, `~/.local/bin`, respecting `MISE_DATA_DIR`/`PYENV_ROOT`/`ASDF_DATA_DIR` when set), and a new `HEADROOM_PYTHON` env override lets operators point straight at their interpreter (mirroring `HEADROOM_URL`). Still shell-free (`execFileSync`). Regression guard: `tests/unit/headroom-detect.test.ts` (5). (thanks @loopyd)
+- **fix(executors):** strip the OpenAI-Codex/Claude-CLI `client_metadata` passthrough field for **NVIDIA** requests (port from 9router#1887, [#6411](https://github.com/diegosouzapw/OmniRoute/pull/6411)). NVIDIA's OpenAI-compatible wrapper rejects it with `400 Unsupported parameter`, the same class already handled for `cerebras`/`mistral`; `nvidia` (executor `default`) was missing from the strip allowlist so Codex/Claude-Code passthrough requests 400'd. Regression guard: `tests/unit/executor-default-strip-client-metadata.test.ts` (+nvidia case). (thanks @phidinhmanh)
+- **fix(translator):** strip the Claude-style `thinking` field for **NVIDIA `z-ai/glm-5.2`** (port from 9router#2023, [#6413](https://github.com/diegosouzapw/OmniRoute/pull/6413)). NVIDIA's OpenAI-compatible wrapper 400s on `thinking` (a Claude-format client routed here leaves a `thinking:{type:"adaptive"}`); the existing strip rule only dropped `reasoning`. Same class already handled for `minimax-m2.7`. Regression guard: `tests/unit/nvidia-minimax-thinking-strip.test.ts` (+glm-5.2 case). (thanks @phidinhmanh)
+- **fix(translator):** suppress the streamed `</think>` close marker for the **Antigravity IDE** client (port from 9router#1061, [#6415](https://github.com/diegosouzapw/OmniRoute/pull/6415)). On thinking-only turns Antigravity rendered a bare `</think>` as the sole visible content, tripping its loop-detection and wasting requests. Antigravity's UA (`vscode/<v> (Antigravity/<v>)`) is added to the marker-suppress allowlist (alongside OpenCode); Claude Code / Cursor still get the marker, and `x-omniroute-thinking-marker: on` force-restores it. Regression guard: `tests/unit/think-close-marker-suppress-5245.test.ts`. (thanks @abdofallah)
+- **fix(executors):** strip nested `reasoning_content` from messages for **Mistral** (port from 9router#1649, [#6417](https://github.com/diegosouzapw/OmniRoute/pull/6417)). Mistral's API returns `422 extra_forbidden` when an assistant message carries `reasoning_content` (replayed thinking from a prior turn, e.g. via the Codex `/responses` path); the generic top-level 400 field-downgrade retry never covered the nested per-message field. `DefaultExecutor` now strips it for provider `mistral` only, so DeepSeek (which requires replayed `reasoning_content`) is unaffected. Regression guard: `tests/unit/mistral-strip-reasoning-content-1649.test.ts`. (thanks @xxy9468615)
+- **fix(executors):** strip the `client_metadata` passthrough field on the **OpenCode** path (port from 9router#1442, [#6418](https://github.com/diegosouzapw/OmniRoute/pull/6418)). OpenCode upstreams (e.g. `kimi-k2.6` via opencode-go) reject it with `400 "Extra inputs are not permitted, field: 'client_metadata'"`; the DefaultExecutor strip only covered cerebras/mistral and `OpencodeExecutor` extends `BaseExecutor` directly, so nothing removed it there. Regression guard: `tests/unit/opencode-strip-client-metadata-1442.test.ts`. (thanks @yanpaing007)
+- **fix(executors):** inject the `reasoning_content` echo for the native **Moonshot Kimi** provider (port from 9router#1480, [#6419](https://github.com/diegosouzapw/OmniRoute/pull/6419)). Kimi (executor `default`) is a thinking-mode upstream that 400s with "reasoning_content must be passed back" when a prior assistant turn lacks it; the placeholder injection was only wired into the OpenCode meta-provider, so direct multi-turn Kimi conversations failed. Scoped to `kimi` (gateway-served models matching the thinking-model name pattern are unaffected). Regression guard: `tests/unit/kimi-native-reasoning-injected-1480.test.ts`. (thanks @2220258345)
+- **fix(executors):** recover from a strict gateway's `context_management: Extra inputs are not permitted` 400 (port from 9router#1468, [#6420](https://github.com/diegosouzapw/OmniRoute/pull/6420)). **Claude Code** always sends a top-level `context_management` field; strict anthropic-compatible gateways reject it. The dedicated context-editing 400-fallback only fired when OmniRoute's own `contextEditing` feature was enabled (default off), so a client-sent field passed through untouched and 400'd. `context_management` is now in the generic reactive field-strip list, so it's stripped-and-retried once regardless of the feature flag (with correct request re-signing for claude-compatible relays). Regression guard: `tests/unit/provider-field-strips.test.ts`. (thanks @ohahe52-dot)
+- **fix(network):** enable **RFC 8305 Happy Eyeballs** (`autoSelectFamily`) on the direct-egress undici dispatcher (port from 9router#1237, [#6423](https://github.com/diegosouzapw/OmniRoute/pull/6423)). When DNS returns both IPv6 (AAAA) and IPv4 (A) and the IPv6 route is broken (e.g. a NAT64 `64:ff9b::` prefix without routing), undici tried IPv6 first and hung until `ETIMEDOUT` (then a 502 + account lockout), even though `curl` reached the same host. The direct dispatcher now races both families and uses whichever connects first. Proxy paths pin family via `proxyTls` and are unaffected. Regression guard: `tests/unit/direct-dispatcher-pipelining-4580.test.ts`. (thanks @adentdk)
+- **fix(combo):** round-robin now advances the rotation pointer past the model that **actually served**, not the eagerly-scheduled one (port from 9router#948, [#6428](https://github.com/diegosouzapw/OmniRoute/pull/6428)). With `stickyLimit: 1` (true round-robin), when the scheduled model failed and a _different_ model served via fallback, the counter had already advanced +1 from the scheduled index — so the next request reused the fallback-served model, degrading round-robin into hot-spotting on whichever model was healthy. The pointer now advances to the served index + 1 (mirroring the sticky-limit>1 path). Session-stickiness (#3825) and distribution are preserved. Regression guard: `tests/unit/combo-rr-fallback-advance-948.test.ts`. (thanks @binsarjr)
+- **fix(sse):** a non-string `model` field is now rejected with a `400` before the resolver, instead of crashing downstream `.toLowerCase()`/`.split()` calls into an empty-body `500` that escapes the error sanitizer ([#6407](https://github.com/diegosouzapw/OmniRoute/issues/6407)). Regression guard: `tests/unit/chat-non-string-model-6407.test.ts`. (thanks @chirag127)
+- **fix(api):** unknown `/api/*` routes now return a JSON `404` (instead of the dashboard HTML shell) and scalar chat params (`model`/`temperature`/etc.) are validated **before** the provider lookup so malformed requests fail fast with a clear `400` ([#6424](https://github.com/diegosouzapw/OmniRoute/issues/6424), [#6412](https://github.com/diegosouzapw/OmniRoute/issues/6412)). Regression guards: `tests/unit/api/api-catchall-json-404.test.ts`, `tests/unit/chat-early-schema-validation-6412.test.ts`. (thanks @chirag127)
+- **fix(api):** `/v1/chat/completions` now rejects a non-JSON `Content-Type` with a `400` before parsing the body ([#6414](https://github.com/diegosouzapw/OmniRoute/issues/6414)). Regression guard: `tests/unit/v1-chat-completions-content-type-6414.test.ts`. (thanks @chirag127)
+- **fix(api):** the `X-OmniRoute-Compression` response header is now echoed on `/v1/chat/completions` and `/v1/completions` ([#6422](https://github.com/diegosouzapw/OmniRoute/issues/6422)). Regression guard: `tests/unit/compression-header-echo-6422.test.ts`. (thanks @chirag127)
+- **fix(api):** concurrent `GET /v1/models` requests are coalesced into a single catalog build ([#6408](https://github.com/diegosouzapw/OmniRoute/issues/6408)). Regression guard: `tests/unit/v1-models-concurrent-6408.test.ts`. (thanks @chirag127)
+- **fix(api):** `/v1/completions` now echoes the requested `body.model` in its JSON + streamed responses ([#6429](https://github.com/diegosouzapw/OmniRoute/pull/6429)). Regression guard: `tests/unit/completions-body-model-echo.test.ts`. (thanks @chirag127)
+- **fix(api):** env-var master keys now see the full `/v1/models` catalog ([#6406](https://github.com/diegosouzapw/OmniRoute/issues/6406)). Regression guard: `tests/unit/models-catalog-envkey-6406.test.ts`. (thanks @chirag127)
+- **fix(api):** non-streaming `/v1/completions` responses now echo `body.model` aligned with the `X-OmniRoute-Model` header ([#6426](https://github.com/diegosouzapw/OmniRoute/issues/6426)). Regression guard: `tests/unit/v1-completions-model-header-match-6426.test.ts`. (thanks @chirag127)
+- **fix(api):** unknown `/v1/*` routes now return a JSON `404 not_found` instead of the Next.js dashboard HTML shell ([#6405](https://github.com/diegosouzapw/OmniRoute/issues/6405)). Regression guard: `tests/unit/api/v1-catchall-json-404.test.ts`. (thanks @chirag127)
+- **fix(api):** the per-connection provider models route now degrades to the shipped catalog when a provider's `/models` endpoint answers with a redirect ([#6267](https://github.com/diegosouzapw/OmniRoute/issues/6267)) — a `qwen-web` import failed with a raw `Redirect blocked … (307)` 503. `safeOutboundFetch` throws `REDIRECT_BLOCKED` on the 307, `getSafeOutboundFetchErrorStatus` maps it to 503, and `buildDiscoveryErrorFallbackResponse` treated every 503 as a hard error — so the non-empty `getModelsByProviderId("qwen-web")` catalog was never surfaced. A models-endpoint redirect is not a fixable-config error (unlike `URL_GUARD_BLOCKED`/`INVALID_URL`, which stay hard errors), so it now falls back to the local/cached catalog before the 503 short-circuit. General fix — covers any config-driven provider that 307s. Regression guard: `tests/unit/provider-models-qwen-web-redirect-6267.test.ts`. (thanks @chirag127)
+- **fix(api):** the per-connection provider models route (MCP `list_models_catalog` + the dashboard import view) now merges USER-ADDED custom models into its response ([#6247](https://github.com/diegosouzapw/OmniRoute/issues/6247)) — custom models live in the `key_value` namespace `customModels`, which the live REST `/api/v1/models` already merges, but `src/app/api/providers/[id]/models/route.ts` never read `getCustomModels`, so custom models were dropped on both the discovery-success and local_catalog paths. They are now folded into the returned model list (deduped by id, stamped `owned_by: provider`), fixing MCP + the dashboard import view in one place. Regression guard: `tests/unit/provider-models-custom-merge-6247.test.ts`. (thanks @RCrushMe)
+
+- **fix(providers):** GitLab Duo tool-calling follow-up turns no longer fail upstream with `422 {"detail":"Validation error"}` (tokens 0/0, rejected pre-inference). The #6234 tool-result-feedback fix serialized the **entire** multi-turn conversation into GitLab's single-file `code_suggestions` (`small_file`) generation endpoint — folded history that turn-N sent as an oversized `current_file.content_above_cursor` **and** duplicated verbatim into `user_instruction`, tripping the AI-Gateway's `small_file` validation guard. The executor now **bounds** that prompt: it keeps system + latest user message + the most-recent tool round (dropping older turns), caps oversized tool results, and stops duplicating the full prompt into `user_instruction` (which now carries only the short latest user message) — while still feeding the most-recent tool result back so the agent continues (`open-sse/executors/gitlab.ts`, [#6220](https://github.com/diegosouzapw/OmniRoute/issues/6220)). The unit test covers the bounding logic; the upstream 422→200 clearing is VPS-only (Hard Rule #18). Regression guard: `tests/unit/gitlab-tool-exchange-bounded-6220.test.ts`.
+
+- **fix(i18n):** the provider-detail (`/dashboard/providers/[id]`) connection-status filter labels no longer render as `__MISSING__:All` / `__MISSING__:Active` / `__MISSING__:Error` / `__MISSING__:Banned` / `__MISSING__:CreditsExhausted` in non-English locales (notably pt-BR) ([#6290](https://github.com/diegosouzapw/OmniRoute/issues/6290)). Root cause was **not** the namespace mismatch the issue guessed — the `providers.filter*` keys resolve correctly in `en.json`; the debt lived in the locale mirrors (`src/i18n/messages/*.json`), where these five keys carried the `__MISSING__:` sync sentinel in ~15 locales and were absent entirely in ~26 others, so next-intl found the key and echoed the sentinel verbatim. All 40 non-English/-Chinese mirrors now ship real translations for the five `providers.filter*` labels. Regression guard: `tests/unit/i18n-provider-filter-keys-6290.test.ts`. (thanks @diegosouzapw)
+
+- **fix(providers):** the `copilot-m365-web` streaming executor now emits `debug`-level WebSocket diagnostics ([#6210](https://github.com/diegosouzapw/OmniRoute/issues/6210)) — the outbound WS URL (with the `access_token` **redacted** via `redactWsUrl()`), handshake success/failure, and each received SignalR frame's `type`/`target`. Previously the streaming path logged nothing, so an empty `content:null` response (the M365 Education / Starter tier symptom fixed in #6234) was undiagnosable even at `APP_LOG_LEVEL=debug`. The change is debug-level and side-effect-free — it does not alter streaming behavior or the frame parser, and the token never reaches the logs. Regression guard: `tests/unit/copilot-m365-web-logging-6210.test.ts` (thanks @qpeyba)
+
+- **fix(resilience):** a round-robin combo no longer returns `503 all upstream accounts are unavailable` when a compatibility-rejected target is actually healthy ([#6238](https://github.com/diegosouzapw/OmniRoute/issues/6238)). `filterTargetsByRequestCompatibility` drops request-incompatible targets (tool/vision/structured-output unsupported, or below the required context window) **before** any availability check runs, and its `compatible.length === 0` safety net only fired when _all_ targets were filtered — not when the kept targets later all turned out runtime-unavailable (circuit-open / cooldown / no credentials). So a combo could 503 while a compat-rejected-but-healthy provider sat unused. `handleRoundRobinCombo` now keeps the compat-rejected set and, when every compat-kept target was skipped without a single real attempt, probes those rejected targets as a **last-resort fallback tier** (via the new pure `open-sse/services/combo/comboCompatFallback.ts`) before crystallizing the 503. Regression guard: `tests/unit/combo-roundrobin-compat-fallback-6238.test.ts`. (thanks @ThongAccount)
+
+- **fix(startup):** best-effort self-heal for a corrupted Turbopack dev cache on Windows ([#6289](https://github.com/diegosouzapw/OmniRoute/issues/6289)). On Windows, `pnpm dev` can fail at startup when Turbopack `mmap`s a persistent-cache SST file and the OS refuses the mapping (`os error 1455` — "paging file too small"), which Turbopack surfaces as a misleading `Module not found: Can't resolve '@/shared/utils/machine'`. This is a **known upstream Turbopack cache-corruption bug — not our code**. The dev launcher (`scripts/dev/run-next.mjs`) now wraps `nextApp.prepare()` and, when it rejects with that signature (`isTurbopackCacheCorruption` in the new `scripts/dev/turbopackCacheHeal.mjs`), purges `.build/next/**/cache/turbopack` and retries **once** with a clear log. **Caveat — best-effort only:** the corruption often surfaces as a runtime overlay rather than a `prepare()` rejection, so this cannot always intercept it; the reliable remedy remains manually deleting the Turbopack cache dir. Regression guard: `tests/unit/turbopack-cache-heal-6289.test.ts`. (thanks @chirag127)
+
+- **fix(providers):** qodercli PAT auth no longer fails with `spawn qodercli ENOENT` on Windows ([#6263](https://github.com/diegosouzapw/OmniRoute/issues/6263)) — `spawnQoderCli` spawned the bare `qodercli` name with `shell:false` and an unenriched env, so the npm `.cmd` wrapper under `%APPDATA%\npm` (a user-PATH directory) was never resolved. It now resolves the absolute `.cmd`/`.exe` path through the existing `getCliRuntimeStatus("qoder")` resolver in `src/shared/services/cliRuntime.ts` (memoized), spawns with `shell` when the target is a `.cmd`/`.bat`, and uses the cliRuntime-enriched env (PATH + PATHEXT + APPDATA); the ENOENT error now lists the searched paths plus the `CLI_QODER_BIN` override. End-to-end spawn on a real Windows host is host-only (Hard Rule #18); the path-resolution logic is unit-tested. Regression guard: `tests/unit/qodercli-windows-resolve-6263.test.ts`. (thanks @chirag127)
+
+- **fix(sse):** the reasoning-token buffer no longer inflates **probe-sized `max_tokens`** ([#6274](https://github.com/diegosouzapw/OmniRoute/issues/6274)) — Claude Code's `/model` capability check sends `max_tokens: 1`, but for a thinking-capable model with a large output cap (e.g. `glm-5.2`) the #3587 headroom heuristic (`max(current + 1000, ceil(current * 1.5))`) rewrote it to `1001` and forwarded that upstream, wasting tokens on a request that was never a genuine reasoning budget. `resolveReasoningBufferedMaxTokens()` (`open-sse/services/reasoningTokenBuffer.ts`) now short-circuits and returns the caller's value verbatim when it is below the new `REASONING_BUFFER_MIN_TRIGGER` (256) threshold — a tiny explicit limit is a probe, not a reasoning request. Real budgets still receive the #3587 headroom unchanged, and the guard runs after the existing capability checks so unknown / non-reasoning models keep returning `null`. Regression guard: `tests/unit/reasoning-token-buffer-6274.test.ts`. (thanks @brightfiscalband)
+
+- **fix(cli):** `omniroute reset-password` now works as a real subcommand, and password resets over piped (non-TTY) stdin actually apply ([#6261](https://github.com/diegosouzapw/OmniRoute/issues/6261), [#6258](https://github.com/diegosouzapw/OmniRoute/issues/6258)). Two coupled defects: (1) **#6261** — `bin/omniroute.mjs` routed everything through Commander with only two pre-Commander bypasses (`--mcp`, `reset-encrypted-columns`), so `omniroute reset-password` was rejected as an unknown command; only the separate `omniroute-reset-password` bin worked, while the docs falsely advertised the subcommand (incl. a bogus "legacy alias still works"). A pre-Commander bypass mirroring `reset-encrypted-columns` now dynamically imports `bin/reset-password.mjs` (which self-executes) before Commander parses; the three doc lines were corrected. (2) **#6258** — `bin/reset-password.mjs` issued two sequential `rl.question` prompts; under piped stdin the second read never settled at EOF, so `main()` never reached `resetManagementPassword` and the reset was a silent no-op (both prompts printed, no success, password unchanged). The CLI now detects non-TTY stdin and reads it once (first line = password, second line = confirm if present, else reused), adds a `--password-stdin` flag (entire stdin is the password, no confirmation), and exits `0` explicitly so the success line always flushes; interactive TTY behavior is unchanged. Regression guard: `tests/unit/reset-password-cli-6261-6258.test.ts` (3). (thanks @chirag127)
+
+- **fix(db):** the mass-migration **safety abort** now tells the operator how to bypass it and stops flooding the log ([#6260](https://github.com/diegosouzapw/OmniRoute/issues/6260)) — after restoring a backup that wiped the migration tracking table, `runMigrations()` threw the abort on every downstream `ensureDbInitialized()`, re-logging the full banner 11+ times, and the message never mentioned the existing `OMNIROUTE_MAX_PENDING_MIGRATIONS` escape hatch. The abort text now appends a bypass hint (set `OMNIROUTE_MAX_PENDING_MIGRATIONS=0` in `server.env` / `DATA_DIR/.env`), and a new `MigrationSafetyAbortError` is memoized so repeated calls in the same process throw the same instance and emit a single concise line instead of the full cascade. Regression guard: `tests/unit/migration-safety-abort-6260.test.ts`. (thanks @chirag127)
+
+- **fix(auth):** importing a **distinct** Codex/ChatGPT OAuth `auth.json` is no longer falsely rejected as "already exists" when it belongs to a different user in the same workspace ([#6301](https://github.com/diegosouzapw/OmniRoute/issues/6301)). `findExistingCodexConnection` (in `src/lib/oauth/utils/codexAuthImport.ts`) deduped **only** on `providerSpecificData.workspaceId === accountId`, where `accountId` is the shared `chatgpt_account_id`/`tokens.account_id` — so two members of the same ChatGPT Team collapsed onto a single connection (409 `duplicate_account`). The id_token's `https://api.openai.com/auth` claim carries a per-user `chatgpt_user_id` alongside the workspace id (the device-flow path already persisted it as `chatgptUserId`, but the import path did not). Now `parseAndValidateCodexAuth` extracts `userId` (`chatgpt_user_id` → `user_id` → JWT `sub`) into `ParsedCodexAuth`, the create/update paths persist `chatgptUserId` in `providerSpecificData` (mirroring `codex.ts`), and dedup keys on `workspaceId` **AND** `chatgptUserId` — with a backward-compat fallback to legacy accountId-only matching when no stored connection for that workspace records a `chatgptUserId`, so genuinely-same accounts still dedup. Regression guard: `tests/unit/codex-auth-import-userid-dedup-6301.test.ts` (4). (thanks @anungma)
+
+- **fix(providers):** importing models for the **venice-web** provider no longer fails with a red "Provider venice-web does not support models listing" ([#6269](https://github.com/diegosouzapw/OmniRoute/issues/6269)). `venice-web` is a web-cookie provider with an executor but no upstream `/v1/models` endpoint and no registry `models`, so the models route fell through to the tail `400`. Mirroring the `jules`/`linkup-search`/`ollama-search` fix (#5569), it now ships a static local catalog entry in `src/lib/providers/staticModels.ts` — seeding the current Venice lineup (`venice-uncensored`, `llama-3.3-70b`, `qwen3-235b`, `qwen3-4b`, `deepseek-r1-671b`; Venice rotates its catalog, see docs.venice.ai/models/overview) — so the route returns `200` with `source:"local_catalog"`, `intentional:true`. Regression guard: `tests/unit/static-models-venice-web-6269.test.ts`. (thanks @chirag127)
+
+- **fix(api):** the specialty model catalogs (`/v1/embeddings`, `/v1/images`, `/v1/music`, `/v1/videos` model lists) are now derived from the **unified catalog filtered by a predicate** (`getSpecialtyModelsResponse`) instead of ad-hoc per-route logic, so they consistently respect active-credential visibility and stay in sync with the main catalog ([#6303](https://github.com/diegosouzapw/OmniRoute/pull/6303)). Regression guard: `tests/unit/specialty-model-catalog-routes.test.ts`. (thanks @makcimbx)
+- **fix(api):** the agent-bridge server route now resolves the MITM manager via a **dynamic `import("@/mitm/manager.runtime")`** so Turbopack does not statically pull the stub (or over-bundle the manager), and the agent-skills generator anchors its output base path with `path.join(process.cwd(), …)` so Turbopack's static analyzer stops tracing the whole project root ([#6329](https://github.com/diegosouzapw/OmniRoute/issues/6329), [#6366](https://github.com/diegosouzapw/OmniRoute/pull/6366)). Regression guard: `tests/unit/agent-bridge-server-route-dynamic-import.test.ts`. (thanks @Iammilansoni)
+- **fix(api):** internal probes (combo-test, cloud-sync verify) now pick a **management-scoped / allow-all API key** instead of naively grabbing `getApiKeys()[0]` — a restricted `self:usage` first row made the probe fail with "Model X is not allowed for this API key" even when the combo path was healthy (`pickApiKeyForInternalUse` in `src/lib/db/apiKeys.ts`). The API-manager model editor also falls back to `/api/models?all=true` when `/v1/models` is catalog-protected ([#6372](https://github.com/diegosouzapw/OmniRoute/pull/6372)). Regression guard: `tests/unit/pick-internal-api-key-6372.test.ts`. (thanks @jmengit)
+- **fix(live-ws):** the Live Dashboard WebSocket server now **rejects on bind failure** (e.g. `EADDRINUSE` when the API bridge already holds the port) instead of letting the error surface as an unhandled `error` event that crash-loops the process — the `error` listener is attached to `wss` (not `server`) and releases the EventBus subscription on a failed start ([#6324](https://github.com/diegosouzapw/OmniRoute/issues/6324)). Regression guard: `tests/unit/live-ws-eaddrinuse-6324.test.ts`. (thanks @vinayakkulkarni)
+- **fix(dashboard):** the Home provider-topology widget now trusts the live provider-metrics snapshot — it uses `topology.errorProvider` and live `activeRequests` directly instead of re-deriving state from a stale `lastErrorAt` or applying a frontend timeout filter, so the topology reflects real-time provider health ([#6322](https://github.com/diegosouzapw/OmniRoute/pull/6322)). Regression guard: `tests/unit/home-provider-topology-live-state.test.ts`. (thanks @xz-dev)
+- **fix(sse):** strip zero-width markers from streamed **tool-call arguments** — a follow-up to [#5857](https://github.com/diegosouzapw/OmniRoute/pull/5857). That PR removed injected zero-width joiners (U+200D) from streamed assistant text/reasoning but deliberately left tool-call argument JSON byte-exact. The request-side obfuscation (`open-sse/services/claudeCodeObfuscation.ts`) injects ZWJ into agent words — including the temp path inside the Bash tool description — and Claude models copy that verbatim into generated commands, which are delivered as tool-call arguments rather than assistant text. As a result the ZWJ survived and corrupted code blocks (e.g. a temp path rendered with an invisible joiner). Now `open-sse/handlers/responseSanitizer.ts` strips zero-width code points from tool-call argument strings at every emit site (OpenAI non-stream/stream chat `tool_calls` + legacy `function_call`, native Responses `function_call` items, the OpenAI→Responses conversion, and the native Responses streaming `response.function_call_arguments.delta/.done` events). Only zero-width code points are removed; JSON structure and all other bytes stay identical (no parse/restringify), so normal arguments remain byte-exact. Regression guard: 6 new cases in `tests/unit/response-sanitizer.test.ts` (suite 50/50).
+
+- **fix(nodejs):** the default app log path now resolves under `DATA_DIR` (`~/.omniroute/logs/application/app.log`) instead of `process.cwd()` ([#6197](https://github.com/diegosouzapw/OmniRoute/issues/6197)) — the globally-installed CLI runs from an arbitrary working directory, so anchoring the default to cwd made file logging silently write to (or no-op under) an unrelated directory, contradicting the documented `.env.example` default. `getAppLogFilePath()` now computes the default lazily via the pure `resolveDataDir()` resolver (honours a per-process `DATA_DIR`, no directory-creation side effect); an explicit `APP_LOG_FILE_PATH` still wins. Regression guard: `tests/unit/logenv-datadir-path-6197.test.ts` (3).
+
+- **fix(docker):** AgentBridge/`startMitm` no longer aborts in containers/headless when the Antigravity-default DNS step can't write `/etc/hosts` ([#6127](https://github.com/diegosouzapw/OmniRoute/issues/6127)), and the privileged command's stderr now reaches `app.log` instead of only a bare exit code hitting the toast ([#6198](https://github.com/diegosouzapw/OmniRoute/issues/6198)). The default DNS step (`addDNSEntry`) was called unguarded while cert install and the two sibling DNS steps were each best-effort — in the runtime Docker image (`USER node`, no `sudo`, read-only `/etc/hosts`) it threw `Command failed with code 1` out of `startMitmInternal` and killed the whole start, discarding the stderr. The three DNS steps are extracted into a best-effort `provisionDnsEntries()` where each failure is logged with the full `err` (stderr included, folded in by `systemCommands.ts`) and never aborts the start. Regression guard: `tests/unit/mitm-dns-graceful-degrade-6127.test.ts` (4).
+
+- **fix(providers):** copilot-m365-web now supports the M365 Education "Starter / OfficeWebIncludedCopilot" tier and no longer returns an empty `content:null` stream ([#6210](https://github.com/diegosouzapw/OmniRoute/issues/6210)). Two gaps: (1) `buildWsUrl()` hardcoded the individual-consumer scenario (`OfficeWebPaidConsumerCopilot`, `isEdu=false`) — the EDU tier is now opt-in via `providerSpecificData.tier="edu"`, emitting `scenario=OfficeWebIncludedCopilot`/`isEdu=true` (the individual path is unchanged); (2) the EDU/GPT-5.5 path streams deltas via `arguments[0].writeAtCursor` (incremental) instead of only `messages[].text` (accumulated snapshots), which the parser dropped — a new `accumulateBotContent()` folds both formats, with `type:2 item.result.message` as a last-resort fallback. Regression guard: `tests/unit/copilot-m365-edu-writeatcursor-6210.test.ts` (10). (thanks @qpeyba)
+
+- **fix(providers):** GitLab Duo executor now feeds tool results back into the prompt instead of looping ([#6220](https://github.com/diegosouzapw/OmniRoute/issues/6220)) — `buildPrompt()` branched only on `system`/`user` and took `userParts.at(-1)`, silently dropping the `assistant{tool_calls}` + `tool{result}` turns the client appended, so the reconstructed prompt was byte-identical to turn 1 and the model re-emitted the same `<tool>` call forever. When a tool exchange is present the full conversation is now serialized, folding each tool result back keyed by its `tool_call_id`; simple conversations keep the legacy shape. Complements the tool_call emission from [#6051](https://github.com/diegosouzapw/OmniRoute/issues/6051) (the `kilo-duplicate` label was a false positive — different, sequential defect). Regression guard: `tests/unit/gitlab-tool-result-feedback-6220.test.ts` (4).
+
+- **fix(providers):** opencode-go/opencode-zen can now synthesize the OpenCode CLI identity headers Cloudflare requires on VPS egress ([#5997](https://github.com/diegosouzapw/OmniRoute/issues/5997)) — on a datacenter VPS, `opencode.ai/zen/go/v1/chat/completions` 403s (HTML challenge) requests lacking CLI identity, while the reporter's control curl proved that `User-Agent: opencode-cli/1.0.0` + `x-opencode-client: cli` + `x-opencode-project: default` + fresh request/session UUIDs succeed. Opt-in via `OPENCODE_SYNTHESIZE_CLI_HEADERS=true` (values overridable via `OPENCODE_GO_USER_AGENT`/`OPENCODE_USER_AGENT`/`OPENCODE_CLIENT`/`OPENCODE_PROJECT`); it fills only headers the client did not already send. Kept **off by default** — the forward-only path is deliberate (fabricating a wrong value risks upstream rejection; a prior dedup regressed with `opencode/local`), so this replaces the fragile local header-injection shim without changing default behavior. Regression guard: `tests/unit/opencode-cli-headers-synthesis-5997.test.ts` (6). (thanks @aleksesipenko)
+- fix(resilience): sticky session affinity now evicts and fails over to another account when the pinned account is exhausted/unavailable (#6219)
+
+- fix(sse): Responses API passthrough now drops internal commentary-phase output before forwarding to clients (gated by RESPONSES_PASSTHROUGH_DROP_COMMENTARY, default on) (#6199)
+
+- **fix(sse):** tool-call function schemas with a root `type: null` are now coerced to `type: "object"` before dispatch ([#6359](https://github.com/diegosouzapw/OmniRoute/issues/6359)) — clients like the Codex app emit `parameters: { type: null, ... }` for some tools, which OpenAI-compatible upstreams reject with `400 Invalid schema for function '...': schema must be a JSON Schema of 'type: "object"', got 'type: null'`, failing the whole request. `toolSchemaSanitizer` already stripped the null; it now re-adds the mandatory root `"object"` type (and empty `properties`/open `additionalProperties` when absent). Combinator roots (`anyOf`/`oneOf`/`allOf`) and explicit root types are left untouched. Regression guard: 5 new cases in `tests/unit/tool-schema-sanitizer.test.mjs`.
+
+- **fix(docker):** AgentBridge no longer fails to start on npm/Electron/VPS installs with "MITM manager stub reached at runtime" ([#6344](https://github.com/diegosouzapw/OmniRoute/issues/6344)) — v3.8.45 flipped the production bundler default to Turbopack, but `next.config.mjs` aliased `@/mitm/manager` to its Docker-only degraded stub **unconditionally**. That was harmless while Docker (which sets the alias intentionally for #3390 graceful degradation) was the sole Turbopack consumer, but once every artifact built with Turbopack the stub shipped to all non-Docker users and `startMitm` threw on the first Agent-Bridge start. The alias is now opt-in via `OMNIROUTE_MITM_STUB=1` (set only by the Dockerfile) through the shared `scripts/build/mitm-stub-flag.mjs` helper; default builds bundle the real manager. Regression guard: `tests/unit/mitm-stub-alias-6344.test.mjs` (4).
+
+- **fix(proxy):** stop the v3.8.44 proxy regression that leaked the real IP and disabled healthy proxies ([#6246](https://github.com/diegosouzapw/OmniRoute/issues/6246)). Two coupled defects from the new health scheduler: (1) **IP leak** — when a proxy assigned to a connection was marked `inactive`, resolution fell through to a **direct** egress instead of blocking, exposing the operator's real IP; (2) **over-deactivation** — the sweep flipped a proxy to `inactive` on the **first** failed probe and counted our own 5s timeout / a probe-target `5xx` as the proxy's fault, so healthy paid proxies vanished from egress selection ("my proxies are not being used anymore"). Fix: the sweep decision is extracted into a pure, network-free `decideProxyHealthAction` (`src/lib/proxyHealth/decision.ts`) — by default the health check now **only counts/logs and never downgrades status** (a proxy is downgraded/removed only with `PROXY_AUTO_REMOVE=true`, after `PROXY_AUTO_REMOVE_AFTER` **consecutive** conclusive failures); probes are classified tri-state so an inconclusive result (our timeout, or a `5xx` from the probe target) never penalizes the proxy, and the probe timeout is raised 5s→15s. Separately, `safeResolveProxy` now **fails closed** via the existing policy: a connection whose assigned proxy is dead is blocked instead of leaking direct (`hasBlockingProxyAssignment`), honoring the explicit `proxy off` toggles and the `PROXY_FAIL_OPEN=true` opt-out. Existing proxies stuck `inactive` by the old behavior need a one-time manual re-activate (the operator owns proxy status). Regression guards: `tests/unit/proxy-health-decide-action-6246.test.ts`, `tests/unit/proxy-assigned-unavailable-6246.test.ts`.
+
+- **fix(proxy):** make "Test All" read-only and add bulk enable/disable ([#6246](https://github.com/diegosouzapw/OmniRoute/issues/6246)). Complements the core fail-closed / scheduler fix (#6296) with the two remaining reporter asks. (1) The **"Test All" button** (`POST /api/settings/proxies/auto-test`) used to flip a proxy to `inactive` on a failed reachability probe; since the egress selector excludes `inactive` proxies, a flaky probe (an unreachable `httpbin.org`, a proxy that blocks `HEAD`, or a slow paid proxy) silently disabled every proxy that failed — "Test All" is now **read-only by default** (only the operator sets a proxy active/inactive; opt back into the legacy test-and-set with `PROXY_HEALTH_AUTO_DEACTIVATE=true`). (2) Adds a **bulk enable/disable** proxies endpoint + toolbar action (`POST /api/settings/proxies/batch-activate`) so an operator can re-activate proxies in one click. Regression guard: `tests/unit/proxy-health-6246.test.ts`. (thanks @tenshiak)
+
+- **chatcore (tools): stop the default 128-tool cap from silently dropping opencode's `task`/MCP tools.** opencode (used as an MCP/agent host) sends a large tool list; when it exceeds the speculative `MAX_TOOLS_LIMIT` (128) default, `truncateToolList` did a blind `tools.slice(0, 128)`, dropping every tool past index 128 — including opencode's built-in `task` tool (subagent launch) and many MCP tools, so models routed through OmniRoute could no longer spawn subagents or reach part of their tools. The cap exists to avoid upstream `400`s for providers with real hard limits (e.g. grok-cli 200), so it is kept for those: detection of the opencode client (`isOpencodeClient` — any `x-opencode-*` header, or `opencode` in the user-agent) now only bypasses the **speculative 128 default**, never a known provider ceiling. Precedence is explicit — a proactive/detected provider limit always truncates (even for opencode); otherwise opencode forwards its full tool list; otherwise the unchanged 128 default applies to every other client. Refactors `getEffectiveToolLimit` into `getKnownToolLimit(provider) ?? DEFAULT_LIMIT` (byte-identical for existing callers) and fixes a cosmetic debug-log that reported the truncated count instead of the original. Regression guard: `tests/unit/tool-limit-detector.test.ts`.
+
+- **fix(mitm):** the macOS MITM-cert install check now matches the system keychain again. `security find-certificate -a -Z` prints the SHA-1 as a colon-less hex string, but the installed-check compared it against `getCertFingerprint()`'s colon-separated form, so the substring match never hit — the cert was reported as not-installed and re-prompted for the sudo install on every run. Fingerprints are now normalized (colons stripped, upper-cased) on both sides via the extracted `macCertOutputHasFingerprint` helper. Regression guard: `tests/unit/mitm-cert-mac-fingerprint.test.ts`. ([#6204](https://github.com/diegosouzapw/OmniRoute/pull/6204), closes [#6134](https://github.com/diegosouzapw/OmniRoute/issues/6134) — thanks @rianonehub)
+
+- **fix(api):** `/v1/messages/count_tokens` now counts `tool_use`, `tool_result` and `thinking` content blocks (and array-form `system` prompts) in the local-estimation path, instead of only `text`. Real agentic conversations keep ~95% of their tokens inside tool results; the previous estimate returned near-zero for them, which silently broke Claude Code's auto-compaction (context grew past the window with no compaction until the upstream API rejected the request). The real provider-side count path is unchanged. Regression guard: `tests/unit/messages-count-tokens-route.test.ts`. ([#6221](https://github.com/diegosouzapw/OmniRoute/pull/6221) — thanks @luweiCN)
+
+- **fix(antigravity):** strip a trailing assistant prefill turn for Vertex Claude models to avoid upstream 400s ([#6114](https://github.com/diegosouzapw/OmniRoute/pull/6114)). Regression guard: `tests/unit/antigravity-claude-prefill-strip.test.ts`. (thanks @anki1kr)
+
+- **fix(security):** the mutable cloud-agent routes (`/api/cloud/credentials/update`, `/api/cloud/models/alias`) now require management auth instead of being treated as public. They were classified as public API routes, so a request without management credentials could update stored cloud-agent credentials and model aliases. They are removed from the public-route set, classified as management routes in the authz pipeline, and gated by `requireManagementAuth`; cloud **read**/auth routes stay public. Regression guards: `tests/unit/cloud-write-auth.test.ts`, `tests/unit/authz/classify.test.ts`, `tests/unit/public-api-routes.test.ts`. ([#6233](https://github.com/diegosouzapw/OmniRoute/pull/6233) — thanks @vittoroliveira-dev)
+
+- **refactor(dashboard):** extract the onboarding-wizard "Open provider details" link target into a pure, unit-tested `buildProviderDetailsHref(connection)` helper. The wizard already routes by `connection.id` (the node UUID) rather than the provider category slug (#6144/#6145); this hardens that behavior behind a tested helper that guards a missing id/connection. Regression guard: `tests/unit/provider-onboarding-href.test.ts`. ([#6166](https://github.com/diegosouzapw/OmniRoute/pull/6166) — thanks @KooshaPari)
+
+- **fix(security):** the doubao synthetic device-id generator now derives its digits via an unbiased crypto-random draw (rejection sampling over `crypto.randomBytes()`) instead of a `% 10` reduction, closing a CodeQL `js/biased-cryptographic-random` finding.
+- **fix(agentSkills):** the GitHub-skills generator now resolves `outputDir` to an absolute path before writing, fixing a regression introduced by #6366 (relative-to-cwd base path) that could write generated skill files to the wrong directory.
+- **fix(security):** `/api/keys/{id}/devices` now checks the HTTP method before auth/validation, returning a `405` for non-GET/DELETE verbs instead of a misleading `401`/`500` (closes a `dast-smoke` QUERY-method finding).
+- **fix(quality):** clear the last 2 heavy quality-gate reds on the release tip (cycle pre-flight).
+- **fix(mitm):** the test suite and CI can never mutate the OS trust store — `OMNIROUTE_SKIP_SYSTEM_TRUST=1` is set globally for tests/CI so `installCert`/`uninstallCert`/`installTproxyCa` skip the privileged OS dispatch ([#6310](https://github.com/diegosouzapw/OmniRoute/pull/6310); full detail is under the [3.8.45] section below — this branch received it via the parallel-cycle sync-back).
+- **fix(api):** `POST /api/github-skills` now Zod-validates its request body; documented the new quality-gate env vars and pinned the merge-integrity GitHub Actions to a commit SHA.
+- **fix(skills):** generate the missing `omni-github-skills` registry entry and align the agent-skills catalog-count tests (follow-up to #6186).
+- **fix(quality):** clear the cycle's 11 net-new ESLint errors and make `validate-release-green` suppressions-aware.
+
+### 📝 Maintenance
+
+- **i18n(it):** add 118 missing Italian (`it`) translations (net-additive — no existing keys dropped, valid JSON), improving Italian UI coverage. ([#6212](https://github.com/diegosouzapw/OmniRoute/pull/6212) — thanks @serverless83)
+- **chore(providers):** remove deprecated **MiMo V2** model entries from the catalogs (xiaomi-mimo, opencode-go, zenmux-free, audio TTS) — the upstream V2 line is superseded by MiMo V2.5; drops `mimo-v2-tts`, `mimo-v2-pro`, `mimo-v2-omni`, `mimo-v2-flash`, `mimo-v2-flash-free` and realigns the provider-catalog tests. ([#6248](https://github.com/diegosouzapw/OmniRoute/pull/6248) — thanks @backryun)
+- **chore(release):** ~50 commits on this branch are v3.8.45 pre-flight/hardening fixes and CI-perf work that landed here via the parallel-cycle sync-back (`sync-next-cycle.mjs`, Hard Rule #21) **after** the `v3.8.45` git tag was cut, and are already fully documented under the **[3.8.45]** section below — listed here only so the per-cycle commit-coverage check (`npm run release:uncovered`) doesn't flag them as gaps. Provider/catalog/UX/backend fixes: #6041, #6078, #6108, #6135, #6148, #6149, #6154, #6158, #6161, #6162, #6163, #6164, #6165, #6170, #6177, #6178, #6181, #6186, #6187, #6191, #6193, #6194, #6195, #6200, #6205, #6208, #6209, #6211, #6213, #6223, #6224, #6225, #6226, #6227, #6228, #6229, #6230, #6235, #6291, #6292. CI/release-pipeline work: #6167, #6203, #6214, #6215, #6218, #6273, #6275, #6283, #6284, #6285, #6300, #6305.
+- **chore(release):** additional zero-ref release-cycle plumbing on this branch, kept out of `release:uncovered` on purpose (no `#N` in the commit subject to cite): opening the v3.8.46 cycle, opening/closing the v3.8.45 cycle, the finalized [3.8.45] CHANGELOG i18n sync-back to 42 mirrors, the v3.8.45 cognitive/cyclomatic and file-size drift rebaselines, ESLint stale-suppression pruning (4,273 → 4,233), and clearing test-masking/docs-all pre-flight reds for v3.8.45.
+
+### ⚡ Performance & Infrastructure
+
+- **perf(release-green):** the pre-flight validator (`scripts/quality/validate-release-green.mjs`) now runs its 4 slow suites (unit / vitest / integration / pack-artifact) **concurrently** via `Promise.all` — pre-flight wall time drops from ~the sum of the suites to ~the slowest one (~30min saved per round; Phase 0 was the nº1 bottleneck of the v3.8.45 release benchmark, 2h54 of 6h34 e2e). Guard: `tests/unit/validate-release-green.test.ts` ("runs the slow suites CONCURRENTLY"). ([#6319](https://github.com/diegosouzapw/OmniRoute/pull/6319))
+- **fix(ci):** `scripts/release/sync-next-cycle.mjs` — two defects found live in its first production run (v3.8.45 Phase 5): (1) the `git()` helper's default 1 MiB `maxBuffer` crashed with `ENOBUFS` on `git show origin/main:CHANGELOG.md` (the CHANGELOG alone is >1 MiB) — widened to 64 MiB; (2) the i18n resync only propagated the `[NEXT]` (TBD) section, leaving the just-shipped finalized section as "— TBD" in all 42 mirrors — it now also syncs `[prevVersion]` bounded by the heading below it (new exported pure helper `versionAfter`). Guards: +5 tests in `tests/unit/sync-next-cycle.test.ts` (8/8). ([#6327](https://github.com/diegosouzapw/OmniRoute/pull/6327))
+- **test(ci):** concurrency-sensitive flaky tests are **quarantined into a serial pass** (`tests/unit/serial/`, `--test-concurrency=1`, appended to every unit runner incl. sharded variants — the serial pass is sharded too so concurrent shard jobs never self-collide). Initial set: `glm-coding-plan-monthly-3580`, `quota-division-blocks`, `provider-health-autopilot`, `combo-health-autopilot` — the class behind the ~28min CI wedges/re-runs (two live 1h+ wedges cancelled during this PR's own validation). Discovery + TIA gates track the new glob; systemic root cause (async logger writing after teardown) tracked in [#6360](https://github.com/diegosouzapw/OmniRoute/issues/6360). Guard: `tests/unit/test-serial-quarantine.test.ts` (4). ([#6347](https://github.com/diegosouzapw/OmniRoute/pull/6347))
+
+### 🙌 Contributors
+
+Thanks to everyone whose work landed in v3.8.46:
+
+| Contributor                                                    | PRs / Issues           |
+| -------------------------------------------------------------- | ---------------------- |
+| [@2220258345](https://github.com/2220258345)                   | direct commit / report |
+| [@abdofallah](https://github.com/abdofallah)                   | direct commit / report |
+| [@adentdk](https://github.com/adentdk)                         | direct commit / report |
+| [@aleksesipenko](https://github.com/aleksesipenko)             | direct commit / report |
+| [@anki1kr](https://github.com/anki1kr)                         | direct commit / report |
+| [@anungma](https://github.com/anungma)                         | direct commit / report |
+| [@arpicato](https://github.com/arpicato)                       | direct commit / report |
+| [@backryun](https://github.com/backryun)                       | #6248                  |
+| [@binsarjr](https://github.com/binsarjr)                       | direct commit / report |
+| [@brightfiscalband](https://github.com/brightfiscalband)       | direct commit / report |
+| [@chirag127](https://github.com/chirag127)                     | #6501, #6506           |
+| [@developerjillur](https://github.com/developerjillur)         | direct commit / report |
+| [@dilneiss](https://github.com/dilneiss)                       | #6499                  |
+| [@dtybnrj](https://github.com/dtybnrj)                         | direct commit / report |
+| [@Forcerecon](https://github.com/Forcerecon)                   | direct commit / report |
+| [@hao3039032](https://github.com/hao3039032)                   | direct commit / report |
+| [@Iammilansoni](https://github.com/Iammilansoni)               | #6150, #6245           |
+| [@jmengit](https://github.com/jmengit)                         | direct commit / report |
+| [@jordansilly77-stack](https://github.com/jordansilly77-stack) | direct commit / report |
+| [@JxnLexn](https://github.com/JxnLexn)                         | direct commit / report |
+| [@KooshaPari](https://github.com/KooshaPari)                   | #6166                  |
+| [@loopyd](https://github.com/loopyd)                           | direct commit / report |
+| [@luweiCN](https://github.com/luweiCN)                         | #6221                  |
+| [@makcimbx](https://github.com/makcimbx)                       | direct commit / report |
+| [@muflifadla38](https://github.com/muflifadla38)               | direct commit / report |
+| [@newnol](https://github.com/newnol)                           | direct commit / report |
+| [@ofekbetzalel](https://github.com/ofekbetzalel)               | direct commit / report |
+| [@ohahe52-dot](https://github.com/ohahe52-dot)                 | direct commit / report |
+| [@phidinhmanh](https://github.com/phidinhmanh)                 | direct commit / report |
+| [@powellnorma](https://github.com/powellnorma)                 | direct commit / report |
+| [@qpeyba](https://github.com/qpeyba)                           | direct commit / report |
+| [@RaviTharuma](https://github.com/RaviTharuma)                 | direct commit / report |
+| [@RCrushMe](https://github.com/RCrushMe)                       | direct commit / report |
+| [@rianonehub](https://github.com/rianonehub)                   | #6134, #6204           |
+| [@serverless83](https://github.com/serverless83)               | #6212                  |
+| [@swingtempo](https://github.com/swingtempo)                   | direct commit / report |
+| [@tenshiak](https://github.com/tenshiak)                       | direct commit / report |
+| [@ThongAccount](https://github.com/ThongAccount)               | direct commit / report |
+| [@UnrealAryan](https://github.com/UnrealAryan)                 | direct commit / report |
+| [@vinayakkulkarni](https://github.com/vinayakkulkarni)         | direct commit / report |
+| [@vittoroliveira-dev](https://github.com/vittoroliveira-dev)   | #6233                  |
+| [@warelik](https://github.com/warelik)                         | direct commit / report |
+| [@xxy9468615](https://github.com/xxy9468615)                   | direct commit / report |
+| [@xz-dev](https://github.com/xz-dev)                           | direct commit / report |
+| [@yanpaing007](https://github.com/yanpaing007)                 | direct commit / report |
+| [@diegosouzapw](https://github.com/diegosouzapw)               | maintainer             |
+
+---
+
 ## [3.8.45] — 2026-07-06
 
 ### ✨ New Features

docs/i18n/id/CHANGELOG.md

@@ -6,6 +6,213 @@
 
 ## [3.8.31] — 2026-06-20
 
+## [3.8.46] — TBD
+
+### ✨ New Features
+
+- **feat(sse):** **hide paid-only models from `auto/*` routing** when `hidePaidModels` is on ([#6512](https://github.com/diegosouzapw/OmniRoute/issues/6512)) — follow-up to #6328/#6495. PR #6495 hid paid-only models from the `GET /v1/models` listing, but `auto/*` combos (`auto/best-coding`, `auto/glm`, …) could still pick a paid-only backend into their candidate pool → a 402/403 at request time. `createVirtualAutoCombo` now filters the candidate pool through the new pure `open-sse/services/autoCombo/paidModelFilter.ts` (`filterPaidOnlyCandidates`), applying the same free-model predicate #6495 uses in `catalog.ts` (`providerHasFreeModels(provider) && isFreeModel(provider, {id})`) whenever `settings.hidePaidModels === true`. Applied before the category/tier/family narrowing, so it covers every `auto/*` combo; an all-paid pool degrades to the existing graceful empty-pool path. **Opt-in — default OFF leaves the pool unchanged** (identity). Regression guard: `tests/unit/autoCombo/paid-model-filter-6512.test.ts` (4, incl. the default-off identity guard).
+- **feat(sse):** **provider-family auto combos** — `auto/glm`, `auto/minimax`, `auto/mimo`, `auto/zai`, `auto/gemma`, `auto/llama`, `auto/gemini` ([#6453](https://github.com/diegosouzapw/OmniRoute/issues/6453)) — new routable ids that materialize an on-demand virtual combo spanning whatever installed backends currently expose that model family, degrading gracefully as backends rotate. A new pure `open-sse/services/autoCombo/modelFamily.ts` (`detectModelFamily`) classifies by model-id prefix for six families; `zai` is instead resolved by provider id (z.ai's hosted API serves the same `glm-*` model ids as every other GLM backend, so `auto/zai` means "route to my z.ai backend specifically" vs `auto/glm`'s "any connected GLM backend"). Reuses the existing `createVirtualAutoCombo` on-demand materialization path (no DB writes) and the `/v1/models` catalog advertising loop. Regression guard: `tests/unit/autoCombo/provider-family-combos.test.ts` (11).
+- **feat(proxy):** native **proxy-pool round-robin / egress IP rotation** ([#6365](https://github.com/diegosouzapw/OmniRoute/issues/6365)) — a scope (global / provider / account) can now hold **multiple** proxies as a pool with a rotation strategy, so outbound requests cycle their egress IP instead of pinning one proxy per scope. Migration `117_proxy_pool_rotation.sql` lifts the `UNIQUE(scope, scope_id)` constraint (rebuild via the canonical rename/copy/drop; existing single assignments become 1-element pools) and adds a `proxy_scope_rotation` companion table holding the per-scope strategy + a persisted monotonic round-robin cursor. Strategies: `round-robin` (default, monotonic cursor — never `Math.random`), `random`, and `sticky-per-N-min`. Resolution (`resolveProxyForScopeFromRegistry` / `resolveProxyForConnectionFromRegistry`) now fetches the alive, position-ordered candidate set (unchanged `PROXY_ALIVE_PREDICATE`) and applies the strategy; an empty / all-dead pool still returns `null` — the #6246 fail-closed guard is **untouched** (never falls through to direct egress). Backend + DB only; dashboard pool-builder UI is a follow-up. Regression guard: `tests/unit/proxy-pool-rotation-6365.test.ts` (8, incl. fail-closed + backward-compat).
+- **feat(providers):** end-to-end **tool/function calling on the native Gemini `/v1beta` endpoint** ([#6222](https://github.com/diegosouzapw/OmniRoute/issues/6222)) — both directions of the Gemini↔OpenAI conversion now preserve tool data (previously silently dropped). Request side: `convertGeminiToInternal` (extracted to its own testable module) maps `tools[].functionDeclarations` → OpenAI `tools`, prior `functionCall` parts → assistant `tool_calls`, and `functionResponse` parts → `tool`-role messages. Response side: `convertOpenAIResponseToGemini` emits `parts[].functionCall {name,args}` from `message.tool_calls`, and the streaming `openAIChunkToGeminiChunk` accumulates fragmented `tool_calls` deltas by index into complete `functionCall` parts. The non-Gemini client paths (Claude, OpenAI-Responses) already preserved tool calls — this closes the gap specific to the native Gemini surface. Regression guard: `tests/unit/v1beta-gemini-tool-calling-6222.test.ts` (6, incl. a streaming SSE round-trip).
+- **feat(providers):** copilot-m365-web **enterprise / work tier** support ([#6334](https://github.com/diegosouzapw/OmniRoute/issues/6334)) — mirrors the EDU-tier pattern (#6210): `M365ConnectionParams` gains an `agent` field, a new opt-in `M365_ENTERPRISE_OVERRIDES` preset (`agent=work`, `scenario=officeweb`, `licenseType=Premium`) applies via `providerSpecificData.tier="enterprise"` (alias `"work"`), and `agent` is also overridable directly via `providerSpecificData.agent`. `buildWsUrl` was hardcoding `agent="web"` (the one enterprise-distinguishing param with no override path), so a Premium work account handshook then returned an empty stream. The individual and EDU paths are untouched. Kilo's dup flag vs #6210 (EDU tier) was a false positive — different tier. Regression guard: `tests/unit/copilot-m365-enterprise-6334.test.ts` (7). End-to-end confirmation on a real Premium work account is a live-VPS validation follow-up (Hard Rule #18). (thanks @Forcerecon)
+- **feat(api):** standardized, provider-agnostic **`effort` + `thinking` request params** ([#6241](https://github.com/diegosouzapw/OmniRoute/issues/6241)) — a thin standardization layer over the existing mature per-provider reasoning plumbing (no provider mapper touched). `providerChatCompletionSchema` gains a canonical `effort` (reusing the shared `none/low/medium/high/xhigh` vocabulary — the UI tiers `extra`/`max` collapse onto `xhigh`) and a boolean `thinking`. A pure `normalizeReasoningRequest` (wired once in `src/sse/handlers/chat.ts`, before any reasoning field is read) folds them onto the fields the translators already consume (`reasoning_effort` / `reasoning.effort` / `thinking`), so they fan out to Anthropic / Gemini / xAI / Responses — an explicit client `reasoning_effort` / object-shaped `thinking` always wins (backward-compatible). `/models` additively exposes `supportsThinking` + `effort_tiers` so the frontend can render the toggles (UI component is a follow-up). Regression guard: `tests/unit/effort-thinking-standardization-6241.test.ts` (12). (thanks @Iammilansoni, @shabeer)
+- **feat(combo):** new **`pipeline` (sequential) combo strategy** ([#6297](https://github.com/diegosouzapw/OmniRoute/issues/6297)) — the 18th routing strategy runs targets **in order**, threading each step's output into the next step's input, with an optional per-step `prompt` (system instruction); only the final step's response is returned. Distinct from `fusion` (parallel fan-out + judge). Implemented as a self-contained `open-sse/services/pipeline.ts` (sibling to `fusion.ts`), dispatched from `combo.ts`; the step list reuses `combo.models` order and reads an optional `prompt` off each target (backward-compatible — ignored by every other strategy). Intermediate steps run non-streaming with tools stripped (complete prose to thread forward); the final step keeps the client's `stream` flag + tools. A failing/empty/unparseable intermediate step fails the whole pipeline explicitly via a sanitized error (never silently swallowed). Kilo's dup flag vs #563 was a false positive (that's model→chain selection; this is a sequential chain). Regression guard: `tests/unit/combo-pipeline-strategy.test.ts` (5). (thanks @ofekbetzalel)
+- **feat(ci):** `check:test-masking` now flags **inline-reimplemented prod conditions** ([#6348](https://github.com/diegosouzapw/OmniRoute/issues/6348)) — a new **report-only** subcheck (v2, 6A.10 family) catches the wrong-shape contract test: a test that recomputes the condition under test inline instead of importing/exercising the real function (the #6216 class, where `=== 500` → `>= 500` stayed green because the test re-implemented the branch). For each added/modified test file it warns when the file textually duplicates a ≥3-token conditional from a production file touched in the same PR **and** does not import the symbol/module owning it, via a pure, fixture-tested `findReimplementedConditions()` with an allowlist mirroring `assertReductionAllowlist`. Report-only for now (does not fail the gate) — to be promoted to blocking after a triage cycle. Regression guard: `tests/unit/check-test-masking.test.ts` (45).
+- **feat(sse):** per-connection routing override (native vs CLIProxyAPI) ([#6339](https://github.com/diegosouzapw/OmniRoute/issues/6339)) — the previously-dead `isCliproxyapiDeepModeEnabled` helper is now wired into `resolveExecutorWithProxy`: a single connection can opt itself into the CLIProxyAPI passthrough executor via `providerSpecificData.cliproxyapiMode="claude-native"`, with precedence **connection override > provider `upstream_proxy_config` mode > default**. `resolveExecutorWithProxy` now receives the resolved connection's `providerSpecificData` (threaded from `chatCore.ts`), so one connection can deep-route while the provider's other connections stay native — no DB schema change (the toggle rides in `providerSpecificData`). Also resolves the same-provider mixing ask in #6340. Regression guard: `tests/unit/chatcore-executor-proxy.test.ts` (9). (thanks @RaviTharuma)
+- **feat(dashboard):** "Add session cookie" modal now shows a prominent **"Open ‹host› →"** link to the provider's own site ([#6268](https://github.com/diegosouzapw/OmniRoute/issues/6268)) — every `-web` cookie-session provider (chatgpt-web, claude-web, gemini-web, kimi-web, lmarena, qwen-web, m365-copilot-web, …) renders a one-click external link (opening the provider's login/home page in a new tab) so operators no longer tab away to retype the URL mid-setup. The host resolves from a pure, unit-tested `resolveWebProviderHost()` (prefers `WEB_COOKIE_PROVIDERS[id].website`, falls back to the registry `baseUrl` origin); non-web providers render exactly as before. Kilo's dup flag vs #6265 (modal-too-small-on-1080p) was a false positive — distinct concern. Regression guard: `tests/unit/resolve-web-provider-host.test.ts` (5). (thanks @chirag127)
+- **feat(providers):** add **DigitalOcean** AI (serverless inference) as an OpenAI-compatible API-key provider ([#6373](https://github.com/diegosouzapw/OmniRoute/pull/6373)) — base `https://inference.do-ai.run/v1`, wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/digitalocean/`, `src/shared/constants/providers/apikey/inference-hosts.ts`). Regression guard: `tests/unit/digitalocean-provider.test.ts`. (thanks @newnol)
+- **feat(providers):** add **Huancheng Public API** (`hcnsec`) as an OpenAI-compatible regional provider ([#6410](https://github.com/diegosouzapw/OmniRoute/pull/6410)) — Xinjiang Huancheng Cybersecurity's public LLM platform (base `https://api.hcnsec.cn/v1`, free credits via daily check-ins), wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/hcnsec/`, `src/shared/constants/providers/apikey/regional.ts`). Regression guard: `tests/unit/hcnsec-provider.test.ts`. (thanks @UnrealAryan)
+- **feat(dashboard):** the web-session credential guide now shows an **"Open {host}" link** ([#6316](https://github.com/diegosouzapw/OmniRoute/pull/6316)) to the provider's sign-in site (derived from the provider `website` via `getProviderWebsiteHost`), so you can jump straight to the page where the cookie/session must be captured. Regression guard: `tests/unit/web-session-provider-link-6316.test.ts`. (thanks @jordansilly77-stack)
+- **feat(cerebras):** add the **Gemma 4 31B** model (`gemma-4-31b`) to the Cerebras registry + pricing table ([#6331](https://github.com/diegosouzapw/OmniRoute/pull/6331)). Regression guard extends `tests/unit/t28-model-catalog-updates.test.ts`. (thanks @backryun)
+- **feat(providers):** add **Yuanbao (web)** as a cookie-session provider ([#6196](https://github.com/diegosouzapw/OmniRoute/issues/6196)) — `yuanbao-web` (Tencent Yuanbao, `yuanbao.tencent.com`) with cookie-only auth (`hy_user`/`hy_token` + public agent id), SSE→OpenAI translation incl. `reasoning_content`, exposing DeepSeek V3/R1 + Hunyuan / Hunyuan-T1. Regression guard: `tests/unit/providers-yuanbao-web.test.ts`. `together-web` was **deferred** (no verifiable web-session endpoint — needs a captured request) and `huggingchat-web` **dropped** (the existing `huggingchat` already is a web-cookie provider). (thanks @chirag127)
+- **feat(providers):** route the built-in **agentrouter** through the dynamic Claude-Code wire image ([#6056](https://github.com/diegosouzapw/OmniRoute/issues/6056)) — a small static allow-set (`CC_WIRE_IMAGE_BUILTINS` in `open-sse/services/ccWireImageBuiltins.ts`), consulted by `isClaudeCodeCompatible` / `isClaudeCodeCompatibleProvider` / `applyFingerprint`, makes agentrouter adopt the CC wire-image headers + fingerprint **while guarding the CC baseUrl/auth branches** so it keeps its own registry `baseUrl` and `x-api-key` auth. Regression guard: `tests/unit/agentrouter-cc-wire-image.test.ts` (asserts the wire image is applied AND agentrouter's baseUrl/auth are preserved). Live WAF-acceptance against agentrouter.org is a VPS validation follow-up (Hard Rule #18).
+- **feat(providers):** **bulk-add API keys for Cloudflare Workers AI** ([#6174](https://github.com/diegosouzapw/OmniRoute/issues/6174)) — `cloudflare-ai` is removed from the bulk-add exclusion list and the bulk parser gains a 3-field `name|accountId|apiKey` mode; the bulk route now builds a **per-entry** `providerSpecificData` so each key carries its own `accountId` (fixing the previous shared-object reuse), and both the create + key-validation paths receive it. Regression guard: `tests/unit/bulk-api-key-parser-cloudflare.test.ts`. (thanks @muflifadla38)
+- **feat(dashboard):** routing/settings UX clarity ([#6147](https://github.com/diegosouzapw/OmniRoute/issues/6147)) — (1) weighted combos show the **effective routing share %** next to each weight when weights don't sum to 100 (`WeightTotalBar.tsx`); (2) the status widget's user-facing **"Cloud Sync" label is renamed** to "Remote Settings Sync" (`CloudSyncStatus.tsx`; internal ids/state untouched); (3) built-in providers gain an **opt-in advanced base-URL override** (`isBaseUrlOverrideEligibleProvider`, hidden behind an "Advanced" toggle, reusing the existing `providerSpecificData.baseUrl` persistence — not globally widened). Regression guard: `tests/unit/routing-settings-ux-6147.test.ts`.
+- **feat(combo):** add an option to **disable session stickiness**, per-combo or globally — round-robin / random combos can rotate to a different connection on every request instead of pinning a whole conversation to one connection by its first-message hash. Resolution precedence per-combo `config.disableSessionStickiness` → global `settings.disableSessionStickiness` → default `false` (preserves the #3825 prompt-cache/504 fix); gates **both** stickiness call sites in `open-sse/services/combo.ts`. Exposed as a global toggle (Combo Defaults) and a per-combo Inherit/on/off control. ([#6168](https://github.com/diegosouzapw/OmniRoute/issues/6168)) Regression guard: `tests/unit/combo-disable-session-stickiness.test.ts`. (thanks @RCrushMe)
+- **feat(docker):** add the `OMNIROUTE_NO_SUDO` env flag for root-less / user-namespaced deployments — the MITM cert-trust command path (`resolveSudoSpawn` in `src/mitm/systemCommands.ts`) now strips the leading `sudo` when the flag is truthy, in addition to the existing root / sudo-missing cases, so the Proxy Agent runs without `sudo` (the operator trusts the CA manually, e.g. via `NODE_EXTRA_CA_CERTS`). Argv-array `spawn` preserved — no shell interpolation (Hard Rule #13). ([#6122](https://github.com/diegosouzapw/OmniRoute/issues/6122)) Regression guard: `tests/unit/mitm-systemCommands-no-sudo.test.ts`. (thanks @powellnorma)
+- **feat(providers):** add **Requesty** as an OpenAI-compatible gateway provider (BYOK, base `https://router.requesty.ai/v1`, ~200 free requests/day) — wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/requesty/`, `src/shared/constants/providers/apikey/gateways.ts`). ([#6120](https://github.com/diegosouzapw/OmniRoute/issues/6120)) Regression guard: `tests/unit/requesty-provider.test.ts`. (thanks @chirag127)
+- **feat(dashboard):** add **configured-only / available-only filters** to the Free Provider Rankings page ([#6150](https://github.com/diegosouzapw/OmniRoute/issues/6150)) — hide providers you haven't configured, or whose connections are all rate-limited / out of quota, via server-side query params (`?configuredOnly` / `?availableOnly` on `GET /api/free-provider-rankings`) backed by a testable lib helper reusing the in-process connection state (no Redis). Both filters default off, so the default view is unchanged; this supersedes the earlier client-side "Configured Only" toggle (#6245) with an available-only dimension and unit-tested logic. Regression guard: `tests/unit/freeProviderRankings-filters.test.ts`.
+
+### 🔧 Bug Fixes
+
+- **fix(dashboard):** adding a second API key connection for the same provider no longer silently overwrites the first — the Add-API-key modal now derives a unique default connection name (`main`, then `main-2`, `main-3`, …) so the backend name-based upsert can't collide (#6499 — thanks @dilneiss).
+- **fix(compression):** the session-dedup engine now also deduplicates a large multi-line block repeated **within a single message** (intra-message dedup), not just across turns; the compression-preview API surfaces a `fallbackReason`, and the fusion panel reports how many models were rate-limited vs failed on total-panel failure (#6501 — thanks @chirag127).
+- **fix(compression):** a stacked-pipeline step naming an unregistered engine now surfaces a `validationErrors` entry instead of silently no-op'ing, so misconfigured pipelines are visible in the preview API (#6506 — thanks @chirag127).
+- **feat(usage):** add a **Codex reset-credit redemption** flow to the Provider Limits UI ([#6361](https://github.com/diegosouzapw/OmniRoute/pull/6361)) — a `useCodexResetCreditRedemption` hook + `/api/usage/codex-reset-credit` route + `codexResetCredits` lib let you redeem banked Codex reset credits from the quota card. Regression guard: `tests/unit/codex-reset-credits.test.ts`. (thanks @JxnLexn)
+- **feat(glm):** add **team-plan quota settings** for `glm-cn` connections ([#6351](https://github.com/diegosouzapw/OmniRoute/pull/6351)) — a dedicated `GlmTeamQuotaFields` form section (team quota id / limits) threaded through the Add/Edit connection modals, persisted via `providerSpecificData`, with the GLM usage service reading the team quota. Regression guards: `tests/unit/glm-team-quota.test.ts`, `provider-specific-data-schema.test.ts`. (thanks @hao3039032)
+- **feat(providers):** add **TinyFish** web-fetch/search support ([#6349](https://github.com/diegosouzapw/OmniRoute/pull/6349)) — a `tinyfish-fetch` executor + `/v1/web/fetch` route + MCP web-fetch tool, registered as a specialty-media provider with request-validation and a search-provider catalog entry. Regression guards: `tests/unit/executor-tinyfish-fetch.test.ts`, `web-fetch-handler.test.ts`, `mcp-web-fetch-tool.test.ts`, `provider-validation-tinyfish.test.ts`. (thanks @dtybnrj)
+- **fix(cli):** `omniroute launch-codex` now spawns `codex.cmd` through a shell on Windows (the npm `.cmd` shim is unresolvable by bare `spawn` → ENOENT), mirroring the qodercli Windows fix (#6263) ([#6312](https://github.com/diegosouzapw/OmniRoute/pull/6312)). Regression guard: `tests/unit/launch-codex-windows-spawn-6312.test.ts`. (thanks @swingtempo)
+- **fix(codex):** isolate the **Spark** quota from the shared Codex quota and stabilize the quota UI ordering / hydration so per-scope limits render consistently ([#6336](https://github.com/diegosouzapw/OmniRoute/pull/6336)). Regression guards: `tests/unit/codex-quota-selection-hydration.test.ts`, `provider-limits-ui.test.ts` + 3 more. (thanks @xz-dev)
+- **feat(api):** add a `hidePaidModels` setting that filters paid-only models out of the `/v1/models` catalog. Regression guard: `tests/unit/models-catalog-hide-paid.test.ts`. (thanks @chirag127)
+- **fix(api-manager):** the fallback model picker now preserves combos instead of dropping them when a primary model is unavailable ([#6443](https://github.com/diegosouzapw/OmniRoute/pull/6443)). Regression guard: `tests/unit/api-manager-page-static.test.ts`. (thanks @jmengit)
+- **fix(providers):** recoverable Antigravity / Cloud-Code (Gemini Code Assist) `403` responses ([#6452](https://github.com/diegosouzapw/OmniRoute/pull/6452)) are now classified as a retryable project-config error instead of a terminal account ban, so a fixable project/API-disabled 403 no longer forces a ~1-year cooldown / full OAuth reconnect. Regression guard: `tests/unit/errorclassifier-antigravity-403.test.ts`. (thanks @developerjillur)
+- **fix(mitm):** `sanitizeHeaders` now redacts `Set-Cookie` response headers so upstream session cookies never leak into logs / diagnostics ([#6451](https://github.com/diegosouzapw/OmniRoute/pull/6451)). Regression guard: `tests/unit/mitm-sanitize-headers.test.ts`. (thanks @developerjillur)
+- **fix(api):** `/api/compression/preview` now accepts `mode: "caveman"` and correctly handles stacked / zero-compression previews ([#6425](https://github.com/diegosouzapw/OmniRoute/issues/6425)). Regression guard: `tests/unit/api/compression-preview-caveman-and-stacked-6425.test.ts`. (thanks @chirag127)
+- **feat(providers):** add **Zed** hosted LLM aggregator as a native-app provider ([#6118](https://github.com/diegosouzapw/OmniRoute/pull/6118)) — OAuth sign-in via the Zed hosted flow, registered through the shared provider registry + executor. Regression guards: `tests/unit/zed-oauth-provider.test.ts`, `zed-import-utils.test.ts`, `zed-docker-detect.test.ts`, `mitm-handler-zed.test.ts`. VPS-validated via live operator login (Hard Rule #18).
+- **fix(oauth):** the Kiro SSO-cache auto-import now **preserves the IDC region** — cross-region Amazon Q / Kiro profiles imported from the SSO cache are no longer collapsed to the default region ([#6113](https://github.com/diegosouzapw/OmniRoute/pull/6113)). Regression guard: `tests/unit/kiro-auto-import-idc-2059.test.ts`. VPS-validated via live operator login (Hard Rule #18).
+- **fix(dashboard):** passthrough model aliases no longer collide when two namespaced model ids share a last segment (port from 9router#1850, [#6431](https://github.com/diegosouzapw/OmniRoute/pull/6431)). `enx/gpt-5.5` and `enx/codebuddy/gpt-5.5` both auto-generated the alias `gpt-5.5`, so the second model could never be added (the UI just alerted "alias already exists"). Aliases are now disambiguated deterministically — bare last segment when free, then parent-qualified (`codebuddy-gpt-5.5`), then a numeric suffix — while re-adding the exact same model id is still blocked. Regression guard: `tests/unit/passthrough-alias-1850.test.ts`. (thanks @arpicato)
+- **fix(translator):** preserve a Gemini `functionResponse` co-located with other parts (another `functionCall`, or trailing `text`) in the same content when translating **Gemini → OpenAI** ([#6376](https://github.com/diegosouzapw/OmniRoute/pull/6376)). `convertGeminiContent()` early-returned the tool message on the first `functionResponse` part, dropping any co-located parts; such contents are now pre-split (one tool message per `functionResponse`, emitted first, plus one message for the remaining parts). Regression guard: `tests/unit/gemini-to-openai-function-response.test.ts`. (thanks @warelik)
+- **fix(headroom):** detect a python interpreter managed by **mise / pyenv / asdf / conda** (port from 9router#2353, [#6382](https://github.com/diegosouzapw/OmniRoute/pull/6382)). Headroom's python probe (`src/lib/headroom/detect.ts`) searched a hardcoded `PATH`, but version managers expose their interpreters via shim dirs that only join `PATH` through interactive-shell activation — which the non-interactive server never runs, so a managed python (≥3.10) was invisible and Headroom reported it missing. The search path now prepends the well-known shim/bin dirs (`~/.local/share/mise/shims`, `~/.pyenv/shims`, `~/.asdf/shims`, `$CONDA_PREFIX/bin`, `~/.local/bin`, respecting `MISE_DATA_DIR`/`PYENV_ROOT`/`ASDF_DATA_DIR` when set), and a new `HEADROOM_PYTHON` env override lets operators point straight at their interpreter (mirroring `HEADROOM_URL`). Still shell-free (`execFileSync`). Regression guard: `tests/unit/headroom-detect.test.ts` (5). (thanks @loopyd)
+- **fix(executors):** strip the OpenAI-Codex/Claude-CLI `client_metadata` passthrough field for **NVIDIA** requests (port from 9router#1887, [#6411](https://github.com/diegosouzapw/OmniRoute/pull/6411)). NVIDIA's OpenAI-compatible wrapper rejects it with `400 Unsupported parameter`, the same class already handled for `cerebras`/`mistral`; `nvidia` (executor `default`) was missing from the strip allowlist so Codex/Claude-Code passthrough requests 400'd. Regression guard: `tests/unit/executor-default-strip-client-metadata.test.ts` (+nvidia case). (thanks @phidinhmanh)
+- **fix(translator):** strip the Claude-style `thinking` field for **NVIDIA `z-ai/glm-5.2`** (port from 9router#2023, [#6413](https://github.com/diegosouzapw/OmniRoute/pull/6413)). NVIDIA's OpenAI-compatible wrapper 400s on `thinking` (a Claude-format client routed here leaves a `thinking:{type:"adaptive"}`); the existing strip rule only dropped `reasoning`. Same class already handled for `minimax-m2.7`. Regression guard: `tests/unit/nvidia-minimax-thinking-strip.test.ts` (+glm-5.2 case). (thanks @phidinhmanh)
+- **fix(translator):** suppress the streamed `</think>` close marker for the **Antigravity IDE** client (port from 9router#1061, [#6415](https://github.com/diegosouzapw/OmniRoute/pull/6415)). On thinking-only turns Antigravity rendered a bare `</think>` as the sole visible content, tripping its loop-detection and wasting requests. Antigravity's UA (`vscode/<v> (Antigravity/<v>)`) is added to the marker-suppress allowlist (alongside OpenCode); Claude Code / Cursor still get the marker, and `x-omniroute-thinking-marker: on` force-restores it. Regression guard: `tests/unit/think-close-marker-suppress-5245.test.ts`. (thanks @abdofallah)
+- **fix(executors):** strip nested `reasoning_content` from messages for **Mistral** (port from 9router#1649, [#6417](https://github.com/diegosouzapw/OmniRoute/pull/6417)). Mistral's API returns `422 extra_forbidden` when an assistant message carries `reasoning_content` (replayed thinking from a prior turn, e.g. via the Codex `/responses` path); the generic top-level 400 field-downgrade retry never covered the nested per-message field. `DefaultExecutor` now strips it for provider `mistral` only, so DeepSeek (which requires replayed `reasoning_content`) is unaffected. Regression guard: `tests/unit/mistral-strip-reasoning-content-1649.test.ts`. (thanks @xxy9468615)
+- **fix(executors):** strip the `client_metadata` passthrough field on the **OpenCode** path (port from 9router#1442, [#6418](https://github.com/diegosouzapw/OmniRoute/pull/6418)). OpenCode upstreams (e.g. `kimi-k2.6` via opencode-go) reject it with `400 "Extra inputs are not permitted, field: 'client_metadata'"`; the DefaultExecutor strip only covered cerebras/mistral and `OpencodeExecutor` extends `BaseExecutor` directly, so nothing removed it there. Regression guard: `tests/unit/opencode-strip-client-metadata-1442.test.ts`. (thanks @yanpaing007)
+- **fix(executors):** inject the `reasoning_content` echo for the native **Moonshot Kimi** provider (port from 9router#1480, [#6419](https://github.com/diegosouzapw/OmniRoute/pull/6419)). Kimi (executor `default`) is a thinking-mode upstream that 400s with "reasoning_content must be passed back" when a prior assistant turn lacks it; the placeholder injection was only wired into the OpenCode meta-provider, so direct multi-turn Kimi conversations failed. Scoped to `kimi` (gateway-served models matching the thinking-model name pattern are unaffected). Regression guard: `tests/unit/kimi-native-reasoning-injected-1480.test.ts`. (thanks @2220258345)
+- **fix(executors):** recover from a strict gateway's `context_management: Extra inputs are not permitted` 400 (port from 9router#1468, [#6420](https://github.com/diegosouzapw/OmniRoute/pull/6420)). **Claude Code** always sends a top-level `context_management` field; strict anthropic-compatible gateways reject it. The dedicated context-editing 400-fallback only fired when OmniRoute's own `contextEditing` feature was enabled (default off), so a client-sent field passed through untouched and 400'd. `context_management` is now in the generic reactive field-strip list, so it's stripped-and-retried once regardless of the feature flag (with correct request re-signing for claude-compatible relays). Regression guard: `tests/unit/provider-field-strips.test.ts`. (thanks @ohahe52-dot)
+- **fix(network):** enable **RFC 8305 Happy Eyeballs** (`autoSelectFamily`) on the direct-egress undici dispatcher (port from 9router#1237, [#6423](https://github.com/diegosouzapw/OmniRoute/pull/6423)). When DNS returns both IPv6 (AAAA) and IPv4 (A) and the IPv6 route is broken (e.g. a NAT64 `64:ff9b::` prefix without routing), undici tried IPv6 first and hung until `ETIMEDOUT` (then a 502 + account lockout), even though `curl` reached the same host. The direct dispatcher now races both families and uses whichever connects first. Proxy paths pin family via `proxyTls` and are unaffected. Regression guard: `tests/unit/direct-dispatcher-pipelining-4580.test.ts`. (thanks @adentdk)
+- **fix(combo):** round-robin now advances the rotation pointer past the model that **actually served**, not the eagerly-scheduled one (port from 9router#948, [#6428](https://github.com/diegosouzapw/OmniRoute/pull/6428)). With `stickyLimit: 1` (true round-robin), when the scheduled model failed and a _different_ model served via fallback, the counter had already advanced +1 from the scheduled index — so the next request reused the fallback-served model, degrading round-robin into hot-spotting on whichever model was healthy. The pointer now advances to the served index + 1 (mirroring the sticky-limit>1 path). Session-stickiness (#3825) and distribution are preserved. Regression guard: `tests/unit/combo-rr-fallback-advance-948.test.ts`. (thanks @binsarjr)
+- **fix(sse):** a non-string `model` field is now rejected with a `400` before the resolver, instead of crashing downstream `.toLowerCase()`/`.split()` calls into an empty-body `500` that escapes the error sanitizer ([#6407](https://github.com/diegosouzapw/OmniRoute/issues/6407)). Regression guard: `tests/unit/chat-non-string-model-6407.test.ts`. (thanks @chirag127)
+- **fix(api):** unknown `/api/*` routes now return a JSON `404` (instead of the dashboard HTML shell) and scalar chat params (`model`/`temperature`/etc.) are validated **before** the provider lookup so malformed requests fail fast with a clear `400` ([#6424](https://github.com/diegosouzapw/OmniRoute/issues/6424), [#6412](https://github.com/diegosouzapw/OmniRoute/issues/6412)). Regression guards: `tests/unit/api/api-catchall-json-404.test.ts`, `tests/unit/chat-early-schema-validation-6412.test.ts`. (thanks @chirag127)
+- **fix(api):** `/v1/chat/completions` now rejects a non-JSON `Content-Type` with a `400` before parsing the body ([#6414](https://github.com/diegosouzapw/OmniRoute/issues/6414)). Regression guard: `tests/unit/v1-chat-completions-content-type-6414.test.ts`. (thanks @chirag127)
+- **fix(api):** the `X-OmniRoute-Compression` response header is now echoed on `/v1/chat/completions` and `/v1/completions` ([#6422](https://github.com/diegosouzapw/OmniRoute/issues/6422)). Regression guard: `tests/unit/compression-header-echo-6422.test.ts`. (thanks @chirag127)
+- **fix(api):** concurrent `GET /v1/models` requests are coalesced into a single catalog build ([#6408](https://github.com/diegosouzapw/OmniRoute/issues/6408)). Regression guard: `tests/unit/v1-models-concurrent-6408.test.ts`. (thanks @chirag127)
+- **fix(api):** `/v1/completions` now echoes the requested `body.model` in its JSON + streamed responses ([#6429](https://github.com/diegosouzapw/OmniRoute/pull/6429)). Regression guard: `tests/unit/completions-body-model-echo.test.ts`. (thanks @chirag127)
+- **fix(api):** env-var master keys now see the full `/v1/models` catalog ([#6406](https://github.com/diegosouzapw/OmniRoute/issues/6406)). Regression guard: `tests/unit/models-catalog-envkey-6406.test.ts`. (thanks @chirag127)
+- **fix(api):** non-streaming `/v1/completions` responses now echo `body.model` aligned with the `X-OmniRoute-Model` header ([#6426](https://github.com/diegosouzapw/OmniRoute/issues/6426)). Regression guard: `tests/unit/v1-completions-model-header-match-6426.test.ts`. (thanks @chirag127)
+- **fix(api):** unknown `/v1/*` routes now return a JSON `404 not_found` instead of the Next.js dashboard HTML shell ([#6405](https://github.com/diegosouzapw/OmniRoute/issues/6405)). Regression guard: `tests/unit/api/v1-catchall-json-404.test.ts`. (thanks @chirag127)
+- **fix(api):** the per-connection provider models route now degrades to the shipped catalog when a provider's `/models` endpoint answers with a redirect ([#6267](https://github.com/diegosouzapw/OmniRoute/issues/6267)) — a `qwen-web` import failed with a raw `Redirect blocked … (307)` 503. `safeOutboundFetch` throws `REDIRECT_BLOCKED` on the 307, `getSafeOutboundFetchErrorStatus` maps it to 503, and `buildDiscoveryErrorFallbackResponse` treated every 503 as a hard error — so the non-empty `getModelsByProviderId("qwen-web")` catalog was never surfaced. A models-endpoint redirect is not a fixable-config error (unlike `URL_GUARD_BLOCKED`/`INVALID_URL`, which stay hard errors), so it now falls back to the local/cached catalog before the 503 short-circuit. General fix — covers any config-driven provider that 307s. Regression guard: `tests/unit/provider-models-qwen-web-redirect-6267.test.ts`. (thanks @chirag127)
+- **fix(api):** the per-connection provider models route (MCP `list_models_catalog` + the dashboard import view) now merges USER-ADDED custom models into its response ([#6247](https://github.com/diegosouzapw/OmniRoute/issues/6247)) — custom models live in the `key_value` namespace `customModels`, which the live REST `/api/v1/models` already merges, but `src/app/api/providers/[id]/models/route.ts` never read `getCustomModels`, so custom models were dropped on both the discovery-success and local_catalog paths. They are now folded into the returned model list (deduped by id, stamped `owned_by: provider`), fixing MCP + the dashboard import view in one place. Regression guard: `tests/unit/provider-models-custom-merge-6247.test.ts`. (thanks @RCrushMe)
+
+- **fix(providers):** GitLab Duo tool-calling follow-up turns no longer fail upstream with `422 {"detail":"Validation error"}` (tokens 0/0, rejected pre-inference). The #6234 tool-result-feedback fix serialized the **entire** multi-turn conversation into GitLab's single-file `code_suggestions` (`small_file`) generation endpoint — folded history that turn-N sent as an oversized `current_file.content_above_cursor` **and** duplicated verbatim into `user_instruction`, tripping the AI-Gateway's `small_file` validation guard. The executor now **bounds** that prompt: it keeps system + latest user message + the most-recent tool round (dropping older turns), caps oversized tool results, and stops duplicating the full prompt into `user_instruction` (which now carries only the short latest user message) — while still feeding the most-recent tool result back so the agent continues (`open-sse/executors/gitlab.ts`, [#6220](https://github.com/diegosouzapw/OmniRoute/issues/6220)). The unit test covers the bounding logic; the upstream 422→200 clearing is VPS-only (Hard Rule #18). Regression guard: `tests/unit/gitlab-tool-exchange-bounded-6220.test.ts`.
+
+- **fix(i18n):** the provider-detail (`/dashboard/providers/[id]`) connection-status filter labels no longer render as `__MISSING__:All` / `__MISSING__:Active` / `__MISSING__:Error` / `__MISSING__:Banned` / `__MISSING__:CreditsExhausted` in non-English locales (notably pt-BR) ([#6290](https://github.com/diegosouzapw/OmniRoute/issues/6290)). Root cause was **not** the namespace mismatch the issue guessed — the `providers.filter*` keys resolve correctly in `en.json`; the debt lived in the locale mirrors (`src/i18n/messages/*.json`), where these five keys carried the `__MISSING__:` sync sentinel in ~15 locales and were absent entirely in ~26 others, so next-intl found the key and echoed the sentinel verbatim. All 40 non-English/-Chinese mirrors now ship real translations for the five `providers.filter*` labels. Regression guard: `tests/unit/i18n-provider-filter-keys-6290.test.ts`. (thanks @diegosouzapw)
+
+- **fix(providers):** the `copilot-m365-web` streaming executor now emits `debug`-level WebSocket diagnostics ([#6210](https://github.com/diegosouzapw/OmniRoute/issues/6210)) — the outbound WS URL (with the `access_token` **redacted** via `redactWsUrl()`), handshake success/failure, and each received SignalR frame's `type`/`target`. Previously the streaming path logged nothing, so an empty `content:null` response (the M365 Education / Starter tier symptom fixed in #6234) was undiagnosable even at `APP_LOG_LEVEL=debug`. The change is debug-level and side-effect-free — it does not alter streaming behavior or the frame parser, and the token never reaches the logs. Regression guard: `tests/unit/copilot-m365-web-logging-6210.test.ts` (thanks @qpeyba)
+
+- **fix(resilience):** a round-robin combo no longer returns `503 all upstream accounts are unavailable` when a compatibility-rejected target is actually healthy ([#6238](https://github.com/diegosouzapw/OmniRoute/issues/6238)). `filterTargetsByRequestCompatibility` drops request-incompatible targets (tool/vision/structured-output unsupported, or below the required context window) **before** any availability check runs, and its `compatible.length === 0` safety net only fired when _all_ targets were filtered — not when the kept targets later all turned out runtime-unavailable (circuit-open / cooldown / no credentials). So a combo could 503 while a compat-rejected-but-healthy provider sat unused. `handleRoundRobinCombo` now keeps the compat-rejected set and, when every compat-kept target was skipped without a single real attempt, probes those rejected targets as a **last-resort fallback tier** (via the new pure `open-sse/services/combo/comboCompatFallback.ts`) before crystallizing the 503. Regression guard: `tests/unit/combo-roundrobin-compat-fallback-6238.test.ts`. (thanks @ThongAccount)
+
+- **fix(startup):** best-effort self-heal for a corrupted Turbopack dev cache on Windows ([#6289](https://github.com/diegosouzapw/OmniRoute/issues/6289)). On Windows, `pnpm dev` can fail at startup when Turbopack `mmap`s a persistent-cache SST file and the OS refuses the mapping (`os error 1455` — "paging file too small"), which Turbopack surfaces as a misleading `Module not found: Can't resolve '@/shared/utils/machine'`. This is a **known upstream Turbopack cache-corruption bug — not our code**. The dev launcher (`scripts/dev/run-next.mjs`) now wraps `nextApp.prepare()` and, when it rejects with that signature (`isTurbopackCacheCorruption` in the new `scripts/dev/turbopackCacheHeal.mjs`), purges `.build/next/**/cache/turbopack` and retries **once** with a clear log. **Caveat — best-effort only:** the corruption often surfaces as a runtime overlay rather than a `prepare()` rejection, so this cannot always intercept it; the reliable remedy remains manually deleting the Turbopack cache dir. Regression guard: `tests/unit/turbopack-cache-heal-6289.test.ts`. (thanks @chirag127)
+
+- **fix(providers):** qodercli PAT auth no longer fails with `spawn qodercli ENOENT` on Windows ([#6263](https://github.com/diegosouzapw/OmniRoute/issues/6263)) — `spawnQoderCli` spawned the bare `qodercli` name with `shell:false` and an unenriched env, so the npm `.cmd` wrapper under `%APPDATA%\npm` (a user-PATH directory) was never resolved. It now resolves the absolute `.cmd`/`.exe` path through the existing `getCliRuntimeStatus("qoder")` resolver in `src/shared/services/cliRuntime.ts` (memoized), spawns with `shell` when the target is a `.cmd`/`.bat`, and uses the cliRuntime-enriched env (PATH + PATHEXT + APPDATA); the ENOENT error now lists the searched paths plus the `CLI_QODER_BIN` override. End-to-end spawn on a real Windows host is host-only (Hard Rule #18); the path-resolution logic is unit-tested. Regression guard: `tests/unit/qodercli-windows-resolve-6263.test.ts`. (thanks @chirag127)
+
+- **fix(sse):** the reasoning-token buffer no longer inflates **probe-sized `max_tokens`** ([#6274](https://github.com/diegosouzapw/OmniRoute/issues/6274)) — Claude Code's `/model` capability check sends `max_tokens: 1`, but for a thinking-capable model with a large output cap (e.g. `glm-5.2`) the #3587 headroom heuristic (`max(current + 1000, ceil(current * 1.5))`) rewrote it to `1001` and forwarded that upstream, wasting tokens on a request that was never a genuine reasoning budget. `resolveReasoningBufferedMaxTokens()` (`open-sse/services/reasoningTokenBuffer.ts`) now short-circuits and returns the caller's value verbatim when it is below the new `REASONING_BUFFER_MIN_TRIGGER` (256) threshold — a tiny explicit limit is a probe, not a reasoning request. Real budgets still receive the #3587 headroom unchanged, and the guard runs after the existing capability checks so unknown / non-reasoning models keep returning `null`. Regression guard: `tests/unit/reasoning-token-buffer-6274.test.ts`. (thanks @brightfiscalband)
+
+- **fix(cli):** `omniroute reset-password` now works as a real subcommand, and password resets over piped (non-TTY) stdin actually apply ([#6261](https://github.com/diegosouzapw/OmniRoute/issues/6261), [#6258](https://github.com/diegosouzapw/OmniRoute/issues/6258)). Two coupled defects: (1) **#6261** — `bin/omniroute.mjs` routed everything through Commander with only two pre-Commander bypasses (`--mcp`, `reset-encrypted-columns`), so `omniroute reset-password` was rejected as an unknown command; only the separate `omniroute-reset-password` bin worked, while the docs falsely advertised the subcommand (incl. a bogus "legacy alias still works"). A pre-Commander bypass mirroring `reset-encrypted-columns` now dynamically imports `bin/reset-password.mjs` (which self-executes) before Commander parses; the three doc lines were corrected. (2) **#6258** — `bin/reset-password.mjs` issued two sequential `rl.question` prompts; under piped stdin the second read never settled at EOF, so `main()` never reached `resetManagementPassword` and the reset was a silent no-op (both prompts printed, no success, password unchanged). The CLI now detects non-TTY stdin and reads it once (first line = password, second line = confirm if present, else reused), adds a `--password-stdin` flag (entire stdin is the password, no confirmation), and exits `0` explicitly so the success line always flushes; interactive TTY behavior is unchanged. Regression guard: `tests/unit/reset-password-cli-6261-6258.test.ts` (3). (thanks @chirag127)
+
+- **fix(db):** the mass-migration **safety abort** now tells the operator how to bypass it and stops flooding the log ([#6260](https://github.com/diegosouzapw/OmniRoute/issues/6260)) — after restoring a backup that wiped the migration tracking table, `runMigrations()` threw the abort on every downstream `ensureDbInitialized()`, re-logging the full banner 11+ times, and the message never mentioned the existing `OMNIROUTE_MAX_PENDING_MIGRATIONS` escape hatch. The abort text now appends a bypass hint (set `OMNIROUTE_MAX_PENDING_MIGRATIONS=0` in `server.env` / `DATA_DIR/.env`), and a new `MigrationSafetyAbortError` is memoized so repeated calls in the same process throw the same instance and emit a single concise line instead of the full cascade. Regression guard: `tests/unit/migration-safety-abort-6260.test.ts`. (thanks @chirag127)
+
+- **fix(auth):** importing a **distinct** Codex/ChatGPT OAuth `auth.json` is no longer falsely rejected as "already exists" when it belongs to a different user in the same workspace ([#6301](https://github.com/diegosouzapw/OmniRoute/issues/6301)). `findExistingCodexConnection` (in `src/lib/oauth/utils/codexAuthImport.ts`) deduped **only** on `providerSpecificData.workspaceId === accountId`, where `accountId` is the shared `chatgpt_account_id`/`tokens.account_id` — so two members of the same ChatGPT Team collapsed onto a single connection (409 `duplicate_account`). The id_token's `https://api.openai.com/auth` claim carries a per-user `chatgpt_user_id` alongside the workspace id (the device-flow path already persisted it as `chatgptUserId`, but the import path did not). Now `parseAndValidateCodexAuth` extracts `userId` (`chatgpt_user_id` → `user_id` → JWT `sub`) into `ParsedCodexAuth`, the create/update paths persist `chatgptUserId` in `providerSpecificData` (mirroring `codex.ts`), and dedup keys on `workspaceId` **AND** `chatgptUserId` — with a backward-compat fallback to legacy accountId-only matching when no stored connection for that workspace records a `chatgptUserId`, so genuinely-same accounts still dedup. Regression guard: `tests/unit/codex-auth-import-userid-dedup-6301.test.ts` (4). (thanks @anungma)
+
+- **fix(providers):** importing models for the **venice-web** provider no longer fails with a red "Provider venice-web does not support models listing" ([#6269](https://github.com/diegosouzapw/OmniRoute/issues/6269)). `venice-web` is a web-cookie provider with an executor but no upstream `/v1/models` endpoint and no registry `models`, so the models route fell through to the tail `400`. Mirroring the `jules`/`linkup-search`/`ollama-search` fix (#5569), it now ships a static local catalog entry in `src/lib/providers/staticModels.ts` — seeding the current Venice lineup (`venice-uncensored`, `llama-3.3-70b`, `qwen3-235b`, `qwen3-4b`, `deepseek-r1-671b`; Venice rotates its catalog, see docs.venice.ai/models/overview) — so the route returns `200` with `source:"local_catalog"`, `intentional:true`. Regression guard: `tests/unit/static-models-venice-web-6269.test.ts`. (thanks @chirag127)
+
+- **fix(api):** the specialty model catalogs (`/v1/embeddings`, `/v1/images`, `/v1/music`, `/v1/videos` model lists) are now derived from the **unified catalog filtered by a predicate** (`getSpecialtyModelsResponse`) instead of ad-hoc per-route logic, so they consistently respect active-credential visibility and stay in sync with the main catalog ([#6303](https://github.com/diegosouzapw/OmniRoute/pull/6303)). Regression guard: `tests/unit/specialty-model-catalog-routes.test.ts`. (thanks @makcimbx)
+- **fix(api):** the agent-bridge server route now resolves the MITM manager via a **dynamic `import("@/mitm/manager.runtime")`** so Turbopack does not statically pull the stub (or over-bundle the manager), and the agent-skills generator anchors its output base path with `path.join(process.cwd(), …)` so Turbopack's static analyzer stops tracing the whole project root ([#6329](https://github.com/diegosouzapw/OmniRoute/issues/6329), [#6366](https://github.com/diegosouzapw/OmniRoute/pull/6366)). Regression guard: `tests/unit/agent-bridge-server-route-dynamic-import.test.ts`. (thanks @Iammilansoni)
+- **fix(api):** internal probes (combo-test, cloud-sync verify) now pick a **management-scoped / allow-all API key** instead of naively grabbing `getApiKeys()[0]` — a restricted `self:usage` first row made the probe fail with "Model X is not allowed for this API key" even when the combo path was healthy (`pickApiKeyForInternalUse` in `src/lib/db/apiKeys.ts`). The API-manager model editor also falls back to `/api/models?all=true` when `/v1/models` is catalog-protected ([#6372](https://github.com/diegosouzapw/OmniRoute/pull/6372)). Regression guard: `tests/unit/pick-internal-api-key-6372.test.ts`. (thanks @jmengit)
+- **fix(live-ws):** the Live Dashboard WebSocket server now **rejects on bind failure** (e.g. `EADDRINUSE` when the API bridge already holds the port) instead of letting the error surface as an unhandled `error` event that crash-loops the process — the `error` listener is attached to `wss` (not `server`) and releases the EventBus subscription on a failed start ([#6324](https://github.com/diegosouzapw/OmniRoute/issues/6324)). Regression guard: `tests/unit/live-ws-eaddrinuse-6324.test.ts`. (thanks @vinayakkulkarni)
+- **fix(dashboard):** the Home provider-topology widget now trusts the live provider-metrics snapshot — it uses `topology.errorProvider` and live `activeRequests` directly instead of re-deriving state from a stale `lastErrorAt` or applying a frontend timeout filter, so the topology reflects real-time provider health ([#6322](https://github.com/diegosouzapw/OmniRoute/pull/6322)). Regression guard: `tests/unit/home-provider-topology-live-state.test.ts`. (thanks @xz-dev)
+- **fix(sse):** strip zero-width markers from streamed **tool-call arguments** — a follow-up to [#5857](https://github.com/diegosouzapw/OmniRoute/pull/5857). That PR removed injected zero-width joiners (U+200D) from streamed assistant text/reasoning but deliberately left tool-call argument JSON byte-exact. The request-side obfuscation (`open-sse/services/claudeCodeObfuscation.ts`) injects ZWJ into agent words — including the temp path inside the Bash tool description — and Claude models copy that verbatim into generated commands, which are delivered as tool-call arguments rather than assistant text. As a result the ZWJ survived and corrupted code blocks (e.g. a temp path rendered with an invisible joiner). Now `open-sse/handlers/responseSanitizer.ts` strips zero-width code points from tool-call argument strings at every emit site (OpenAI non-stream/stream chat `tool_calls` + legacy `function_call`, native Responses `function_call` items, the OpenAI→Responses conversion, and the native Responses streaming `response.function_call_arguments.delta/.done` events). Only zero-width code points are removed; JSON structure and all other bytes stay identical (no parse/restringify), so normal arguments remain byte-exact. Regression guard: 6 new cases in `tests/unit/response-sanitizer.test.ts` (suite 50/50).
+
+- **fix(nodejs):** the default app log path now resolves under `DATA_DIR` (`~/.omniroute/logs/application/app.log`) instead of `process.cwd()` ([#6197](https://github.com/diegosouzapw/OmniRoute/issues/6197)) — the globally-installed CLI runs from an arbitrary working directory, so anchoring the default to cwd made file logging silently write to (or no-op under) an unrelated directory, contradicting the documented `.env.example` default. `getAppLogFilePath()` now computes the default lazily via the pure `resolveDataDir()` resolver (honours a per-process `DATA_DIR`, no directory-creation side effect); an explicit `APP_LOG_FILE_PATH` still wins. Regression guard: `tests/unit/logenv-datadir-path-6197.test.ts` (3).
+
+- **fix(docker):** AgentBridge/`startMitm` no longer aborts in containers/headless when the Antigravity-default DNS step can't write `/etc/hosts` ([#6127](https://github.com/diegosouzapw/OmniRoute/issues/6127)), and the privileged command's stderr now reaches `app.log` instead of only a bare exit code hitting the toast ([#6198](https://github.com/diegosouzapw/OmniRoute/issues/6198)). The default DNS step (`addDNSEntry`) was called unguarded while cert install and the two sibling DNS steps were each best-effort — in the runtime Docker image (`USER node`, no `sudo`, read-only `/etc/hosts`) it threw `Command failed with code 1` out of `startMitmInternal` and killed the whole start, discarding the stderr. The three DNS steps are extracted into a best-effort `provisionDnsEntries()` where each failure is logged with the full `err` (stderr included, folded in by `systemCommands.ts`) and never aborts the start. Regression guard: `tests/unit/mitm-dns-graceful-degrade-6127.test.ts` (4).
+
+- **fix(providers):** copilot-m365-web now supports the M365 Education "Starter / OfficeWebIncludedCopilot" tier and no longer returns an empty `content:null` stream ([#6210](https://github.com/diegosouzapw/OmniRoute/issues/6210)). Two gaps: (1) `buildWsUrl()` hardcoded the individual-consumer scenario (`OfficeWebPaidConsumerCopilot`, `isEdu=false`) — the EDU tier is now opt-in via `providerSpecificData.tier="edu"`, emitting `scenario=OfficeWebIncludedCopilot`/`isEdu=true` (the individual path is unchanged); (2) the EDU/GPT-5.5 path streams deltas via `arguments[0].writeAtCursor` (incremental) instead of only `messages[].text` (accumulated snapshots), which the parser dropped — a new `accumulateBotContent()` folds both formats, with `type:2 item.result.message` as a last-resort fallback. Regression guard: `tests/unit/copilot-m365-edu-writeatcursor-6210.test.ts` (10). (thanks @qpeyba)
+
+- **fix(providers):** GitLab Duo executor now feeds tool results back into the prompt instead of looping ([#6220](https://github.com/diegosouzapw/OmniRoute/issues/6220)) — `buildPrompt()` branched only on `system`/`user` and took `userParts.at(-1)`, silently dropping the `assistant{tool_calls}` + `tool{result}` turns the client appended, so the reconstructed prompt was byte-identical to turn 1 and the model re-emitted the same `<tool>` call forever. When a tool exchange is present the full conversation is now serialized, folding each tool result back keyed by its `tool_call_id`; simple conversations keep the legacy shape. Complements the tool_call emission from [#6051](https://github.com/diegosouzapw/OmniRoute/issues/6051) (the `kilo-duplicate` label was a false positive — different, sequential defect). Regression guard: `tests/unit/gitlab-tool-result-feedback-6220.test.ts` (4).
+
+- **fix(providers):** opencode-go/opencode-zen can now synthesize the OpenCode CLI identity headers Cloudflare requires on VPS egress ([#5997](https://github.com/diegosouzapw/OmniRoute/issues/5997)) — on a datacenter VPS, `opencode.ai/zen/go/v1/chat/completions` 403s (HTML challenge) requests lacking CLI identity, while the reporter's control curl proved that `User-Agent: opencode-cli/1.0.0` + `x-opencode-client: cli` + `x-opencode-project: default` + fresh request/session UUIDs succeed. Opt-in via `OPENCODE_SYNTHESIZE_CLI_HEADERS=true` (values overridable via `OPENCODE_GO_USER_AGENT`/`OPENCODE_USER_AGENT`/`OPENCODE_CLIENT`/`OPENCODE_PROJECT`); it fills only headers the client did not already send. Kept **off by default** — the forward-only path is deliberate (fabricating a wrong value risks upstream rejection; a prior dedup regressed with `opencode/local`), so this replaces the fragile local header-injection shim without changing default behavior. Regression guard: `tests/unit/opencode-cli-headers-synthesis-5997.test.ts` (6). (thanks @aleksesipenko)
+- fix(resilience): sticky session affinity now evicts and fails over to another account when the pinned account is exhausted/unavailable (#6219)
+
+- fix(sse): Responses API passthrough now drops internal commentary-phase output before forwarding to clients (gated by RESPONSES_PASSTHROUGH_DROP_COMMENTARY, default on) (#6199)
+
+- **fix(sse):** tool-call function schemas with a root `type: null` are now coerced to `type: "object"` before dispatch ([#6359](https://github.com/diegosouzapw/OmniRoute/issues/6359)) — clients like the Codex app emit `parameters: { type: null, ... }` for some tools, which OpenAI-compatible upstreams reject with `400 Invalid schema for function '...': schema must be a JSON Schema of 'type: "object"', got 'type: null'`, failing the whole request. `toolSchemaSanitizer` already stripped the null; it now re-adds the mandatory root `"object"` type (and empty `properties`/open `additionalProperties` when absent). Combinator roots (`anyOf`/`oneOf`/`allOf`) and explicit root types are left untouched. Regression guard: 5 new cases in `tests/unit/tool-schema-sanitizer.test.mjs`.
+
+- **fix(docker):** AgentBridge no longer fails to start on npm/Electron/VPS installs with "MITM manager stub reached at runtime" ([#6344](https://github.com/diegosouzapw/OmniRoute/issues/6344)) — v3.8.45 flipped the production bundler default to Turbopack, but `next.config.mjs` aliased `@/mitm/manager` to its Docker-only degraded stub **unconditionally**. That was harmless while Docker (which sets the alias intentionally for #3390 graceful degradation) was the sole Turbopack consumer, but once every artifact built with Turbopack the stub shipped to all non-Docker users and `startMitm` threw on the first Agent-Bridge start. The alias is now opt-in via `OMNIROUTE_MITM_STUB=1` (set only by the Dockerfile) through the shared `scripts/build/mitm-stub-flag.mjs` helper; default builds bundle the real manager. Regression guard: `tests/unit/mitm-stub-alias-6344.test.mjs` (4).
+
+- **fix(proxy):** stop the v3.8.44 proxy regression that leaked the real IP and disabled healthy proxies ([#6246](https://github.com/diegosouzapw/OmniRoute/issues/6246)). Two coupled defects from the new health scheduler: (1) **IP leak** — when a proxy assigned to a connection was marked `inactive`, resolution fell through to a **direct** egress instead of blocking, exposing the operator's real IP; (2) **over-deactivation** — the sweep flipped a proxy to `inactive` on the **first** failed probe and counted our own 5s timeout / a probe-target `5xx` as the proxy's fault, so healthy paid proxies vanished from egress selection ("my proxies are not being used anymore"). Fix: the sweep decision is extracted into a pure, network-free `decideProxyHealthAction` (`src/lib/proxyHealth/decision.ts`) — by default the health check now **only counts/logs and never downgrades status** (a proxy is downgraded/removed only with `PROXY_AUTO_REMOVE=true`, after `PROXY_AUTO_REMOVE_AFTER` **consecutive** conclusive failures); probes are classified tri-state so an inconclusive result (our timeout, or a `5xx` from the probe target) never penalizes the proxy, and the probe timeout is raised 5s→15s. Separately, `safeResolveProxy` now **fails closed** via the existing policy: a connection whose assigned proxy is dead is blocked instead of leaking direct (`hasBlockingProxyAssignment`), honoring the explicit `proxy off` toggles and the `PROXY_FAIL_OPEN=true` opt-out. Existing proxies stuck `inactive` by the old behavior need a one-time manual re-activate (the operator owns proxy status). Regression guards: `tests/unit/proxy-health-decide-action-6246.test.ts`, `tests/unit/proxy-assigned-unavailable-6246.test.ts`.
+
+- **fix(proxy):** make "Test All" read-only and add bulk enable/disable ([#6246](https://github.com/diegosouzapw/OmniRoute/issues/6246)). Complements the core fail-closed / scheduler fix (#6296) with the two remaining reporter asks. (1) The **"Test All" button** (`POST /api/settings/proxies/auto-test`) used to flip a proxy to `inactive` on a failed reachability probe; since the egress selector excludes `inactive` proxies, a flaky probe (an unreachable `httpbin.org`, a proxy that blocks `HEAD`, or a slow paid proxy) silently disabled every proxy that failed — "Test All" is now **read-only by default** (only the operator sets a proxy active/inactive; opt back into the legacy test-and-set with `PROXY_HEALTH_AUTO_DEACTIVATE=true`). (2) Adds a **bulk enable/disable** proxies endpoint + toolbar action (`POST /api/settings/proxies/batch-activate`) so an operator can re-activate proxies in one click. Regression guard: `tests/unit/proxy-health-6246.test.ts`. (thanks @tenshiak)
+
+- **chatcore (tools): stop the default 128-tool cap from silently dropping opencode's `task`/MCP tools.** opencode (used as an MCP/agent host) sends a large tool list; when it exceeds the speculative `MAX_TOOLS_LIMIT` (128) default, `truncateToolList` did a blind `tools.slice(0, 128)`, dropping every tool past index 128 — including opencode's built-in `task` tool (subagent launch) and many MCP tools, so models routed through OmniRoute could no longer spawn subagents or reach part of their tools. The cap exists to avoid upstream `400`s for providers with real hard limits (e.g. grok-cli 200), so it is kept for those: detection of the opencode client (`isOpencodeClient` — any `x-opencode-*` header, or `opencode` in the user-agent) now only bypasses the **speculative 128 default**, never a known provider ceiling. Precedence is explicit — a proactive/detected provider limit always truncates (even for opencode); otherwise opencode forwards its full tool list; otherwise the unchanged 128 default applies to every other client. Refactors `getEffectiveToolLimit` into `getKnownToolLimit(provider) ?? DEFAULT_LIMIT` (byte-identical for existing callers) and fixes a cosmetic debug-log that reported the truncated count instead of the original. Regression guard: `tests/unit/tool-limit-detector.test.ts`.
+
+- **fix(mitm):** the macOS MITM-cert install check now matches the system keychain again. `security find-certificate -a -Z` prints the SHA-1 as a colon-less hex string, but the installed-check compared it against `getCertFingerprint()`'s colon-separated form, so the substring match never hit — the cert was reported as not-installed and re-prompted for the sudo install on every run. Fingerprints are now normalized (colons stripped, upper-cased) on both sides via the extracted `macCertOutputHasFingerprint` helper. Regression guard: `tests/unit/mitm-cert-mac-fingerprint.test.ts`. ([#6204](https://github.com/diegosouzapw/OmniRoute/pull/6204), closes [#6134](https://github.com/diegosouzapw/OmniRoute/issues/6134) — thanks @rianonehub)
+
+- **fix(api):** `/v1/messages/count_tokens` now counts `tool_use`, `tool_result` and `thinking` content blocks (and array-form `system` prompts) in the local-estimation path, instead of only `text`. Real agentic conversations keep ~95% of their tokens inside tool results; the previous estimate returned near-zero for them, which silently broke Claude Code's auto-compaction (context grew past the window with no compaction until the upstream API rejected the request). The real provider-side count path is unchanged. Regression guard: `tests/unit/messages-count-tokens-route.test.ts`. ([#6221](https://github.com/diegosouzapw/OmniRoute/pull/6221) — thanks @luweiCN)
+
+- **fix(antigravity):** strip a trailing assistant prefill turn for Vertex Claude models to avoid upstream 400s ([#6114](https://github.com/diegosouzapw/OmniRoute/pull/6114)). Regression guard: `tests/unit/antigravity-claude-prefill-strip.test.ts`. (thanks @anki1kr)
+
+- **fix(security):** the mutable cloud-agent routes (`/api/cloud/credentials/update`, `/api/cloud/models/alias`) now require management auth instead of being treated as public. They were classified as public API routes, so a request without management credentials could update stored cloud-agent credentials and model aliases. They are removed from the public-route set, classified as management routes in the authz pipeline, and gated by `requireManagementAuth`; cloud **read**/auth routes stay public. Regression guards: `tests/unit/cloud-write-auth.test.ts`, `tests/unit/authz/classify.test.ts`, `tests/unit/public-api-routes.test.ts`. ([#6233](https://github.com/diegosouzapw/OmniRoute/pull/6233) — thanks @vittoroliveira-dev)
+
+- **refactor(dashboard):** extract the onboarding-wizard "Open provider details" link target into a pure, unit-tested `buildProviderDetailsHref(connection)` helper. The wizard already routes by `connection.id` (the node UUID) rather than the provider category slug (#6144/#6145); this hardens that behavior behind a tested helper that guards a missing id/connection. Regression guard: `tests/unit/provider-onboarding-href.test.ts`. ([#6166](https://github.com/diegosouzapw/OmniRoute/pull/6166) — thanks @KooshaPari)
+
+- **fix(security):** the doubao synthetic device-id generator now derives its digits via an unbiased crypto-random draw (rejection sampling over `crypto.randomBytes()`) instead of a `% 10` reduction, closing a CodeQL `js/biased-cryptographic-random` finding.
+- **fix(agentSkills):** the GitHub-skills generator now resolves `outputDir` to an absolute path before writing, fixing a regression introduced by #6366 (relative-to-cwd base path) that could write generated skill files to the wrong directory.
+- **fix(security):** `/api/keys/{id}/devices` now checks the HTTP method before auth/validation, returning a `405` for non-GET/DELETE verbs instead of a misleading `401`/`500` (closes a `dast-smoke` QUERY-method finding).
+- **fix(quality):** clear the last 2 heavy quality-gate reds on the release tip (cycle pre-flight).
+- **fix(mitm):** the test suite and CI can never mutate the OS trust store — `OMNIROUTE_SKIP_SYSTEM_TRUST=1` is set globally for tests/CI so `installCert`/`uninstallCert`/`installTproxyCa` skip the privileged OS dispatch ([#6310](https://github.com/diegosouzapw/OmniRoute/pull/6310); full detail is under the [3.8.45] section below — this branch received it via the parallel-cycle sync-back).
+- **fix(api):** `POST /api/github-skills` now Zod-validates its request body; documented the new quality-gate env vars and pinned the merge-integrity GitHub Actions to a commit SHA.
+- **fix(skills):** generate the missing `omni-github-skills` registry entry and align the agent-skills catalog-count tests (follow-up to #6186).
+- **fix(quality):** clear the cycle's 11 net-new ESLint errors and make `validate-release-green` suppressions-aware.
+
+### 📝 Maintenance
+
+- **i18n(it):** add 118 missing Italian (`it`) translations (net-additive — no existing keys dropped, valid JSON), improving Italian UI coverage. ([#6212](https://github.com/diegosouzapw/OmniRoute/pull/6212) — thanks @serverless83)
+- **chore(providers):** remove deprecated **MiMo V2** model entries from the catalogs (xiaomi-mimo, opencode-go, zenmux-free, audio TTS) — the upstream V2 line is superseded by MiMo V2.5; drops `mimo-v2-tts`, `mimo-v2-pro`, `mimo-v2-omni`, `mimo-v2-flash`, `mimo-v2-flash-free` and realigns the provider-catalog tests. ([#6248](https://github.com/diegosouzapw/OmniRoute/pull/6248) — thanks @backryun)
+- **chore(release):** ~50 commits on this branch are v3.8.45 pre-flight/hardening fixes and CI-perf work that landed here via the parallel-cycle sync-back (`sync-next-cycle.mjs`, Hard Rule #21) **after** the `v3.8.45` git tag was cut, and are already fully documented under the **[3.8.45]** section below — listed here only so the per-cycle commit-coverage check (`npm run release:uncovered`) doesn't flag them as gaps. Provider/catalog/UX/backend fixes: #6041, #6078, #6108, #6135, #6148, #6149, #6154, #6158, #6161, #6162, #6163, #6164, #6165, #6170, #6177, #6178, #6181, #6186, #6187, #6191, #6193, #6194, #6195, #6200, #6205, #6208, #6209, #6211, #6213, #6223, #6224, #6225, #6226, #6227, #6228, #6229, #6230, #6235, #6291, #6292. CI/release-pipeline work: #6167, #6203, #6214, #6215, #6218, #6273, #6275, #6283, #6284, #6285, #6300, #6305.
+- **chore(release):** additional zero-ref release-cycle plumbing on this branch, kept out of `release:uncovered` on purpose (no `#N` in the commit subject to cite): opening the v3.8.46 cycle, opening/closing the v3.8.45 cycle, the finalized [3.8.45] CHANGELOG i18n sync-back to 42 mirrors, the v3.8.45 cognitive/cyclomatic and file-size drift rebaselines, ESLint stale-suppression pruning (4,273 → 4,233), and clearing test-masking/docs-all pre-flight reds for v3.8.45.
+
+### ⚡ Performance & Infrastructure
+
+- **perf(release-green):** the pre-flight validator (`scripts/quality/validate-release-green.mjs`) now runs its 4 slow suites (unit / vitest / integration / pack-artifact) **concurrently** via `Promise.all` — pre-flight wall time drops from ~the sum of the suites to ~the slowest one (~30min saved per round; Phase 0 was the nº1 bottleneck of the v3.8.45 release benchmark, 2h54 of 6h34 e2e). Guard: `tests/unit/validate-release-green.test.ts` ("runs the slow suites CONCURRENTLY"). ([#6319](https://github.com/diegosouzapw/OmniRoute/pull/6319))
+- **fix(ci):** `scripts/release/sync-next-cycle.mjs` — two defects found live in its first production run (v3.8.45 Phase 5): (1) the `git()` helper's default 1 MiB `maxBuffer` crashed with `ENOBUFS` on `git show origin/main:CHANGELOG.md` (the CHANGELOG alone is >1 MiB) — widened to 64 MiB; (2) the i18n resync only propagated the `[NEXT]` (TBD) section, leaving the just-shipped finalized section as "— TBD" in all 42 mirrors — it now also syncs `[prevVersion]` bounded by the heading below it (new exported pure helper `versionAfter`). Guards: +5 tests in `tests/unit/sync-next-cycle.test.ts` (8/8). ([#6327](https://github.com/diegosouzapw/OmniRoute/pull/6327))
+- **test(ci):** concurrency-sensitive flaky tests are **quarantined into a serial pass** (`tests/unit/serial/`, `--test-concurrency=1`, appended to every unit runner incl. sharded variants — the serial pass is sharded too so concurrent shard jobs never self-collide). Initial set: `glm-coding-plan-monthly-3580`, `quota-division-blocks`, `provider-health-autopilot`, `combo-health-autopilot` — the class behind the ~28min CI wedges/re-runs (two live 1h+ wedges cancelled during this PR's own validation). Discovery + TIA gates track the new glob; systemic root cause (async logger writing after teardown) tracked in [#6360](https://github.com/diegosouzapw/OmniRoute/issues/6360). Guard: `tests/unit/test-serial-quarantine.test.ts` (4). ([#6347](https://github.com/diegosouzapw/OmniRoute/pull/6347))
+
+### 🙌 Contributors
+
+Thanks to everyone whose work landed in v3.8.46:
+
+| Contributor                                                    | PRs / Issues           |
+| -------------------------------------------------------------- | ---------------------- |
+| [@2220258345](https://github.com/2220258345)                   | direct commit / report |
+| [@abdofallah](https://github.com/abdofallah)                   | direct commit / report |
+| [@adentdk](https://github.com/adentdk)                         | direct commit / report |
+| [@aleksesipenko](https://github.com/aleksesipenko)             | direct commit / report |
+| [@anki1kr](https://github.com/anki1kr)                         | direct commit / report |
+| [@anungma](https://github.com/anungma)                         | direct commit / report |
+| [@arpicato](https://github.com/arpicato)                       | direct commit / report |
+| [@backryun](https://github.com/backryun)                       | #6248                  |
+| [@binsarjr](https://github.com/binsarjr)                       | direct commit / report |
+| [@brightfiscalband](https://github.com/brightfiscalband)       | direct commit / report |
+| [@chirag127](https://github.com/chirag127)                     | #6501, #6506           |
+| [@developerjillur](https://github.com/developerjillur)         | direct commit / report |
+| [@dilneiss](https://github.com/dilneiss)                       | #6499                  |
+| [@dtybnrj](https://github.com/dtybnrj)                         | direct commit / report |
+| [@Forcerecon](https://github.com/Forcerecon)                   | direct commit / report |
+| [@hao3039032](https://github.com/hao3039032)                   | direct commit / report |
+| [@Iammilansoni](https://github.com/Iammilansoni)               | #6150, #6245           |
+| [@jmengit](https://github.com/jmengit)                         | direct commit / report |
+| [@jordansilly77-stack](https://github.com/jordansilly77-stack) | direct commit / report |
+| [@JxnLexn](https://github.com/JxnLexn)                         | direct commit / report |
+| [@KooshaPari](https://github.com/KooshaPari)                   | #6166                  |
+| [@loopyd](https://github.com/loopyd)                           | direct commit / report |
+| [@luweiCN](https://github.com/luweiCN)                         | #6221                  |
+| [@makcimbx](https://github.com/makcimbx)                       | direct commit / report |
+| [@muflifadla38](https://github.com/muflifadla38)               | direct commit / report |
+| [@newnol](https://github.com/newnol)                           | direct commit / report |
+| [@ofekbetzalel](https://github.com/ofekbetzalel)               | direct commit / report |
+| [@ohahe52-dot](https://github.com/ohahe52-dot)                 | direct commit / report |
+| [@phidinhmanh](https://github.com/phidinhmanh)                 | direct commit / report |
+| [@powellnorma](https://github.com/powellnorma)                 | direct commit / report |
+| [@qpeyba](https://github.com/qpeyba)                           | direct commit / report |
+| [@RaviTharuma](https://github.com/RaviTharuma)                 | direct commit / report |
+| [@RCrushMe](https://github.com/RCrushMe)                       | direct commit / report |
+| [@rianonehub](https://github.com/rianonehub)                   | #6134, #6204           |
+| [@serverless83](https://github.com/serverless83)               | #6212                  |
+| [@swingtempo](https://github.com/swingtempo)                   | direct commit / report |
+| [@tenshiak](https://github.com/tenshiak)                       | direct commit / report |
+| [@ThongAccount](https://github.com/ThongAccount)               | direct commit / report |
+| [@UnrealAryan](https://github.com/UnrealAryan)                 | direct commit / report |
+| [@vinayakkulkarni](https://github.com/vinayakkulkarni)         | direct commit / report |
+| [@vittoroliveira-dev](https://github.com/vittoroliveira-dev)   | #6233                  |
+| [@warelik](https://github.com/warelik)                         | direct commit / report |
+| [@xxy9468615](https://github.com/xxy9468615)                   | direct commit / report |
+| [@xz-dev](https://github.com/xz-dev)                           | direct commit / report |
+| [@yanpaing007](https://github.com/yanpaing007)                 | direct commit / report |
+| [@diegosouzapw](https://github.com/diegosouzapw)               | maintainer             |
+
+---
+
 ## [3.8.45] — 2026-07-06
 
 ### ✨ New Features

docs/i18n/in/CHANGELOG.md

@@ -6,6 +6,213 @@
 
 ## [3.8.31] — 2026-06-20
 
+## [3.8.46] — TBD
+
+### ✨ New Features
+
+- **feat(sse):** **hide paid-only models from `auto/*` routing** when `hidePaidModels` is on ([#6512](https://github.com/diegosouzapw/OmniRoute/issues/6512)) — follow-up to #6328/#6495. PR #6495 hid paid-only models from the `GET /v1/models` listing, but `auto/*` combos (`auto/best-coding`, `auto/glm`, …) could still pick a paid-only backend into their candidate pool → a 402/403 at request time. `createVirtualAutoCombo` now filters the candidate pool through the new pure `open-sse/services/autoCombo/paidModelFilter.ts` (`filterPaidOnlyCandidates`), applying the same free-model predicate #6495 uses in `catalog.ts` (`providerHasFreeModels(provider) && isFreeModel(provider, {id})`) whenever `settings.hidePaidModels === true`. Applied before the category/tier/family narrowing, so it covers every `auto/*` combo; an all-paid pool degrades to the existing graceful empty-pool path. **Opt-in — default OFF leaves the pool unchanged** (identity). Regression guard: `tests/unit/autoCombo/paid-model-filter-6512.test.ts` (4, incl. the default-off identity guard).
+- **feat(sse):** **provider-family auto combos** — `auto/glm`, `auto/minimax`, `auto/mimo`, `auto/zai`, `auto/gemma`, `auto/llama`, `auto/gemini` ([#6453](https://github.com/diegosouzapw/OmniRoute/issues/6453)) — new routable ids that materialize an on-demand virtual combo spanning whatever installed backends currently expose that model family, degrading gracefully as backends rotate. A new pure `open-sse/services/autoCombo/modelFamily.ts` (`detectModelFamily`) classifies by model-id prefix for six families; `zai` is instead resolved by provider id (z.ai's hosted API serves the same `glm-*` model ids as every other GLM backend, so `auto/zai` means "route to my z.ai backend specifically" vs `auto/glm`'s "any connected GLM backend"). Reuses the existing `createVirtualAutoCombo` on-demand materialization path (no DB writes) and the `/v1/models` catalog advertising loop. Regression guard: `tests/unit/autoCombo/provider-family-combos.test.ts` (11).
+- **feat(proxy):** native **proxy-pool round-robin / egress IP rotation** ([#6365](https://github.com/diegosouzapw/OmniRoute/issues/6365)) — a scope (global / provider / account) can now hold **multiple** proxies as a pool with a rotation strategy, so outbound requests cycle their egress IP instead of pinning one proxy per scope. Migration `117_proxy_pool_rotation.sql` lifts the `UNIQUE(scope, scope_id)` constraint (rebuild via the canonical rename/copy/drop; existing single assignments become 1-element pools) and adds a `proxy_scope_rotation` companion table holding the per-scope strategy + a persisted monotonic round-robin cursor. Strategies: `round-robin` (default, monotonic cursor — never `Math.random`), `random`, and `sticky-per-N-min`. Resolution (`resolveProxyForScopeFromRegistry` / `resolveProxyForConnectionFromRegistry`) now fetches the alive, position-ordered candidate set (unchanged `PROXY_ALIVE_PREDICATE`) and applies the strategy; an empty / all-dead pool still returns `null` — the #6246 fail-closed guard is **untouched** (never falls through to direct egress). Backend + DB only; dashboard pool-builder UI is a follow-up. Regression guard: `tests/unit/proxy-pool-rotation-6365.test.ts` (8, incl. fail-closed + backward-compat).
+- **feat(providers):** end-to-end **tool/function calling on the native Gemini `/v1beta` endpoint** ([#6222](https://github.com/diegosouzapw/OmniRoute/issues/6222)) — both directions of the Gemini↔OpenAI conversion now preserve tool data (previously silently dropped). Request side: `convertGeminiToInternal` (extracted to its own testable module) maps `tools[].functionDeclarations` → OpenAI `tools`, prior `functionCall` parts → assistant `tool_calls`, and `functionResponse` parts → `tool`-role messages. Response side: `convertOpenAIResponseToGemini` emits `parts[].functionCall {name,args}` from `message.tool_calls`, and the streaming `openAIChunkToGeminiChunk` accumulates fragmented `tool_calls` deltas by index into complete `functionCall` parts. The non-Gemini client paths (Claude, OpenAI-Responses) already preserved tool calls — this closes the gap specific to the native Gemini surface. Regression guard: `tests/unit/v1beta-gemini-tool-calling-6222.test.ts` (6, incl. a streaming SSE round-trip).
+- **feat(providers):** copilot-m365-web **enterprise / work tier** support ([#6334](https://github.com/diegosouzapw/OmniRoute/issues/6334)) — mirrors the EDU-tier pattern (#6210): `M365ConnectionParams` gains an `agent` field, a new opt-in `M365_ENTERPRISE_OVERRIDES` preset (`agent=work`, `scenario=officeweb`, `licenseType=Premium`) applies via `providerSpecificData.tier="enterprise"` (alias `"work"`), and `agent` is also overridable directly via `providerSpecificData.agent`. `buildWsUrl` was hardcoding `agent="web"` (the one enterprise-distinguishing param with no override path), so a Premium work account handshook then returned an empty stream. The individual and EDU paths are untouched. Kilo's dup flag vs #6210 (EDU tier) was a false positive — different tier. Regression guard: `tests/unit/copilot-m365-enterprise-6334.test.ts` (7). End-to-end confirmation on a real Premium work account is a live-VPS validation follow-up (Hard Rule #18). (thanks @Forcerecon)
+- **feat(api):** standardized, provider-agnostic **`effort` + `thinking` request params** ([#6241](https://github.com/diegosouzapw/OmniRoute/issues/6241)) — a thin standardization layer over the existing mature per-provider reasoning plumbing (no provider mapper touched). `providerChatCompletionSchema` gains a canonical `effort` (reusing the shared `none/low/medium/high/xhigh` vocabulary — the UI tiers `extra`/`max` collapse onto `xhigh`) and a boolean `thinking`. A pure `normalizeReasoningRequest` (wired once in `src/sse/handlers/chat.ts`, before any reasoning field is read) folds them onto the fields the translators already consume (`reasoning_effort` / `reasoning.effort` / `thinking`), so they fan out to Anthropic / Gemini / xAI / Responses — an explicit client `reasoning_effort` / object-shaped `thinking` always wins (backward-compatible). `/models` additively exposes `supportsThinking` + `effort_tiers` so the frontend can render the toggles (UI component is a follow-up). Regression guard: `tests/unit/effort-thinking-standardization-6241.test.ts` (12). (thanks @Iammilansoni, @shabeer)
+- **feat(combo):** new **`pipeline` (sequential) combo strategy** ([#6297](https://github.com/diegosouzapw/OmniRoute/issues/6297)) — the 18th routing strategy runs targets **in order**, threading each step's output into the next step's input, with an optional per-step `prompt` (system instruction); only the final step's response is returned. Distinct from `fusion` (parallel fan-out + judge). Implemented as a self-contained `open-sse/services/pipeline.ts` (sibling to `fusion.ts`), dispatched from `combo.ts`; the step list reuses `combo.models` order and reads an optional `prompt` off each target (backward-compatible — ignored by every other strategy). Intermediate steps run non-streaming with tools stripped (complete prose to thread forward); the final step keeps the client's `stream` flag + tools. A failing/empty/unparseable intermediate step fails the whole pipeline explicitly via a sanitized error (never silently swallowed). Kilo's dup flag vs #563 was a false positive (that's model→chain selection; this is a sequential chain). Regression guard: `tests/unit/combo-pipeline-strategy.test.ts` (5). (thanks @ofekbetzalel)
+- **feat(ci):** `check:test-masking` now flags **inline-reimplemented prod conditions** ([#6348](https://github.com/diegosouzapw/OmniRoute/issues/6348)) — a new **report-only** subcheck (v2, 6A.10 family) catches the wrong-shape contract test: a test that recomputes the condition under test inline instead of importing/exercising the real function (the #6216 class, where `=== 500` → `>= 500` stayed green because the test re-implemented the branch). For each added/modified test file it warns when the file textually duplicates a ≥3-token conditional from a production file touched in the same PR **and** does not import the symbol/module owning it, via a pure, fixture-tested `findReimplementedConditions()` with an allowlist mirroring `assertReductionAllowlist`. Report-only for now (does not fail the gate) — to be promoted to blocking after a triage cycle. Regression guard: `tests/unit/check-test-masking.test.ts` (45).
+- **feat(sse):** per-connection routing override (native vs CLIProxyAPI) ([#6339](https://github.com/diegosouzapw/OmniRoute/issues/6339)) — the previously-dead `isCliproxyapiDeepModeEnabled` helper is now wired into `resolveExecutorWithProxy`: a single connection can opt itself into the CLIProxyAPI passthrough executor via `providerSpecificData.cliproxyapiMode="claude-native"`, with precedence **connection override > provider `upstream_proxy_config` mode > default**. `resolveExecutorWithProxy` now receives the resolved connection's `providerSpecificData` (threaded from `chatCore.ts`), so one connection can deep-route while the provider's other connections stay native — no DB schema change (the toggle rides in `providerSpecificData`). Also resolves the same-provider mixing ask in #6340. Regression guard: `tests/unit/chatcore-executor-proxy.test.ts` (9). (thanks @RaviTharuma)
+- **feat(dashboard):** "Add session cookie" modal now shows a prominent **"Open ‹host› →"** link to the provider's own site ([#6268](https://github.com/diegosouzapw/OmniRoute/issues/6268)) — every `-web` cookie-session provider (chatgpt-web, claude-web, gemini-web, kimi-web, lmarena, qwen-web, m365-copilot-web, …) renders a one-click external link (opening the provider's login/home page in a new tab) so operators no longer tab away to retype the URL mid-setup. The host resolves from a pure, unit-tested `resolveWebProviderHost()` (prefers `WEB_COOKIE_PROVIDERS[id].website`, falls back to the registry `baseUrl` origin); non-web providers render exactly as before. Kilo's dup flag vs #6265 (modal-too-small-on-1080p) was a false positive — distinct concern. Regression guard: `tests/unit/resolve-web-provider-host.test.ts` (5). (thanks @chirag127)
+- **feat(providers):** add **DigitalOcean** AI (serverless inference) as an OpenAI-compatible API-key provider ([#6373](https://github.com/diegosouzapw/OmniRoute/pull/6373)) — base `https://inference.do-ai.run/v1`, wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/digitalocean/`, `src/shared/constants/providers/apikey/inference-hosts.ts`). Regression guard: `tests/unit/digitalocean-provider.test.ts`. (thanks @newnol)
+- **feat(providers):** add **Huancheng Public API** (`hcnsec`) as an OpenAI-compatible regional provider ([#6410](https://github.com/diegosouzapw/OmniRoute/pull/6410)) — Xinjiang Huancheng Cybersecurity's public LLM platform (base `https://api.hcnsec.cn/v1`, free credits via daily check-ins), wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/hcnsec/`, `src/shared/constants/providers/apikey/regional.ts`). Regression guard: `tests/unit/hcnsec-provider.test.ts`. (thanks @UnrealAryan)
+- **feat(dashboard):** the web-session credential guide now shows an **"Open {host}" link** ([#6316](https://github.com/diegosouzapw/OmniRoute/pull/6316)) to the provider's sign-in site (derived from the provider `website` via `getProviderWebsiteHost`), so you can jump straight to the page where the cookie/session must be captured. Regression guard: `tests/unit/web-session-provider-link-6316.test.ts`. (thanks @jordansilly77-stack)
+- **feat(cerebras):** add the **Gemma 4 31B** model (`gemma-4-31b`) to the Cerebras registry + pricing table ([#6331](https://github.com/diegosouzapw/OmniRoute/pull/6331)). Regression guard extends `tests/unit/t28-model-catalog-updates.test.ts`. (thanks @backryun)
+- **feat(providers):** add **Yuanbao (web)** as a cookie-session provider ([#6196](https://github.com/diegosouzapw/OmniRoute/issues/6196)) — `yuanbao-web` (Tencent Yuanbao, `yuanbao.tencent.com`) with cookie-only auth (`hy_user`/`hy_token` + public agent id), SSE→OpenAI translation incl. `reasoning_content`, exposing DeepSeek V3/R1 + Hunyuan / Hunyuan-T1. Regression guard: `tests/unit/providers-yuanbao-web.test.ts`. `together-web` was **deferred** (no verifiable web-session endpoint — needs a captured request) and `huggingchat-web` **dropped** (the existing `huggingchat` already is a web-cookie provider). (thanks @chirag127)
+- **feat(providers):** route the built-in **agentrouter** through the dynamic Claude-Code wire image ([#6056](https://github.com/diegosouzapw/OmniRoute/issues/6056)) — a small static allow-set (`CC_WIRE_IMAGE_BUILTINS` in `open-sse/services/ccWireImageBuiltins.ts`), consulted by `isClaudeCodeCompatible` / `isClaudeCodeCompatibleProvider` / `applyFingerprint`, makes agentrouter adopt the CC wire-image headers + fingerprint **while guarding the CC baseUrl/auth branches** so it keeps its own registry `baseUrl` and `x-api-key` auth. Regression guard: `tests/unit/agentrouter-cc-wire-image.test.ts` (asserts the wire image is applied AND agentrouter's baseUrl/auth are preserved). Live WAF-acceptance against agentrouter.org is a VPS validation follow-up (Hard Rule #18).
+- **feat(providers):** **bulk-add API keys for Cloudflare Workers AI** ([#6174](https://github.com/diegosouzapw/OmniRoute/issues/6174)) — `cloudflare-ai` is removed from the bulk-add exclusion list and the bulk parser gains a 3-field `name|accountId|apiKey` mode; the bulk route now builds a **per-entry** `providerSpecificData` so each key carries its own `accountId` (fixing the previous shared-object reuse), and both the create + key-validation paths receive it. Regression guard: `tests/unit/bulk-api-key-parser-cloudflare.test.ts`. (thanks @muflifadla38)
+- **feat(dashboard):** routing/settings UX clarity ([#6147](https://github.com/diegosouzapw/OmniRoute/issues/6147)) — (1) weighted combos show the **effective routing share %** next to each weight when weights don't sum to 100 (`WeightTotalBar.tsx`); (2) the status widget's user-facing **"Cloud Sync" label is renamed** to "Remote Settings Sync" (`CloudSyncStatus.tsx`; internal ids/state untouched); (3) built-in providers gain an **opt-in advanced base-URL override** (`isBaseUrlOverrideEligibleProvider`, hidden behind an "Advanced" toggle, reusing the existing `providerSpecificData.baseUrl` persistence — not globally widened). Regression guard: `tests/unit/routing-settings-ux-6147.test.ts`.
+- **feat(combo):** add an option to **disable session stickiness**, per-combo or globally — round-robin / random combos can rotate to a different connection on every request instead of pinning a whole conversation to one connection by its first-message hash. Resolution precedence per-combo `config.disableSessionStickiness` → global `settings.disableSessionStickiness` → default `false` (preserves the #3825 prompt-cache/504 fix); gates **both** stickiness call sites in `open-sse/services/combo.ts`. Exposed as a global toggle (Combo Defaults) and a per-combo Inherit/on/off control. ([#6168](https://github.com/diegosouzapw/OmniRoute/issues/6168)) Regression guard: `tests/unit/combo-disable-session-stickiness.test.ts`. (thanks @RCrushMe)
+- **feat(docker):** add the `OMNIROUTE_NO_SUDO` env flag for root-less / user-namespaced deployments — the MITM cert-trust command path (`resolveSudoSpawn` in `src/mitm/systemCommands.ts`) now strips the leading `sudo` when the flag is truthy, in addition to the existing root / sudo-missing cases, so the Proxy Agent runs without `sudo` (the operator trusts the CA manually, e.g. via `NODE_EXTRA_CA_CERTS`). Argv-array `spawn` preserved — no shell interpolation (Hard Rule #13). ([#6122](https://github.com/diegosouzapw/OmniRoute/issues/6122)) Regression guard: `tests/unit/mitm-systemCommands-no-sudo.test.ts`. (thanks @powellnorma)
+- **feat(providers):** add **Requesty** as an OpenAI-compatible gateway provider (BYOK, base `https://router.requesty.ai/v1`, ~200 free requests/day) — wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/requesty/`, `src/shared/constants/providers/apikey/gateways.ts`). ([#6120](https://github.com/diegosouzapw/OmniRoute/issues/6120)) Regression guard: `tests/unit/requesty-provider.test.ts`. (thanks @chirag127)
+- **feat(dashboard):** add **configured-only / available-only filters** to the Free Provider Rankings page ([#6150](https://github.com/diegosouzapw/OmniRoute/issues/6150)) — hide providers you haven't configured, or whose connections are all rate-limited / out of quota, via server-side query params (`?configuredOnly` / `?availableOnly` on `GET /api/free-provider-rankings`) backed by a testable lib helper reusing the in-process connection state (no Redis). Both filters default off, so the default view is unchanged; this supersedes the earlier client-side "Configured Only" toggle (#6245) with an available-only dimension and unit-tested logic. Regression guard: `tests/unit/freeProviderRankings-filters.test.ts`.
+
+### 🔧 Bug Fixes
+
+- **fix(dashboard):** adding a second API key connection for the same provider no longer silently overwrites the first — the Add-API-key modal now derives a unique default connection name (`main`, then `main-2`, `main-3`, …) so the backend name-based upsert can't collide (#6499 — thanks @dilneiss).
+- **fix(compression):** the session-dedup engine now also deduplicates a large multi-line block repeated **within a single message** (intra-message dedup), not just across turns; the compression-preview API surfaces a `fallbackReason`, and the fusion panel reports how many models were rate-limited vs failed on total-panel failure (#6501 — thanks @chirag127).
+- **fix(compression):** a stacked-pipeline step naming an unregistered engine now surfaces a `validationErrors` entry instead of silently no-op'ing, so misconfigured pipelines are visible in the preview API (#6506 — thanks @chirag127).
+- **feat(usage):** add a **Codex reset-credit redemption** flow to the Provider Limits UI ([#6361](https://github.com/diegosouzapw/OmniRoute/pull/6361)) — a `useCodexResetCreditRedemption` hook + `/api/usage/codex-reset-credit` route + `codexResetCredits` lib let you redeem banked Codex reset credits from the quota card. Regression guard: `tests/unit/codex-reset-credits.test.ts`. (thanks @JxnLexn)
+- **feat(glm):** add **team-plan quota settings** for `glm-cn` connections ([#6351](https://github.com/diegosouzapw/OmniRoute/pull/6351)) — a dedicated `GlmTeamQuotaFields` form section (team quota id / limits) threaded through the Add/Edit connection modals, persisted via `providerSpecificData`, with the GLM usage service reading the team quota. Regression guards: `tests/unit/glm-team-quota.test.ts`, `provider-specific-data-schema.test.ts`. (thanks @hao3039032)
+- **feat(providers):** add **TinyFish** web-fetch/search support ([#6349](https://github.com/diegosouzapw/OmniRoute/pull/6349)) — a `tinyfish-fetch` executor + `/v1/web/fetch` route + MCP web-fetch tool, registered as a specialty-media provider with request-validation and a search-provider catalog entry. Regression guards: `tests/unit/executor-tinyfish-fetch.test.ts`, `web-fetch-handler.test.ts`, `mcp-web-fetch-tool.test.ts`, `provider-validation-tinyfish.test.ts`. (thanks @dtybnrj)
+- **fix(cli):** `omniroute launch-codex` now spawns `codex.cmd` through a shell on Windows (the npm `.cmd` shim is unresolvable by bare `spawn` → ENOENT), mirroring the qodercli Windows fix (#6263) ([#6312](https://github.com/diegosouzapw/OmniRoute/pull/6312)). Regression guard: `tests/unit/launch-codex-windows-spawn-6312.test.ts`. (thanks @swingtempo)
+- **fix(codex):** isolate the **Spark** quota from the shared Codex quota and stabilize the quota UI ordering / hydration so per-scope limits render consistently ([#6336](https://github.com/diegosouzapw/OmniRoute/pull/6336)). Regression guards: `tests/unit/codex-quota-selection-hydration.test.ts`, `provider-limits-ui.test.ts` + 3 more. (thanks @xz-dev)
+- **feat(api):** add a `hidePaidModels` setting that filters paid-only models out of the `/v1/models` catalog. Regression guard: `tests/unit/models-catalog-hide-paid.test.ts`. (thanks @chirag127)
+- **fix(api-manager):** the fallback model picker now preserves combos instead of dropping them when a primary model is unavailable ([#6443](https://github.com/diegosouzapw/OmniRoute/pull/6443)). Regression guard: `tests/unit/api-manager-page-static.test.ts`. (thanks @jmengit)
+- **fix(providers):** recoverable Antigravity / Cloud-Code (Gemini Code Assist) `403` responses ([#6452](https://github.com/diegosouzapw/OmniRoute/pull/6452)) are now classified as a retryable project-config error instead of a terminal account ban, so a fixable project/API-disabled 403 no longer forces a ~1-year cooldown / full OAuth reconnect. Regression guard: `tests/unit/errorclassifier-antigravity-403.test.ts`. (thanks @developerjillur)
+- **fix(mitm):** `sanitizeHeaders` now redacts `Set-Cookie` response headers so upstream session cookies never leak into logs / diagnostics ([#6451](https://github.com/diegosouzapw/OmniRoute/pull/6451)). Regression guard: `tests/unit/mitm-sanitize-headers.test.ts`. (thanks @developerjillur)
+- **fix(api):** `/api/compression/preview` now accepts `mode: "caveman"` and correctly handles stacked / zero-compression previews ([#6425](https://github.com/diegosouzapw/OmniRoute/issues/6425)). Regression guard: `tests/unit/api/compression-preview-caveman-and-stacked-6425.test.ts`. (thanks @chirag127)
+- **feat(providers):** add **Zed** hosted LLM aggregator as a native-app provider ([#6118](https://github.com/diegosouzapw/OmniRoute/pull/6118)) — OAuth sign-in via the Zed hosted flow, registered through the shared provider registry + executor. Regression guards: `tests/unit/zed-oauth-provider.test.ts`, `zed-import-utils.test.ts`, `zed-docker-detect.test.ts`, `mitm-handler-zed.test.ts`. VPS-validated via live operator login (Hard Rule #18).
+- **fix(oauth):** the Kiro SSO-cache auto-import now **preserves the IDC region** — cross-region Amazon Q / Kiro profiles imported from the SSO cache are no longer collapsed to the default region ([#6113](https://github.com/diegosouzapw/OmniRoute/pull/6113)). Regression guard: `tests/unit/kiro-auto-import-idc-2059.test.ts`. VPS-validated via live operator login (Hard Rule #18).
+- **fix(dashboard):** passthrough model aliases no longer collide when two namespaced model ids share a last segment (port from 9router#1850, [#6431](https://github.com/diegosouzapw/OmniRoute/pull/6431)). `enx/gpt-5.5` and `enx/codebuddy/gpt-5.5` both auto-generated the alias `gpt-5.5`, so the second model could never be added (the UI just alerted "alias already exists"). Aliases are now disambiguated deterministically — bare last segment when free, then parent-qualified (`codebuddy-gpt-5.5`), then a numeric suffix — while re-adding the exact same model id is still blocked. Regression guard: `tests/unit/passthrough-alias-1850.test.ts`. (thanks @arpicato)
+- **fix(translator):** preserve a Gemini `functionResponse` co-located with other parts (another `functionCall`, or trailing `text`) in the same content when translating **Gemini → OpenAI** ([#6376](https://github.com/diegosouzapw/OmniRoute/pull/6376)). `convertGeminiContent()` early-returned the tool message on the first `functionResponse` part, dropping any co-located parts; such contents are now pre-split (one tool message per `functionResponse`, emitted first, plus one message for the remaining parts). Regression guard: `tests/unit/gemini-to-openai-function-response.test.ts`. (thanks @warelik)
+- **fix(headroom):** detect a python interpreter managed by **mise / pyenv / asdf / conda** (port from 9router#2353, [#6382](https://github.com/diegosouzapw/OmniRoute/pull/6382)). Headroom's python probe (`src/lib/headroom/detect.ts`) searched a hardcoded `PATH`, but version managers expose their interpreters via shim dirs that only join `PATH` through interactive-shell activation — which the non-interactive server never runs, so a managed python (≥3.10) was invisible and Headroom reported it missing. The search path now prepends the well-known shim/bin dirs (`~/.local/share/mise/shims`, `~/.pyenv/shims`, `~/.asdf/shims`, `$CONDA_PREFIX/bin`, `~/.local/bin`, respecting `MISE_DATA_DIR`/`PYENV_ROOT`/`ASDF_DATA_DIR` when set), and a new `HEADROOM_PYTHON` env override lets operators point straight at their interpreter (mirroring `HEADROOM_URL`). Still shell-free (`execFileSync`). Regression guard: `tests/unit/headroom-detect.test.ts` (5). (thanks @loopyd)
+- **fix(executors):** strip the OpenAI-Codex/Claude-CLI `client_metadata` passthrough field for **NVIDIA** requests (port from 9router#1887, [#6411](https://github.com/diegosouzapw/OmniRoute/pull/6411)). NVIDIA's OpenAI-compatible wrapper rejects it with `400 Unsupported parameter`, the same class already handled for `cerebras`/`mistral`; `nvidia` (executor `default`) was missing from the strip allowlist so Codex/Claude-Code passthrough requests 400'd. Regression guard: `tests/unit/executor-default-strip-client-metadata.test.ts` (+nvidia case). (thanks @phidinhmanh)
+- **fix(translator):** strip the Claude-style `thinking` field for **NVIDIA `z-ai/glm-5.2`** (port from 9router#2023, [#6413](https://github.com/diegosouzapw/OmniRoute/pull/6413)). NVIDIA's OpenAI-compatible wrapper 400s on `thinking` (a Claude-format client routed here leaves a `thinking:{type:"adaptive"}`); the existing strip rule only dropped `reasoning`. Same class already handled for `minimax-m2.7`. Regression guard: `tests/unit/nvidia-minimax-thinking-strip.test.ts` (+glm-5.2 case). (thanks @phidinhmanh)
+- **fix(translator):** suppress the streamed `</think>` close marker for the **Antigravity IDE** client (port from 9router#1061, [#6415](https://github.com/diegosouzapw/OmniRoute/pull/6415)). On thinking-only turns Antigravity rendered a bare `</think>` as the sole visible content, tripping its loop-detection and wasting requests. Antigravity's UA (`vscode/<v> (Antigravity/<v>)`) is added to the marker-suppress allowlist (alongside OpenCode); Claude Code / Cursor still get the marker, and `x-omniroute-thinking-marker: on` force-restores it. Regression guard: `tests/unit/think-close-marker-suppress-5245.test.ts`. (thanks @abdofallah)
+- **fix(executors):** strip nested `reasoning_content` from messages for **Mistral** (port from 9router#1649, [#6417](https://github.com/diegosouzapw/OmniRoute/pull/6417)). Mistral's API returns `422 extra_forbidden` when an assistant message carries `reasoning_content` (replayed thinking from a prior turn, e.g. via the Codex `/responses` path); the generic top-level 400 field-downgrade retry never covered the nested per-message field. `DefaultExecutor` now strips it for provider `mistral` only, so DeepSeek (which requires replayed `reasoning_content`) is unaffected. Regression guard: `tests/unit/mistral-strip-reasoning-content-1649.test.ts`. (thanks @xxy9468615)
+- **fix(executors):** strip the `client_metadata` passthrough field on the **OpenCode** path (port from 9router#1442, [#6418](https://github.com/diegosouzapw/OmniRoute/pull/6418)). OpenCode upstreams (e.g. `kimi-k2.6` via opencode-go) reject it with `400 "Extra inputs are not permitted, field: 'client_metadata'"`; the DefaultExecutor strip only covered cerebras/mistral and `OpencodeExecutor` extends `BaseExecutor` directly, so nothing removed it there. Regression guard: `tests/unit/opencode-strip-client-metadata-1442.test.ts`. (thanks @yanpaing007)
+- **fix(executors):** inject the `reasoning_content` echo for the native **Moonshot Kimi** provider (port from 9router#1480, [#6419](https://github.com/diegosouzapw/OmniRoute/pull/6419)). Kimi (executor `default`) is a thinking-mode upstream that 400s with "reasoning_content must be passed back" when a prior assistant turn lacks it; the placeholder injection was only wired into the OpenCode meta-provider, so direct multi-turn Kimi conversations failed. Scoped to `kimi` (gateway-served models matching the thinking-model name pattern are unaffected). Regression guard: `tests/unit/kimi-native-reasoning-injected-1480.test.ts`. (thanks @2220258345)
+- **fix(executors):** recover from a strict gateway's `context_management: Extra inputs are not permitted` 400 (port from 9router#1468, [#6420](https://github.com/diegosouzapw/OmniRoute/pull/6420)). **Claude Code** always sends a top-level `context_management` field; strict anthropic-compatible gateways reject it. The dedicated context-editing 400-fallback only fired when OmniRoute's own `contextEditing` feature was enabled (default off), so a client-sent field passed through untouched and 400'd. `context_management` is now in the generic reactive field-strip list, so it's stripped-and-retried once regardless of the feature flag (with correct request re-signing for claude-compatible relays). Regression guard: `tests/unit/provider-field-strips.test.ts`. (thanks @ohahe52-dot)
+- **fix(network):** enable **RFC 8305 Happy Eyeballs** (`autoSelectFamily`) on the direct-egress undici dispatcher (port from 9router#1237, [#6423](https://github.com/diegosouzapw/OmniRoute/pull/6423)). When DNS returns both IPv6 (AAAA) and IPv4 (A) and the IPv6 route is broken (e.g. a NAT64 `64:ff9b::` prefix without routing), undici tried IPv6 first and hung until `ETIMEDOUT` (then a 502 + account lockout), even though `curl` reached the same host. The direct dispatcher now races both families and uses whichever connects first. Proxy paths pin family via `proxyTls` and are unaffected. Regression guard: `tests/unit/direct-dispatcher-pipelining-4580.test.ts`. (thanks @adentdk)
+- **fix(combo):** round-robin now advances the rotation pointer past the model that **actually served**, not the eagerly-scheduled one (port from 9router#948, [#6428](https://github.com/diegosouzapw/OmniRoute/pull/6428)). With `stickyLimit: 1` (true round-robin), when the scheduled model failed and a _different_ model served via fallback, the counter had already advanced +1 from the scheduled index — so the next request reused the fallback-served model, degrading round-robin into hot-spotting on whichever model was healthy. The pointer now advances to the served index + 1 (mirroring the sticky-limit>1 path). Session-stickiness (#3825) and distribution are preserved. Regression guard: `tests/unit/combo-rr-fallback-advance-948.test.ts`. (thanks @binsarjr)
+- **fix(sse):** a non-string `model` field is now rejected with a `400` before the resolver, instead of crashing downstream `.toLowerCase()`/`.split()` calls into an empty-body `500` that escapes the error sanitizer ([#6407](https://github.com/diegosouzapw/OmniRoute/issues/6407)). Regression guard: `tests/unit/chat-non-string-model-6407.test.ts`. (thanks @chirag127)
+- **fix(api):** unknown `/api/*` routes now return a JSON `404` (instead of the dashboard HTML shell) and scalar chat params (`model`/`temperature`/etc.) are validated **before** the provider lookup so malformed requests fail fast with a clear `400` ([#6424](https://github.com/diegosouzapw/OmniRoute/issues/6424), [#6412](https://github.com/diegosouzapw/OmniRoute/issues/6412)). Regression guards: `tests/unit/api/api-catchall-json-404.test.ts`, `tests/unit/chat-early-schema-validation-6412.test.ts`. (thanks @chirag127)
+- **fix(api):** `/v1/chat/completions` now rejects a non-JSON `Content-Type` with a `400` before parsing the body ([#6414](https://github.com/diegosouzapw/OmniRoute/issues/6414)). Regression guard: `tests/unit/v1-chat-completions-content-type-6414.test.ts`. (thanks @chirag127)
+- **fix(api):** the `X-OmniRoute-Compression` response header is now echoed on `/v1/chat/completions` and `/v1/completions` ([#6422](https://github.com/diegosouzapw/OmniRoute/issues/6422)). Regression guard: `tests/unit/compression-header-echo-6422.test.ts`. (thanks @chirag127)
+- **fix(api):** concurrent `GET /v1/models` requests are coalesced into a single catalog build ([#6408](https://github.com/diegosouzapw/OmniRoute/issues/6408)). Regression guard: `tests/unit/v1-models-concurrent-6408.test.ts`. (thanks @chirag127)
+- **fix(api):** `/v1/completions` now echoes the requested `body.model` in its JSON + streamed responses ([#6429](https://github.com/diegosouzapw/OmniRoute/pull/6429)). Regression guard: `tests/unit/completions-body-model-echo.test.ts`. (thanks @chirag127)
+- **fix(api):** env-var master keys now see the full `/v1/models` catalog ([#6406](https://github.com/diegosouzapw/OmniRoute/issues/6406)). Regression guard: `tests/unit/models-catalog-envkey-6406.test.ts`. (thanks @chirag127)
+- **fix(api):** non-streaming `/v1/completions` responses now echo `body.model` aligned with the `X-OmniRoute-Model` header ([#6426](https://github.com/diegosouzapw/OmniRoute/issues/6426)). Regression guard: `tests/unit/v1-completions-model-header-match-6426.test.ts`. (thanks @chirag127)
+- **fix(api):** unknown `/v1/*` routes now return a JSON `404 not_found` instead of the Next.js dashboard HTML shell ([#6405](https://github.com/diegosouzapw/OmniRoute/issues/6405)). Regression guard: `tests/unit/api/v1-catchall-json-404.test.ts`. (thanks @chirag127)
+- **fix(api):** the per-connection provider models route now degrades to the shipped catalog when a provider's `/models` endpoint answers with a redirect ([#6267](https://github.com/diegosouzapw/OmniRoute/issues/6267)) — a `qwen-web` import failed with a raw `Redirect blocked … (307)` 503. `safeOutboundFetch` throws `REDIRECT_BLOCKED` on the 307, `getSafeOutboundFetchErrorStatus` maps it to 503, and `buildDiscoveryErrorFallbackResponse` treated every 503 as a hard error — so the non-empty `getModelsByProviderId("qwen-web")` catalog was never surfaced. A models-endpoint redirect is not a fixable-config error (unlike `URL_GUARD_BLOCKED`/`INVALID_URL`, which stay hard errors), so it now falls back to the local/cached catalog before the 503 short-circuit. General fix — covers any config-driven provider that 307s. Regression guard: `tests/unit/provider-models-qwen-web-redirect-6267.test.ts`. (thanks @chirag127)
+- **fix(api):** the per-connection provider models route (MCP `list_models_catalog` + the dashboard import view) now merges USER-ADDED custom models into its response ([#6247](https://github.com/diegosouzapw/OmniRoute/issues/6247)) — custom models live in the `key_value` namespace `customModels`, which the live REST `/api/v1/models` already merges, but `src/app/api/providers/[id]/models/route.ts` never read `getCustomModels`, so custom models were dropped on both the discovery-success and local_catalog paths. They are now folded into the returned model list (deduped by id, stamped `owned_by: provider`), fixing MCP + the dashboard import view in one place. Regression guard: `tests/unit/provider-models-custom-merge-6247.test.ts`. (thanks @RCrushMe)
+
+- **fix(providers):** GitLab Duo tool-calling follow-up turns no longer fail upstream with `422 {"detail":"Validation error"}` (tokens 0/0, rejected pre-inference). The #6234 tool-result-feedback fix serialized the **entire** multi-turn conversation into GitLab's single-file `code_suggestions` (`small_file`) generation endpoint — folded history that turn-N sent as an oversized `current_file.content_above_cursor` **and** duplicated verbatim into `user_instruction`, tripping the AI-Gateway's `small_file` validation guard. The executor now **bounds** that prompt: it keeps system + latest user message + the most-recent tool round (dropping older turns), caps oversized tool results, and stops duplicating the full prompt into `user_instruction` (which now carries only the short latest user message) — while still feeding the most-recent tool result back so the agent continues (`open-sse/executors/gitlab.ts`, [#6220](https://github.com/diegosouzapw/OmniRoute/issues/6220)). The unit test covers the bounding logic; the upstream 422→200 clearing is VPS-only (Hard Rule #18). Regression guard: `tests/unit/gitlab-tool-exchange-bounded-6220.test.ts`.
+
+- **fix(i18n):** the provider-detail (`/dashboard/providers/[id]`) connection-status filter labels no longer render as `__MISSING__:All` / `__MISSING__:Active` / `__MISSING__:Error` / `__MISSING__:Banned` / `__MISSING__:CreditsExhausted` in non-English locales (notably pt-BR) ([#6290](https://github.com/diegosouzapw/OmniRoute/issues/6290)). Root cause was **not** the namespace mismatch the issue guessed — the `providers.filter*` keys resolve correctly in `en.json`; the debt lived in the locale mirrors (`src/i18n/messages/*.json`), where these five keys carried the `__MISSING__:` sync sentinel in ~15 locales and were absent entirely in ~26 others, so next-intl found the key and echoed the sentinel verbatim. All 40 non-English/-Chinese mirrors now ship real translations for the five `providers.filter*` labels. Regression guard: `tests/unit/i18n-provider-filter-keys-6290.test.ts`. (thanks @diegosouzapw)
+
+- **fix(providers):** the `copilot-m365-web` streaming executor now emits `debug`-level WebSocket diagnostics ([#6210](https://github.com/diegosouzapw/OmniRoute/issues/6210)) — the outbound WS URL (with the `access_token` **redacted** via `redactWsUrl()`), handshake success/failure, and each received SignalR frame's `type`/`target`. Previously the streaming path logged nothing, so an empty `content:null` response (the M365 Education / Starter tier symptom fixed in #6234) was undiagnosable even at `APP_LOG_LEVEL=debug`. The change is debug-level and side-effect-free — it does not alter streaming behavior or the frame parser, and the token never reaches the logs. Regression guard: `tests/unit/copilot-m365-web-logging-6210.test.ts` (thanks @qpeyba)
+
+- **fix(resilience):** a round-robin combo no longer returns `503 all upstream accounts are unavailable` when a compatibility-rejected target is actually healthy ([#6238](https://github.com/diegosouzapw/OmniRoute/issues/6238)). `filterTargetsByRequestCompatibility` drops request-incompatible targets (tool/vision/structured-output unsupported, or below the required context window) **before** any availability check runs, and its `compatible.length === 0` safety net only fired when _all_ targets were filtered — not when the kept targets later all turned out runtime-unavailable (circuit-open / cooldown / no credentials). So a combo could 503 while a compat-rejected-but-healthy provider sat unused. `handleRoundRobinCombo` now keeps the compat-rejected set and, when every compat-kept target was skipped without a single real attempt, probes those rejected targets as a **last-resort fallback tier** (via the new pure `open-sse/services/combo/comboCompatFallback.ts`) before crystallizing the 503. Regression guard: `tests/unit/combo-roundrobin-compat-fallback-6238.test.ts`. (thanks @ThongAccount)
+
+- **fix(startup):** best-effort self-heal for a corrupted Turbopack dev cache on Windows ([#6289](https://github.com/diegosouzapw/OmniRoute/issues/6289)). On Windows, `pnpm dev` can fail at startup when Turbopack `mmap`s a persistent-cache SST file and the OS refuses the mapping (`os error 1455` — "paging file too small"), which Turbopack surfaces as a misleading `Module not found: Can't resolve '@/shared/utils/machine'`. This is a **known upstream Turbopack cache-corruption bug — not our code**. The dev launcher (`scripts/dev/run-next.mjs`) now wraps `nextApp.prepare()` and, when it rejects with that signature (`isTurbopackCacheCorruption` in the new `scripts/dev/turbopackCacheHeal.mjs`), purges `.build/next/**/cache/turbopack` and retries **once** with a clear log. **Caveat — best-effort only:** the corruption often surfaces as a runtime overlay rather than a `prepare()` rejection, so this cannot always intercept it; the reliable remedy remains manually deleting the Turbopack cache dir. Regression guard: `tests/unit/turbopack-cache-heal-6289.test.ts`. (thanks @chirag127)
+
+- **fix(providers):** qodercli PAT auth no longer fails with `spawn qodercli ENOENT` on Windows ([#6263](https://github.com/diegosouzapw/OmniRoute/issues/6263)) — `spawnQoderCli` spawned the bare `qodercli` name with `shell:false` and an unenriched env, so the npm `.cmd` wrapper under `%APPDATA%\npm` (a user-PATH directory) was never resolved. It now resolves the absolute `.cmd`/`.exe` path through the existing `getCliRuntimeStatus("qoder")` resolver in `src/shared/services/cliRuntime.ts` (memoized), spawns with `shell` when the target is a `.cmd`/`.bat`, and uses the cliRuntime-enriched env (PATH + PATHEXT + APPDATA); the ENOENT error now lists the searched paths plus the `CLI_QODER_BIN` override. End-to-end spawn on a real Windows host is host-only (Hard Rule #18); the path-resolution logic is unit-tested. Regression guard: `tests/unit/qodercli-windows-resolve-6263.test.ts`. (thanks @chirag127)
+
+- **fix(sse):** the reasoning-token buffer no longer inflates **probe-sized `max_tokens`** ([#6274](https://github.com/diegosouzapw/OmniRoute/issues/6274)) — Claude Code's `/model` capability check sends `max_tokens: 1`, but for a thinking-capable model with a large output cap (e.g. `glm-5.2`) the #3587 headroom heuristic (`max(current + 1000, ceil(current * 1.5))`) rewrote it to `1001` and forwarded that upstream, wasting tokens on a request that was never a genuine reasoning budget. `resolveReasoningBufferedMaxTokens()` (`open-sse/services/reasoningTokenBuffer.ts`) now short-circuits and returns the caller's value verbatim when it is below the new `REASONING_BUFFER_MIN_TRIGGER` (256) threshold — a tiny explicit limit is a probe, not a reasoning request. Real budgets still receive the #3587 headroom unchanged, and the guard runs after the existing capability checks so unknown / non-reasoning models keep returning `null`. Regression guard: `tests/unit/reasoning-token-buffer-6274.test.ts`. (thanks @brightfiscalband)
+
+- **fix(cli):** `omniroute reset-password` now works as a real subcommand, and password resets over piped (non-TTY) stdin actually apply ([#6261](https://github.com/diegosouzapw/OmniRoute/issues/6261), [#6258](https://github.com/diegosouzapw/OmniRoute/issues/6258)). Two coupled defects: (1) **#6261** — `bin/omniroute.mjs` routed everything through Commander with only two pre-Commander bypasses (`--mcp`, `reset-encrypted-columns`), so `omniroute reset-password` was rejected as an unknown command; only the separate `omniroute-reset-password` bin worked, while the docs falsely advertised the subcommand (incl. a bogus "legacy alias still works"). A pre-Commander bypass mirroring `reset-encrypted-columns` now dynamically imports `bin/reset-password.mjs` (which self-executes) before Commander parses; the three doc lines were corrected. (2) **#6258** — `bin/reset-password.mjs` issued two sequential `rl.question` prompts; under piped stdin the second read never settled at EOF, so `main()` never reached `resetManagementPassword` and the reset was a silent no-op (both prompts printed, no success, password unchanged). The CLI now detects non-TTY stdin and reads it once (first line = password, second line = confirm if present, else reused), adds a `--password-stdin` flag (entire stdin is the password, no confirmation), and exits `0` explicitly so the success line always flushes; interactive TTY behavior is unchanged. Regression guard: `tests/unit/reset-password-cli-6261-6258.test.ts` (3). (thanks @chirag127)
+
+- **fix(db):** the mass-migration **safety abort** now tells the operator how to bypass it and stops flooding the log ([#6260](https://github.com/diegosouzapw/OmniRoute/issues/6260)) — after restoring a backup that wiped the migration tracking table, `runMigrations()` threw the abort on every downstream `ensureDbInitialized()`, re-logging the full banner 11+ times, and the message never mentioned the existing `OMNIROUTE_MAX_PENDING_MIGRATIONS` escape hatch. The abort text now appends a bypass hint (set `OMNIROUTE_MAX_PENDING_MIGRATIONS=0` in `server.env` / `DATA_DIR/.env`), and a new `MigrationSafetyAbortError` is memoized so repeated calls in the same process throw the same instance and emit a single concise line instead of the full cascade. Regression guard: `tests/unit/migration-safety-abort-6260.test.ts`. (thanks @chirag127)
+
+- **fix(auth):** importing a **distinct** Codex/ChatGPT OAuth `auth.json` is no longer falsely rejected as "already exists" when it belongs to a different user in the same workspace ([#6301](https://github.com/diegosouzapw/OmniRoute/issues/6301)). `findExistingCodexConnection` (in `src/lib/oauth/utils/codexAuthImport.ts`) deduped **only** on `providerSpecificData.workspaceId === accountId`, where `accountId` is the shared `chatgpt_account_id`/`tokens.account_id` — so two members of the same ChatGPT Team collapsed onto a single connection (409 `duplicate_account`). The id_token's `https://api.openai.com/auth` claim carries a per-user `chatgpt_user_id` alongside the workspace id (the device-flow path already persisted it as `chatgptUserId`, but the import path did not). Now `parseAndValidateCodexAuth` extracts `userId` (`chatgpt_user_id` → `user_id` → JWT `sub`) into `ParsedCodexAuth`, the create/update paths persist `chatgptUserId` in `providerSpecificData` (mirroring `codex.ts`), and dedup keys on `workspaceId` **AND** `chatgptUserId` — with a backward-compat fallback to legacy accountId-only matching when no stored connection for that workspace records a `chatgptUserId`, so genuinely-same accounts still dedup. Regression guard: `tests/unit/codex-auth-import-userid-dedup-6301.test.ts` (4). (thanks @anungma)
+
+- **fix(providers):** importing models for the **venice-web** provider no longer fails with a red "Provider venice-web does not support models listing" ([#6269](https://github.com/diegosouzapw/OmniRoute/issues/6269)). `venice-web` is a web-cookie provider with an executor but no upstream `/v1/models` endpoint and no registry `models`, so the models route fell through to the tail `400`. Mirroring the `jules`/`linkup-search`/`ollama-search` fix (#5569), it now ships a static local catalog entry in `src/lib/providers/staticModels.ts` — seeding the current Venice lineup (`venice-uncensored`, `llama-3.3-70b`, `qwen3-235b`, `qwen3-4b`, `deepseek-r1-671b`; Venice rotates its catalog, see docs.venice.ai/models/overview) — so the route returns `200` with `source:"local_catalog"`, `intentional:true`. Regression guard: `tests/unit/static-models-venice-web-6269.test.ts`. (thanks @chirag127)
+
+- **fix(api):** the specialty model catalogs (`/v1/embeddings`, `/v1/images`, `/v1/music`, `/v1/videos` model lists) are now derived from the **unified catalog filtered by a predicate** (`getSpecialtyModelsResponse`) instead of ad-hoc per-route logic, so they consistently respect active-credential visibility and stay in sync with the main catalog ([#6303](https://github.com/diegosouzapw/OmniRoute/pull/6303)). Regression guard: `tests/unit/specialty-model-catalog-routes.test.ts`. (thanks @makcimbx)
+- **fix(api):** the agent-bridge server route now resolves the MITM manager via a **dynamic `import("@/mitm/manager.runtime")`** so Turbopack does not statically pull the stub (or over-bundle the manager), and the agent-skills generator anchors its output base path with `path.join(process.cwd(), …)` so Turbopack's static analyzer stops tracing the whole project root ([#6329](https://github.com/diegosouzapw/OmniRoute/issues/6329), [#6366](https://github.com/diegosouzapw/OmniRoute/pull/6366)). Regression guard: `tests/unit/agent-bridge-server-route-dynamic-import.test.ts`. (thanks @Iammilansoni)
+- **fix(api):** internal probes (combo-test, cloud-sync verify) now pick a **management-scoped / allow-all API key** instead of naively grabbing `getApiKeys()[0]` — a restricted `self:usage` first row made the probe fail with "Model X is not allowed for this API key" even when the combo path was healthy (`pickApiKeyForInternalUse` in `src/lib/db/apiKeys.ts`). The API-manager model editor also falls back to `/api/models?all=true` when `/v1/models` is catalog-protected ([#6372](https://github.com/diegosouzapw/OmniRoute/pull/6372)). Regression guard: `tests/unit/pick-internal-api-key-6372.test.ts`. (thanks @jmengit)
+- **fix(live-ws):** the Live Dashboard WebSocket server now **rejects on bind failure** (e.g. `EADDRINUSE` when the API bridge already holds the port) instead of letting the error surface as an unhandled `error` event that crash-loops the process — the `error` listener is attached to `wss` (not `server`) and releases the EventBus subscription on a failed start ([#6324](https://github.com/diegosouzapw/OmniRoute/issues/6324)). Regression guard: `tests/unit/live-ws-eaddrinuse-6324.test.ts`. (thanks @vinayakkulkarni)
+- **fix(dashboard):** the Home provider-topology widget now trusts the live provider-metrics snapshot — it uses `topology.errorProvider` and live `activeRequests` directly instead of re-deriving state from a stale `lastErrorAt` or applying a frontend timeout filter, so the topology reflects real-time provider health ([#6322](https://github.com/diegosouzapw/OmniRoute/pull/6322)). Regression guard: `tests/unit/home-provider-topology-live-state.test.ts`. (thanks @xz-dev)
+- **fix(sse):** strip zero-width markers from streamed **tool-call arguments** — a follow-up to [#5857](https://github.com/diegosouzapw/OmniRoute/pull/5857). That PR removed injected zero-width joiners (U+200D) from streamed assistant text/reasoning but deliberately left tool-call argument JSON byte-exact. The request-side obfuscation (`open-sse/services/claudeCodeObfuscation.ts`) injects ZWJ into agent words — including the temp path inside the Bash tool description — and Claude models copy that verbatim into generated commands, which are delivered as tool-call arguments rather than assistant text. As a result the ZWJ survived and corrupted code blocks (e.g. a temp path rendered with an invisible joiner). Now `open-sse/handlers/responseSanitizer.ts` strips zero-width code points from tool-call argument strings at every emit site (OpenAI non-stream/stream chat `tool_calls` + legacy `function_call`, native Responses `function_call` items, the OpenAI→Responses conversion, and the native Responses streaming `response.function_call_arguments.delta/.done` events). Only zero-width code points are removed; JSON structure and all other bytes stay identical (no parse/restringify), so normal arguments remain byte-exact. Regression guard: 6 new cases in `tests/unit/response-sanitizer.test.ts` (suite 50/50).
+
+- **fix(nodejs):** the default app log path now resolves under `DATA_DIR` (`~/.omniroute/logs/application/app.log`) instead of `process.cwd()` ([#6197](https://github.com/diegosouzapw/OmniRoute/issues/6197)) — the globally-installed CLI runs from an arbitrary working directory, so anchoring the default to cwd made file logging silently write to (or no-op under) an unrelated directory, contradicting the documented `.env.example` default. `getAppLogFilePath()` now computes the default lazily via the pure `resolveDataDir()` resolver (honours a per-process `DATA_DIR`, no directory-creation side effect); an explicit `APP_LOG_FILE_PATH` still wins. Regression guard: `tests/unit/logenv-datadir-path-6197.test.ts` (3).
+
+- **fix(docker):** AgentBridge/`startMitm` no longer aborts in containers/headless when the Antigravity-default DNS step can't write `/etc/hosts` ([#6127](https://github.com/diegosouzapw/OmniRoute/issues/6127)), and the privileged command's stderr now reaches `app.log` instead of only a bare exit code hitting the toast ([#6198](https://github.com/diegosouzapw/OmniRoute/issues/6198)). The default DNS step (`addDNSEntry`) was called unguarded while cert install and the two sibling DNS steps were each best-effort — in the runtime Docker image (`USER node`, no `sudo`, read-only `/etc/hosts`) it threw `Command failed with code 1` out of `startMitmInternal` and killed the whole start, discarding the stderr. The three DNS steps are extracted into a best-effort `provisionDnsEntries()` where each failure is logged with the full `err` (stderr included, folded in by `systemCommands.ts`) and never aborts the start. Regression guard: `tests/unit/mitm-dns-graceful-degrade-6127.test.ts` (4).
+
+- **fix(providers):** copilot-m365-web now supports the M365 Education "Starter / OfficeWebIncludedCopilot" tier and no longer returns an empty `content:null` stream ([#6210](https://github.com/diegosouzapw/OmniRoute/issues/6210)). Two gaps: (1) `buildWsUrl()` hardcoded the individual-consumer scenario (`OfficeWebPaidConsumerCopilot`, `isEdu=false`) — the EDU tier is now opt-in via `providerSpecificData.tier="edu"`, emitting `scenario=OfficeWebIncludedCopilot`/`isEdu=true` (the individual path is unchanged); (2) the EDU/GPT-5.5 path streams deltas via `arguments[0].writeAtCursor` (incremental) instead of only `messages[].text` (accumulated snapshots), which the parser dropped — a new `accumulateBotContent()` folds both formats, with `type:2 item.result.message` as a last-resort fallback. Regression guard: `tests/unit/copilot-m365-edu-writeatcursor-6210.test.ts` (10). (thanks @qpeyba)
+
+- **fix(providers):** GitLab Duo executor now feeds tool results back into the prompt instead of looping ([#6220](https://github.com/diegosouzapw/OmniRoute/issues/6220)) — `buildPrompt()` branched only on `system`/`user` and took `userParts.at(-1)`, silently dropping the `assistant{tool_calls}` + `tool{result}` turns the client appended, so the reconstructed prompt was byte-identical to turn 1 and the model re-emitted the same `<tool>` call forever. When a tool exchange is present the full conversation is now serialized, folding each tool result back keyed by its `tool_call_id`; simple conversations keep the legacy shape. Complements the tool_call emission from [#6051](https://github.com/diegosouzapw/OmniRoute/issues/6051) (the `kilo-duplicate` label was a false positive — different, sequential defect). Regression guard: `tests/unit/gitlab-tool-result-feedback-6220.test.ts` (4).
+
+- **fix(providers):** opencode-go/opencode-zen can now synthesize the OpenCode CLI identity headers Cloudflare requires on VPS egress ([#5997](https://github.com/diegosouzapw/OmniRoute/issues/5997)) — on a datacenter VPS, `opencode.ai/zen/go/v1/chat/completions` 403s (HTML challenge) requests lacking CLI identity, while the reporter's control curl proved that `User-Agent: opencode-cli/1.0.0` + `x-opencode-client: cli` + `x-opencode-project: default` + fresh request/session UUIDs succeed. Opt-in via `OPENCODE_SYNTHESIZE_CLI_HEADERS=true` (values overridable via `OPENCODE_GO_USER_AGENT`/`OPENCODE_USER_AGENT`/`OPENCODE_CLIENT`/`OPENCODE_PROJECT`); it fills only headers the client did not already send. Kept **off by default** — the forward-only path is deliberate (fabricating a wrong value risks upstream rejection; a prior dedup regressed with `opencode/local`), so this replaces the fragile local header-injection shim without changing default behavior. Regression guard: `tests/unit/opencode-cli-headers-synthesis-5997.test.ts` (6). (thanks @aleksesipenko)
+- fix(resilience): sticky session affinity now evicts and fails over to another account when the pinned account is exhausted/unavailable (#6219)
+
+- fix(sse): Responses API passthrough now drops internal commentary-phase output before forwarding to clients (gated by RESPONSES_PASSTHROUGH_DROP_COMMENTARY, default on) (#6199)
+
+- **fix(sse):** tool-call function schemas with a root `type: null` are now coerced to `type: "object"` before dispatch ([#6359](https://github.com/diegosouzapw/OmniRoute/issues/6359)) — clients like the Codex app emit `parameters: { type: null, ... }` for some tools, which OpenAI-compatible upstreams reject with `400 Invalid schema for function '...': schema must be a JSON Schema of 'type: "object"', got 'type: null'`, failing the whole request. `toolSchemaSanitizer` already stripped the null; it now re-adds the mandatory root `"object"` type (and empty `properties`/open `additionalProperties` when absent). Combinator roots (`anyOf`/`oneOf`/`allOf`) and explicit root types are left untouched. Regression guard: 5 new cases in `tests/unit/tool-schema-sanitizer.test.mjs`.
+
+- **fix(docker):** AgentBridge no longer fails to start on npm/Electron/VPS installs with "MITM manager stub reached at runtime" ([#6344](https://github.com/diegosouzapw/OmniRoute/issues/6344)) — v3.8.45 flipped the production bundler default to Turbopack, but `next.config.mjs` aliased `@/mitm/manager` to its Docker-only degraded stub **unconditionally**. That was harmless while Docker (which sets the alias intentionally for #3390 graceful degradation) was the sole Turbopack consumer, but once every artifact built with Turbopack the stub shipped to all non-Docker users and `startMitm` threw on the first Agent-Bridge start. The alias is now opt-in via `OMNIROUTE_MITM_STUB=1` (set only by the Dockerfile) through the shared `scripts/build/mitm-stub-flag.mjs` helper; default builds bundle the real manager. Regression guard: `tests/unit/mitm-stub-alias-6344.test.mjs` (4).
+
+- **fix(proxy):** stop the v3.8.44 proxy regression that leaked the real IP and disabled healthy proxies ([#6246](https://github.com/diegosouzapw/OmniRoute/issues/6246)). Two coupled defects from the new health scheduler: (1) **IP leak** — when a proxy assigned to a connection was marked `inactive`, resolution fell through to a **direct** egress instead of blocking, exposing the operator's real IP; (2) **over-deactivation** — the sweep flipped a proxy to `inactive` on the **first** failed probe and counted our own 5s timeout / a probe-target `5xx` as the proxy's fault, so healthy paid proxies vanished from egress selection ("my proxies are not being used anymore"). Fix: the sweep decision is extracted into a pure, network-free `decideProxyHealthAction` (`src/lib/proxyHealth/decision.ts`) — by default the health check now **only counts/logs and never downgrades status** (a proxy is downgraded/removed only with `PROXY_AUTO_REMOVE=true`, after `PROXY_AUTO_REMOVE_AFTER` **consecutive** conclusive failures); probes are classified tri-state so an inconclusive result (our timeout, or a `5xx` from the probe target) never penalizes the proxy, and the probe timeout is raised 5s→15s. Separately, `safeResolveProxy` now **fails closed** via the existing policy: a connection whose assigned proxy is dead is blocked instead of leaking direct (`hasBlockingProxyAssignment`), honoring the explicit `proxy off` toggles and the `PROXY_FAIL_OPEN=true` opt-out. Existing proxies stuck `inactive` by the old behavior need a one-time manual re-activate (the operator owns proxy status). Regression guards: `tests/unit/proxy-health-decide-action-6246.test.ts`, `tests/unit/proxy-assigned-unavailable-6246.test.ts`.
+
+- **fix(proxy):** make "Test All" read-only and add bulk enable/disable ([#6246](https://github.com/diegosouzapw/OmniRoute/issues/6246)). Complements the core fail-closed / scheduler fix (#6296) with the two remaining reporter asks. (1) The **"Test All" button** (`POST /api/settings/proxies/auto-test`) used to flip a proxy to `inactive` on a failed reachability probe; since the egress selector excludes `inactive` proxies, a flaky probe (an unreachable `httpbin.org`, a proxy that blocks `HEAD`, or a slow paid proxy) silently disabled every proxy that failed — "Test All" is now **read-only by default** (only the operator sets a proxy active/inactive; opt back into the legacy test-and-set with `PROXY_HEALTH_AUTO_DEACTIVATE=true`). (2) Adds a **bulk enable/disable** proxies endpoint + toolbar action (`POST /api/settings/proxies/batch-activate`) so an operator can re-activate proxies in one click. Regression guard: `tests/unit/proxy-health-6246.test.ts`. (thanks @tenshiak)
+
+- **chatcore (tools): stop the default 128-tool cap from silently dropping opencode's `task`/MCP tools.** opencode (used as an MCP/agent host) sends a large tool list; when it exceeds the speculative `MAX_TOOLS_LIMIT` (128) default, `truncateToolList` did a blind `tools.slice(0, 128)`, dropping every tool past index 128 — including opencode's built-in `task` tool (subagent launch) and many MCP tools, so models routed through OmniRoute could no longer spawn subagents or reach part of their tools. The cap exists to avoid upstream `400`s for providers with real hard limits (e.g. grok-cli 200), so it is kept for those: detection of the opencode client (`isOpencodeClient` — any `x-opencode-*` header, or `opencode` in the user-agent) now only bypasses the **speculative 128 default**, never a known provider ceiling. Precedence is explicit — a proactive/detected provider limit always truncates (even for opencode); otherwise opencode forwards its full tool list; otherwise the unchanged 128 default applies to every other client. Refactors `getEffectiveToolLimit` into `getKnownToolLimit(provider) ?? DEFAULT_LIMIT` (byte-identical for existing callers) and fixes a cosmetic debug-log that reported the truncated count instead of the original. Regression guard: `tests/unit/tool-limit-detector.test.ts`.
+
+- **fix(mitm):** the macOS MITM-cert install check now matches the system keychain again. `security find-certificate -a -Z` prints the SHA-1 as a colon-less hex string, but the installed-check compared it against `getCertFingerprint()`'s colon-separated form, so the substring match never hit — the cert was reported as not-installed and re-prompted for the sudo install on every run. Fingerprints are now normalized (colons stripped, upper-cased) on both sides via the extracted `macCertOutputHasFingerprint` helper. Regression guard: `tests/unit/mitm-cert-mac-fingerprint.test.ts`. ([#6204](https://github.com/diegosouzapw/OmniRoute/pull/6204), closes [#6134](https://github.com/diegosouzapw/OmniRoute/issues/6134) — thanks @rianonehub)
+
+- **fix(api):** `/v1/messages/count_tokens` now counts `tool_use`, `tool_result` and `thinking` content blocks (and array-form `system` prompts) in the local-estimation path, instead of only `text`. Real agentic conversations keep ~95% of their tokens inside tool results; the previous estimate returned near-zero for them, which silently broke Claude Code's auto-compaction (context grew past the window with no compaction until the upstream API rejected the request). The real provider-side count path is unchanged. Regression guard: `tests/unit/messages-count-tokens-route.test.ts`. ([#6221](https://github.com/diegosouzapw/OmniRoute/pull/6221) — thanks @luweiCN)
+
+- **fix(antigravity):** strip a trailing assistant prefill turn for Vertex Claude models to avoid upstream 400s ([#6114](https://github.com/diegosouzapw/OmniRoute/pull/6114)). Regression guard: `tests/unit/antigravity-claude-prefill-strip.test.ts`. (thanks @anki1kr)
+
+- **fix(security):** the mutable cloud-agent routes (`/api/cloud/credentials/update`, `/api/cloud/models/alias`) now require management auth instead of being treated as public. They were classified as public API routes, so a request without management credentials could update stored cloud-agent credentials and model aliases. They are removed from the public-route set, classified as management routes in the authz pipeline, and gated by `requireManagementAuth`; cloud **read**/auth routes stay public. Regression guards: `tests/unit/cloud-write-auth.test.ts`, `tests/unit/authz/classify.test.ts`, `tests/unit/public-api-routes.test.ts`. ([#6233](https://github.com/diegosouzapw/OmniRoute/pull/6233) — thanks @vittoroliveira-dev)
+
+- **refactor(dashboard):** extract the onboarding-wizard "Open provider details" link target into a pure, unit-tested `buildProviderDetailsHref(connection)` helper. The wizard already routes by `connection.id` (the node UUID) rather than the provider category slug (#6144/#6145); this hardens that behavior behind a tested helper that guards a missing id/connection. Regression guard: `tests/unit/provider-onboarding-href.test.ts`. ([#6166](https://github.com/diegosouzapw/OmniRoute/pull/6166) — thanks @KooshaPari)
+
+- **fix(security):** the doubao synthetic device-id generator now derives its digits via an unbiased crypto-random draw (rejection sampling over `crypto.randomBytes()`) instead of a `% 10` reduction, closing a CodeQL `js/biased-cryptographic-random` finding.
+- **fix(agentSkills):** the GitHub-skills generator now resolves `outputDir` to an absolute path before writing, fixing a regression introduced by #6366 (relative-to-cwd base path) that could write generated skill files to the wrong directory.
+- **fix(security):** `/api/keys/{id}/devices` now checks the HTTP method before auth/validation, returning a `405` for non-GET/DELETE verbs instead of a misleading `401`/`500` (closes a `dast-smoke` QUERY-method finding).
+- **fix(quality):** clear the last 2 heavy quality-gate reds on the release tip (cycle pre-flight).
+- **fix(mitm):** the test suite and CI can never mutate the OS trust store — `OMNIROUTE_SKIP_SYSTEM_TRUST=1` is set globally for tests/CI so `installCert`/`uninstallCert`/`installTproxyCa` skip the privileged OS dispatch ([#6310](https://github.com/diegosouzapw/OmniRoute/pull/6310); full detail is under the [3.8.45] section below — this branch received it via the parallel-cycle sync-back).
+- **fix(api):** `POST /api/github-skills` now Zod-validates its request body; documented the new quality-gate env vars and pinned the merge-integrity GitHub Actions to a commit SHA.
+- **fix(skills):** generate the missing `omni-github-skills` registry entry and align the agent-skills catalog-count tests (follow-up to #6186).
+- **fix(quality):** clear the cycle's 11 net-new ESLint errors and make `validate-release-green` suppressions-aware.
+
+### 📝 Maintenance
+
+- **i18n(it):** add 118 missing Italian (`it`) translations (net-additive — no existing keys dropped, valid JSON), improving Italian UI coverage. ([#6212](https://github.com/diegosouzapw/OmniRoute/pull/6212) — thanks @serverless83)
+- **chore(providers):** remove deprecated **MiMo V2** model entries from the catalogs (xiaomi-mimo, opencode-go, zenmux-free, audio TTS) — the upstream V2 line is superseded by MiMo V2.5; drops `mimo-v2-tts`, `mimo-v2-pro`, `mimo-v2-omni`, `mimo-v2-flash`, `mimo-v2-flash-free` and realigns the provider-catalog tests. ([#6248](https://github.com/diegosouzapw/OmniRoute/pull/6248) — thanks @backryun)
+- **chore(release):** ~50 commits on this branch are v3.8.45 pre-flight/hardening fixes and CI-perf work that landed here via the parallel-cycle sync-back (`sync-next-cycle.mjs`, Hard Rule #21) **after** the `v3.8.45` git tag was cut, and are already fully documented under the **[3.8.45]** section below — listed here only so the per-cycle commit-coverage check (`npm run release:uncovered`) doesn't flag them as gaps. Provider/catalog/UX/backend fixes: #6041, #6078, #6108, #6135, #6148, #6149, #6154, #6158, #6161, #6162, #6163, #6164, #6165, #6170, #6177, #6178, #6181, #6186, #6187, #6191, #6193, #6194, #6195, #6200, #6205, #6208, #6209, #6211, #6213, #6223, #6224, #6225, #6226, #6227, #6228, #6229, #6230, #6235, #6291, #6292. CI/release-pipeline work: #6167, #6203, #6214, #6215, #6218, #6273, #6275, #6283, #6284, #6285, #6300, #6305.
+- **chore(release):** additional zero-ref release-cycle plumbing on this branch, kept out of `release:uncovered` on purpose (no `#N` in the commit subject to cite): opening the v3.8.46 cycle, opening/closing the v3.8.45 cycle, the finalized [3.8.45] CHANGELOG i18n sync-back to 42 mirrors, the v3.8.45 cognitive/cyclomatic and file-size drift rebaselines, ESLint stale-suppression pruning (4,273 → 4,233), and clearing test-masking/docs-all pre-flight reds for v3.8.45.
+
+### ⚡ Performance & Infrastructure
+
+- **perf(release-green):** the pre-flight validator (`scripts/quality/validate-release-green.mjs`) now runs its 4 slow suites (unit / vitest / integration / pack-artifact) **concurrently** via `Promise.all` — pre-flight wall time drops from ~the sum of the suites to ~the slowest one (~30min saved per round; Phase 0 was the nº1 bottleneck of the v3.8.45 release benchmark, 2h54 of 6h34 e2e). Guard: `tests/unit/validate-release-green.test.ts` ("runs the slow suites CONCURRENTLY"). ([#6319](https://github.com/diegosouzapw/OmniRoute/pull/6319))
+- **fix(ci):** `scripts/release/sync-next-cycle.mjs` — two defects found live in its first production run (v3.8.45 Phase 5): (1) the `git()` helper's default 1 MiB `maxBuffer` crashed with `ENOBUFS` on `git show origin/main:CHANGELOG.md` (the CHANGELOG alone is >1 MiB) — widened to 64 MiB; (2) the i18n resync only propagated the `[NEXT]` (TBD) section, leaving the just-shipped finalized section as "— TBD" in all 42 mirrors — it now also syncs `[prevVersion]` bounded by the heading below it (new exported pure helper `versionAfter`). Guards: +5 tests in `tests/unit/sync-next-cycle.test.ts` (8/8). ([#6327](https://github.com/diegosouzapw/OmniRoute/pull/6327))
+- **test(ci):** concurrency-sensitive flaky tests are **quarantined into a serial pass** (`tests/unit/serial/`, `--test-concurrency=1`, appended to every unit runner incl. sharded variants — the serial pass is sharded too so concurrent shard jobs never self-collide). Initial set: `glm-coding-plan-monthly-3580`, `quota-division-blocks`, `provider-health-autopilot`, `combo-health-autopilot` — the class behind the ~28min CI wedges/re-runs (two live 1h+ wedges cancelled during this PR's own validation). Discovery + TIA gates track the new glob; systemic root cause (async logger writing after teardown) tracked in [#6360](https://github.com/diegosouzapw/OmniRoute/issues/6360). Guard: `tests/unit/test-serial-quarantine.test.ts` (4). ([#6347](https://github.com/diegosouzapw/OmniRoute/pull/6347))
+
+### 🙌 Contributors
+
+Thanks to everyone whose work landed in v3.8.46:
+
+| Contributor                                                    | PRs / Issues           |
+| -------------------------------------------------------------- | ---------------------- |
+| [@2220258345](https://github.com/2220258345)                   | direct commit / report |
+| [@abdofallah](https://github.com/abdofallah)                   | direct commit / report |
+| [@adentdk](https://github.com/adentdk)                         | direct commit / report |
+| [@aleksesipenko](https://github.com/aleksesipenko)             | direct commit / report |
+| [@anki1kr](https://github.com/anki1kr)                         | direct commit / report |
+| [@anungma](https://github.com/anungma)                         | direct commit / report |
+| [@arpicato](https://github.com/arpicato)                       | direct commit / report |
+| [@backryun](https://github.com/backryun)                       | #6248                  |
+| [@binsarjr](https://github.com/binsarjr)                       | direct commit / report |
+| [@brightfiscalband](https://github.com/brightfiscalband)       | direct commit / report |
+| [@chirag127](https://github.com/chirag127)                     | #6501, #6506           |
+| [@developerjillur](https://github.com/developerjillur)         | direct commit / report |
+| [@dilneiss](https://github.com/dilneiss)                       | #6499                  |
+| [@dtybnrj](https://github.com/dtybnrj)                         | direct commit / report |
+| [@Forcerecon](https://github.com/Forcerecon)                   | direct commit / report |
+| [@hao3039032](https://github.com/hao3039032)                   | direct commit / report |
+| [@Iammilansoni](https://github.com/Iammilansoni)               | #6150, #6245           |
+| [@jmengit](https://github.com/jmengit)                         | direct commit / report |
+| [@jordansilly77-stack](https://github.com/jordansilly77-stack) | direct commit / report |
+| [@JxnLexn](https://github.com/JxnLexn)                         | direct commit / report |
+| [@KooshaPari](https://github.com/KooshaPari)                   | #6166                  |
+| [@loopyd](https://github.com/loopyd)                           | direct commit / report |
+| [@luweiCN](https://github.com/luweiCN)                         | #6221                  |
+| [@makcimbx](https://github.com/makcimbx)                       | direct commit / report |
+| [@muflifadla38](https://github.com/muflifadla38)               | direct commit / report |
+| [@newnol](https://github.com/newnol)                           | direct commit / report |
+| [@ofekbetzalel](https://github.com/ofekbetzalel)               | direct commit / report |
+| [@ohahe52-dot](https://github.com/ohahe52-dot)                 | direct commit / report |
+| [@phidinhmanh](https://github.com/phidinhmanh)                 | direct commit / report |
+| [@powellnorma](https://github.com/powellnorma)                 | direct commit / report |
+| [@qpeyba](https://github.com/qpeyba)                           | direct commit / report |
+| [@RaviTharuma](https://github.com/RaviTharuma)                 | direct commit / report |
+| [@RCrushMe](https://github.com/RCrushMe)                       | direct commit / report |
+| [@rianonehub](https://github.com/rianonehub)                   | #6134, #6204           |
+| [@serverless83](https://github.com/serverless83)               | #6212                  |
+| [@swingtempo](https://github.com/swingtempo)                   | direct commit / report |
+| [@tenshiak](https://github.com/tenshiak)                       | direct commit / report |
+| [@ThongAccount](https://github.com/ThongAccount)               | direct commit / report |
+| [@UnrealAryan](https://github.com/UnrealAryan)                 | direct commit / report |
+| [@vinayakkulkarni](https://github.com/vinayakkulkarni)         | direct commit / report |
+| [@vittoroliveira-dev](https://github.com/vittoroliveira-dev)   | #6233                  |
+| [@warelik](https://github.com/warelik)                         | direct commit / report |
+| [@xxy9468615](https://github.com/xxy9468615)                   | direct commit / report |
+| [@xz-dev](https://github.com/xz-dev)                           | direct commit / report |
+| [@yanpaing007](https://github.com/yanpaing007)                 | direct commit / report |
+| [@diegosouzapw](https://github.com/diegosouzapw)               | maintainer             |
+
+---
+
 ## [3.8.45] — 2026-07-06
 
 ### ✨ New Features

docs/i18n/it/CHANGELOG.md

@@ -6,6 +6,213 @@
 
 ## [3.8.31] — 2026-06-20
 
+## [3.8.46] — TBD
+
+### ✨ New Features
+
+- **feat(sse):** **hide paid-only models from `auto/*` routing** when `hidePaidModels` is on ([#6512](https://github.com/diegosouzapw/OmniRoute/issues/6512)) — follow-up to #6328/#6495. PR #6495 hid paid-only models from the `GET /v1/models` listing, but `auto/*` combos (`auto/best-coding`, `auto/glm`, …) could still pick a paid-only backend into their candidate pool → a 402/403 at request time. `createVirtualAutoCombo` now filters the candidate pool through the new pure `open-sse/services/autoCombo/paidModelFilter.ts` (`filterPaidOnlyCandidates`), applying the same free-model predicate #6495 uses in `catalog.ts` (`providerHasFreeModels(provider) && isFreeModel(provider, {id})`) whenever `settings.hidePaidModels === true`. Applied before the category/tier/family narrowing, so it covers every `auto/*` combo; an all-paid pool degrades to the existing graceful empty-pool path. **Opt-in — default OFF leaves the pool unchanged** (identity). Regression guard: `tests/unit/autoCombo/paid-model-filter-6512.test.ts` (4, incl. the default-off identity guard).
+- **feat(sse):** **provider-family auto combos** — `auto/glm`, `auto/minimax`, `auto/mimo`, `auto/zai`, `auto/gemma`, `auto/llama`, `auto/gemini` ([#6453](https://github.com/diegosouzapw/OmniRoute/issues/6453)) — new routable ids that materialize an on-demand virtual combo spanning whatever installed backends currently expose that model family, degrading gracefully as backends rotate. A new pure `open-sse/services/autoCombo/modelFamily.ts` (`detectModelFamily`) classifies by model-id prefix for six families; `zai` is instead resolved by provider id (z.ai's hosted API serves the same `glm-*` model ids as every other GLM backend, so `auto/zai` means "route to my z.ai backend specifically" vs `auto/glm`'s "any connected GLM backend"). Reuses the existing `createVirtualAutoCombo` on-demand materialization path (no DB writes) and the `/v1/models` catalog advertising loop. Regression guard: `tests/unit/autoCombo/provider-family-combos.test.ts` (11).
+- **feat(proxy):** native **proxy-pool round-robin / egress IP rotation** ([#6365](https://github.com/diegosouzapw/OmniRoute/issues/6365)) — a scope (global / provider / account) can now hold **multiple** proxies as a pool with a rotation strategy, so outbound requests cycle their egress IP instead of pinning one proxy per scope. Migration `117_proxy_pool_rotation.sql` lifts the `UNIQUE(scope, scope_id)` constraint (rebuild via the canonical rename/copy/drop; existing single assignments become 1-element pools) and adds a `proxy_scope_rotation` companion table holding the per-scope strategy + a persisted monotonic round-robin cursor. Strategies: `round-robin` (default, monotonic cursor — never `Math.random`), `random`, and `sticky-per-N-min`. Resolution (`resolveProxyForScopeFromRegistry` / `resolveProxyForConnectionFromRegistry`) now fetches the alive, position-ordered candidate set (unchanged `PROXY_ALIVE_PREDICATE`) and applies the strategy; an empty / all-dead pool still returns `null` — the #6246 fail-closed guard is **untouched** (never falls through to direct egress). Backend + DB only; dashboard pool-builder UI is a follow-up. Regression guard: `tests/unit/proxy-pool-rotation-6365.test.ts` (8, incl. fail-closed + backward-compat).
+- **feat(providers):** end-to-end **tool/function calling on the native Gemini `/v1beta` endpoint** ([#6222](https://github.com/diegosouzapw/OmniRoute/issues/6222)) — both directions of the Gemini↔OpenAI conversion now preserve tool data (previously silently dropped). Request side: `convertGeminiToInternal` (extracted to its own testable module) maps `tools[].functionDeclarations` → OpenAI `tools`, prior `functionCall` parts → assistant `tool_calls`, and `functionResponse` parts → `tool`-role messages. Response side: `convertOpenAIResponseToGemini` emits `parts[].functionCall {name,args}` from `message.tool_calls`, and the streaming `openAIChunkToGeminiChunk` accumulates fragmented `tool_calls` deltas by index into complete `functionCall` parts. The non-Gemini client paths (Claude, OpenAI-Responses) already preserved tool calls — this closes the gap specific to the native Gemini surface. Regression guard: `tests/unit/v1beta-gemini-tool-calling-6222.test.ts` (6, incl. a streaming SSE round-trip).
+- **feat(providers):** copilot-m365-web **enterprise / work tier** support ([#6334](https://github.com/diegosouzapw/OmniRoute/issues/6334)) — mirrors the EDU-tier pattern (#6210): `M365ConnectionParams` gains an `agent` field, a new opt-in `M365_ENTERPRISE_OVERRIDES` preset (`agent=work`, `scenario=officeweb`, `licenseType=Premium`) applies via `providerSpecificData.tier="enterprise"` (alias `"work"`), and `agent` is also overridable directly via `providerSpecificData.agent`. `buildWsUrl` was hardcoding `agent="web"` (the one enterprise-distinguishing param with no override path), so a Premium work account handshook then returned an empty stream. The individual and EDU paths are untouched. Kilo's dup flag vs #6210 (EDU tier) was a false positive — different tier. Regression guard: `tests/unit/copilot-m365-enterprise-6334.test.ts` (7). End-to-end confirmation on a real Premium work account is a live-VPS validation follow-up (Hard Rule #18). (thanks @Forcerecon)
+- **feat(api):** standardized, provider-agnostic **`effort` + `thinking` request params** ([#6241](https://github.com/diegosouzapw/OmniRoute/issues/6241)) — a thin standardization layer over the existing mature per-provider reasoning plumbing (no provider mapper touched). `providerChatCompletionSchema` gains a canonical `effort` (reusing the shared `none/low/medium/high/xhigh` vocabulary — the UI tiers `extra`/`max` collapse onto `xhigh`) and a boolean `thinking`. A pure `normalizeReasoningRequest` (wired once in `src/sse/handlers/chat.ts`, before any reasoning field is read) folds them onto the fields the translators already consume (`reasoning_effort` / `reasoning.effort` / `thinking`), so they fan out to Anthropic / Gemini / xAI / Responses — an explicit client `reasoning_effort` / object-shaped `thinking` always wins (backward-compatible). `/models` additively exposes `supportsThinking` + `effort_tiers` so the frontend can render the toggles (UI component is a follow-up). Regression guard: `tests/unit/effort-thinking-standardization-6241.test.ts` (12). (thanks @Iammilansoni, @shabeer)
+- **feat(combo):** new **`pipeline` (sequential) combo strategy** ([#6297](https://github.com/diegosouzapw/OmniRoute/issues/6297)) — the 18th routing strategy runs targets **in order**, threading each step's output into the next step's input, with an optional per-step `prompt` (system instruction); only the final step's response is returned. Distinct from `fusion` (parallel fan-out + judge). Implemented as a self-contained `open-sse/services/pipeline.ts` (sibling to `fusion.ts`), dispatched from `combo.ts`; the step list reuses `combo.models` order and reads an optional `prompt` off each target (backward-compatible — ignored by every other strategy). Intermediate steps run non-streaming with tools stripped (complete prose to thread forward); the final step keeps the client's `stream` flag + tools. A failing/empty/unparseable intermediate step fails the whole pipeline explicitly via a sanitized error (never silently swallowed). Kilo's dup flag vs #563 was a false positive (that's model→chain selection; this is a sequential chain). Regression guard: `tests/unit/combo-pipeline-strategy.test.ts` (5). (thanks @ofekbetzalel)
+- **feat(ci):** `check:test-masking` now flags **inline-reimplemented prod conditions** ([#6348](https://github.com/diegosouzapw/OmniRoute/issues/6348)) — a new **report-only** subcheck (v2, 6A.10 family) catches the wrong-shape contract test: a test that recomputes the condition under test inline instead of importing/exercising the real function (the #6216 class, where `=== 500` → `>= 500` stayed green because the test re-implemented the branch). For each added/modified test file it warns when the file textually duplicates a ≥3-token conditional from a production file touched in the same PR **and** does not import the symbol/module owning it, via a pure, fixture-tested `findReimplementedConditions()` with an allowlist mirroring `assertReductionAllowlist`. Report-only for now (does not fail the gate) — to be promoted to blocking after a triage cycle. Regression guard: `tests/unit/check-test-masking.test.ts` (45).
+- **feat(sse):** per-connection routing override (native vs CLIProxyAPI) ([#6339](https://github.com/diegosouzapw/OmniRoute/issues/6339)) — the previously-dead `isCliproxyapiDeepModeEnabled` helper is now wired into `resolveExecutorWithProxy`: a single connection can opt itself into the CLIProxyAPI passthrough executor via `providerSpecificData.cliproxyapiMode="claude-native"`, with precedence **connection override > provider `upstream_proxy_config` mode > default**. `resolveExecutorWithProxy` now receives the resolved connection's `providerSpecificData` (threaded from `chatCore.ts`), so one connection can deep-route while the provider's other connections stay native — no DB schema change (the toggle rides in `providerSpecificData`). Also resolves the same-provider mixing ask in #6340. Regression guard: `tests/unit/chatcore-executor-proxy.test.ts` (9). (thanks @RaviTharuma)
+- **feat(dashboard):** "Add session cookie" modal now shows a prominent **"Open ‹host› →"** link to the provider's own site ([#6268](https://github.com/diegosouzapw/OmniRoute/issues/6268)) — every `-web` cookie-session provider (chatgpt-web, claude-web, gemini-web, kimi-web, lmarena, qwen-web, m365-copilot-web, …) renders a one-click external link (opening the provider's login/home page in a new tab) so operators no longer tab away to retype the URL mid-setup. The host resolves from a pure, unit-tested `resolveWebProviderHost()` (prefers `WEB_COOKIE_PROVIDERS[id].website`, falls back to the registry `baseUrl` origin); non-web providers render exactly as before. Kilo's dup flag vs #6265 (modal-too-small-on-1080p) was a false positive — distinct concern. Regression guard: `tests/unit/resolve-web-provider-host.test.ts` (5). (thanks @chirag127)
+- **feat(providers):** add **DigitalOcean** AI (serverless inference) as an OpenAI-compatible API-key provider ([#6373](https://github.com/diegosouzapw/OmniRoute/pull/6373)) — base `https://inference.do-ai.run/v1`, wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/digitalocean/`, `src/shared/constants/providers/apikey/inference-hosts.ts`). Regression guard: `tests/unit/digitalocean-provider.test.ts`. (thanks @newnol)
+- **feat(providers):** add **Huancheng Public API** (`hcnsec`) as an OpenAI-compatible regional provider ([#6410](https://github.com/diegosouzapw/OmniRoute/pull/6410)) — Xinjiang Huancheng Cybersecurity's public LLM platform (base `https://api.hcnsec.cn/v1`, free credits via daily check-ins), wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/hcnsec/`, `src/shared/constants/providers/apikey/regional.ts`). Regression guard: `tests/unit/hcnsec-provider.test.ts`. (thanks @UnrealAryan)
+- **feat(dashboard):** the web-session credential guide now shows an **"Open {host}" link** ([#6316](https://github.com/diegosouzapw/OmniRoute/pull/6316)) to the provider's sign-in site (derived from the provider `website` via `getProviderWebsiteHost`), so you can jump straight to the page where the cookie/session must be captured. Regression guard: `tests/unit/web-session-provider-link-6316.test.ts`. (thanks @jordansilly77-stack)
+- **feat(cerebras):** add the **Gemma 4 31B** model (`gemma-4-31b`) to the Cerebras registry + pricing table ([#6331](https://github.com/diegosouzapw/OmniRoute/pull/6331)). Regression guard extends `tests/unit/t28-model-catalog-updates.test.ts`. (thanks @backryun)
+- **feat(providers):** add **Yuanbao (web)** as a cookie-session provider ([#6196](https://github.com/diegosouzapw/OmniRoute/issues/6196)) — `yuanbao-web` (Tencent Yuanbao, `yuanbao.tencent.com`) with cookie-only auth (`hy_user`/`hy_token` + public agent id), SSE→OpenAI translation incl. `reasoning_content`, exposing DeepSeek V3/R1 + Hunyuan / Hunyuan-T1. Regression guard: `tests/unit/providers-yuanbao-web.test.ts`. `together-web` was **deferred** (no verifiable web-session endpoint — needs a captured request) and `huggingchat-web` **dropped** (the existing `huggingchat` already is a web-cookie provider). (thanks @chirag127)
+- **feat(providers):** route the built-in **agentrouter** through the dynamic Claude-Code wire image ([#6056](https://github.com/diegosouzapw/OmniRoute/issues/6056)) — a small static allow-set (`CC_WIRE_IMAGE_BUILTINS` in `open-sse/services/ccWireImageBuiltins.ts`), consulted by `isClaudeCodeCompatible` / `isClaudeCodeCompatibleProvider` / `applyFingerprint`, makes agentrouter adopt the CC wire-image headers + fingerprint **while guarding the CC baseUrl/auth branches** so it keeps its own registry `baseUrl` and `x-api-key` auth. Regression guard: `tests/unit/agentrouter-cc-wire-image.test.ts` (asserts the wire image is applied AND agentrouter's baseUrl/auth are preserved). Live WAF-acceptance against agentrouter.org is a VPS validation follow-up (Hard Rule #18).
+- **feat(providers):** **bulk-add API keys for Cloudflare Workers AI** ([#6174](https://github.com/diegosouzapw/OmniRoute/issues/6174)) — `cloudflare-ai` is removed from the bulk-add exclusion list and the bulk parser gains a 3-field `name|accountId|apiKey` mode; the bulk route now builds a **per-entry** `providerSpecificData` so each key carries its own `accountId` (fixing the previous shared-object reuse), and both the create + key-validation paths receive it. Regression guard: `tests/unit/bulk-api-key-parser-cloudflare.test.ts`. (thanks @muflifadla38)
+- **feat(dashboard):** routing/settings UX clarity ([#6147](https://github.com/diegosouzapw/OmniRoute/issues/6147)) — (1) weighted combos show the **effective routing share %** next to each weight when weights don't sum to 100 (`WeightTotalBar.tsx`); (2) the status widget's user-facing **"Cloud Sync" label is renamed** to "Remote Settings Sync" (`CloudSyncStatus.tsx`; internal ids/state untouched); (3) built-in providers gain an **opt-in advanced base-URL override** (`isBaseUrlOverrideEligibleProvider`, hidden behind an "Advanced" toggle, reusing the existing `providerSpecificData.baseUrl` persistence — not globally widened). Regression guard: `tests/unit/routing-settings-ux-6147.test.ts`.
+- **feat(combo):** add an option to **disable session stickiness**, per-combo or globally — round-robin / random combos can rotate to a different connection on every request instead of pinning a whole conversation to one connection by its first-message hash. Resolution precedence per-combo `config.disableSessionStickiness` → global `settings.disableSessionStickiness` → default `false` (preserves the #3825 prompt-cache/504 fix); gates **both** stickiness call sites in `open-sse/services/combo.ts`. Exposed as a global toggle (Combo Defaults) and a per-combo Inherit/on/off control. ([#6168](https://github.com/diegosouzapw/OmniRoute/issues/6168)) Regression guard: `tests/unit/combo-disable-session-stickiness.test.ts`. (thanks @RCrushMe)
+- **feat(docker):** add the `OMNIROUTE_NO_SUDO` env flag for root-less / user-namespaced deployments — the MITM cert-trust command path (`resolveSudoSpawn` in `src/mitm/systemCommands.ts`) now strips the leading `sudo` when the flag is truthy, in addition to the existing root / sudo-missing cases, so the Proxy Agent runs without `sudo` (the operator trusts the CA manually, e.g. via `NODE_EXTRA_CA_CERTS`). Argv-array `spawn` preserved — no shell interpolation (Hard Rule #13). ([#6122](https://github.com/diegosouzapw/OmniRoute/issues/6122)) Regression guard: `tests/unit/mitm-systemCommands-no-sudo.test.ts`. (thanks @powellnorma)
+- **feat(providers):** add **Requesty** as an OpenAI-compatible gateway provider (BYOK, base `https://router.requesty.ai/v1`, ~200 free requests/day) — wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/requesty/`, `src/shared/constants/providers/apikey/gateways.ts`). ([#6120](https://github.com/diegosouzapw/OmniRoute/issues/6120)) Regression guard: `tests/unit/requesty-provider.test.ts`. (thanks @chirag127)
+- **feat(dashboard):** add **configured-only / available-only filters** to the Free Provider Rankings page ([#6150](https://github.com/diegosouzapw/OmniRoute/issues/6150)) — hide providers you haven't configured, or whose connections are all rate-limited / out of quota, via server-side query params (`?configuredOnly` / `?availableOnly` on `GET /api/free-provider-rankings`) backed by a testable lib helper reusing the in-process connection state (no Redis). Both filters default off, so the default view is unchanged; this supersedes the earlier client-side "Configured Only" toggle (#6245) with an available-only dimension and unit-tested logic. Regression guard: `tests/unit/freeProviderRankings-filters.test.ts`.
+
+### 🔧 Bug Fixes
+
+- **fix(dashboard):** adding a second API key connection for the same provider no longer silently overwrites the first — the Add-API-key modal now derives a unique default connection name (`main`, then `main-2`, `main-3`, …) so the backend name-based upsert can't collide (#6499 — thanks @dilneiss).
+- **fix(compression):** the session-dedup engine now also deduplicates a large multi-line block repeated **within a single message** (intra-message dedup), not just across turns; the compression-preview API surfaces a `fallbackReason`, and the fusion panel reports how many models were rate-limited vs failed on total-panel failure (#6501 — thanks @chirag127).
+- **fix(compression):** a stacked-pipeline step naming an unregistered engine now surfaces a `validationErrors` entry instead of silently no-op'ing, so misconfigured pipelines are visible in the preview API (#6506 — thanks @chirag127).
+- **feat(usage):** add a **Codex reset-credit redemption** flow to the Provider Limits UI ([#6361](https://github.com/diegosouzapw/OmniRoute/pull/6361)) — a `useCodexResetCreditRedemption` hook + `/api/usage/codex-reset-credit` route + `codexResetCredits` lib let you redeem banked Codex reset credits from the quota card. Regression guard: `tests/unit/codex-reset-credits.test.ts`. (thanks @JxnLexn)
+- **feat(glm):** add **team-plan quota settings** for `glm-cn` connections ([#6351](https://github.com/diegosouzapw/OmniRoute/pull/6351)) — a dedicated `GlmTeamQuotaFields` form section (team quota id / limits) threaded through the Add/Edit connection modals, persisted via `providerSpecificData`, with the GLM usage service reading the team quota. Regression guards: `tests/unit/glm-team-quota.test.ts`, `provider-specific-data-schema.test.ts`. (thanks @hao3039032)
+- **feat(providers):** add **TinyFish** web-fetch/search support ([#6349](https://github.com/diegosouzapw/OmniRoute/pull/6349)) — a `tinyfish-fetch` executor + `/v1/web/fetch` route + MCP web-fetch tool, registered as a specialty-media provider with request-validation and a search-provider catalog entry. Regression guards: `tests/unit/executor-tinyfish-fetch.test.ts`, `web-fetch-handler.test.ts`, `mcp-web-fetch-tool.test.ts`, `provider-validation-tinyfish.test.ts`. (thanks @dtybnrj)
+- **fix(cli):** `omniroute launch-codex` now spawns `codex.cmd` through a shell on Windows (the npm `.cmd` shim is unresolvable by bare `spawn` → ENOENT), mirroring the qodercli Windows fix (#6263) ([#6312](https://github.com/diegosouzapw/OmniRoute/pull/6312)). Regression guard: `tests/unit/launch-codex-windows-spawn-6312.test.ts`. (thanks @swingtempo)
+- **fix(codex):** isolate the **Spark** quota from the shared Codex quota and stabilize the quota UI ordering / hydration so per-scope limits render consistently ([#6336](https://github.com/diegosouzapw/OmniRoute/pull/6336)). Regression guards: `tests/unit/codex-quota-selection-hydration.test.ts`, `provider-limits-ui.test.ts` + 3 more. (thanks @xz-dev)
+- **feat(api):** add a `hidePaidModels` setting that filters paid-only models out of the `/v1/models` catalog. Regression guard: `tests/unit/models-catalog-hide-paid.test.ts`. (thanks @chirag127)
+- **fix(api-manager):** the fallback model picker now preserves combos instead of dropping them when a primary model is unavailable ([#6443](https://github.com/diegosouzapw/OmniRoute/pull/6443)). Regression guard: `tests/unit/api-manager-page-static.test.ts`. (thanks @jmengit)
+- **fix(providers):** recoverable Antigravity / Cloud-Code (Gemini Code Assist) `403` responses ([#6452](https://github.com/diegosouzapw/OmniRoute/pull/6452)) are now classified as a retryable project-config error instead of a terminal account ban, so a fixable project/API-disabled 403 no longer forces a ~1-year cooldown / full OAuth reconnect. Regression guard: `tests/unit/errorclassifier-antigravity-403.test.ts`. (thanks @developerjillur)
+- **fix(mitm):** `sanitizeHeaders` now redacts `Set-Cookie` response headers so upstream session cookies never leak into logs / diagnostics ([#6451](https://github.com/diegosouzapw/OmniRoute/pull/6451)). Regression guard: `tests/unit/mitm-sanitize-headers.test.ts`. (thanks @developerjillur)
+- **fix(api):** `/api/compression/preview` now accepts `mode: "caveman"` and correctly handles stacked / zero-compression previews ([#6425](https://github.com/diegosouzapw/OmniRoute/issues/6425)). Regression guard: `tests/unit/api/compression-preview-caveman-and-stacked-6425.test.ts`. (thanks @chirag127)
+- **feat(providers):** add **Zed** hosted LLM aggregator as a native-app provider ([#6118](https://github.com/diegosouzapw/OmniRoute/pull/6118)) — OAuth sign-in via the Zed hosted flow, registered through the shared provider registry + executor. Regression guards: `tests/unit/zed-oauth-provider.test.ts`, `zed-import-utils.test.ts`, `zed-docker-detect.test.ts`, `mitm-handler-zed.test.ts`. VPS-validated via live operator login (Hard Rule #18).
+- **fix(oauth):** the Kiro SSO-cache auto-import now **preserves the IDC region** — cross-region Amazon Q / Kiro profiles imported from the SSO cache are no longer collapsed to the default region ([#6113](https://github.com/diegosouzapw/OmniRoute/pull/6113)). Regression guard: `tests/unit/kiro-auto-import-idc-2059.test.ts`. VPS-validated via live operator login (Hard Rule #18).
+- **fix(dashboard):** passthrough model aliases no longer collide when two namespaced model ids share a last segment (port from 9router#1850, [#6431](https://github.com/diegosouzapw/OmniRoute/pull/6431)). `enx/gpt-5.5` and `enx/codebuddy/gpt-5.5` both auto-generated the alias `gpt-5.5`, so the second model could never be added (the UI just alerted "alias already exists"). Aliases are now disambiguated deterministically — bare last segment when free, then parent-qualified (`codebuddy-gpt-5.5`), then a numeric suffix — while re-adding the exact same model id is still blocked. Regression guard: `tests/unit/passthrough-alias-1850.test.ts`. (thanks @arpicato)
+- **fix(translator):** preserve a Gemini `functionResponse` co-located with other parts (another `functionCall`, or trailing `text`) in the same content when translating **Gemini → OpenAI** ([#6376](https://github.com/diegosouzapw/OmniRoute/pull/6376)). `convertGeminiContent()` early-returned the tool message on the first `functionResponse` part, dropping any co-located parts; such contents are now pre-split (one tool message per `functionResponse`, emitted first, plus one message for the remaining parts). Regression guard: `tests/unit/gemini-to-openai-function-response.test.ts`. (thanks @warelik)
+- **fix(headroom):** detect a python interpreter managed by **mise / pyenv / asdf / conda** (port from 9router#2353, [#6382](https://github.com/diegosouzapw/OmniRoute/pull/6382)). Headroom's python probe (`src/lib/headroom/detect.ts`) searched a hardcoded `PATH`, but version managers expose their interpreters via shim dirs that only join `PATH` through interactive-shell activation — which the non-interactive server never runs, so a managed python (≥3.10) was invisible and Headroom reported it missing. The search path now prepends the well-known shim/bin dirs (`~/.local/share/mise/shims`, `~/.pyenv/shims`, `~/.asdf/shims`, `$CONDA_PREFIX/bin`, `~/.local/bin`, respecting `MISE_DATA_DIR`/`PYENV_ROOT`/`ASDF_DATA_DIR` when set), and a new `HEADROOM_PYTHON` env override lets operators point straight at their interpreter (mirroring `HEADROOM_URL`). Still shell-free (`execFileSync`). Regression guard: `tests/unit/headroom-detect.test.ts` (5). (thanks @loopyd)
+- **fix(executors):** strip the OpenAI-Codex/Claude-CLI `client_metadata` passthrough field for **NVIDIA** requests (port from 9router#1887, [#6411](https://github.com/diegosouzapw/OmniRoute/pull/6411)). NVIDIA's OpenAI-compatible wrapper rejects it with `400 Unsupported parameter`, the same class already handled for `cerebras`/`mistral`; `nvidia` (executor `default`) was missing from the strip allowlist so Codex/Claude-Code passthrough requests 400'd. Regression guard: `tests/unit/executor-default-strip-client-metadata.test.ts` (+nvidia case). (thanks @phidinhmanh)
+- **fix(translator):** strip the Claude-style `thinking` field for **NVIDIA `z-ai/glm-5.2`** (port from 9router#2023, [#6413](https://github.com/diegosouzapw/OmniRoute/pull/6413)). NVIDIA's OpenAI-compatible wrapper 400s on `thinking` (a Claude-format client routed here leaves a `thinking:{type:"adaptive"}`); the existing strip rule only dropped `reasoning`. Same class already handled for `minimax-m2.7`. Regression guard: `tests/unit/nvidia-minimax-thinking-strip.test.ts` (+glm-5.2 case). (thanks @phidinhmanh)
+- **fix(translator):** suppress the streamed `</think>` close marker for the **Antigravity IDE** client (port from 9router#1061, [#6415](https://github.com/diegosouzapw/OmniRoute/pull/6415)). On thinking-only turns Antigravity rendered a bare `</think>` as the sole visible content, tripping its loop-detection and wasting requests. Antigravity's UA (`vscode/<v> (Antigravity/<v>)`) is added to the marker-suppress allowlist (alongside OpenCode); Claude Code / Cursor still get the marker, and `x-omniroute-thinking-marker: on` force-restores it. Regression guard: `tests/unit/think-close-marker-suppress-5245.test.ts`. (thanks @abdofallah)
+- **fix(executors):** strip nested `reasoning_content` from messages for **Mistral** (port from 9router#1649, [#6417](https://github.com/diegosouzapw/OmniRoute/pull/6417)). Mistral's API returns `422 extra_forbidden` when an assistant message carries `reasoning_content` (replayed thinking from a prior turn, e.g. via the Codex `/responses` path); the generic top-level 400 field-downgrade retry never covered the nested per-message field. `DefaultExecutor` now strips it for provider `mistral` only, so DeepSeek (which requires replayed `reasoning_content`) is unaffected. Regression guard: `tests/unit/mistral-strip-reasoning-content-1649.test.ts`. (thanks @xxy9468615)
+- **fix(executors):** strip the `client_metadata` passthrough field on the **OpenCode** path (port from 9router#1442, [#6418](https://github.com/diegosouzapw/OmniRoute/pull/6418)). OpenCode upstreams (e.g. `kimi-k2.6` via opencode-go) reject it with `400 "Extra inputs are not permitted, field: 'client_metadata'"`; the DefaultExecutor strip only covered cerebras/mistral and `OpencodeExecutor` extends `BaseExecutor` directly, so nothing removed it there. Regression guard: `tests/unit/opencode-strip-client-metadata-1442.test.ts`. (thanks @yanpaing007)
+- **fix(executors):** inject the `reasoning_content` echo for the native **Moonshot Kimi** provider (port from 9router#1480, [#6419](https://github.com/diegosouzapw/OmniRoute/pull/6419)). Kimi (executor `default`) is a thinking-mode upstream that 400s with "reasoning_content must be passed back" when a prior assistant turn lacks it; the placeholder injection was only wired into the OpenCode meta-provider, so direct multi-turn Kimi conversations failed. Scoped to `kimi` (gateway-served models matching the thinking-model name pattern are unaffected). Regression guard: `tests/unit/kimi-native-reasoning-injected-1480.test.ts`. (thanks @2220258345)
+- **fix(executors):** recover from a strict gateway's `context_management: Extra inputs are not permitted` 400 (port from 9router#1468, [#6420](https://github.com/diegosouzapw/OmniRoute/pull/6420)). **Claude Code** always sends a top-level `context_management` field; strict anthropic-compatible gateways reject it. The dedicated context-editing 400-fallback only fired when OmniRoute's own `contextEditing` feature was enabled (default off), so a client-sent field passed through untouched and 400'd. `context_management` is now in the generic reactive field-strip list, so it's stripped-and-retried once regardless of the feature flag (with correct request re-signing for claude-compatible relays). Regression guard: `tests/unit/provider-field-strips.test.ts`. (thanks @ohahe52-dot)
+- **fix(network):** enable **RFC 8305 Happy Eyeballs** (`autoSelectFamily`) on the direct-egress undici dispatcher (port from 9router#1237, [#6423](https://github.com/diegosouzapw/OmniRoute/pull/6423)). When DNS returns both IPv6 (AAAA) and IPv4 (A) and the IPv6 route is broken (e.g. a NAT64 `64:ff9b::` prefix without routing), undici tried IPv6 first and hung until `ETIMEDOUT` (then a 502 + account lockout), even though `curl` reached the same host. The direct dispatcher now races both families and uses whichever connects first. Proxy paths pin family via `proxyTls` and are unaffected. Regression guard: `tests/unit/direct-dispatcher-pipelining-4580.test.ts`. (thanks @adentdk)
+- **fix(combo):** round-robin now advances the rotation pointer past the model that **actually served**, not the eagerly-scheduled one (port from 9router#948, [#6428](https://github.com/diegosouzapw/OmniRoute/pull/6428)). With `stickyLimit: 1` (true round-robin), when the scheduled model failed and a _different_ model served via fallback, the counter had already advanced +1 from the scheduled index — so the next request reused the fallback-served model, degrading round-robin into hot-spotting on whichever model was healthy. The pointer now advances to the served index + 1 (mirroring the sticky-limit>1 path). Session-stickiness (#3825) and distribution are preserved. Regression guard: `tests/unit/combo-rr-fallback-advance-948.test.ts`. (thanks @binsarjr)
+- **fix(sse):** a non-string `model` field is now rejected with a `400` before the resolver, instead of crashing downstream `.toLowerCase()`/`.split()` calls into an empty-body `500` that escapes the error sanitizer ([#6407](https://github.com/diegosouzapw/OmniRoute/issues/6407)). Regression guard: `tests/unit/chat-non-string-model-6407.test.ts`. (thanks @chirag127)
+- **fix(api):** unknown `/api/*` routes now return a JSON `404` (instead of the dashboard HTML shell) and scalar chat params (`model`/`temperature`/etc.) are validated **before** the provider lookup so malformed requests fail fast with a clear `400` ([#6424](https://github.com/diegosouzapw/OmniRoute/issues/6424), [#6412](https://github.com/diegosouzapw/OmniRoute/issues/6412)). Regression guards: `tests/unit/api/api-catchall-json-404.test.ts`, `tests/unit/chat-early-schema-validation-6412.test.ts`. (thanks @chirag127)
+- **fix(api):** `/v1/chat/completions` now rejects a non-JSON `Content-Type` with a `400` before parsing the body ([#6414](https://github.com/diegosouzapw/OmniRoute/issues/6414)). Regression guard: `tests/unit/v1-chat-completions-content-type-6414.test.ts`. (thanks @chirag127)
+- **fix(api):** the `X-OmniRoute-Compression` response header is now echoed on `/v1/chat/completions` and `/v1/completions` ([#6422](https://github.com/diegosouzapw/OmniRoute/issues/6422)). Regression guard: `tests/unit/compression-header-echo-6422.test.ts`. (thanks @chirag127)
+- **fix(api):** concurrent `GET /v1/models` requests are coalesced into a single catalog build ([#6408](https://github.com/diegosouzapw/OmniRoute/issues/6408)). Regression guard: `tests/unit/v1-models-concurrent-6408.test.ts`. (thanks @chirag127)
+- **fix(api):** `/v1/completions` now echoes the requested `body.model` in its JSON + streamed responses ([#6429](https://github.com/diegosouzapw/OmniRoute/pull/6429)). Regression guard: `tests/unit/completions-body-model-echo.test.ts`. (thanks @chirag127)
+- **fix(api):** env-var master keys now see the full `/v1/models` catalog ([#6406](https://github.com/diegosouzapw/OmniRoute/issues/6406)). Regression guard: `tests/unit/models-catalog-envkey-6406.test.ts`. (thanks @chirag127)
+- **fix(api):** non-streaming `/v1/completions` responses now echo `body.model` aligned with the `X-OmniRoute-Model` header ([#6426](https://github.com/diegosouzapw/OmniRoute/issues/6426)). Regression guard: `tests/unit/v1-completions-model-header-match-6426.test.ts`. (thanks @chirag127)
+- **fix(api):** unknown `/v1/*` routes now return a JSON `404 not_found` instead of the Next.js dashboard HTML shell ([#6405](https://github.com/diegosouzapw/OmniRoute/issues/6405)). Regression guard: `tests/unit/api/v1-catchall-json-404.test.ts`. (thanks @chirag127)
+- **fix(api):** the per-connection provider models route now degrades to the shipped catalog when a provider's `/models` endpoint answers with a redirect ([#6267](https://github.com/diegosouzapw/OmniRoute/issues/6267)) — a `qwen-web` import failed with a raw `Redirect blocked … (307)` 503. `safeOutboundFetch` throws `REDIRECT_BLOCKED` on the 307, `getSafeOutboundFetchErrorStatus` maps it to 503, and `buildDiscoveryErrorFallbackResponse` treated every 503 as a hard error — so the non-empty `getModelsByProviderId("qwen-web")` catalog was never surfaced. A models-endpoint redirect is not a fixable-config error (unlike `URL_GUARD_BLOCKED`/`INVALID_URL`, which stay hard errors), so it now falls back to the local/cached catalog before the 503 short-circuit. General fix — covers any config-driven provider that 307s. Regression guard: `tests/unit/provider-models-qwen-web-redirect-6267.test.ts`. (thanks @chirag127)
+- **fix(api):** the per-connection provider models route (MCP `list_models_catalog` + the dashboard import view) now merges USER-ADDED custom models into its response ([#6247](https://github.com/diegosouzapw/OmniRoute/issues/6247)) — custom models live in the `key_value` namespace `customModels`, which the live REST `/api/v1/models` already merges, but `src/app/api/providers/[id]/models/route.ts` never read `getCustomModels`, so custom models were dropped on both the discovery-success and local_catalog paths. They are now folded into the returned model list (deduped by id, stamped `owned_by: provider`), fixing MCP + the dashboard import view in one place. Regression guard: `tests/unit/provider-models-custom-merge-6247.test.ts`. (thanks @RCrushMe)
+
+- **fix(providers):** GitLab Duo tool-calling follow-up turns no longer fail upstream with `422 {"detail":"Validation error"}` (tokens 0/0, rejected pre-inference). The #6234 tool-result-feedback fix serialized the **entire** multi-turn conversation into GitLab's single-file `code_suggestions` (`small_file`) generation endpoint — folded history that turn-N sent as an oversized `current_file.content_above_cursor` **and** duplicated verbatim into `user_instruction`, tripping the AI-Gateway's `small_file` validation guard. The executor now **bounds** that prompt: it keeps system + latest user message + the most-recent tool round (dropping older turns), caps oversized tool results, and stops duplicating the full prompt into `user_instruction` (which now carries only the short latest user message) — while still feeding the most-recent tool result back so the agent continues (`open-sse/executors/gitlab.ts`, [#6220](https://github.com/diegosouzapw/OmniRoute/issues/6220)). The unit test covers the bounding logic; the upstream 422→200 clearing is VPS-only (Hard Rule #18). Regression guard: `tests/unit/gitlab-tool-exchange-bounded-6220.test.ts`.
+
+- **fix(i18n):** the provider-detail (`/dashboard/providers/[id]`) connection-status filter labels no longer render as `__MISSING__:All` / `__MISSING__:Active` / `__MISSING__:Error` / `__MISSING__:Banned` / `__MISSING__:CreditsExhausted` in non-English locales (notably pt-BR) ([#6290](https://github.com/diegosouzapw/OmniRoute/issues/6290)). Root cause was **not** the namespace mismatch the issue guessed — the `providers.filter*` keys resolve correctly in `en.json`; the debt lived in the locale mirrors (`src/i18n/messages/*.json`), where these five keys carried the `__MISSING__:` sync sentinel in ~15 locales and were absent entirely in ~26 others, so next-intl found the key and echoed the sentinel verbatim. All 40 non-English/-Chinese mirrors now ship real translations for the five `providers.filter*` labels. Regression guard: `tests/unit/i18n-provider-filter-keys-6290.test.ts`. (thanks @diegosouzapw)
+
+- **fix(providers):** the `copilot-m365-web` streaming executor now emits `debug`-level WebSocket diagnostics ([#6210](https://github.com/diegosouzapw/OmniRoute/issues/6210)) — the outbound WS URL (with the `access_token` **redacted** via `redactWsUrl()`), handshake success/failure, and each received SignalR frame's `type`/`target`. Previously the streaming path logged nothing, so an empty `content:null` response (the M365 Education / Starter tier symptom fixed in #6234) was undiagnosable even at `APP_LOG_LEVEL=debug`. The change is debug-level and side-effect-free — it does not alter streaming behavior or the frame parser, and the token never reaches the logs. Regression guard: `tests/unit/copilot-m365-web-logging-6210.test.ts` (thanks @qpeyba)
+
+- **fix(resilience):** a round-robin combo no longer returns `503 all upstream accounts are unavailable` when a compatibility-rejected target is actually healthy ([#6238](https://github.com/diegosouzapw/OmniRoute/issues/6238)). `filterTargetsByRequestCompatibility` drops request-incompatible targets (tool/vision/structured-output unsupported, or below the required context window) **before** any availability check runs, and its `compatible.length === 0` safety net only fired when _all_ targets were filtered — not when the kept targets later all turned out runtime-unavailable (circuit-open / cooldown / no credentials). So a combo could 503 while a compat-rejected-but-healthy provider sat unused. `handleRoundRobinCombo` now keeps the compat-rejected set and, when every compat-kept target was skipped without a single real attempt, probes those rejected targets as a **last-resort fallback tier** (via the new pure `open-sse/services/combo/comboCompatFallback.ts`) before crystallizing the 503. Regression guard: `tests/unit/combo-roundrobin-compat-fallback-6238.test.ts`. (thanks @ThongAccount)
+
+- **fix(startup):** best-effort self-heal for a corrupted Turbopack dev cache on Windows ([#6289](https://github.com/diegosouzapw/OmniRoute/issues/6289)). On Windows, `pnpm dev` can fail at startup when Turbopack `mmap`s a persistent-cache SST file and the OS refuses the mapping (`os error 1455` — "paging file too small"), which Turbopack surfaces as a misleading `Module not found: Can't resolve '@/shared/utils/machine'`. This is a **known upstream Turbopack cache-corruption bug — not our code**. The dev launcher (`scripts/dev/run-next.mjs`) now wraps `nextApp.prepare()` and, when it rejects with that signature (`isTurbopackCacheCorruption` in the new `scripts/dev/turbopackCacheHeal.mjs`), purges `.build/next/**/cache/turbopack` and retries **once** with a clear log. **Caveat — best-effort only:** the corruption often surfaces as a runtime overlay rather than a `prepare()` rejection, so this cannot always intercept it; the reliable remedy remains manually deleting the Turbopack cache dir. Regression guard: `tests/unit/turbopack-cache-heal-6289.test.ts`. (thanks @chirag127)
+
+- **fix(providers):** qodercli PAT auth no longer fails with `spawn qodercli ENOENT` on Windows ([#6263](https://github.com/diegosouzapw/OmniRoute/issues/6263)) — `spawnQoderCli` spawned the bare `qodercli` name with `shell:false` and an unenriched env, so the npm `.cmd` wrapper under `%APPDATA%\npm` (a user-PATH directory) was never resolved. It now resolves the absolute `.cmd`/`.exe` path through the existing `getCliRuntimeStatus("qoder")` resolver in `src/shared/services/cliRuntime.ts` (memoized), spawns with `shell` when the target is a `.cmd`/`.bat`, and uses the cliRuntime-enriched env (PATH + PATHEXT + APPDATA); the ENOENT error now lists the searched paths plus the `CLI_QODER_BIN` override. End-to-end spawn on a real Windows host is host-only (Hard Rule #18); the path-resolution logic is unit-tested. Regression guard: `tests/unit/qodercli-windows-resolve-6263.test.ts`. (thanks @chirag127)
+
+- **fix(sse):** the reasoning-token buffer no longer inflates **probe-sized `max_tokens`** ([#6274](https://github.com/diegosouzapw/OmniRoute/issues/6274)) — Claude Code's `/model` capability check sends `max_tokens: 1`, but for a thinking-capable model with a large output cap (e.g. `glm-5.2`) the #3587 headroom heuristic (`max(current + 1000, ceil(current * 1.5))`) rewrote it to `1001` and forwarded that upstream, wasting tokens on a request that was never a genuine reasoning budget. `resolveReasoningBufferedMaxTokens()` (`open-sse/services/reasoningTokenBuffer.ts`) now short-circuits and returns the caller's value verbatim when it is below the new `REASONING_BUFFER_MIN_TRIGGER` (256) threshold — a tiny explicit limit is a probe, not a reasoning request. Real budgets still receive the #3587 headroom unchanged, and the guard runs after the existing capability checks so unknown / non-reasoning models keep returning `null`. Regression guard: `tests/unit/reasoning-token-buffer-6274.test.ts`. (thanks @brightfiscalband)
+
+- **fix(cli):** `omniroute reset-password` now works as a real subcommand, and password resets over piped (non-TTY) stdin actually apply ([#6261](https://github.com/diegosouzapw/OmniRoute/issues/6261), [#6258](https://github.com/diegosouzapw/OmniRoute/issues/6258)). Two coupled defects: (1) **#6261** — `bin/omniroute.mjs` routed everything through Commander with only two pre-Commander bypasses (`--mcp`, `reset-encrypted-columns`), so `omniroute reset-password` was rejected as an unknown command; only the separate `omniroute-reset-password` bin worked, while the docs falsely advertised the subcommand (incl. a bogus "legacy alias still works"). A pre-Commander bypass mirroring `reset-encrypted-columns` now dynamically imports `bin/reset-password.mjs` (which self-executes) before Commander parses; the three doc lines were corrected. (2) **#6258** — `bin/reset-password.mjs` issued two sequential `rl.question` prompts; under piped stdin the second read never settled at EOF, so `main()` never reached `resetManagementPassword` and the reset was a silent no-op (both prompts printed, no success, password unchanged). The CLI now detects non-TTY stdin and reads it once (first line = password, second line = confirm if present, else reused), adds a `--password-stdin` flag (entire stdin is the password, no confirmation), and exits `0` explicitly so the success line always flushes; interactive TTY behavior is unchanged. Regression guard: `tests/unit/reset-password-cli-6261-6258.test.ts` (3). (thanks @chirag127)
+
+- **fix(db):** the mass-migration **safety abort** now tells the operator how to bypass it and stops flooding the log ([#6260](https://github.com/diegosouzapw/OmniRoute/issues/6260)) — after restoring a backup that wiped the migration tracking table, `runMigrations()` threw the abort on every downstream `ensureDbInitialized()`, re-logging the full banner 11+ times, and the message never mentioned the existing `OMNIROUTE_MAX_PENDING_MIGRATIONS` escape hatch. The abort text now appends a bypass hint (set `OMNIROUTE_MAX_PENDING_MIGRATIONS=0` in `server.env` / `DATA_DIR/.env`), and a new `MigrationSafetyAbortError` is memoized so repeated calls in the same process throw the same instance and emit a single concise line instead of the full cascade. Regression guard: `tests/unit/migration-safety-abort-6260.test.ts`. (thanks @chirag127)
+
+- **fix(auth):** importing a **distinct** Codex/ChatGPT OAuth `auth.json` is no longer falsely rejected as "already exists" when it belongs to a different user in the same workspace ([#6301](https://github.com/diegosouzapw/OmniRoute/issues/6301)). `findExistingCodexConnection` (in `src/lib/oauth/utils/codexAuthImport.ts`) deduped **only** on `providerSpecificData.workspaceId === accountId`, where `accountId` is the shared `chatgpt_account_id`/`tokens.account_id` — so two members of the same ChatGPT Team collapsed onto a single connection (409 `duplicate_account`). The id_token's `https://api.openai.com/auth` claim carries a per-user `chatgpt_user_id` alongside the workspace id (the device-flow path already persisted it as `chatgptUserId`, but the import path did not). Now `parseAndValidateCodexAuth` extracts `userId` (`chatgpt_user_id` → `user_id` → JWT `sub`) into `ParsedCodexAuth`, the create/update paths persist `chatgptUserId` in `providerSpecificData` (mirroring `codex.ts`), and dedup keys on `workspaceId` **AND** `chatgptUserId` — with a backward-compat fallback to legacy accountId-only matching when no stored connection for that workspace records a `chatgptUserId`, so genuinely-same accounts still dedup. Regression guard: `tests/unit/codex-auth-import-userid-dedup-6301.test.ts` (4). (thanks @anungma)
+
+- **fix(providers):** importing models for the **venice-web** provider no longer fails with a red "Provider venice-web does not support models listing" ([#6269](https://github.com/diegosouzapw/OmniRoute/issues/6269)). `venice-web` is a web-cookie provider with an executor but no upstream `/v1/models` endpoint and no registry `models`, so the models route fell through to the tail `400`. Mirroring the `jules`/`linkup-search`/`ollama-search` fix (#5569), it now ships a static local catalog entry in `src/lib/providers/staticModels.ts` — seeding the current Venice lineup (`venice-uncensored`, `llama-3.3-70b`, `qwen3-235b`, `qwen3-4b`, `deepseek-r1-671b`; Venice rotates its catalog, see docs.venice.ai/models/overview) — so the route returns `200` with `source:"local_catalog"`, `intentional:true`. Regression guard: `tests/unit/static-models-venice-web-6269.test.ts`. (thanks @chirag127)
+
+- **fix(api):** the specialty model catalogs (`/v1/embeddings`, `/v1/images`, `/v1/music`, `/v1/videos` model lists) are now derived from the **unified catalog filtered by a predicate** (`getSpecialtyModelsResponse`) instead of ad-hoc per-route logic, so they consistently respect active-credential visibility and stay in sync with the main catalog ([#6303](https://github.com/diegosouzapw/OmniRoute/pull/6303)). Regression guard: `tests/unit/specialty-model-catalog-routes.test.ts`. (thanks @makcimbx)
+- **fix(api):** the agent-bridge server route now resolves the MITM manager via a **dynamic `import("@/mitm/manager.runtime")`** so Turbopack does not statically pull the stub (or over-bundle the manager), and the agent-skills generator anchors its output base path with `path.join(process.cwd(), …)` so Turbopack's static analyzer stops tracing the whole project root ([#6329](https://github.com/diegosouzapw/OmniRoute/issues/6329), [#6366](https://github.com/diegosouzapw/OmniRoute/pull/6366)). Regression guard: `tests/unit/agent-bridge-server-route-dynamic-import.test.ts`. (thanks @Iammilansoni)
+- **fix(api):** internal probes (combo-test, cloud-sync verify) now pick a **management-scoped / allow-all API key** instead of naively grabbing `getApiKeys()[0]` — a restricted `self:usage` first row made the probe fail with "Model X is not allowed for this API key" even when the combo path was healthy (`pickApiKeyForInternalUse` in `src/lib/db/apiKeys.ts`). The API-manager model editor also falls back to `/api/models?all=true` when `/v1/models` is catalog-protected ([#6372](https://github.com/diegosouzapw/OmniRoute/pull/6372)). Regression guard: `tests/unit/pick-internal-api-key-6372.test.ts`. (thanks @jmengit)
+- **fix(live-ws):** the Live Dashboard WebSocket server now **rejects on bind failure** (e.g. `EADDRINUSE` when the API bridge already holds the port) instead of letting the error surface as an unhandled `error` event that crash-loops the process — the `error` listener is attached to `wss` (not `server`) and releases the EventBus subscription on a failed start ([#6324](https://github.com/diegosouzapw/OmniRoute/issues/6324)). Regression guard: `tests/unit/live-ws-eaddrinuse-6324.test.ts`. (thanks @vinayakkulkarni)
+- **fix(dashboard):** the Home provider-topology widget now trusts the live provider-metrics snapshot — it uses `topology.errorProvider` and live `activeRequests` directly instead of re-deriving state from a stale `lastErrorAt` or applying a frontend timeout filter, so the topology reflects real-time provider health ([#6322](https://github.com/diegosouzapw/OmniRoute/pull/6322)). Regression guard: `tests/unit/home-provider-topology-live-state.test.ts`. (thanks @xz-dev)
+- **fix(sse):** strip zero-width markers from streamed **tool-call arguments** — a follow-up to [#5857](https://github.com/diegosouzapw/OmniRoute/pull/5857). That PR removed injected zero-width joiners (U+200D) from streamed assistant text/reasoning but deliberately left tool-call argument JSON byte-exact. The request-side obfuscation (`open-sse/services/claudeCodeObfuscation.ts`) injects ZWJ into agent words — including the temp path inside the Bash tool description — and Claude models copy that verbatim into generated commands, which are delivered as tool-call arguments rather than assistant text. As a result the ZWJ survived and corrupted code blocks (e.g. a temp path rendered with an invisible joiner). Now `open-sse/handlers/responseSanitizer.ts` strips zero-width code points from tool-call argument strings at every emit site (OpenAI non-stream/stream chat `tool_calls` + legacy `function_call`, native Responses `function_call` items, the OpenAI→Responses conversion, and the native Responses streaming `response.function_call_arguments.delta/.done` events). Only zero-width code points are removed; JSON structure and all other bytes stay identical (no parse/restringify), so normal arguments remain byte-exact. Regression guard: 6 new cases in `tests/unit/response-sanitizer.test.ts` (suite 50/50).
+
+- **fix(nodejs):** the default app log path now resolves under `DATA_DIR` (`~/.omniroute/logs/application/app.log`) instead of `process.cwd()` ([#6197](https://github.com/diegosouzapw/OmniRoute/issues/6197)) — the globally-installed CLI runs from an arbitrary working directory, so anchoring the default to cwd made file logging silently write to (or no-op under) an unrelated directory, contradicting the documented `.env.example` default. `getAppLogFilePath()` now computes the default lazily via the pure `resolveDataDir()` resolver (honours a per-process `DATA_DIR`, no directory-creation side effect); an explicit `APP_LOG_FILE_PATH` still wins. Regression guard: `tests/unit/logenv-datadir-path-6197.test.ts` (3).
+
+- **fix(docker):** AgentBridge/`startMitm` no longer aborts in containers/headless when the Antigravity-default DNS step can't write `/etc/hosts` ([#6127](https://github.com/diegosouzapw/OmniRoute/issues/6127)), and the privileged command's stderr now reaches `app.log` instead of only a bare exit code hitting the toast ([#6198](https://github.com/diegosouzapw/OmniRoute/issues/6198)). The default DNS step (`addDNSEntry`) was called unguarded while cert install and the two sibling DNS steps were each best-effort — in the runtime Docker image (`USER node`, no `sudo`, read-only `/etc/hosts`) it threw `Command failed with code 1` out of `startMitmInternal` and killed the whole start, discarding the stderr. The three DNS steps are extracted into a best-effort `provisionDnsEntries()` where each failure is logged with the full `err` (stderr included, folded in by `systemCommands.ts`) and never aborts the start. Regression guard: `tests/unit/mitm-dns-graceful-degrade-6127.test.ts` (4).
+
+- **fix(providers):** copilot-m365-web now supports the M365 Education "Starter / OfficeWebIncludedCopilot" tier and no longer returns an empty `content:null` stream ([#6210](https://github.com/diegosouzapw/OmniRoute/issues/6210)). Two gaps: (1) `buildWsUrl()` hardcoded the individual-consumer scenario (`OfficeWebPaidConsumerCopilot`, `isEdu=false`) — the EDU tier is now opt-in via `providerSpecificData.tier="edu"`, emitting `scenario=OfficeWebIncludedCopilot`/`isEdu=true` (the individual path is unchanged); (2) the EDU/GPT-5.5 path streams deltas via `arguments[0].writeAtCursor` (incremental) instead of only `messages[].text` (accumulated snapshots), which the parser dropped — a new `accumulateBotContent()` folds both formats, with `type:2 item.result.message` as a last-resort fallback. Regression guard: `tests/unit/copilot-m365-edu-writeatcursor-6210.test.ts` (10). (thanks @qpeyba)
+
+- **fix(providers):** GitLab Duo executor now feeds tool results back into the prompt instead of looping ([#6220](https://github.com/diegosouzapw/OmniRoute/issues/6220)) — `buildPrompt()` branched only on `system`/`user` and took `userParts.at(-1)`, silently dropping the `assistant{tool_calls}` + `tool{result}` turns the client appended, so the reconstructed prompt was byte-identical to turn 1 and the model re-emitted the same `<tool>` call forever. When a tool exchange is present the full conversation is now serialized, folding each tool result back keyed by its `tool_call_id`; simple conversations keep the legacy shape. Complements the tool_call emission from [#6051](https://github.com/diegosouzapw/OmniRoute/issues/6051) (the `kilo-duplicate` label was a false positive — different, sequential defect). Regression guard: `tests/unit/gitlab-tool-result-feedback-6220.test.ts` (4).
+
+- **fix(providers):** opencode-go/opencode-zen can now synthesize the OpenCode CLI identity headers Cloudflare requires on VPS egress ([#5997](https://github.com/diegosouzapw/OmniRoute/issues/5997)) — on a datacenter VPS, `opencode.ai/zen/go/v1/chat/completions` 403s (HTML challenge) requests lacking CLI identity, while the reporter's control curl proved that `User-Agent: opencode-cli/1.0.0` + `x-opencode-client: cli` + `x-opencode-project: default` + fresh request/session UUIDs succeed. Opt-in via `OPENCODE_SYNTHESIZE_CLI_HEADERS=true` (values overridable via `OPENCODE_GO_USER_AGENT`/`OPENCODE_USER_AGENT`/`OPENCODE_CLIENT`/`OPENCODE_PROJECT`); it fills only headers the client did not already send. Kept **off by default** — the forward-only path is deliberate (fabricating a wrong value risks upstream rejection; a prior dedup regressed with `opencode/local`), so this replaces the fragile local header-injection shim without changing default behavior. Regression guard: `tests/unit/opencode-cli-headers-synthesis-5997.test.ts` (6). (thanks @aleksesipenko)
+- fix(resilience): sticky session affinity now evicts and fails over to another account when the pinned account is exhausted/unavailable (#6219)
+
+- fix(sse): Responses API passthrough now drops internal commentary-phase output before forwarding to clients (gated by RESPONSES_PASSTHROUGH_DROP_COMMENTARY, default on) (#6199)
+
+- **fix(sse):** tool-call function schemas with a root `type: null` are now coerced to `type: "object"` before dispatch ([#6359](https://github.com/diegosouzapw/OmniRoute/issues/6359)) — clients like the Codex app emit `parameters: { type: null, ... }` for some tools, which OpenAI-compatible upstreams reject with `400 Invalid schema for function '...': schema must be a JSON Schema of 'type: "object"', got 'type: null'`, failing the whole request. `toolSchemaSanitizer` already stripped the null; it now re-adds the mandatory root `"object"` type (and empty `properties`/open `additionalProperties` when absent). Combinator roots (`anyOf`/`oneOf`/`allOf`) and explicit root types are left untouched. Regression guard: 5 new cases in `tests/unit/tool-schema-sanitizer.test.mjs`.
+
+- **fix(docker):** AgentBridge no longer fails to start on npm/Electron/VPS installs with "MITM manager stub reached at runtime" ([#6344](https://github.com/diegosouzapw/OmniRoute/issues/6344)) — v3.8.45 flipped the production bundler default to Turbopack, but `next.config.mjs` aliased `@/mitm/manager` to its Docker-only degraded stub **unconditionally**. That was harmless while Docker (which sets the alias intentionally for #3390 graceful degradation) was the sole Turbopack consumer, but once every artifact built with Turbopack the stub shipped to all non-Docker users and `startMitm` threw on the first Agent-Bridge start. The alias is now opt-in via `OMNIROUTE_MITM_STUB=1` (set only by the Dockerfile) through the shared `scripts/build/mitm-stub-flag.mjs` helper; default builds bundle the real manager. Regression guard: `tests/unit/mitm-stub-alias-6344.test.mjs` (4).
+
+- **fix(proxy):** stop the v3.8.44 proxy regression that leaked the real IP and disabled healthy proxies ([#6246](https://github.com/diegosouzapw/OmniRoute/issues/6246)). Two coupled defects from the new health scheduler: (1) **IP leak** — when a proxy assigned to a connection was marked `inactive`, resolution fell through to a **direct** egress instead of blocking, exposing the operator's real IP; (2) **over-deactivation** — the sweep flipped a proxy to `inactive` on the **first** failed probe and counted our own 5s timeout / a probe-target `5xx` as the proxy's fault, so healthy paid proxies vanished from egress selection ("my proxies are not being used anymore"). Fix: the sweep decision is extracted into a pure, network-free `decideProxyHealthAction` (`src/lib/proxyHealth/decision.ts`) — by default the health check now **only counts/logs and never downgrades status** (a proxy is downgraded/removed only with `PROXY_AUTO_REMOVE=true`, after `PROXY_AUTO_REMOVE_AFTER` **consecutive** conclusive failures); probes are classified tri-state so an inconclusive result (our timeout, or a `5xx` from the probe target) never penalizes the proxy, and the probe timeout is raised 5s→15s. Separately, `safeResolveProxy` now **fails closed** via the existing policy: a connection whose assigned proxy is dead is blocked instead of leaking direct (`hasBlockingProxyAssignment`), honoring the explicit `proxy off` toggles and the `PROXY_FAIL_OPEN=true` opt-out. Existing proxies stuck `inactive` by the old behavior need a one-time manual re-activate (the operator owns proxy status). Regression guards: `tests/unit/proxy-health-decide-action-6246.test.ts`, `tests/unit/proxy-assigned-unavailable-6246.test.ts`.
+
+- **fix(proxy):** make "Test All" read-only and add bulk enable/disable ([#6246](https://github.com/diegosouzapw/OmniRoute/issues/6246)). Complements the core fail-closed / scheduler fix (#6296) with the two remaining reporter asks. (1) The **"Test All" button** (`POST /api/settings/proxies/auto-test`) used to flip a proxy to `inactive` on a failed reachability probe; since the egress selector excludes `inactive` proxies, a flaky probe (an unreachable `httpbin.org`, a proxy that blocks `HEAD`, or a slow paid proxy) silently disabled every proxy that failed — "Test All" is now **read-only by default** (only the operator sets a proxy active/inactive; opt back into the legacy test-and-set with `PROXY_HEALTH_AUTO_DEACTIVATE=true`). (2) Adds a **bulk enable/disable** proxies endpoint + toolbar action (`POST /api/settings/proxies/batch-activate`) so an operator can re-activate proxies in one click. Regression guard: `tests/unit/proxy-health-6246.test.ts`. (thanks @tenshiak)
+
+- **chatcore (tools): stop the default 128-tool cap from silently dropping opencode's `task`/MCP tools.** opencode (used as an MCP/agent host) sends a large tool list; when it exceeds the speculative `MAX_TOOLS_LIMIT` (128) default, `truncateToolList` did a blind `tools.slice(0, 128)`, dropping every tool past index 128 — including opencode's built-in `task` tool (subagent launch) and many MCP tools, so models routed through OmniRoute could no longer spawn subagents or reach part of their tools. The cap exists to avoid upstream `400`s for providers with real hard limits (e.g. grok-cli 200), so it is kept for those: detection of the opencode client (`isOpencodeClient` — any `x-opencode-*` header, or `opencode` in the user-agent) now only bypasses the **speculative 128 default**, never a known provider ceiling. Precedence is explicit — a proactive/detected provider limit always truncates (even for opencode); otherwise opencode forwards its full tool list; otherwise the unchanged 128 default applies to every other client. Refactors `getEffectiveToolLimit` into `getKnownToolLimit(provider) ?? DEFAULT_LIMIT` (byte-identical for existing callers) and fixes a cosmetic debug-log that reported the truncated count instead of the original. Regression guard: `tests/unit/tool-limit-detector.test.ts`.
+
+- **fix(mitm):** the macOS MITM-cert install check now matches the system keychain again. `security find-certificate -a -Z` prints the SHA-1 as a colon-less hex string, but the installed-check compared it against `getCertFingerprint()`'s colon-separated form, so the substring match never hit — the cert was reported as not-installed and re-prompted for the sudo install on every run. Fingerprints are now normalized (colons stripped, upper-cased) on both sides via the extracted `macCertOutputHasFingerprint` helper. Regression guard: `tests/unit/mitm-cert-mac-fingerprint.test.ts`. ([#6204](https://github.com/diegosouzapw/OmniRoute/pull/6204), closes [#6134](https://github.com/diegosouzapw/OmniRoute/issues/6134) — thanks @rianonehub)
+
+- **fix(api):** `/v1/messages/count_tokens` now counts `tool_use`, `tool_result` and `thinking` content blocks (and array-form `system` prompts) in the local-estimation path, instead of only `text`. Real agentic conversations keep ~95% of their tokens inside tool results; the previous estimate returned near-zero for them, which silently broke Claude Code's auto-compaction (context grew past the window with no compaction until the upstream API rejected the request). The real provider-side count path is unchanged. Regression guard: `tests/unit/messages-count-tokens-route.test.ts`. ([#6221](https://github.com/diegosouzapw/OmniRoute/pull/6221) — thanks @luweiCN)
+
+- **fix(antigravity):** strip a trailing assistant prefill turn for Vertex Claude models to avoid upstream 400s ([#6114](https://github.com/diegosouzapw/OmniRoute/pull/6114)). Regression guard: `tests/unit/antigravity-claude-prefill-strip.test.ts`. (thanks @anki1kr)
+
+- **fix(security):** the mutable cloud-agent routes (`/api/cloud/credentials/update`, `/api/cloud/models/alias`) now require management auth instead of being treated as public. They were classified as public API routes, so a request without management credentials could update stored cloud-agent credentials and model aliases. They are removed from the public-route set, classified as management routes in the authz pipeline, and gated by `requireManagementAuth`; cloud **read**/auth routes stay public. Regression guards: `tests/unit/cloud-write-auth.test.ts`, `tests/unit/authz/classify.test.ts`, `tests/unit/public-api-routes.test.ts`. ([#6233](https://github.com/diegosouzapw/OmniRoute/pull/6233) — thanks @vittoroliveira-dev)
+
+- **refactor(dashboard):** extract the onboarding-wizard "Open provider details" link target into a pure, unit-tested `buildProviderDetailsHref(connection)` helper. The wizard already routes by `connection.id` (the node UUID) rather than the provider category slug (#6144/#6145); this hardens that behavior behind a tested helper that guards a missing id/connection. Regression guard: `tests/unit/provider-onboarding-href.test.ts`. ([#6166](https://github.com/diegosouzapw/OmniRoute/pull/6166) — thanks @KooshaPari)
+
+- **fix(security):** the doubao synthetic device-id generator now derives its digits via an unbiased crypto-random draw (rejection sampling over `crypto.randomBytes()`) instead of a `% 10` reduction, closing a CodeQL `js/biased-cryptographic-random` finding.
+- **fix(agentSkills):** the GitHub-skills generator now resolves `outputDir` to an absolute path before writing, fixing a regression introduced by #6366 (relative-to-cwd base path) that could write generated skill files to the wrong directory.
+- **fix(security):** `/api/keys/{id}/devices` now checks the HTTP method before auth/validation, returning a `405` for non-GET/DELETE verbs instead of a misleading `401`/`500` (closes a `dast-smoke` QUERY-method finding).
+- **fix(quality):** clear the last 2 heavy quality-gate reds on the release tip (cycle pre-flight).
+- **fix(mitm):** the test suite and CI can never mutate the OS trust store — `OMNIROUTE_SKIP_SYSTEM_TRUST=1` is set globally for tests/CI so `installCert`/`uninstallCert`/`installTproxyCa` skip the privileged OS dispatch ([#6310](https://github.com/diegosouzapw/OmniRoute/pull/6310); full detail is under the [3.8.45] section below — this branch received it via the parallel-cycle sync-back).
+- **fix(api):** `POST /api/github-skills` now Zod-validates its request body; documented the new quality-gate env vars and pinned the merge-integrity GitHub Actions to a commit SHA.
+- **fix(skills):** generate the missing `omni-github-skills` registry entry and align the agent-skills catalog-count tests (follow-up to #6186).
+- **fix(quality):** clear the cycle's 11 net-new ESLint errors and make `validate-release-green` suppressions-aware.
+
+### 📝 Maintenance
+
+- **i18n(it):** add 118 missing Italian (`it`) translations (net-additive — no existing keys dropped, valid JSON), improving Italian UI coverage. ([#6212](https://github.com/diegosouzapw/OmniRoute/pull/6212) — thanks @serverless83)
+- **chore(providers):** remove deprecated **MiMo V2** model entries from the catalogs (xiaomi-mimo, opencode-go, zenmux-free, audio TTS) — the upstream V2 line is superseded by MiMo V2.5; drops `mimo-v2-tts`, `mimo-v2-pro`, `mimo-v2-omni`, `mimo-v2-flash`, `mimo-v2-flash-free` and realigns the provider-catalog tests. ([#6248](https://github.com/diegosouzapw/OmniRoute/pull/6248) — thanks @backryun)
+- **chore(release):** ~50 commits on this branch are v3.8.45 pre-flight/hardening fixes and CI-perf work that landed here via the parallel-cycle sync-back (`sync-next-cycle.mjs`, Hard Rule #21) **after** the `v3.8.45` git tag was cut, and are already fully documented under the **[3.8.45]** section below — listed here only so the per-cycle commit-coverage check (`npm run release:uncovered`) doesn't flag them as gaps. Provider/catalog/UX/backend fixes: #6041, #6078, #6108, #6135, #6148, #6149, #6154, #6158, #6161, #6162, #6163, #6164, #6165, #6170, #6177, #6178, #6181, #6186, #6187, #6191, #6193, #6194, #6195, #6200, #6205, #6208, #6209, #6211, #6213, #6223, #6224, #6225, #6226, #6227, #6228, #6229, #6230, #6235, #6291, #6292. CI/release-pipeline work: #6167, #6203, #6214, #6215, #6218, #6273, #6275, #6283, #6284, #6285, #6300, #6305.
+- **chore(release):** additional zero-ref release-cycle plumbing on this branch, kept out of `release:uncovered` on purpose (no `#N` in the commit subject to cite): opening the v3.8.46 cycle, opening/closing the v3.8.45 cycle, the finalized [3.8.45] CHANGELOG i18n sync-back to 42 mirrors, the v3.8.45 cognitive/cyclomatic and file-size drift rebaselines, ESLint stale-suppression pruning (4,273 → 4,233), and clearing test-masking/docs-all pre-flight reds for v3.8.45.
+
+### ⚡ Performance & Infrastructure
+
+- **perf(release-green):** the pre-flight validator (`scripts/quality/validate-release-green.mjs`) now runs its 4 slow suites (unit / vitest / integration / pack-artifact) **concurrently** via `Promise.all` — pre-flight wall time drops from ~the sum of the suites to ~the slowest one (~30min saved per round; Phase 0 was the nº1 bottleneck of the v3.8.45 release benchmark, 2h54 of 6h34 e2e). Guard: `tests/unit/validate-release-green.test.ts` ("runs the slow suites CONCURRENTLY"). ([#6319](https://github.com/diegosouzapw/OmniRoute/pull/6319))
+- **fix(ci):** `scripts/release/sync-next-cycle.mjs` — two defects found live in its first production run (v3.8.45 Phase 5): (1) the `git()` helper's default 1 MiB `maxBuffer` crashed with `ENOBUFS` on `git show origin/main:CHANGELOG.md` (the CHANGELOG alone is >1 MiB) — widened to 64 MiB; (2) the i18n resync only propagated the `[NEXT]` (TBD) section, leaving the just-shipped finalized section as "— TBD" in all 42 mirrors — it now also syncs `[prevVersion]` bounded by the heading below it (new exported pure helper `versionAfter`). Guards: +5 tests in `tests/unit/sync-next-cycle.test.ts` (8/8). ([#6327](https://github.com/diegosouzapw/OmniRoute/pull/6327))
+- **test(ci):** concurrency-sensitive flaky tests are **quarantined into a serial pass** (`tests/unit/serial/`, `--test-concurrency=1`, appended to every unit runner incl. sharded variants — the serial pass is sharded too so concurrent shard jobs never self-collide). Initial set: `glm-coding-plan-monthly-3580`, `quota-division-blocks`, `provider-health-autopilot`, `combo-health-autopilot` — the class behind the ~28min CI wedges/re-runs (two live 1h+ wedges cancelled during this PR's own validation). Discovery + TIA gates track the new glob; systemic root cause (async logger writing after teardown) tracked in [#6360](https://github.com/diegosouzapw/OmniRoute/issues/6360). Guard: `tests/unit/test-serial-quarantine.test.ts` (4). ([#6347](https://github.com/diegosouzapw/OmniRoute/pull/6347))
+
+### 🙌 Contributors
+
+Thanks to everyone whose work landed in v3.8.46:
+
+| Contributor                                                    | PRs / Issues           |
+| -------------------------------------------------------------- | ---------------------- |
+| [@2220258345](https://github.com/2220258345)                   | direct commit / report |
+| [@abdofallah](https://github.com/abdofallah)                   | direct commit / report |
+| [@adentdk](https://github.com/adentdk)                         | direct commit / report |
+| [@aleksesipenko](https://github.com/aleksesipenko)             | direct commit / report |
+| [@anki1kr](https://github.com/anki1kr)                         | direct commit / report |
+| [@anungma](https://github.com/anungma)                         | direct commit / report |
+| [@arpicato](https://github.com/arpicato)                       | direct commit / report |
+| [@backryun](https://github.com/backryun)                       | #6248                  |
+| [@binsarjr](https://github.com/binsarjr)                       | direct commit / report |
+| [@brightfiscalband](https://github.com/brightfiscalband)       | direct commit / report |
+| [@chirag127](https://github.com/chirag127)                     | #6501, #6506           |
+| [@developerjillur](https://github.com/developerjillur)         | direct commit / report |
+| [@dilneiss](https://github.com/dilneiss)                       | #6499                  |
+| [@dtybnrj](https://github.com/dtybnrj)                         | direct commit / report |
+| [@Forcerecon](https://github.com/Forcerecon)                   | direct commit / report |
+| [@hao3039032](https://github.com/hao3039032)                   | direct commit / report |
+| [@Iammilansoni](https://github.com/Iammilansoni)               | #6150, #6245           |
+| [@jmengit](https://github.com/jmengit)                         | direct commit / report |
+| [@jordansilly77-stack](https://github.com/jordansilly77-stack) | direct commit / report |
+| [@JxnLexn](https://github.com/JxnLexn)                         | direct commit / report |
+| [@KooshaPari](https://github.com/KooshaPari)                   | #6166                  |
+| [@loopyd](https://github.com/loopyd)                           | direct commit / report |
+| [@luweiCN](https://github.com/luweiCN)                         | #6221                  |
+| [@makcimbx](https://github.com/makcimbx)                       | direct commit / report |
+| [@muflifadla38](https://github.com/muflifadla38)               | direct commit / report |
+| [@newnol](https://github.com/newnol)                           | direct commit / report |
+| [@ofekbetzalel](https://github.com/ofekbetzalel)               | direct commit / report |
+| [@ohahe52-dot](https://github.com/ohahe52-dot)                 | direct commit / report |
+| [@phidinhmanh](https://github.com/phidinhmanh)                 | direct commit / report |
+| [@powellnorma](https://github.com/powellnorma)                 | direct commit / report |
+| [@qpeyba](https://github.com/qpeyba)                           | direct commit / report |
+| [@RaviTharuma](https://github.com/RaviTharuma)                 | direct commit / report |
+| [@RCrushMe](https://github.com/RCrushMe)                       | direct commit / report |
+| [@rianonehub](https://github.com/rianonehub)                   | #6134, #6204           |
+| [@serverless83](https://github.com/serverless83)               | #6212                  |
+| [@swingtempo](https://github.com/swingtempo)                   | direct commit / report |
+| [@tenshiak](https://github.com/tenshiak)                       | direct commit / report |
+| [@ThongAccount](https://github.com/ThongAccount)               | direct commit / report |
+| [@UnrealAryan](https://github.com/UnrealAryan)                 | direct commit / report |
+| [@vinayakkulkarni](https://github.com/vinayakkulkarni)         | direct commit / report |
+| [@vittoroliveira-dev](https://github.com/vittoroliveira-dev)   | #6233                  |
+| [@warelik](https://github.com/warelik)                         | direct commit / report |
+| [@xxy9468615](https://github.com/xxy9468615)                   | direct commit / report |
+| [@xz-dev](https://github.com/xz-dev)                           | direct commit / report |
+| [@yanpaing007](https://github.com/yanpaing007)                 | direct commit / report |
+| [@diegosouzapw](https://github.com/diegosouzapw)               | maintainer             |
+
+---
+
 ## [3.8.45] — 2026-07-06
 
 ### ✨ New Features

docs/i18n/ja/CHANGELOG.md

@@ -6,6 +6,213 @@
 
 ## [3.8.31] — 2026-06-20
 
+## [3.8.46] — TBD
+
+### ✨ New Features
+
+- **feat(sse):** **hide paid-only models from `auto/*` routing** when `hidePaidModels` is on ([#6512](https://github.com/diegosouzapw/OmniRoute/issues/6512)) — follow-up to #6328/#6495. PR #6495 hid paid-only models from the `GET /v1/models` listing, but `auto/*` combos (`auto/best-coding`, `auto/glm`, …) could still pick a paid-only backend into their candidate pool → a 402/403 at request time. `createVirtualAutoCombo` now filters the candidate pool through the new pure `open-sse/services/autoCombo/paidModelFilter.ts` (`filterPaidOnlyCandidates`), applying the same free-model predicate #6495 uses in `catalog.ts` (`providerHasFreeModels(provider) && isFreeModel(provider, {id})`) whenever `settings.hidePaidModels === true`. Applied before the category/tier/family narrowing, so it covers every `auto/*` combo; an all-paid pool degrades to the existing graceful empty-pool path. **Opt-in — default OFF leaves the pool unchanged** (identity). Regression guard: `tests/unit/autoCombo/paid-model-filter-6512.test.ts` (4, incl. the default-off identity guard).
+- **feat(sse):** **provider-family auto combos** — `auto/glm`, `auto/minimax`, `auto/mimo`, `auto/zai`, `auto/gemma`, `auto/llama`, `auto/gemini` ([#6453](https://github.com/diegosouzapw/OmniRoute/issues/6453)) — new routable ids that materialize an on-demand virtual combo spanning whatever installed backends currently expose that model family, degrading gracefully as backends rotate. A new pure `open-sse/services/autoCombo/modelFamily.ts` (`detectModelFamily`) classifies by model-id prefix for six families; `zai` is instead resolved by provider id (z.ai's hosted API serves the same `glm-*` model ids as every other GLM backend, so `auto/zai` means "route to my z.ai backend specifically" vs `auto/glm`'s "any connected GLM backend"). Reuses the existing `createVirtualAutoCombo` on-demand materialization path (no DB writes) and the `/v1/models` catalog advertising loop. Regression guard: `tests/unit/autoCombo/provider-family-combos.test.ts` (11).
+- **feat(proxy):** native **proxy-pool round-robin / egress IP rotation** ([#6365](https://github.com/diegosouzapw/OmniRoute/issues/6365)) — a scope (global / provider / account) can now hold **multiple** proxies as a pool with a rotation strategy, so outbound requests cycle their egress IP instead of pinning one proxy per scope. Migration `117_proxy_pool_rotation.sql` lifts the `UNIQUE(scope, scope_id)` constraint (rebuild via the canonical rename/copy/drop; existing single assignments become 1-element pools) and adds a `proxy_scope_rotation` companion table holding the per-scope strategy + a persisted monotonic round-robin cursor. Strategies: `round-robin` (default, monotonic cursor — never `Math.random`), `random`, and `sticky-per-N-min`. Resolution (`resolveProxyForScopeFromRegistry` / `resolveProxyForConnectionFromRegistry`) now fetches the alive, position-ordered candidate set (unchanged `PROXY_ALIVE_PREDICATE`) and applies the strategy; an empty / all-dead pool still returns `null` — the #6246 fail-closed guard is **untouched** (never falls through to direct egress). Backend + DB only; dashboard pool-builder UI is a follow-up. Regression guard: `tests/unit/proxy-pool-rotation-6365.test.ts` (8, incl. fail-closed + backward-compat).
+- **feat(providers):** end-to-end **tool/function calling on the native Gemini `/v1beta` endpoint** ([#6222](https://github.com/diegosouzapw/OmniRoute/issues/6222)) — both directions of the Gemini↔OpenAI conversion now preserve tool data (previously silently dropped). Request side: `convertGeminiToInternal` (extracted to its own testable module) maps `tools[].functionDeclarations` → OpenAI `tools`, prior `functionCall` parts → assistant `tool_calls`, and `functionResponse` parts → `tool`-role messages. Response side: `convertOpenAIResponseToGemini` emits `parts[].functionCall {name,args}` from `message.tool_calls`, and the streaming `openAIChunkToGeminiChunk` accumulates fragmented `tool_calls` deltas by index into complete `functionCall` parts. The non-Gemini client paths (Claude, OpenAI-Responses) already preserved tool calls — this closes the gap specific to the native Gemini surface. Regression guard: `tests/unit/v1beta-gemini-tool-calling-6222.test.ts` (6, incl. a streaming SSE round-trip).
+- **feat(providers):** copilot-m365-web **enterprise / work tier** support ([#6334](https://github.com/diegosouzapw/OmniRoute/issues/6334)) — mirrors the EDU-tier pattern (#6210): `M365ConnectionParams` gains an `agent` field, a new opt-in `M365_ENTERPRISE_OVERRIDES` preset (`agent=work`, `scenario=officeweb`, `licenseType=Premium`) applies via `providerSpecificData.tier="enterprise"` (alias `"work"`), and `agent` is also overridable directly via `providerSpecificData.agent`. `buildWsUrl` was hardcoding `agent="web"` (the one enterprise-distinguishing param with no override path), so a Premium work account handshook then returned an empty stream. The individual and EDU paths are untouched. Kilo's dup flag vs #6210 (EDU tier) was a false positive — different tier. Regression guard: `tests/unit/copilot-m365-enterprise-6334.test.ts` (7). End-to-end confirmation on a real Premium work account is a live-VPS validation follow-up (Hard Rule #18). (thanks @Forcerecon)
+- **feat(api):** standardized, provider-agnostic **`effort` + `thinking` request params** ([#6241](https://github.com/diegosouzapw/OmniRoute/issues/6241)) — a thin standardization layer over the existing mature per-provider reasoning plumbing (no provider mapper touched). `providerChatCompletionSchema` gains a canonical `effort` (reusing the shared `none/low/medium/high/xhigh` vocabulary — the UI tiers `extra`/`max` collapse onto `xhigh`) and a boolean `thinking`. A pure `normalizeReasoningRequest` (wired once in `src/sse/handlers/chat.ts`, before any reasoning field is read) folds them onto the fields the translators already consume (`reasoning_effort` / `reasoning.effort` / `thinking`), so they fan out to Anthropic / Gemini / xAI / Responses — an explicit client `reasoning_effort` / object-shaped `thinking` always wins (backward-compatible). `/models` additively exposes `supportsThinking` + `effort_tiers` so the frontend can render the toggles (UI component is a follow-up). Regression guard: `tests/unit/effort-thinking-standardization-6241.test.ts` (12). (thanks @Iammilansoni, @shabeer)
+- **feat(combo):** new **`pipeline` (sequential) combo strategy** ([#6297](https://github.com/diegosouzapw/OmniRoute/issues/6297)) — the 18th routing strategy runs targets **in order**, threading each step's output into the next step's input, with an optional per-step `prompt` (system instruction); only the final step's response is returned. Distinct from `fusion` (parallel fan-out + judge). Implemented as a self-contained `open-sse/services/pipeline.ts` (sibling to `fusion.ts`), dispatched from `combo.ts`; the step list reuses `combo.models` order and reads an optional `prompt` off each target (backward-compatible — ignored by every other strategy). Intermediate steps run non-streaming with tools stripped (complete prose to thread forward); the final step keeps the client's `stream` flag + tools. A failing/empty/unparseable intermediate step fails the whole pipeline explicitly via a sanitized error (never silently swallowed). Kilo's dup flag vs #563 was a false positive (that's model→chain selection; this is a sequential chain). Regression guard: `tests/unit/combo-pipeline-strategy.test.ts` (5). (thanks @ofekbetzalel)
+- **feat(ci):** `check:test-masking` now flags **inline-reimplemented prod conditions** ([#6348](https://github.com/diegosouzapw/OmniRoute/issues/6348)) — a new **report-only** subcheck (v2, 6A.10 family) catches the wrong-shape contract test: a test that recomputes the condition under test inline instead of importing/exercising the real function (the #6216 class, where `=== 500` → `>= 500` stayed green because the test re-implemented the branch). For each added/modified test file it warns when the file textually duplicates a ≥3-token conditional from a production file touched in the same PR **and** does not import the symbol/module owning it, via a pure, fixture-tested `findReimplementedConditions()` with an allowlist mirroring `assertReductionAllowlist`. Report-only for now (does not fail the gate) — to be promoted to blocking after a triage cycle. Regression guard: `tests/unit/check-test-masking.test.ts` (45).
+- **feat(sse):** per-connection routing override (native vs CLIProxyAPI) ([#6339](https://github.com/diegosouzapw/OmniRoute/issues/6339)) — the previously-dead `isCliproxyapiDeepModeEnabled` helper is now wired into `resolveExecutorWithProxy`: a single connection can opt itself into the CLIProxyAPI passthrough executor via `providerSpecificData.cliproxyapiMode="claude-native"`, with precedence **connection override > provider `upstream_proxy_config` mode > default**. `resolveExecutorWithProxy` now receives the resolved connection's `providerSpecificData` (threaded from `chatCore.ts`), so one connection can deep-route while the provider's other connections stay native — no DB schema change (the toggle rides in `providerSpecificData`). Also resolves the same-provider mixing ask in #6340. Regression guard: `tests/unit/chatcore-executor-proxy.test.ts` (9). (thanks @RaviTharuma)
+- **feat(dashboard):** "Add session cookie" modal now shows a prominent **"Open ‹host› →"** link to the provider's own site ([#6268](https://github.com/diegosouzapw/OmniRoute/issues/6268)) — every `-web` cookie-session provider (chatgpt-web, claude-web, gemini-web, kimi-web, lmarena, qwen-web, m365-copilot-web, …) renders a one-click external link (opening the provider's login/home page in a new tab) so operators no longer tab away to retype the URL mid-setup. The host resolves from a pure, unit-tested `resolveWebProviderHost()` (prefers `WEB_COOKIE_PROVIDERS[id].website`, falls back to the registry `baseUrl` origin); non-web providers render exactly as before. Kilo's dup flag vs #6265 (modal-too-small-on-1080p) was a false positive — distinct concern. Regression guard: `tests/unit/resolve-web-provider-host.test.ts` (5). (thanks @chirag127)
+- **feat(providers):** add **DigitalOcean** AI (serverless inference) as an OpenAI-compatible API-key provider ([#6373](https://github.com/diegosouzapw/OmniRoute/pull/6373)) — base `https://inference.do-ai.run/v1`, wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/digitalocean/`, `src/shared/constants/providers/apikey/inference-hosts.ts`). Regression guard: `tests/unit/digitalocean-provider.test.ts`. (thanks @newnol)
+- **feat(providers):** add **Huancheng Public API** (`hcnsec`) as an OpenAI-compatible regional provider ([#6410](https://github.com/diegosouzapw/OmniRoute/pull/6410)) — Xinjiang Huancheng Cybersecurity's public LLM platform (base `https://api.hcnsec.cn/v1`, free credits via daily check-ins), wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/hcnsec/`, `src/shared/constants/providers/apikey/regional.ts`). Regression guard: `tests/unit/hcnsec-provider.test.ts`. (thanks @UnrealAryan)
+- **feat(dashboard):** the web-session credential guide now shows an **"Open {host}" link** ([#6316](https://github.com/diegosouzapw/OmniRoute/pull/6316)) to the provider's sign-in site (derived from the provider `website` via `getProviderWebsiteHost`), so you can jump straight to the page where the cookie/session must be captured. Regression guard: `tests/unit/web-session-provider-link-6316.test.ts`. (thanks @jordansilly77-stack)
+- **feat(cerebras):** add the **Gemma 4 31B** model (`gemma-4-31b`) to the Cerebras registry + pricing table ([#6331](https://github.com/diegosouzapw/OmniRoute/pull/6331)). Regression guard extends `tests/unit/t28-model-catalog-updates.test.ts`. (thanks @backryun)
+- **feat(providers):** add **Yuanbao (web)** as a cookie-session provider ([#6196](https://github.com/diegosouzapw/OmniRoute/issues/6196)) — `yuanbao-web` (Tencent Yuanbao, `yuanbao.tencent.com`) with cookie-only auth (`hy_user`/`hy_token` + public agent id), SSE→OpenAI translation incl. `reasoning_content`, exposing DeepSeek V3/R1 + Hunyuan / Hunyuan-T1. Regression guard: `tests/unit/providers-yuanbao-web.test.ts`. `together-web` was **deferred** (no verifiable web-session endpoint — needs a captured request) and `huggingchat-web` **dropped** (the existing `huggingchat` already is a web-cookie provider). (thanks @chirag127)
+- **feat(providers):** route the built-in **agentrouter** through the dynamic Claude-Code wire image ([#6056](https://github.com/diegosouzapw/OmniRoute/issues/6056)) — a small static allow-set (`CC_WIRE_IMAGE_BUILTINS` in `open-sse/services/ccWireImageBuiltins.ts`), consulted by `isClaudeCodeCompatible` / `isClaudeCodeCompatibleProvider` / `applyFingerprint`, makes agentrouter adopt the CC wire-image headers + fingerprint **while guarding the CC baseUrl/auth branches** so it keeps its own registry `baseUrl` and `x-api-key` auth. Regression guard: `tests/unit/agentrouter-cc-wire-image.test.ts` (asserts the wire image is applied AND agentrouter's baseUrl/auth are preserved). Live WAF-acceptance against agentrouter.org is a VPS validation follow-up (Hard Rule #18).
+- **feat(providers):** **bulk-add API keys for Cloudflare Workers AI** ([#6174](https://github.com/diegosouzapw/OmniRoute/issues/6174)) — `cloudflare-ai` is removed from the bulk-add exclusion list and the bulk parser gains a 3-field `name|accountId|apiKey` mode; the bulk route now builds a **per-entry** `providerSpecificData` so each key carries its own `accountId` (fixing the previous shared-object reuse), and both the create + key-validation paths receive it. Regression guard: `tests/unit/bulk-api-key-parser-cloudflare.test.ts`. (thanks @muflifadla38)
+- **feat(dashboard):** routing/settings UX clarity ([#6147](https://github.com/diegosouzapw/OmniRoute/issues/6147)) — (1) weighted combos show the **effective routing share %** next to each weight when weights don't sum to 100 (`WeightTotalBar.tsx`); (2) the status widget's user-facing **"Cloud Sync" label is renamed** to "Remote Settings Sync" (`CloudSyncStatus.tsx`; internal ids/state untouched); (3) built-in providers gain an **opt-in advanced base-URL override** (`isBaseUrlOverrideEligibleProvider`, hidden behind an "Advanced" toggle, reusing the existing `providerSpecificData.baseUrl` persistence — not globally widened). Regression guard: `tests/unit/routing-settings-ux-6147.test.ts`.
+- **feat(combo):** add an option to **disable session stickiness**, per-combo or globally — round-robin / random combos can rotate to a different connection on every request instead of pinning a whole conversation to one connection by its first-message hash. Resolution precedence per-combo `config.disableSessionStickiness` → global `settings.disableSessionStickiness` → default `false` (preserves the #3825 prompt-cache/504 fix); gates **both** stickiness call sites in `open-sse/services/combo.ts`. Exposed as a global toggle (Combo Defaults) and a per-combo Inherit/on/off control. ([#6168](https://github.com/diegosouzapw/OmniRoute/issues/6168)) Regression guard: `tests/unit/combo-disable-session-stickiness.test.ts`. (thanks @RCrushMe)
+- **feat(docker):** add the `OMNIROUTE_NO_SUDO` env flag for root-less / user-namespaced deployments — the MITM cert-trust command path (`resolveSudoSpawn` in `src/mitm/systemCommands.ts`) now strips the leading `sudo` when the flag is truthy, in addition to the existing root / sudo-missing cases, so the Proxy Agent runs without `sudo` (the operator trusts the CA manually, e.g. via `NODE_EXTRA_CA_CERTS`). Argv-array `spawn` preserved — no shell interpolation (Hard Rule #13). ([#6122](https://github.com/diegosouzapw/OmniRoute/issues/6122)) Regression guard: `tests/unit/mitm-systemCommands-no-sudo.test.ts`. (thanks @powellnorma)
+- **feat(providers):** add **Requesty** as an OpenAI-compatible gateway provider (BYOK, base `https://router.requesty.ai/v1`, ~200 free requests/day) — wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/requesty/`, `src/shared/constants/providers/apikey/gateways.ts`). ([#6120](https://github.com/diegosouzapw/OmniRoute/issues/6120)) Regression guard: `tests/unit/requesty-provider.test.ts`. (thanks @chirag127)
+- **feat(dashboard):** add **configured-only / available-only filters** to the Free Provider Rankings page ([#6150](https://github.com/diegosouzapw/OmniRoute/issues/6150)) — hide providers you haven't configured, or whose connections are all rate-limited / out of quota, via server-side query params (`?configuredOnly` / `?availableOnly` on `GET /api/free-provider-rankings`) backed by a testable lib helper reusing the in-process connection state (no Redis). Both filters default off, so the default view is unchanged; this supersedes the earlier client-side "Configured Only" toggle (#6245) with an available-only dimension and unit-tested logic. Regression guard: `tests/unit/freeProviderRankings-filters.test.ts`.
+
+### 🔧 Bug Fixes
+
+- **fix(dashboard):** adding a second API key connection for the same provider no longer silently overwrites the first — the Add-API-key modal now derives a unique default connection name (`main`, then `main-2`, `main-3`, …) so the backend name-based upsert can't collide (#6499 — thanks @dilneiss).
+- **fix(compression):** the session-dedup engine now also deduplicates a large multi-line block repeated **within a single message** (intra-message dedup), not just across turns; the compression-preview API surfaces a `fallbackReason`, and the fusion panel reports how many models were rate-limited vs failed on total-panel failure (#6501 — thanks @chirag127).
+- **fix(compression):** a stacked-pipeline step naming an unregistered engine now surfaces a `validationErrors` entry instead of silently no-op'ing, so misconfigured pipelines are visible in the preview API (#6506 — thanks @chirag127).
+- **feat(usage):** add a **Codex reset-credit redemption** flow to the Provider Limits UI ([#6361](https://github.com/diegosouzapw/OmniRoute/pull/6361)) — a `useCodexResetCreditRedemption` hook + `/api/usage/codex-reset-credit` route + `codexResetCredits` lib let you redeem banked Codex reset credits from the quota card. Regression guard: `tests/unit/codex-reset-credits.test.ts`. (thanks @JxnLexn)
+- **feat(glm):** add **team-plan quota settings** for `glm-cn` connections ([#6351](https://github.com/diegosouzapw/OmniRoute/pull/6351)) — a dedicated `GlmTeamQuotaFields` form section (team quota id / limits) threaded through the Add/Edit connection modals, persisted via `providerSpecificData`, with the GLM usage service reading the team quota. Regression guards: `tests/unit/glm-team-quota.test.ts`, `provider-specific-data-schema.test.ts`. (thanks @hao3039032)
+- **feat(providers):** add **TinyFish** web-fetch/search support ([#6349](https://github.com/diegosouzapw/OmniRoute/pull/6349)) — a `tinyfish-fetch` executor + `/v1/web/fetch` route + MCP web-fetch tool, registered as a specialty-media provider with request-validation and a search-provider catalog entry. Regression guards: `tests/unit/executor-tinyfish-fetch.test.ts`, `web-fetch-handler.test.ts`, `mcp-web-fetch-tool.test.ts`, `provider-validation-tinyfish.test.ts`. (thanks @dtybnrj)
+- **fix(cli):** `omniroute launch-codex` now spawns `codex.cmd` through a shell on Windows (the npm `.cmd` shim is unresolvable by bare `spawn` → ENOENT), mirroring the qodercli Windows fix (#6263) ([#6312](https://github.com/diegosouzapw/OmniRoute/pull/6312)). Regression guard: `tests/unit/launch-codex-windows-spawn-6312.test.ts`. (thanks @swingtempo)
+- **fix(codex):** isolate the **Spark** quota from the shared Codex quota and stabilize the quota UI ordering / hydration so per-scope limits render consistently ([#6336](https://github.com/diegosouzapw/OmniRoute/pull/6336)). Regression guards: `tests/unit/codex-quota-selection-hydration.test.ts`, `provider-limits-ui.test.ts` + 3 more. (thanks @xz-dev)
+- **feat(api):** add a `hidePaidModels` setting that filters paid-only models out of the `/v1/models` catalog. Regression guard: `tests/unit/models-catalog-hide-paid.test.ts`. (thanks @chirag127)
+- **fix(api-manager):** the fallback model picker now preserves combos instead of dropping them when a primary model is unavailable ([#6443](https://github.com/diegosouzapw/OmniRoute/pull/6443)). Regression guard: `tests/unit/api-manager-page-static.test.ts`. (thanks @jmengit)
+- **fix(providers):** recoverable Antigravity / Cloud-Code (Gemini Code Assist) `403` responses ([#6452](https://github.com/diegosouzapw/OmniRoute/pull/6452)) are now classified as a retryable project-config error instead of a terminal account ban, so a fixable project/API-disabled 403 no longer forces a ~1-year cooldown / full OAuth reconnect. Regression guard: `tests/unit/errorclassifier-antigravity-403.test.ts`. (thanks @developerjillur)
+- **fix(mitm):** `sanitizeHeaders` now redacts `Set-Cookie` response headers so upstream session cookies never leak into logs / diagnostics ([#6451](https://github.com/diegosouzapw/OmniRoute/pull/6451)). Regression guard: `tests/unit/mitm-sanitize-headers.test.ts`. (thanks @developerjillur)
+- **fix(api):** `/api/compression/preview` now accepts `mode: "caveman"` and correctly handles stacked / zero-compression previews ([#6425](https://github.com/diegosouzapw/OmniRoute/issues/6425)). Regression guard: `tests/unit/api/compression-preview-caveman-and-stacked-6425.test.ts`. (thanks @chirag127)
+- **feat(providers):** add **Zed** hosted LLM aggregator as a native-app provider ([#6118](https://github.com/diegosouzapw/OmniRoute/pull/6118)) — OAuth sign-in via the Zed hosted flow, registered through the shared provider registry + executor. Regression guards: `tests/unit/zed-oauth-provider.test.ts`, `zed-import-utils.test.ts`, `zed-docker-detect.test.ts`, `mitm-handler-zed.test.ts`. VPS-validated via live operator login (Hard Rule #18).
+- **fix(oauth):** the Kiro SSO-cache auto-import now **preserves the IDC region** — cross-region Amazon Q / Kiro profiles imported from the SSO cache are no longer collapsed to the default region ([#6113](https://github.com/diegosouzapw/OmniRoute/pull/6113)). Regression guard: `tests/unit/kiro-auto-import-idc-2059.test.ts`. VPS-validated via live operator login (Hard Rule #18).
+- **fix(dashboard):** passthrough model aliases no longer collide when two namespaced model ids share a last segment (port from 9router#1850, [#6431](https://github.com/diegosouzapw/OmniRoute/pull/6431)). `enx/gpt-5.5` and `enx/codebuddy/gpt-5.5` both auto-generated the alias `gpt-5.5`, so the second model could never be added (the UI just alerted "alias already exists"). Aliases are now disambiguated deterministically — bare last segment when free, then parent-qualified (`codebuddy-gpt-5.5`), then a numeric suffix — while re-adding the exact same model id is still blocked. Regression guard: `tests/unit/passthrough-alias-1850.test.ts`. (thanks @arpicato)
+- **fix(translator):** preserve a Gemini `functionResponse` co-located with other parts (another `functionCall`, or trailing `text`) in the same content when translating **Gemini → OpenAI** ([#6376](https://github.com/diegosouzapw/OmniRoute/pull/6376)). `convertGeminiContent()` early-returned the tool message on the first `functionResponse` part, dropping any co-located parts; such contents are now pre-split (one tool message per `functionResponse`, emitted first, plus one message for the remaining parts). Regression guard: `tests/unit/gemini-to-openai-function-response.test.ts`. (thanks @warelik)
+- **fix(headroom):** detect a python interpreter managed by **mise / pyenv / asdf / conda** (port from 9router#2353, [#6382](https://github.com/diegosouzapw/OmniRoute/pull/6382)). Headroom's python probe (`src/lib/headroom/detect.ts`) searched a hardcoded `PATH`, but version managers expose their interpreters via shim dirs that only join `PATH` through interactive-shell activation — which the non-interactive server never runs, so a managed python (≥3.10) was invisible and Headroom reported it missing. The search path now prepends the well-known shim/bin dirs (`~/.local/share/mise/shims`, `~/.pyenv/shims`, `~/.asdf/shims`, `$CONDA_PREFIX/bin`, `~/.local/bin`, respecting `MISE_DATA_DIR`/`PYENV_ROOT`/`ASDF_DATA_DIR` when set), and a new `HEADROOM_PYTHON` env override lets operators point straight at their interpreter (mirroring `HEADROOM_URL`). Still shell-free (`execFileSync`). Regression guard: `tests/unit/headroom-detect.test.ts` (5). (thanks @loopyd)
+- **fix(executors):** strip the OpenAI-Codex/Claude-CLI `client_metadata` passthrough field for **NVIDIA** requests (port from 9router#1887, [#6411](https://github.com/diegosouzapw/OmniRoute/pull/6411)). NVIDIA's OpenAI-compatible wrapper rejects it with `400 Unsupported parameter`, the same class already handled for `cerebras`/`mistral`; `nvidia` (executor `default`) was missing from the strip allowlist so Codex/Claude-Code passthrough requests 400'd. Regression guard: `tests/unit/executor-default-strip-client-metadata.test.ts` (+nvidia case). (thanks @phidinhmanh)
+- **fix(translator):** strip the Claude-style `thinking` field for **NVIDIA `z-ai/glm-5.2`** (port from 9router#2023, [#6413](https://github.com/diegosouzapw/OmniRoute/pull/6413)). NVIDIA's OpenAI-compatible wrapper 400s on `thinking` (a Claude-format client routed here leaves a `thinking:{type:"adaptive"}`); the existing strip rule only dropped `reasoning`. Same class already handled for `minimax-m2.7`. Regression guard: `tests/unit/nvidia-minimax-thinking-strip.test.ts` (+glm-5.2 case). (thanks @phidinhmanh)
+- **fix(translator):** suppress the streamed `</think>` close marker for the **Antigravity IDE** client (port from 9router#1061, [#6415](https://github.com/diegosouzapw/OmniRoute/pull/6415)). On thinking-only turns Antigravity rendered a bare `</think>` as the sole visible content, tripping its loop-detection and wasting requests. Antigravity's UA (`vscode/<v> (Antigravity/<v>)`) is added to the marker-suppress allowlist (alongside OpenCode); Claude Code / Cursor still get the marker, and `x-omniroute-thinking-marker: on` force-restores it. Regression guard: `tests/unit/think-close-marker-suppress-5245.test.ts`. (thanks @abdofallah)
+- **fix(executors):** strip nested `reasoning_content` from messages for **Mistral** (port from 9router#1649, [#6417](https://github.com/diegosouzapw/OmniRoute/pull/6417)). Mistral's API returns `422 extra_forbidden` when an assistant message carries `reasoning_content` (replayed thinking from a prior turn, e.g. via the Codex `/responses` path); the generic top-level 400 field-downgrade retry never covered the nested per-message field. `DefaultExecutor` now strips it for provider `mistral` only, so DeepSeek (which requires replayed `reasoning_content`) is unaffected. Regression guard: `tests/unit/mistral-strip-reasoning-content-1649.test.ts`. (thanks @xxy9468615)
+- **fix(executors):** strip the `client_metadata` passthrough field on the **OpenCode** path (port from 9router#1442, [#6418](https://github.com/diegosouzapw/OmniRoute/pull/6418)). OpenCode upstreams (e.g. `kimi-k2.6` via opencode-go) reject it with `400 "Extra inputs are not permitted, field: 'client_metadata'"`; the DefaultExecutor strip only covered cerebras/mistral and `OpencodeExecutor` extends `BaseExecutor` directly, so nothing removed it there. Regression guard: `tests/unit/opencode-strip-client-metadata-1442.test.ts`. (thanks @yanpaing007)
+- **fix(executors):** inject the `reasoning_content` echo for the native **Moonshot Kimi** provider (port from 9router#1480, [#6419](https://github.com/diegosouzapw/OmniRoute/pull/6419)). Kimi (executor `default`) is a thinking-mode upstream that 400s with "reasoning_content must be passed back" when a prior assistant turn lacks it; the placeholder injection was only wired into the OpenCode meta-provider, so direct multi-turn Kimi conversations failed. Scoped to `kimi` (gateway-served models matching the thinking-model name pattern are unaffected). Regression guard: `tests/unit/kimi-native-reasoning-injected-1480.test.ts`. (thanks @2220258345)
+- **fix(executors):** recover from a strict gateway's `context_management: Extra inputs are not permitted` 400 (port from 9router#1468, [#6420](https://github.com/diegosouzapw/OmniRoute/pull/6420)). **Claude Code** always sends a top-level `context_management` field; strict anthropic-compatible gateways reject it. The dedicated context-editing 400-fallback only fired when OmniRoute's own `contextEditing` feature was enabled (default off), so a client-sent field passed through untouched and 400'd. `context_management` is now in the generic reactive field-strip list, so it's stripped-and-retried once regardless of the feature flag (with correct request re-signing for claude-compatible relays). Regression guard: `tests/unit/provider-field-strips.test.ts`. (thanks @ohahe52-dot)
+- **fix(network):** enable **RFC 8305 Happy Eyeballs** (`autoSelectFamily`) on the direct-egress undici dispatcher (port from 9router#1237, [#6423](https://github.com/diegosouzapw/OmniRoute/pull/6423)). When DNS returns both IPv6 (AAAA) and IPv4 (A) and the IPv6 route is broken (e.g. a NAT64 `64:ff9b::` prefix without routing), undici tried IPv6 first and hung until `ETIMEDOUT` (then a 502 + account lockout), even though `curl` reached the same host. The direct dispatcher now races both families and uses whichever connects first. Proxy paths pin family via `proxyTls` and are unaffected. Regression guard: `tests/unit/direct-dispatcher-pipelining-4580.test.ts`. (thanks @adentdk)
+- **fix(combo):** round-robin now advances the rotation pointer past the model that **actually served**, not the eagerly-scheduled one (port from 9router#948, [#6428](https://github.com/diegosouzapw/OmniRoute/pull/6428)). With `stickyLimit: 1` (true round-robin), when the scheduled model failed and a _different_ model served via fallback, the counter had already advanced +1 from the scheduled index — so the next request reused the fallback-served model, degrading round-robin into hot-spotting on whichever model was healthy. The pointer now advances to the served index + 1 (mirroring the sticky-limit>1 path). Session-stickiness (#3825) and distribution are preserved. Regression guard: `tests/unit/combo-rr-fallback-advance-948.test.ts`. (thanks @binsarjr)
+- **fix(sse):** a non-string `model` field is now rejected with a `400` before the resolver, instead of crashing downstream `.toLowerCase()`/`.split()` calls into an empty-body `500` that escapes the error sanitizer ([#6407](https://github.com/diegosouzapw/OmniRoute/issues/6407)). Regression guard: `tests/unit/chat-non-string-model-6407.test.ts`. (thanks @chirag127)
+- **fix(api):** unknown `/api/*` routes now return a JSON `404` (instead of the dashboard HTML shell) and scalar chat params (`model`/`temperature`/etc.) are validated **before** the provider lookup so malformed requests fail fast with a clear `400` ([#6424](https://github.com/diegosouzapw/OmniRoute/issues/6424), [#6412](https://github.com/diegosouzapw/OmniRoute/issues/6412)). Regression guards: `tests/unit/api/api-catchall-json-404.test.ts`, `tests/unit/chat-early-schema-validation-6412.test.ts`. (thanks @chirag127)
+- **fix(api):** `/v1/chat/completions` now rejects a non-JSON `Content-Type` with a `400` before parsing the body ([#6414](https://github.com/diegosouzapw/OmniRoute/issues/6414)). Regression guard: `tests/unit/v1-chat-completions-content-type-6414.test.ts`. (thanks @chirag127)
+- **fix(api):** the `X-OmniRoute-Compression` response header is now echoed on `/v1/chat/completions` and `/v1/completions` ([#6422](https://github.com/diegosouzapw/OmniRoute/issues/6422)). Regression guard: `tests/unit/compression-header-echo-6422.test.ts`. (thanks @chirag127)
+- **fix(api):** concurrent `GET /v1/models` requests are coalesced into a single catalog build ([#6408](https://github.com/diegosouzapw/OmniRoute/issues/6408)). Regression guard: `tests/unit/v1-models-concurrent-6408.test.ts`. (thanks @chirag127)
+- **fix(api):** `/v1/completions` now echoes the requested `body.model` in its JSON + streamed responses ([#6429](https://github.com/diegosouzapw/OmniRoute/pull/6429)). Regression guard: `tests/unit/completions-body-model-echo.test.ts`. (thanks @chirag127)
+- **fix(api):** env-var master keys now see the full `/v1/models` catalog ([#6406](https://github.com/diegosouzapw/OmniRoute/issues/6406)). Regression guard: `tests/unit/models-catalog-envkey-6406.test.ts`. (thanks @chirag127)
+- **fix(api):** non-streaming `/v1/completions` responses now echo `body.model` aligned with the `X-OmniRoute-Model` header ([#6426](https://github.com/diegosouzapw/OmniRoute/issues/6426)). Regression guard: `tests/unit/v1-completions-model-header-match-6426.test.ts`. (thanks @chirag127)
+- **fix(api):** unknown `/v1/*` routes now return a JSON `404 not_found` instead of the Next.js dashboard HTML shell ([#6405](https://github.com/diegosouzapw/OmniRoute/issues/6405)). Regression guard: `tests/unit/api/v1-catchall-json-404.test.ts`. (thanks @chirag127)
+- **fix(api):** the per-connection provider models route now degrades to the shipped catalog when a provider's `/models` endpoint answers with a redirect ([#6267](https://github.com/diegosouzapw/OmniRoute/issues/6267)) — a `qwen-web` import failed with a raw `Redirect blocked … (307)` 503. `safeOutboundFetch` throws `REDIRECT_BLOCKED` on the 307, `getSafeOutboundFetchErrorStatus` maps it to 503, and `buildDiscoveryErrorFallbackResponse` treated every 503 as a hard error — so the non-empty `getModelsByProviderId("qwen-web")` catalog was never surfaced. A models-endpoint redirect is not a fixable-config error (unlike `URL_GUARD_BLOCKED`/`INVALID_URL`, which stay hard errors), so it now falls back to the local/cached catalog before the 503 short-circuit. General fix — covers any config-driven provider that 307s. Regression guard: `tests/unit/provider-models-qwen-web-redirect-6267.test.ts`. (thanks @chirag127)
+- **fix(api):** the per-connection provider models route (MCP `list_models_catalog` + the dashboard import view) now merges USER-ADDED custom models into its response ([#6247](https://github.com/diegosouzapw/OmniRoute/issues/6247)) — custom models live in the `key_value` namespace `customModels`, which the live REST `/api/v1/models` already merges, but `src/app/api/providers/[id]/models/route.ts` never read `getCustomModels`, so custom models were dropped on both the discovery-success and local_catalog paths. They are now folded into the returned model list (deduped by id, stamped `owned_by: provider`), fixing MCP + the dashboard import view in one place. Regression guard: `tests/unit/provider-models-custom-merge-6247.test.ts`. (thanks @RCrushMe)
+
+- **fix(providers):** GitLab Duo tool-calling follow-up turns no longer fail upstream with `422 {"detail":"Validation error"}` (tokens 0/0, rejected pre-inference). The #6234 tool-result-feedback fix serialized the **entire** multi-turn conversation into GitLab's single-file `code_suggestions` (`small_file`) generation endpoint — folded history that turn-N sent as an oversized `current_file.content_above_cursor` **and** duplicated verbatim into `user_instruction`, tripping the AI-Gateway's `small_file` validation guard. The executor now **bounds** that prompt: it keeps system + latest user message + the most-recent tool round (dropping older turns), caps oversized tool results, and stops duplicating the full prompt into `user_instruction` (which now carries only the short latest user message) — while still feeding the most-recent tool result back so the agent continues (`open-sse/executors/gitlab.ts`, [#6220](https://github.com/diegosouzapw/OmniRoute/issues/6220)). The unit test covers the bounding logic; the upstream 422→200 clearing is VPS-only (Hard Rule #18). Regression guard: `tests/unit/gitlab-tool-exchange-bounded-6220.test.ts`.
+
+- **fix(i18n):** the provider-detail (`/dashboard/providers/[id]`) connection-status filter labels no longer render as `__MISSING__:All` / `__MISSING__:Active` / `__MISSING__:Error` / `__MISSING__:Banned` / `__MISSING__:CreditsExhausted` in non-English locales (notably pt-BR) ([#6290](https://github.com/diegosouzapw/OmniRoute/issues/6290)). Root cause was **not** the namespace mismatch the issue guessed — the `providers.filter*` keys resolve correctly in `en.json`; the debt lived in the locale mirrors (`src/i18n/messages/*.json`), where these five keys carried the `__MISSING__:` sync sentinel in ~15 locales and were absent entirely in ~26 others, so next-intl found the key and echoed the sentinel verbatim. All 40 non-English/-Chinese mirrors now ship real translations for the five `providers.filter*` labels. Regression guard: `tests/unit/i18n-provider-filter-keys-6290.test.ts`. (thanks @diegosouzapw)
+
+- **fix(providers):** the `copilot-m365-web` streaming executor now emits `debug`-level WebSocket diagnostics ([#6210](https://github.com/diegosouzapw/OmniRoute/issues/6210)) — the outbound WS URL (with the `access_token` **redacted** via `redactWsUrl()`), handshake success/failure, and each received SignalR frame's `type`/`target`. Previously the streaming path logged nothing, so an empty `content:null` response (the M365 Education / Starter tier symptom fixed in #6234) was undiagnosable even at `APP_LOG_LEVEL=debug`. The change is debug-level and side-effect-free — it does not alter streaming behavior or the frame parser, and the token never reaches the logs. Regression guard: `tests/unit/copilot-m365-web-logging-6210.test.ts` (thanks @qpeyba)
+
+- **fix(resilience):** a round-robin combo no longer returns `503 all upstream accounts are unavailable` when a compatibility-rejected target is actually healthy ([#6238](https://github.com/diegosouzapw/OmniRoute/issues/6238)). `filterTargetsByRequestCompatibility` drops request-incompatible targets (tool/vision/structured-output unsupported, or below the required context window) **before** any availability check runs, and its `compatible.length === 0` safety net only fired when _all_ targets were filtered — not when the kept targets later all turned out runtime-unavailable (circuit-open / cooldown / no credentials). So a combo could 503 while a compat-rejected-but-healthy provider sat unused. `handleRoundRobinCombo` now keeps the compat-rejected set and, when every compat-kept target was skipped without a single real attempt, probes those rejected targets as a **last-resort fallback tier** (via the new pure `open-sse/services/combo/comboCompatFallback.ts`) before crystallizing the 503. Regression guard: `tests/unit/combo-roundrobin-compat-fallback-6238.test.ts`. (thanks @ThongAccount)
+
+- **fix(startup):** best-effort self-heal for a corrupted Turbopack dev cache on Windows ([#6289](https://github.com/diegosouzapw/OmniRoute/issues/6289)). On Windows, `pnpm dev` can fail at startup when Turbopack `mmap`s a persistent-cache SST file and the OS refuses the mapping (`os error 1455` — "paging file too small"), which Turbopack surfaces as a misleading `Module not found: Can't resolve '@/shared/utils/machine'`. This is a **known upstream Turbopack cache-corruption bug — not our code**. The dev launcher (`scripts/dev/run-next.mjs`) now wraps `nextApp.prepare()` and, when it rejects with that signature (`isTurbopackCacheCorruption` in the new `scripts/dev/turbopackCacheHeal.mjs`), purges `.build/next/**/cache/turbopack` and retries **once** with a clear log. **Caveat — best-effort only:** the corruption often surfaces as a runtime overlay rather than a `prepare()` rejection, so this cannot always intercept it; the reliable remedy remains manually deleting the Turbopack cache dir. Regression guard: `tests/unit/turbopack-cache-heal-6289.test.ts`. (thanks @chirag127)
+
+- **fix(providers):** qodercli PAT auth no longer fails with `spawn qodercli ENOENT` on Windows ([#6263](https://github.com/diegosouzapw/OmniRoute/issues/6263)) — `spawnQoderCli` spawned the bare `qodercli` name with `shell:false` and an unenriched env, so the npm `.cmd` wrapper under `%APPDATA%\npm` (a user-PATH directory) was never resolved. It now resolves the absolute `.cmd`/`.exe` path through the existing `getCliRuntimeStatus("qoder")` resolver in `src/shared/services/cliRuntime.ts` (memoized), spawns with `shell` when the target is a `.cmd`/`.bat`, and uses the cliRuntime-enriched env (PATH + PATHEXT + APPDATA); the ENOENT error now lists the searched paths plus the `CLI_QODER_BIN` override. End-to-end spawn on a real Windows host is host-only (Hard Rule #18); the path-resolution logic is unit-tested. Regression guard: `tests/unit/qodercli-windows-resolve-6263.test.ts`. (thanks @chirag127)
+
+- **fix(sse):** the reasoning-token buffer no longer inflates **probe-sized `max_tokens`** ([#6274](https://github.com/diegosouzapw/OmniRoute/issues/6274)) — Claude Code's `/model` capability check sends `max_tokens: 1`, but for a thinking-capable model with a large output cap (e.g. `glm-5.2`) the #3587 headroom heuristic (`max(current + 1000, ceil(current * 1.5))`) rewrote it to `1001` and forwarded that upstream, wasting tokens on a request that was never a genuine reasoning budget. `resolveReasoningBufferedMaxTokens()` (`open-sse/services/reasoningTokenBuffer.ts`) now short-circuits and returns the caller's value verbatim when it is below the new `REASONING_BUFFER_MIN_TRIGGER` (256) threshold — a tiny explicit limit is a probe, not a reasoning request. Real budgets still receive the #3587 headroom unchanged, and the guard runs after the existing capability checks so unknown / non-reasoning models keep returning `null`. Regression guard: `tests/unit/reasoning-token-buffer-6274.test.ts`. (thanks @brightfiscalband)
+
+- **fix(cli):** `omniroute reset-password` now works as a real subcommand, and password resets over piped (non-TTY) stdin actually apply ([#6261](https://github.com/diegosouzapw/OmniRoute/issues/6261), [#6258](https://github.com/diegosouzapw/OmniRoute/issues/6258)). Two coupled defects: (1) **#6261** — `bin/omniroute.mjs` routed everything through Commander with only two pre-Commander bypasses (`--mcp`, `reset-encrypted-columns`), so `omniroute reset-password` was rejected as an unknown command; only the separate `omniroute-reset-password` bin worked, while the docs falsely advertised the subcommand (incl. a bogus "legacy alias still works"). A pre-Commander bypass mirroring `reset-encrypted-columns` now dynamically imports `bin/reset-password.mjs` (which self-executes) before Commander parses; the three doc lines were corrected. (2) **#6258** — `bin/reset-password.mjs` issued two sequential `rl.question` prompts; under piped stdin the second read never settled at EOF, so `main()` never reached `resetManagementPassword` and the reset was a silent no-op (both prompts printed, no success, password unchanged). The CLI now detects non-TTY stdin and reads it once (first line = password, second line = confirm if present, else reused), adds a `--password-stdin` flag (entire stdin is the password, no confirmation), and exits `0` explicitly so the success line always flushes; interactive TTY behavior is unchanged. Regression guard: `tests/unit/reset-password-cli-6261-6258.test.ts` (3). (thanks @chirag127)
+
+- **fix(db):** the mass-migration **safety abort** now tells the operator how to bypass it and stops flooding the log ([#6260](https://github.com/diegosouzapw/OmniRoute/issues/6260)) — after restoring a backup that wiped the migration tracking table, `runMigrations()` threw the abort on every downstream `ensureDbInitialized()`, re-logging the full banner 11+ times, and the message never mentioned the existing `OMNIROUTE_MAX_PENDING_MIGRATIONS` escape hatch. The abort text now appends a bypass hint (set `OMNIROUTE_MAX_PENDING_MIGRATIONS=0` in `server.env` / `DATA_DIR/.env`), and a new `MigrationSafetyAbortError` is memoized so repeated calls in the same process throw the same instance and emit a single concise line instead of the full cascade. Regression guard: `tests/unit/migration-safety-abort-6260.test.ts`. (thanks @chirag127)
+
+- **fix(auth):** importing a **distinct** Codex/ChatGPT OAuth `auth.json` is no longer falsely rejected as "already exists" when it belongs to a different user in the same workspace ([#6301](https://github.com/diegosouzapw/OmniRoute/issues/6301)). `findExistingCodexConnection` (in `src/lib/oauth/utils/codexAuthImport.ts`) deduped **only** on `providerSpecificData.workspaceId === accountId`, where `accountId` is the shared `chatgpt_account_id`/`tokens.account_id` — so two members of the same ChatGPT Team collapsed onto a single connection (409 `duplicate_account`). The id_token's `https://api.openai.com/auth` claim carries a per-user `chatgpt_user_id` alongside the workspace id (the device-flow path already persisted it as `chatgptUserId`, but the import path did not). Now `parseAndValidateCodexAuth` extracts `userId` (`chatgpt_user_id` → `user_id` → JWT `sub`) into `ParsedCodexAuth`, the create/update paths persist `chatgptUserId` in `providerSpecificData` (mirroring `codex.ts`), and dedup keys on `workspaceId` **AND** `chatgptUserId` — with a backward-compat fallback to legacy accountId-only matching when no stored connection for that workspace records a `chatgptUserId`, so genuinely-same accounts still dedup. Regression guard: `tests/unit/codex-auth-import-userid-dedup-6301.test.ts` (4). (thanks @anungma)
+
+- **fix(providers):** importing models for the **venice-web** provider no longer fails with a red "Provider venice-web does not support models listing" ([#6269](https://github.com/diegosouzapw/OmniRoute/issues/6269)). `venice-web` is a web-cookie provider with an executor but no upstream `/v1/models` endpoint and no registry `models`, so the models route fell through to the tail `400`. Mirroring the `jules`/`linkup-search`/`ollama-search` fix (#5569), it now ships a static local catalog entry in `src/lib/providers/staticModels.ts` — seeding the current Venice lineup (`venice-uncensored`, `llama-3.3-70b`, `qwen3-235b`, `qwen3-4b`, `deepseek-r1-671b`; Venice rotates its catalog, see docs.venice.ai/models/overview) — so the route returns `200` with `source:"local_catalog"`, `intentional:true`. Regression guard: `tests/unit/static-models-venice-web-6269.test.ts`. (thanks @chirag127)
+
+- **fix(api):** the specialty model catalogs (`/v1/embeddings`, `/v1/images`, `/v1/music`, `/v1/videos` model lists) are now derived from the **unified catalog filtered by a predicate** (`getSpecialtyModelsResponse`) instead of ad-hoc per-route logic, so they consistently respect active-credential visibility and stay in sync with the main catalog ([#6303](https://github.com/diegosouzapw/OmniRoute/pull/6303)). Regression guard: `tests/unit/specialty-model-catalog-routes.test.ts`. (thanks @makcimbx)
+- **fix(api):** the agent-bridge server route now resolves the MITM manager via a **dynamic `import("@/mitm/manager.runtime")`** so Turbopack does not statically pull the stub (or over-bundle the manager), and the agent-skills generator anchors its output base path with `path.join(process.cwd(), …)` so Turbopack's static analyzer stops tracing the whole project root ([#6329](https://github.com/diegosouzapw/OmniRoute/issues/6329), [#6366](https://github.com/diegosouzapw/OmniRoute/pull/6366)). Regression guard: `tests/unit/agent-bridge-server-route-dynamic-import.test.ts`. (thanks @Iammilansoni)
+- **fix(api):** internal probes (combo-test, cloud-sync verify) now pick a **management-scoped / allow-all API key** instead of naively grabbing `getApiKeys()[0]` — a restricted `self:usage` first row made the probe fail with "Model X is not allowed for this API key" even when the combo path was healthy (`pickApiKeyForInternalUse` in `src/lib/db/apiKeys.ts`). The API-manager model editor also falls back to `/api/models?all=true` when `/v1/models` is catalog-protected ([#6372](https://github.com/diegosouzapw/OmniRoute/pull/6372)). Regression guard: `tests/unit/pick-internal-api-key-6372.test.ts`. (thanks @jmengit)
+- **fix(live-ws):** the Live Dashboard WebSocket server now **rejects on bind failure** (e.g. `EADDRINUSE` when the API bridge already holds the port) instead of letting the error surface as an unhandled `error` event that crash-loops the process — the `error` listener is attached to `wss` (not `server`) and releases the EventBus subscription on a failed start ([#6324](https://github.com/diegosouzapw/OmniRoute/issues/6324)). Regression guard: `tests/unit/live-ws-eaddrinuse-6324.test.ts`. (thanks @vinayakkulkarni)
+- **fix(dashboard):** the Home provider-topology widget now trusts the live provider-metrics snapshot — it uses `topology.errorProvider` and live `activeRequests` directly instead of re-deriving state from a stale `lastErrorAt` or applying a frontend timeout filter, so the topology reflects real-time provider health ([#6322](https://github.com/diegosouzapw/OmniRoute/pull/6322)). Regression guard: `tests/unit/home-provider-topology-live-state.test.ts`. (thanks @xz-dev)
+- **fix(sse):** strip zero-width markers from streamed **tool-call arguments** — a follow-up to [#5857](https://github.com/diegosouzapw/OmniRoute/pull/5857). That PR removed injected zero-width joiners (U+200D) from streamed assistant text/reasoning but deliberately left tool-call argument JSON byte-exact. The request-side obfuscation (`open-sse/services/claudeCodeObfuscation.ts`) injects ZWJ into agent words — including the temp path inside the Bash tool description — and Claude models copy that verbatim into generated commands, which are delivered as tool-call arguments rather than assistant text. As a result the ZWJ survived and corrupted code blocks (e.g. a temp path rendered with an invisible joiner). Now `open-sse/handlers/responseSanitizer.ts` strips zero-width code points from tool-call argument strings at every emit site (OpenAI non-stream/stream chat `tool_calls` + legacy `function_call`, native Responses `function_call` items, the OpenAI→Responses conversion, and the native Responses streaming `response.function_call_arguments.delta/.done` events). Only zero-width code points are removed; JSON structure and all other bytes stay identical (no parse/restringify), so normal arguments remain byte-exact. Regression guard: 6 new cases in `tests/unit/response-sanitizer.test.ts` (suite 50/50).
+
+- **fix(nodejs):** the default app log path now resolves under `DATA_DIR` (`~/.omniroute/logs/application/app.log`) instead of `process.cwd()` ([#6197](https://github.com/diegosouzapw/OmniRoute/issues/6197)) — the globally-installed CLI runs from an arbitrary working directory, so anchoring the default to cwd made file logging silently write to (or no-op under) an unrelated directory, contradicting the documented `.env.example` default. `getAppLogFilePath()` now computes the default lazily via the pure `resolveDataDir()` resolver (honours a per-process `DATA_DIR`, no directory-creation side effect); an explicit `APP_LOG_FILE_PATH` still wins. Regression guard: `tests/unit/logenv-datadir-path-6197.test.ts` (3).
+
+- **fix(docker):** AgentBridge/`startMitm` no longer aborts in containers/headless when the Antigravity-default DNS step can't write `/etc/hosts` ([#6127](https://github.com/diegosouzapw/OmniRoute/issues/6127)), and the privileged command's stderr now reaches `app.log` instead of only a bare exit code hitting the toast ([#6198](https://github.com/diegosouzapw/OmniRoute/issues/6198)). The default DNS step (`addDNSEntry`) was called unguarded while cert install and the two sibling DNS steps were each best-effort — in the runtime Docker image (`USER node`, no `sudo`, read-only `/etc/hosts`) it threw `Command failed with code 1` out of `startMitmInternal` and killed the whole start, discarding the stderr. The three DNS steps are extracted into a best-effort `provisionDnsEntries()` where each failure is logged with the full `err` (stderr included, folded in by `systemCommands.ts`) and never aborts the start. Regression guard: `tests/unit/mitm-dns-graceful-degrade-6127.test.ts` (4).
+
+- **fix(providers):** copilot-m365-web now supports the M365 Education "Starter / OfficeWebIncludedCopilot" tier and no longer returns an empty `content:null` stream ([#6210](https://github.com/diegosouzapw/OmniRoute/issues/6210)). Two gaps: (1) `buildWsUrl()` hardcoded the individual-consumer scenario (`OfficeWebPaidConsumerCopilot`, `isEdu=false`) — the EDU tier is now opt-in via `providerSpecificData.tier="edu"`, emitting `scenario=OfficeWebIncludedCopilot`/`isEdu=true` (the individual path is unchanged); (2) the EDU/GPT-5.5 path streams deltas via `arguments[0].writeAtCursor` (incremental) instead of only `messages[].text` (accumulated snapshots), which the parser dropped — a new `accumulateBotContent()` folds both formats, with `type:2 item.result.message` as a last-resort fallback. Regression guard: `tests/unit/copilot-m365-edu-writeatcursor-6210.test.ts` (10). (thanks @qpeyba)
+
+- **fix(providers):** GitLab Duo executor now feeds tool results back into the prompt instead of looping ([#6220](https://github.com/diegosouzapw/OmniRoute/issues/6220)) — `buildPrompt()` branched only on `system`/`user` and took `userParts.at(-1)`, silently dropping the `assistant{tool_calls}` + `tool{result}` turns the client appended, so the reconstructed prompt was byte-identical to turn 1 and the model re-emitted the same `<tool>` call forever. When a tool exchange is present the full conversation is now serialized, folding each tool result back keyed by its `tool_call_id`; simple conversations keep the legacy shape. Complements the tool_call emission from [#6051](https://github.com/diegosouzapw/OmniRoute/issues/6051) (the `kilo-duplicate` label was a false positive — different, sequential defect). Regression guard: `tests/unit/gitlab-tool-result-feedback-6220.test.ts` (4).
+
+- **fix(providers):** opencode-go/opencode-zen can now synthesize the OpenCode CLI identity headers Cloudflare requires on VPS egress ([#5997](https://github.com/diegosouzapw/OmniRoute/issues/5997)) — on a datacenter VPS, `opencode.ai/zen/go/v1/chat/completions` 403s (HTML challenge) requests lacking CLI identity, while the reporter's control curl proved that `User-Agent: opencode-cli/1.0.0` + `x-opencode-client: cli` + `x-opencode-project: default` + fresh request/session UUIDs succeed. Opt-in via `OPENCODE_SYNTHESIZE_CLI_HEADERS=true` (values overridable via `OPENCODE_GO_USER_AGENT`/`OPENCODE_USER_AGENT`/`OPENCODE_CLIENT`/`OPENCODE_PROJECT`); it fills only headers the client did not already send. Kept **off by default** — the forward-only path is deliberate (fabricating a wrong value risks upstream rejection; a prior dedup regressed with `opencode/local`), so this replaces the fragile local header-injection shim without changing default behavior. Regression guard: `tests/unit/opencode-cli-headers-synthesis-5997.test.ts` (6). (thanks @aleksesipenko)
+- fix(resilience): sticky session affinity now evicts and fails over to another account when the pinned account is exhausted/unavailable (#6219)
+
+- fix(sse): Responses API passthrough now drops internal commentary-phase output before forwarding to clients (gated by RESPONSES_PASSTHROUGH_DROP_COMMENTARY, default on) (#6199)
+
+- **fix(sse):** tool-call function schemas with a root `type: null` are now coerced to `type: "object"` before dispatch ([#6359](https://github.com/diegosouzapw/OmniRoute/issues/6359)) — clients like the Codex app emit `parameters: { type: null, ... }` for some tools, which OpenAI-compatible upstreams reject with `400 Invalid schema for function '...': schema must be a JSON Schema of 'type: "object"', got 'type: null'`, failing the whole request. `toolSchemaSanitizer` already stripped the null; it now re-adds the mandatory root `"object"` type (and empty `properties`/open `additionalProperties` when absent). Combinator roots (`anyOf`/`oneOf`/`allOf`) and explicit root types are left untouched. Regression guard: 5 new cases in `tests/unit/tool-schema-sanitizer.test.mjs`.
+
+- **fix(docker):** AgentBridge no longer fails to start on npm/Electron/VPS installs with "MITM manager stub reached at runtime" ([#6344](https://github.com/diegosouzapw/OmniRoute/issues/6344)) — v3.8.45 flipped the production bundler default to Turbopack, but `next.config.mjs` aliased `@/mitm/manager` to its Docker-only degraded stub **unconditionally**. That was harmless while Docker (which sets the alias intentionally for #3390 graceful degradation) was the sole Turbopack consumer, but once every artifact built with Turbopack the stub shipped to all non-Docker users and `startMitm` threw on the first Agent-Bridge start. The alias is now opt-in via `OMNIROUTE_MITM_STUB=1` (set only by the Dockerfile) through the shared `scripts/build/mitm-stub-flag.mjs` helper; default builds bundle the real manager. Regression guard: `tests/unit/mitm-stub-alias-6344.test.mjs` (4).
+
+- **fix(proxy):** stop the v3.8.44 proxy regression that leaked the real IP and disabled healthy proxies ([#6246](https://github.com/diegosouzapw/OmniRoute/issues/6246)). Two coupled defects from the new health scheduler: (1) **IP leak** — when a proxy assigned to a connection was marked `inactive`, resolution fell through to a **direct** egress instead of blocking, exposing the operator's real IP; (2) **over-deactivation** — the sweep flipped a proxy to `inactive` on the **first** failed probe and counted our own 5s timeout / a probe-target `5xx` as the proxy's fault, so healthy paid proxies vanished from egress selection ("my proxies are not being used anymore"). Fix: the sweep decision is extracted into a pure, network-free `decideProxyHealthAction` (`src/lib/proxyHealth/decision.ts`) — by default the health check now **only counts/logs and never downgrades status** (a proxy is downgraded/removed only with `PROXY_AUTO_REMOVE=true`, after `PROXY_AUTO_REMOVE_AFTER` **consecutive** conclusive failures); probes are classified tri-state so an inconclusive result (our timeout, or a `5xx` from the probe target) never penalizes the proxy, and the probe timeout is raised 5s→15s. Separately, `safeResolveProxy` now **fails closed** via the existing policy: a connection whose assigned proxy is dead is blocked instead of leaking direct (`hasBlockingProxyAssignment`), honoring the explicit `proxy off` toggles and the `PROXY_FAIL_OPEN=true` opt-out. Existing proxies stuck `inactive` by the old behavior need a one-time manual re-activate (the operator owns proxy status). Regression guards: `tests/unit/proxy-health-decide-action-6246.test.ts`, `tests/unit/proxy-assigned-unavailable-6246.test.ts`.
+
+- **fix(proxy):** make "Test All" read-only and add bulk enable/disable ([#6246](https://github.com/diegosouzapw/OmniRoute/issues/6246)). Complements the core fail-closed / scheduler fix (#6296) with the two remaining reporter asks. (1) The **"Test All" button** (`POST /api/settings/proxies/auto-test`) used to flip a proxy to `inactive` on a failed reachability probe; since the egress selector excludes `inactive` proxies, a flaky probe (an unreachable `httpbin.org`, a proxy that blocks `HEAD`, or a slow paid proxy) silently disabled every proxy that failed — "Test All" is now **read-only by default** (only the operator sets a proxy active/inactive; opt back into the legacy test-and-set with `PROXY_HEALTH_AUTO_DEACTIVATE=true`). (2) Adds a **bulk enable/disable** proxies endpoint + toolbar action (`POST /api/settings/proxies/batch-activate`) so an operator can re-activate proxies in one click. Regression guard: `tests/unit/proxy-health-6246.test.ts`. (thanks @tenshiak)
+
+- **chatcore (tools): stop the default 128-tool cap from silently dropping opencode's `task`/MCP tools.** opencode (used as an MCP/agent host) sends a large tool list; when it exceeds the speculative `MAX_TOOLS_LIMIT` (128) default, `truncateToolList` did a blind `tools.slice(0, 128)`, dropping every tool past index 128 — including opencode's built-in `task` tool (subagent launch) and many MCP tools, so models routed through OmniRoute could no longer spawn subagents or reach part of their tools. The cap exists to avoid upstream `400`s for providers with real hard limits (e.g. grok-cli 200), so it is kept for those: detection of the opencode client (`isOpencodeClient` — any `x-opencode-*` header, or `opencode` in the user-agent) now only bypasses the **speculative 128 default**, never a known provider ceiling. Precedence is explicit — a proactive/detected provider limit always truncates (even for opencode); otherwise opencode forwards its full tool list; otherwise the unchanged 128 default applies to every other client. Refactors `getEffectiveToolLimit` into `getKnownToolLimit(provider) ?? DEFAULT_LIMIT` (byte-identical for existing callers) and fixes a cosmetic debug-log that reported the truncated count instead of the original. Regression guard: `tests/unit/tool-limit-detector.test.ts`.
+
+- **fix(mitm):** the macOS MITM-cert install check now matches the system keychain again. `security find-certificate -a -Z` prints the SHA-1 as a colon-less hex string, but the installed-check compared it against `getCertFingerprint()`'s colon-separated form, so the substring match never hit — the cert was reported as not-installed and re-prompted for the sudo install on every run. Fingerprints are now normalized (colons stripped, upper-cased) on both sides via the extracted `macCertOutputHasFingerprint` helper. Regression guard: `tests/unit/mitm-cert-mac-fingerprint.test.ts`. ([#6204](https://github.com/diegosouzapw/OmniRoute/pull/6204), closes [#6134](https://github.com/diegosouzapw/OmniRoute/issues/6134) — thanks @rianonehub)
+
+- **fix(api):** `/v1/messages/count_tokens` now counts `tool_use`, `tool_result` and `thinking` content blocks (and array-form `system` prompts) in the local-estimation path, instead of only `text`. Real agentic conversations keep ~95% of their tokens inside tool results; the previous estimate returned near-zero for them, which silently broke Claude Code's auto-compaction (context grew past the window with no compaction until the upstream API rejected the request). The real provider-side count path is unchanged. Regression guard: `tests/unit/messages-count-tokens-route.test.ts`. ([#6221](https://github.com/diegosouzapw/OmniRoute/pull/6221) — thanks @luweiCN)
+
+- **fix(antigravity):** strip a trailing assistant prefill turn for Vertex Claude models to avoid upstream 400s ([#6114](https://github.com/diegosouzapw/OmniRoute/pull/6114)). Regression guard: `tests/unit/antigravity-claude-prefill-strip.test.ts`. (thanks @anki1kr)
+
+- **fix(security):** the mutable cloud-agent routes (`/api/cloud/credentials/update`, `/api/cloud/models/alias`) now require management auth instead of being treated as public. They were classified as public API routes, so a request without management credentials could update stored cloud-agent credentials and model aliases. They are removed from the public-route set, classified as management routes in the authz pipeline, and gated by `requireManagementAuth`; cloud **read**/auth routes stay public. Regression guards: `tests/unit/cloud-write-auth.test.ts`, `tests/unit/authz/classify.test.ts`, `tests/unit/public-api-routes.test.ts`. ([#6233](https://github.com/diegosouzapw/OmniRoute/pull/6233) — thanks @vittoroliveira-dev)
+
+- **refactor(dashboard):** extract the onboarding-wizard "Open provider details" link target into a pure, unit-tested `buildProviderDetailsHref(connection)` helper. The wizard already routes by `connection.id` (the node UUID) rather than the provider category slug (#6144/#6145); this hardens that behavior behind a tested helper that guards a missing id/connection. Regression guard: `tests/unit/provider-onboarding-href.test.ts`. ([#6166](https://github.com/diegosouzapw/OmniRoute/pull/6166) — thanks @KooshaPari)
+
+- **fix(security):** the doubao synthetic device-id generator now derives its digits via an unbiased crypto-random draw (rejection sampling over `crypto.randomBytes()`) instead of a `% 10` reduction, closing a CodeQL `js/biased-cryptographic-random` finding.
+- **fix(agentSkills):** the GitHub-skills generator now resolves `outputDir` to an absolute path before writing, fixing a regression introduced by #6366 (relative-to-cwd base path) that could write generated skill files to the wrong directory.
+- **fix(security):** `/api/keys/{id}/devices` now checks the HTTP method before auth/validation, returning a `405` for non-GET/DELETE verbs instead of a misleading `401`/`500` (closes a `dast-smoke` QUERY-method finding).
+- **fix(quality):** clear the last 2 heavy quality-gate reds on the release tip (cycle pre-flight).
+- **fix(mitm):** the test suite and CI can never mutate the OS trust store — `OMNIROUTE_SKIP_SYSTEM_TRUST=1` is set globally for tests/CI so `installCert`/`uninstallCert`/`installTproxyCa` skip the privileged OS dispatch ([#6310](https://github.com/diegosouzapw/OmniRoute/pull/6310); full detail is under the [3.8.45] section below — this branch received it via the parallel-cycle sync-back).
+- **fix(api):** `POST /api/github-skills` now Zod-validates its request body; documented the new quality-gate env vars and pinned the merge-integrity GitHub Actions to a commit SHA.
+- **fix(skills):** generate the missing `omni-github-skills` registry entry and align the agent-skills catalog-count tests (follow-up to #6186).
+- **fix(quality):** clear the cycle's 11 net-new ESLint errors and make `validate-release-green` suppressions-aware.
+
+### 📝 Maintenance
+
+- **i18n(it):** add 118 missing Italian (`it`) translations (net-additive — no existing keys dropped, valid JSON), improving Italian UI coverage. ([#6212](https://github.com/diegosouzapw/OmniRoute/pull/6212) — thanks @serverless83)
+- **chore(providers):** remove deprecated **MiMo V2** model entries from the catalogs (xiaomi-mimo, opencode-go, zenmux-free, audio TTS) — the upstream V2 line is superseded by MiMo V2.5; drops `mimo-v2-tts`, `mimo-v2-pro`, `mimo-v2-omni`, `mimo-v2-flash`, `mimo-v2-flash-free` and realigns the provider-catalog tests. ([#6248](https://github.com/diegosouzapw/OmniRoute/pull/6248) — thanks @backryun)
+- **chore(release):** ~50 commits on this branch are v3.8.45 pre-flight/hardening fixes and CI-perf work that landed here via the parallel-cycle sync-back (`sync-next-cycle.mjs`, Hard Rule #21) **after** the `v3.8.45` git tag was cut, and are already fully documented under the **[3.8.45]** section below — listed here only so the per-cycle commit-coverage check (`npm run release:uncovered`) doesn't flag them as gaps. Provider/catalog/UX/backend fixes: #6041, #6078, #6108, #6135, #6148, #6149, #6154, #6158, #6161, #6162, #6163, #6164, #6165, #6170, #6177, #6178, #6181, #6186, #6187, #6191, #6193, #6194, #6195, #6200, #6205, #6208, #6209, #6211, #6213, #6223, #6224, #6225, #6226, #6227, #6228, #6229, #6230, #6235, #6291, #6292. CI/release-pipeline work: #6167, #6203, #6214, #6215, #6218, #6273, #6275, #6283, #6284, #6285, #6300, #6305.
+- **chore(release):** additional zero-ref release-cycle plumbing on this branch, kept out of `release:uncovered` on purpose (no `#N` in the commit subject to cite): opening the v3.8.46 cycle, opening/closing the v3.8.45 cycle, the finalized [3.8.45] CHANGELOG i18n sync-back to 42 mirrors, the v3.8.45 cognitive/cyclomatic and file-size drift rebaselines, ESLint stale-suppression pruning (4,273 → 4,233), and clearing test-masking/docs-all pre-flight reds for v3.8.45.
+
+### ⚡ Performance & Infrastructure
+
+- **perf(release-green):** the pre-flight validator (`scripts/quality/validate-release-green.mjs`) now runs its 4 slow suites (unit / vitest / integration / pack-artifact) **concurrently** via `Promise.all` — pre-flight wall time drops from ~the sum of the suites to ~the slowest one (~30min saved per round; Phase 0 was the nº1 bottleneck of the v3.8.45 release benchmark, 2h54 of 6h34 e2e). Guard: `tests/unit/validate-release-green.test.ts` ("runs the slow suites CONCURRENTLY"). ([#6319](https://github.com/diegosouzapw/OmniRoute/pull/6319))
+- **fix(ci):** `scripts/release/sync-next-cycle.mjs` — two defects found live in its first production run (v3.8.45 Phase 5): (1) the `git()` helper's default 1 MiB `maxBuffer` crashed with `ENOBUFS` on `git show origin/main:CHANGELOG.md` (the CHANGELOG alone is >1 MiB) — widened to 64 MiB; (2) the i18n resync only propagated the `[NEXT]` (TBD) section, leaving the just-shipped finalized section as "— TBD" in all 42 mirrors — it now also syncs `[prevVersion]` bounded by the heading below it (new exported pure helper `versionAfter`). Guards: +5 tests in `tests/unit/sync-next-cycle.test.ts` (8/8). ([#6327](https://github.com/diegosouzapw/OmniRoute/pull/6327))
+- **test(ci):** concurrency-sensitive flaky tests are **quarantined into a serial pass** (`tests/unit/serial/`, `--test-concurrency=1`, appended to every unit runner incl. sharded variants — the serial pass is sharded too so concurrent shard jobs never self-collide). Initial set: `glm-coding-plan-monthly-3580`, `quota-division-blocks`, `provider-health-autopilot`, `combo-health-autopilot` — the class behind the ~28min CI wedges/re-runs (two live 1h+ wedges cancelled during this PR's own validation). Discovery + TIA gates track the new glob; systemic root cause (async logger writing after teardown) tracked in [#6360](https://github.com/diegosouzapw/OmniRoute/issues/6360). Guard: `tests/unit/test-serial-quarantine.test.ts` (4). ([#6347](https://github.com/diegosouzapw/OmniRoute/pull/6347))
+
+### 🙌 Contributors
+
+Thanks to everyone whose work landed in v3.8.46:
+
+| Contributor                                                    | PRs / Issues           |
+| -------------------------------------------------------------- | ---------------------- |
+| [@2220258345](https://github.com/2220258345)                   | direct commit / report |
+| [@abdofallah](https://github.com/abdofallah)                   | direct commit / report |
+| [@adentdk](https://github.com/adentdk)                         | direct commit / report |
+| [@aleksesipenko](https://github.com/aleksesipenko)             | direct commit / report |
+| [@anki1kr](https://github.com/anki1kr)                         | direct commit / report |
+| [@anungma](https://github.com/anungma)                         | direct commit / report |
+| [@arpicato](https://github.com/arpicato)                       | direct commit / report |
+| [@backryun](https://github.com/backryun)                       | #6248                  |
+| [@binsarjr](https://github.com/binsarjr)                       | direct commit / report |
+| [@brightfiscalband](https://github.com/brightfiscalband)       | direct commit / report |
+| [@chirag127](https://github.com/chirag127)                     | #6501, #6506           |
+| [@developerjillur](https://github.com/developerjillur)         | direct commit / report |
+| [@dilneiss](https://github.com/dilneiss)                       | #6499                  |
+| [@dtybnrj](https://github.com/dtybnrj)                         | direct commit / report |
+| [@Forcerecon](https://github.com/Forcerecon)                   | direct commit / report |
+| [@hao3039032](https://github.com/hao3039032)                   | direct commit / report |
+| [@Iammilansoni](https://github.com/Iammilansoni)               | #6150, #6245           |
+| [@jmengit](https://github.com/jmengit)                         | direct commit / report |
+| [@jordansilly77-stack](https://github.com/jordansilly77-stack) | direct commit / report |
+| [@JxnLexn](https://github.com/JxnLexn)                         | direct commit / report |
+| [@KooshaPari](https://github.com/KooshaPari)                   | #6166                  |
+| [@loopyd](https://github.com/loopyd)                           | direct commit / report |
+| [@luweiCN](https://github.com/luweiCN)                         | #6221                  |
+| [@makcimbx](https://github.com/makcimbx)                       | direct commit / report |
+| [@muflifadla38](https://github.com/muflifadla38)               | direct commit / report |
+| [@newnol](https://github.com/newnol)                           | direct commit / report |
+| [@ofekbetzalel](https://github.com/ofekbetzalel)               | direct commit / report |
+| [@ohahe52-dot](https://github.com/ohahe52-dot)                 | direct commit / report |
+| [@phidinhmanh](https://github.com/phidinhmanh)                 | direct commit / report |
+| [@powellnorma](https://github.com/powellnorma)                 | direct commit / report |
+| [@qpeyba](https://github.com/qpeyba)                           | direct commit / report |
+| [@RaviTharuma](https://github.com/RaviTharuma)                 | direct commit / report |
+| [@RCrushMe](https://github.com/RCrushMe)                       | direct commit / report |
+| [@rianonehub](https://github.com/rianonehub)                   | #6134, #6204           |
+| [@serverless83](https://github.com/serverless83)               | #6212                  |
+| [@swingtempo](https://github.com/swingtempo)                   | direct commit / report |
+| [@tenshiak](https://github.com/tenshiak)                       | direct commit / report |
+| [@ThongAccount](https://github.com/ThongAccount)               | direct commit / report |
+| [@UnrealAryan](https://github.com/UnrealAryan)                 | direct commit / report |
+| [@vinayakkulkarni](https://github.com/vinayakkulkarni)         | direct commit / report |
+| [@vittoroliveira-dev](https://github.com/vittoroliveira-dev)   | #6233                  |
+| [@warelik](https://github.com/warelik)                         | direct commit / report |
+| [@xxy9468615](https://github.com/xxy9468615)                   | direct commit / report |
+| [@xz-dev](https://github.com/xz-dev)                           | direct commit / report |
+| [@yanpaing007](https://github.com/yanpaing007)                 | direct commit / report |
+| [@diegosouzapw](https://github.com/diegosouzapw)               | maintainer             |
+
+---
+
 ## [3.8.45] — 2026-07-06
 
 ### ✨ New Features

docs/i18n/ko/CHANGELOG.md

@@ -6,6 +6,213 @@
 
 ## [3.8.31] — 2026-06-20
 
+## [3.8.46] — TBD
+
+### ✨ New Features
+
+- **feat(sse):** **hide paid-only models from `auto/*` routing** when `hidePaidModels` is on ([#6512](https://github.com/diegosouzapw/OmniRoute/issues/6512)) — follow-up to #6328/#6495. PR #6495 hid paid-only models from the `GET /v1/models` listing, but `auto/*` combos (`auto/best-coding`, `auto/glm`, …) could still pick a paid-only backend into their candidate pool → a 402/403 at request time. `createVirtualAutoCombo` now filters the candidate pool through the new pure `open-sse/services/autoCombo/paidModelFilter.ts` (`filterPaidOnlyCandidates`), applying the same free-model predicate #6495 uses in `catalog.ts` (`providerHasFreeModels(provider) && isFreeModel(provider, {id})`) whenever `settings.hidePaidModels === true`. Applied before the category/tier/family narrowing, so it covers every `auto/*` combo; an all-paid pool degrades to the existing graceful empty-pool path. **Opt-in — default OFF leaves the pool unchanged** (identity). Regression guard: `tests/unit/autoCombo/paid-model-filter-6512.test.ts` (4, incl. the default-off identity guard).
+- **feat(sse):** **provider-family auto combos** — `auto/glm`, `auto/minimax`, `auto/mimo`, `auto/zai`, `auto/gemma`, `auto/llama`, `auto/gemini` ([#6453](https://github.com/diegosouzapw/OmniRoute/issues/6453)) — new routable ids that materialize an on-demand virtual combo spanning whatever installed backends currently expose that model family, degrading gracefully as backends rotate. A new pure `open-sse/services/autoCombo/modelFamily.ts` (`detectModelFamily`) classifies by model-id prefix for six families; `zai` is instead resolved by provider id (z.ai's hosted API serves the same `glm-*` model ids as every other GLM backend, so `auto/zai` means "route to my z.ai backend specifically" vs `auto/glm`'s "any connected GLM backend"). Reuses the existing `createVirtualAutoCombo` on-demand materialization path (no DB writes) and the `/v1/models` catalog advertising loop. Regression guard: `tests/unit/autoCombo/provider-family-combos.test.ts` (11).
+- **feat(proxy):** native **proxy-pool round-robin / egress IP rotation** ([#6365](https://github.com/diegosouzapw/OmniRoute/issues/6365)) — a scope (global / provider / account) can now hold **multiple** proxies as a pool with a rotation strategy, so outbound requests cycle their egress IP instead of pinning one proxy per scope. Migration `117_proxy_pool_rotation.sql` lifts the `UNIQUE(scope, scope_id)` constraint (rebuild via the canonical rename/copy/drop; existing single assignments become 1-element pools) and adds a `proxy_scope_rotation` companion table holding the per-scope strategy + a persisted monotonic round-robin cursor. Strategies: `round-robin` (default, monotonic cursor — never `Math.random`), `random`, and `sticky-per-N-min`. Resolution (`resolveProxyForScopeFromRegistry` / `resolveProxyForConnectionFromRegistry`) now fetches the alive, position-ordered candidate set (unchanged `PROXY_ALIVE_PREDICATE`) and applies the strategy; an empty / all-dead pool still returns `null` — the #6246 fail-closed guard is **untouched** (never falls through to direct egress). Backend + DB only; dashboard pool-builder UI is a follow-up. Regression guard: `tests/unit/proxy-pool-rotation-6365.test.ts` (8, incl. fail-closed + backward-compat).
+- **feat(providers):** end-to-end **tool/function calling on the native Gemini `/v1beta` endpoint** ([#6222](https://github.com/diegosouzapw/OmniRoute/issues/6222)) — both directions of the Gemini↔OpenAI conversion now preserve tool data (previously silently dropped). Request side: `convertGeminiToInternal` (extracted to its own testable module) maps `tools[].functionDeclarations` → OpenAI `tools`, prior `functionCall` parts → assistant `tool_calls`, and `functionResponse` parts → `tool`-role messages. Response side: `convertOpenAIResponseToGemini` emits `parts[].functionCall {name,args}` from `message.tool_calls`, and the streaming `openAIChunkToGeminiChunk` accumulates fragmented `tool_calls` deltas by index into complete `functionCall` parts. The non-Gemini client paths (Claude, OpenAI-Responses) already preserved tool calls — this closes the gap specific to the native Gemini surface. Regression guard: `tests/unit/v1beta-gemini-tool-calling-6222.test.ts` (6, incl. a streaming SSE round-trip).
+- **feat(providers):** copilot-m365-web **enterprise / work tier** support ([#6334](https://github.com/diegosouzapw/OmniRoute/issues/6334)) — mirrors the EDU-tier pattern (#6210): `M365ConnectionParams` gains an `agent` field, a new opt-in `M365_ENTERPRISE_OVERRIDES` preset (`agent=work`, `scenario=officeweb`, `licenseType=Premium`) applies via `providerSpecificData.tier="enterprise"` (alias `"work"`), and `agent` is also overridable directly via `providerSpecificData.agent`. `buildWsUrl` was hardcoding `agent="web"` (the one enterprise-distinguishing param with no override path), so a Premium work account handshook then returned an empty stream. The individual and EDU paths are untouched. Kilo's dup flag vs #6210 (EDU tier) was a false positive — different tier. Regression guard: `tests/unit/copilot-m365-enterprise-6334.test.ts` (7). End-to-end confirmation on a real Premium work account is a live-VPS validation follow-up (Hard Rule #18). (thanks @Forcerecon)
+- **feat(api):** standardized, provider-agnostic **`effort` + `thinking` request params** ([#6241](https://github.com/diegosouzapw/OmniRoute/issues/6241)) — a thin standardization layer over the existing mature per-provider reasoning plumbing (no provider mapper touched). `providerChatCompletionSchema` gains a canonical `effort` (reusing the shared `none/low/medium/high/xhigh` vocabulary — the UI tiers `extra`/`max` collapse onto `xhigh`) and a boolean `thinking`. A pure `normalizeReasoningRequest` (wired once in `src/sse/handlers/chat.ts`, before any reasoning field is read) folds them onto the fields the translators already consume (`reasoning_effort` / `reasoning.effort` / `thinking`), so they fan out to Anthropic / Gemini / xAI / Responses — an explicit client `reasoning_effort` / object-shaped `thinking` always wins (backward-compatible). `/models` additively exposes `supportsThinking` + `effort_tiers` so the frontend can render the toggles (UI component is a follow-up). Regression guard: `tests/unit/effort-thinking-standardization-6241.test.ts` (12). (thanks @Iammilansoni, @shabeer)
+- **feat(combo):** new **`pipeline` (sequential) combo strategy** ([#6297](https://github.com/diegosouzapw/OmniRoute/issues/6297)) — the 18th routing strategy runs targets **in order**, threading each step's output into the next step's input, with an optional per-step `prompt` (system instruction); only the final step's response is returned. Distinct from `fusion` (parallel fan-out + judge). Implemented as a self-contained `open-sse/services/pipeline.ts` (sibling to `fusion.ts`), dispatched from `combo.ts`; the step list reuses `combo.models` order and reads an optional `prompt` off each target (backward-compatible — ignored by every other strategy). Intermediate steps run non-streaming with tools stripped (complete prose to thread forward); the final step keeps the client's `stream` flag + tools. A failing/empty/unparseable intermediate step fails the whole pipeline explicitly via a sanitized error (never silently swallowed). Kilo's dup flag vs #563 was a false positive (that's model→chain selection; this is a sequential chain). Regression guard: `tests/unit/combo-pipeline-strategy.test.ts` (5). (thanks @ofekbetzalel)
+- **feat(ci):** `check:test-masking` now flags **inline-reimplemented prod conditions** ([#6348](https://github.com/diegosouzapw/OmniRoute/issues/6348)) — a new **report-only** subcheck (v2, 6A.10 family) catches the wrong-shape contract test: a test that recomputes the condition under test inline instead of importing/exercising the real function (the #6216 class, where `=== 500` → `>= 500` stayed green because the test re-implemented the branch). For each added/modified test file it warns when the file textually duplicates a ≥3-token conditional from a production file touched in the same PR **and** does not import the symbol/module owning it, via a pure, fixture-tested `findReimplementedConditions()` with an allowlist mirroring `assertReductionAllowlist`. Report-only for now (does not fail the gate) — to be promoted to blocking after a triage cycle. Regression guard: `tests/unit/check-test-masking.test.ts` (45).
+- **feat(sse):** per-connection routing override (native vs CLIProxyAPI) ([#6339](https://github.com/diegosouzapw/OmniRoute/issues/6339)) — the previously-dead `isCliproxyapiDeepModeEnabled` helper is now wired into `resolveExecutorWithProxy`: a single connection can opt itself into the CLIProxyAPI passthrough executor via `providerSpecificData.cliproxyapiMode="claude-native"`, with precedence **connection override > provider `upstream_proxy_config` mode > default**. `resolveExecutorWithProxy` now receives the resolved connection's `providerSpecificData` (threaded from `chatCore.ts`), so one connection can deep-route while the provider's other connections stay native — no DB schema change (the toggle rides in `providerSpecificData`). Also resolves the same-provider mixing ask in #6340. Regression guard: `tests/unit/chatcore-executor-proxy.test.ts` (9). (thanks @RaviTharuma)
+- **feat(dashboard):** "Add session cookie" modal now shows a prominent **"Open ‹host› →"** link to the provider's own site ([#6268](https://github.com/diegosouzapw/OmniRoute/issues/6268)) — every `-web` cookie-session provider (chatgpt-web, claude-web, gemini-web, kimi-web, lmarena, qwen-web, m365-copilot-web, …) renders a one-click external link (opening the provider's login/home page in a new tab) so operators no longer tab away to retype the URL mid-setup. The host resolves from a pure, unit-tested `resolveWebProviderHost()` (prefers `WEB_COOKIE_PROVIDERS[id].website`, falls back to the registry `baseUrl` origin); non-web providers render exactly as before. Kilo's dup flag vs #6265 (modal-too-small-on-1080p) was a false positive — distinct concern. Regression guard: `tests/unit/resolve-web-provider-host.test.ts` (5). (thanks @chirag127)
+- **feat(providers):** add **DigitalOcean** AI (serverless inference) as an OpenAI-compatible API-key provider ([#6373](https://github.com/diegosouzapw/OmniRoute/pull/6373)) — base `https://inference.do-ai.run/v1`, wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/digitalocean/`, `src/shared/constants/providers/apikey/inference-hosts.ts`). Regression guard: `tests/unit/digitalocean-provider.test.ts`. (thanks @newnol)
+- **feat(providers):** add **Huancheng Public API** (`hcnsec`) as an OpenAI-compatible regional provider ([#6410](https://github.com/diegosouzapw/OmniRoute/pull/6410)) — Xinjiang Huancheng Cybersecurity's public LLM platform (base `https://api.hcnsec.cn/v1`, free credits via daily check-ins), wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/hcnsec/`, `src/shared/constants/providers/apikey/regional.ts`). Regression guard: `tests/unit/hcnsec-provider.test.ts`. (thanks @UnrealAryan)
+- **feat(dashboard):** the web-session credential guide now shows an **"Open {host}" link** ([#6316](https://github.com/diegosouzapw/OmniRoute/pull/6316)) to the provider's sign-in site (derived from the provider `website` via `getProviderWebsiteHost`), so you can jump straight to the page where the cookie/session must be captured. Regression guard: `tests/unit/web-session-provider-link-6316.test.ts`. (thanks @jordansilly77-stack)
+- **feat(cerebras):** add the **Gemma 4 31B** model (`gemma-4-31b`) to the Cerebras registry + pricing table ([#6331](https://github.com/diegosouzapw/OmniRoute/pull/6331)). Regression guard extends `tests/unit/t28-model-catalog-updates.test.ts`. (thanks @backryun)
+- **feat(providers):** add **Yuanbao (web)** as a cookie-session provider ([#6196](https://github.com/diegosouzapw/OmniRoute/issues/6196)) — `yuanbao-web` (Tencent Yuanbao, `yuanbao.tencent.com`) with cookie-only auth (`hy_user`/`hy_token` + public agent id), SSE→OpenAI translation incl. `reasoning_content`, exposing DeepSeek V3/R1 + Hunyuan / Hunyuan-T1. Regression guard: `tests/unit/providers-yuanbao-web.test.ts`. `together-web` was **deferred** (no verifiable web-session endpoint — needs a captured request) and `huggingchat-web` **dropped** (the existing `huggingchat` already is a web-cookie provider). (thanks @chirag127)
+- **feat(providers):** route the built-in **agentrouter** through the dynamic Claude-Code wire image ([#6056](https://github.com/diegosouzapw/OmniRoute/issues/6056)) — a small static allow-set (`CC_WIRE_IMAGE_BUILTINS` in `open-sse/services/ccWireImageBuiltins.ts`), consulted by `isClaudeCodeCompatible` / `isClaudeCodeCompatibleProvider` / `applyFingerprint`, makes agentrouter adopt the CC wire-image headers + fingerprint **while guarding the CC baseUrl/auth branches** so it keeps its own registry `baseUrl` and `x-api-key` auth. Regression guard: `tests/unit/agentrouter-cc-wire-image.test.ts` (asserts the wire image is applied AND agentrouter's baseUrl/auth are preserved). Live WAF-acceptance against agentrouter.org is a VPS validation follow-up (Hard Rule #18).
+- **feat(providers):** **bulk-add API keys for Cloudflare Workers AI** ([#6174](https://github.com/diegosouzapw/OmniRoute/issues/6174)) — `cloudflare-ai` is removed from the bulk-add exclusion list and the bulk parser gains a 3-field `name|accountId|apiKey` mode; the bulk route now builds a **per-entry** `providerSpecificData` so each key carries its own `accountId` (fixing the previous shared-object reuse), and both the create + key-validation paths receive it. Regression guard: `tests/unit/bulk-api-key-parser-cloudflare.test.ts`. (thanks @muflifadla38)
+- **feat(dashboard):** routing/settings UX clarity ([#6147](https://github.com/diegosouzapw/OmniRoute/issues/6147)) — (1) weighted combos show the **effective routing share %** next to each weight when weights don't sum to 100 (`WeightTotalBar.tsx`); (2) the status widget's user-facing **"Cloud Sync" label is renamed** to "Remote Settings Sync" (`CloudSyncStatus.tsx`; internal ids/state untouched); (3) built-in providers gain an **opt-in advanced base-URL override** (`isBaseUrlOverrideEligibleProvider`, hidden behind an "Advanced" toggle, reusing the existing `providerSpecificData.baseUrl` persistence — not globally widened). Regression guard: `tests/unit/routing-settings-ux-6147.test.ts`.
+- **feat(combo):** add an option to **disable session stickiness**, per-combo or globally — round-robin / random combos can rotate to a different connection on every request instead of pinning a whole conversation to one connection by its first-message hash. Resolution precedence per-combo `config.disableSessionStickiness` → global `settings.disableSessionStickiness` → default `false` (preserves the #3825 prompt-cache/504 fix); gates **both** stickiness call sites in `open-sse/services/combo.ts`. Exposed as a global toggle (Combo Defaults) and a per-combo Inherit/on/off control. ([#6168](https://github.com/diegosouzapw/OmniRoute/issues/6168)) Regression guard: `tests/unit/combo-disable-session-stickiness.test.ts`. (thanks @RCrushMe)
+- **feat(docker):** add the `OMNIROUTE_NO_SUDO` env flag for root-less / user-namespaced deployments — the MITM cert-trust command path (`resolveSudoSpawn` in `src/mitm/systemCommands.ts`) now strips the leading `sudo` when the flag is truthy, in addition to the existing root / sudo-missing cases, so the Proxy Agent runs without `sudo` (the operator trusts the CA manually, e.g. via `NODE_EXTRA_CA_CERTS`). Argv-array `spawn` preserved — no shell interpolation (Hard Rule #13). ([#6122](https://github.com/diegosouzapw/OmniRoute/issues/6122)) Regression guard: `tests/unit/mitm-systemCommands-no-sudo.test.ts`. (thanks @powellnorma)
+- **feat(providers):** add **Requesty** as an OpenAI-compatible gateway provider (BYOK, base `https://router.requesty.ai/v1`, ~200 free requests/day) — wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/requesty/`, `src/shared/constants/providers/apikey/gateways.ts`). ([#6120](https://github.com/diegosouzapw/OmniRoute/issues/6120)) Regression guard: `tests/unit/requesty-provider.test.ts`. (thanks @chirag127)
+- **feat(dashboard):** add **configured-only / available-only filters** to the Free Provider Rankings page ([#6150](https://github.com/diegosouzapw/OmniRoute/issues/6150)) — hide providers you haven't configured, or whose connections are all rate-limited / out of quota, via server-side query params (`?configuredOnly` / `?availableOnly` on `GET /api/free-provider-rankings`) backed by a testable lib helper reusing the in-process connection state (no Redis). Both filters default off, so the default view is unchanged; this supersedes the earlier client-side "Configured Only" toggle (#6245) with an available-only dimension and unit-tested logic. Regression guard: `tests/unit/freeProviderRankings-filters.test.ts`.
+
+### 🔧 Bug Fixes
+
+- **fix(dashboard):** adding a second API key connection for the same provider no longer silently overwrites the first — the Add-API-key modal now derives a unique default connection name (`main`, then `main-2`, `main-3`, …) so the backend name-based upsert can't collide (#6499 — thanks @dilneiss).
+- **fix(compression):** the session-dedup engine now also deduplicates a large multi-line block repeated **within a single message** (intra-message dedup), not just across turns; the compression-preview API surfaces a `fallbackReason`, and the fusion panel reports how many models were rate-limited vs failed on total-panel failure (#6501 — thanks @chirag127).
+- **fix(compression):** a stacked-pipeline step naming an unregistered engine now surfaces a `validationErrors` entry instead of silently no-op'ing, so misconfigured pipelines are visible in the preview API (#6506 — thanks @chirag127).
+- **feat(usage):** add a **Codex reset-credit redemption** flow to the Provider Limits UI ([#6361](https://github.com/diegosouzapw/OmniRoute/pull/6361)) — a `useCodexResetCreditRedemption` hook + `/api/usage/codex-reset-credit` route + `codexResetCredits` lib let you redeem banked Codex reset credits from the quota card. Regression guard: `tests/unit/codex-reset-credits.test.ts`. (thanks @JxnLexn)
+- **feat(glm):** add **team-plan quota settings** for `glm-cn` connections ([#6351](https://github.com/diegosouzapw/OmniRoute/pull/6351)) — a dedicated `GlmTeamQuotaFields` form section (team quota id / limits) threaded through the Add/Edit connection modals, persisted via `providerSpecificData`, with the GLM usage service reading the team quota. Regression guards: `tests/unit/glm-team-quota.test.ts`, `provider-specific-data-schema.test.ts`. (thanks @hao3039032)
+- **feat(providers):** add **TinyFish** web-fetch/search support ([#6349](https://github.com/diegosouzapw/OmniRoute/pull/6349)) — a `tinyfish-fetch` executor + `/v1/web/fetch` route + MCP web-fetch tool, registered as a specialty-media provider with request-validation and a search-provider catalog entry. Regression guards: `tests/unit/executor-tinyfish-fetch.test.ts`, `web-fetch-handler.test.ts`, `mcp-web-fetch-tool.test.ts`, `provider-validation-tinyfish.test.ts`. (thanks @dtybnrj)
+- **fix(cli):** `omniroute launch-codex` now spawns `codex.cmd` through a shell on Windows (the npm `.cmd` shim is unresolvable by bare `spawn` → ENOENT), mirroring the qodercli Windows fix (#6263) ([#6312](https://github.com/diegosouzapw/OmniRoute/pull/6312)). Regression guard: `tests/unit/launch-codex-windows-spawn-6312.test.ts`. (thanks @swingtempo)
+- **fix(codex):** isolate the **Spark** quota from the shared Codex quota and stabilize the quota UI ordering / hydration so per-scope limits render consistently ([#6336](https://github.com/diegosouzapw/OmniRoute/pull/6336)). Regression guards: `tests/unit/codex-quota-selection-hydration.test.ts`, `provider-limits-ui.test.ts` + 3 more. (thanks @xz-dev)
+- **feat(api):** add a `hidePaidModels` setting that filters paid-only models out of the `/v1/models` catalog. Regression guard: `tests/unit/models-catalog-hide-paid.test.ts`. (thanks @chirag127)
+- **fix(api-manager):** the fallback model picker now preserves combos instead of dropping them when a primary model is unavailable ([#6443](https://github.com/diegosouzapw/OmniRoute/pull/6443)). Regression guard: `tests/unit/api-manager-page-static.test.ts`. (thanks @jmengit)
+- **fix(providers):** recoverable Antigravity / Cloud-Code (Gemini Code Assist) `403` responses ([#6452](https://github.com/diegosouzapw/OmniRoute/pull/6452)) are now classified as a retryable project-config error instead of a terminal account ban, so a fixable project/API-disabled 403 no longer forces a ~1-year cooldown / full OAuth reconnect. Regression guard: `tests/unit/errorclassifier-antigravity-403.test.ts`. (thanks @developerjillur)
+- **fix(mitm):** `sanitizeHeaders` now redacts `Set-Cookie` response headers so upstream session cookies never leak into logs / diagnostics ([#6451](https://github.com/diegosouzapw/OmniRoute/pull/6451)). Regression guard: `tests/unit/mitm-sanitize-headers.test.ts`. (thanks @developerjillur)
+- **fix(api):** `/api/compression/preview` now accepts `mode: "caveman"` and correctly handles stacked / zero-compression previews ([#6425](https://github.com/diegosouzapw/OmniRoute/issues/6425)). Regression guard: `tests/unit/api/compression-preview-caveman-and-stacked-6425.test.ts`. (thanks @chirag127)
+- **feat(providers):** add **Zed** hosted LLM aggregator as a native-app provider ([#6118](https://github.com/diegosouzapw/OmniRoute/pull/6118)) — OAuth sign-in via the Zed hosted flow, registered through the shared provider registry + executor. Regression guards: `tests/unit/zed-oauth-provider.test.ts`, `zed-import-utils.test.ts`, `zed-docker-detect.test.ts`, `mitm-handler-zed.test.ts`. VPS-validated via live operator login (Hard Rule #18).
+- **fix(oauth):** the Kiro SSO-cache auto-import now **preserves the IDC region** — cross-region Amazon Q / Kiro profiles imported from the SSO cache are no longer collapsed to the default region ([#6113](https://github.com/diegosouzapw/OmniRoute/pull/6113)). Regression guard: `tests/unit/kiro-auto-import-idc-2059.test.ts`. VPS-validated via live operator login (Hard Rule #18).
+- **fix(dashboard):** passthrough model aliases no longer collide when two namespaced model ids share a last segment (port from 9router#1850, [#6431](https://github.com/diegosouzapw/OmniRoute/pull/6431)). `enx/gpt-5.5` and `enx/codebuddy/gpt-5.5` both auto-generated the alias `gpt-5.5`, so the second model could never be added (the UI just alerted "alias already exists"). Aliases are now disambiguated deterministically — bare last segment when free, then parent-qualified (`codebuddy-gpt-5.5`), then a numeric suffix — while re-adding the exact same model id is still blocked. Regression guard: `tests/unit/passthrough-alias-1850.test.ts`. (thanks @arpicato)
+- **fix(translator):** preserve a Gemini `functionResponse` co-located with other parts (another `functionCall`, or trailing `text`) in the same content when translating **Gemini → OpenAI** ([#6376](https://github.com/diegosouzapw/OmniRoute/pull/6376)). `convertGeminiContent()` early-returned the tool message on the first `functionResponse` part, dropping any co-located parts; such contents are now pre-split (one tool message per `functionResponse`, emitted first, plus one message for the remaining parts). Regression guard: `tests/unit/gemini-to-openai-function-response.test.ts`. (thanks @warelik)
+- **fix(headroom):** detect a python interpreter managed by **mise / pyenv / asdf / conda** (port from 9router#2353, [#6382](https://github.com/diegosouzapw/OmniRoute/pull/6382)). Headroom's python probe (`src/lib/headroom/detect.ts`) searched a hardcoded `PATH`, but version managers expose their interpreters via shim dirs that only join `PATH` through interactive-shell activation — which the non-interactive server never runs, so a managed python (≥3.10) was invisible and Headroom reported it missing. The search path now prepends the well-known shim/bin dirs (`~/.local/share/mise/shims`, `~/.pyenv/shims`, `~/.asdf/shims`, `$CONDA_PREFIX/bin`, `~/.local/bin`, respecting `MISE_DATA_DIR`/`PYENV_ROOT`/`ASDF_DATA_DIR` when set), and a new `HEADROOM_PYTHON` env override lets operators point straight at their interpreter (mirroring `HEADROOM_URL`). Still shell-free (`execFileSync`). Regression guard: `tests/unit/headroom-detect.test.ts` (5). (thanks @loopyd)
+- **fix(executors):** strip the OpenAI-Codex/Claude-CLI `client_metadata` passthrough field for **NVIDIA** requests (port from 9router#1887, [#6411](https://github.com/diegosouzapw/OmniRoute/pull/6411)). NVIDIA's OpenAI-compatible wrapper rejects it with `400 Unsupported parameter`, the same class already handled for `cerebras`/`mistral`; `nvidia` (executor `default`) was missing from the strip allowlist so Codex/Claude-Code passthrough requests 400'd. Regression guard: `tests/unit/executor-default-strip-client-metadata.test.ts` (+nvidia case). (thanks @phidinhmanh)
+- **fix(translator):** strip the Claude-style `thinking` field for **NVIDIA `z-ai/glm-5.2`** (port from 9router#2023, [#6413](https://github.com/diegosouzapw/OmniRoute/pull/6413)). NVIDIA's OpenAI-compatible wrapper 400s on `thinking` (a Claude-format client routed here leaves a `thinking:{type:"adaptive"}`); the existing strip rule only dropped `reasoning`. Same class already handled for `minimax-m2.7`. Regression guard: `tests/unit/nvidia-minimax-thinking-strip.test.ts` (+glm-5.2 case). (thanks @phidinhmanh)
+- **fix(translator):** suppress the streamed `</think>` close marker for the **Antigravity IDE** client (port from 9router#1061, [#6415](https://github.com/diegosouzapw/OmniRoute/pull/6415)). On thinking-only turns Antigravity rendered a bare `</think>` as the sole visible content, tripping its loop-detection and wasting requests. Antigravity's UA (`vscode/<v> (Antigravity/<v>)`) is added to the marker-suppress allowlist (alongside OpenCode); Claude Code / Cursor still get the marker, and `x-omniroute-thinking-marker: on` force-restores it. Regression guard: `tests/unit/think-close-marker-suppress-5245.test.ts`. (thanks @abdofallah)
+- **fix(executors):** strip nested `reasoning_content` from messages for **Mistral** (port from 9router#1649, [#6417](https://github.com/diegosouzapw/OmniRoute/pull/6417)). Mistral's API returns `422 extra_forbidden` when an assistant message carries `reasoning_content` (replayed thinking from a prior turn, e.g. via the Codex `/responses` path); the generic top-level 400 field-downgrade retry never covered the nested per-message field. `DefaultExecutor` now strips it for provider `mistral` only, so DeepSeek (which requires replayed `reasoning_content`) is unaffected. Regression guard: `tests/unit/mistral-strip-reasoning-content-1649.test.ts`. (thanks @xxy9468615)
+- **fix(executors):** strip the `client_metadata` passthrough field on the **OpenCode** path (port from 9router#1442, [#6418](https://github.com/diegosouzapw/OmniRoute/pull/6418)). OpenCode upstreams (e.g. `kimi-k2.6` via opencode-go) reject it with `400 "Extra inputs are not permitted, field: 'client_metadata'"`; the DefaultExecutor strip only covered cerebras/mistral and `OpencodeExecutor` extends `BaseExecutor` directly, so nothing removed it there. Regression guard: `tests/unit/opencode-strip-client-metadata-1442.test.ts`. (thanks @yanpaing007)
+- **fix(executors):** inject the `reasoning_content` echo for the native **Moonshot Kimi** provider (port from 9router#1480, [#6419](https://github.com/diegosouzapw/OmniRoute/pull/6419)). Kimi (executor `default`) is a thinking-mode upstream that 400s with "reasoning_content must be passed back" when a prior assistant turn lacks it; the placeholder injection was only wired into the OpenCode meta-provider, so direct multi-turn Kimi conversations failed. Scoped to `kimi` (gateway-served models matching the thinking-model name pattern are unaffected). Regression guard: `tests/unit/kimi-native-reasoning-injected-1480.test.ts`. (thanks @2220258345)
+- **fix(executors):** recover from a strict gateway's `context_management: Extra inputs are not permitted` 400 (port from 9router#1468, [#6420](https://github.com/diegosouzapw/OmniRoute/pull/6420)). **Claude Code** always sends a top-level `context_management` field; strict anthropic-compatible gateways reject it. The dedicated context-editing 400-fallback only fired when OmniRoute's own `contextEditing` feature was enabled (default off), so a client-sent field passed through untouched and 400'd. `context_management` is now in the generic reactive field-strip list, so it's stripped-and-retried once regardless of the feature flag (with correct request re-signing for claude-compatible relays). Regression guard: `tests/unit/provider-field-strips.test.ts`. (thanks @ohahe52-dot)
+- **fix(network):** enable **RFC 8305 Happy Eyeballs** (`autoSelectFamily`) on the direct-egress undici dispatcher (port from 9router#1237, [#6423](https://github.com/diegosouzapw/OmniRoute/pull/6423)). When DNS returns both IPv6 (AAAA) and IPv4 (A) and the IPv6 route is broken (e.g. a NAT64 `64:ff9b::` prefix without routing), undici tried IPv6 first and hung until `ETIMEDOUT` (then a 502 + account lockout), even though `curl` reached the same host. The direct dispatcher now races both families and uses whichever connects first. Proxy paths pin family via `proxyTls` and are unaffected. Regression guard: `tests/unit/direct-dispatcher-pipelining-4580.test.ts`. (thanks @adentdk)
+- **fix(combo):** round-robin now advances the rotation pointer past the model that **actually served**, not the eagerly-scheduled one (port from 9router#948, [#6428](https://github.com/diegosouzapw/OmniRoute/pull/6428)). With `stickyLimit: 1` (true round-robin), when the scheduled model failed and a _different_ model served via fallback, the counter had already advanced +1 from the scheduled index — so the next request reused the fallback-served model, degrading round-robin into hot-spotting on whichever model was healthy. The pointer now advances to the served index + 1 (mirroring the sticky-limit>1 path). Session-stickiness (#3825) and distribution are preserved. Regression guard: `tests/unit/combo-rr-fallback-advance-948.test.ts`. (thanks @binsarjr)
+- **fix(sse):** a non-string `model` field is now rejected with a `400` before the resolver, instead of crashing downstream `.toLowerCase()`/`.split()` calls into an empty-body `500` that escapes the error sanitizer ([#6407](https://github.com/diegosouzapw/OmniRoute/issues/6407)). Regression guard: `tests/unit/chat-non-string-model-6407.test.ts`. (thanks @chirag127)
+- **fix(api):** unknown `/api/*` routes now return a JSON `404` (instead of the dashboard HTML shell) and scalar chat params (`model`/`temperature`/etc.) are validated **before** the provider lookup so malformed requests fail fast with a clear `400` ([#6424](https://github.com/diegosouzapw/OmniRoute/issues/6424), [#6412](https://github.com/diegosouzapw/OmniRoute/issues/6412)). Regression guards: `tests/unit/api/api-catchall-json-404.test.ts`, `tests/unit/chat-early-schema-validation-6412.test.ts`. (thanks @chirag127)
+- **fix(api):** `/v1/chat/completions` now rejects a non-JSON `Content-Type` with a `400` before parsing the body ([#6414](https://github.com/diegosouzapw/OmniRoute/issues/6414)). Regression guard: `tests/unit/v1-chat-completions-content-type-6414.test.ts`. (thanks @chirag127)
+- **fix(api):** the `X-OmniRoute-Compression` response header is now echoed on `/v1/chat/completions` and `/v1/completions` ([#6422](https://github.com/diegosouzapw/OmniRoute/issues/6422)). Regression guard: `tests/unit/compression-header-echo-6422.test.ts`. (thanks @chirag127)
+- **fix(api):** concurrent `GET /v1/models` requests are coalesced into a single catalog build ([#6408](https://github.com/diegosouzapw/OmniRoute/issues/6408)). Regression guard: `tests/unit/v1-models-concurrent-6408.test.ts`. (thanks @chirag127)
+- **fix(api):** `/v1/completions` now echoes the requested `body.model` in its JSON + streamed responses ([#6429](https://github.com/diegosouzapw/OmniRoute/pull/6429)). Regression guard: `tests/unit/completions-body-model-echo.test.ts`. (thanks @chirag127)
+- **fix(api):** env-var master keys now see the full `/v1/models` catalog ([#6406](https://github.com/diegosouzapw/OmniRoute/issues/6406)). Regression guard: `tests/unit/models-catalog-envkey-6406.test.ts`. (thanks @chirag127)
+- **fix(api):** non-streaming `/v1/completions` responses now echo `body.model` aligned with the `X-OmniRoute-Model` header ([#6426](https://github.com/diegosouzapw/OmniRoute/issues/6426)). Regression guard: `tests/unit/v1-completions-model-header-match-6426.test.ts`. (thanks @chirag127)
+- **fix(api):** unknown `/v1/*` routes now return a JSON `404 not_found` instead of the Next.js dashboard HTML shell ([#6405](https://github.com/diegosouzapw/OmniRoute/issues/6405)). Regression guard: `tests/unit/api/v1-catchall-json-404.test.ts`. (thanks @chirag127)
+- **fix(api):** the per-connection provider models route now degrades to the shipped catalog when a provider's `/models` endpoint answers with a redirect ([#6267](https://github.com/diegosouzapw/OmniRoute/issues/6267)) — a `qwen-web` import failed with a raw `Redirect blocked … (307)` 503. `safeOutboundFetch` throws `REDIRECT_BLOCKED` on the 307, `getSafeOutboundFetchErrorStatus` maps it to 503, and `buildDiscoveryErrorFallbackResponse` treated every 503 as a hard error — so the non-empty `getModelsByProviderId("qwen-web")` catalog was never surfaced. A models-endpoint redirect is not a fixable-config error (unlike `URL_GUARD_BLOCKED`/`INVALID_URL`, which stay hard errors), so it now falls back to the local/cached catalog before the 503 short-circuit. General fix — covers any config-driven provider that 307s. Regression guard: `tests/unit/provider-models-qwen-web-redirect-6267.test.ts`. (thanks @chirag127)
+- **fix(api):** the per-connection provider models route (MCP `list_models_catalog` + the dashboard import view) now merges USER-ADDED custom models into its response ([#6247](https://github.com/diegosouzapw/OmniRoute/issues/6247)) — custom models live in the `key_value` namespace `customModels`, which the live REST `/api/v1/models` already merges, but `src/app/api/providers/[id]/models/route.ts` never read `getCustomModels`, so custom models were dropped on both the discovery-success and local_catalog paths. They are now folded into the returned model list (deduped by id, stamped `owned_by: provider`), fixing MCP + the dashboard import view in one place. Regression guard: `tests/unit/provider-models-custom-merge-6247.test.ts`. (thanks @RCrushMe)
+
+- **fix(providers):** GitLab Duo tool-calling follow-up turns no longer fail upstream with `422 {"detail":"Validation error"}` (tokens 0/0, rejected pre-inference). The #6234 tool-result-feedback fix serialized the **entire** multi-turn conversation into GitLab's single-file `code_suggestions` (`small_file`) generation endpoint — folded history that turn-N sent as an oversized `current_file.content_above_cursor` **and** duplicated verbatim into `user_instruction`, tripping the AI-Gateway's `small_file` validation guard. The executor now **bounds** that prompt: it keeps system + latest user message + the most-recent tool round (dropping older turns), caps oversized tool results, and stops duplicating the full prompt into `user_instruction` (which now carries only the short latest user message) — while still feeding the most-recent tool result back so the agent continues (`open-sse/executors/gitlab.ts`, [#6220](https://github.com/diegosouzapw/OmniRoute/issues/6220)). The unit test covers the bounding logic; the upstream 422→200 clearing is VPS-only (Hard Rule #18). Regression guard: `tests/unit/gitlab-tool-exchange-bounded-6220.test.ts`.
+
+- **fix(i18n):** the provider-detail (`/dashboard/providers/[id]`) connection-status filter labels no longer render as `__MISSING__:All` / `__MISSING__:Active` / `__MISSING__:Error` / `__MISSING__:Banned` / `__MISSING__:CreditsExhausted` in non-English locales (notably pt-BR) ([#6290](https://github.com/diegosouzapw/OmniRoute/issues/6290)). Root cause was **not** the namespace mismatch the issue guessed — the `providers.filter*` keys resolve correctly in `en.json`; the debt lived in the locale mirrors (`src/i18n/messages/*.json`), where these five keys carried the `__MISSING__:` sync sentinel in ~15 locales and were absent entirely in ~26 others, so next-intl found the key and echoed the sentinel verbatim. All 40 non-English/-Chinese mirrors now ship real translations for the five `providers.filter*` labels. Regression guard: `tests/unit/i18n-provider-filter-keys-6290.test.ts`. (thanks @diegosouzapw)
+
+- **fix(providers):** the `copilot-m365-web` streaming executor now emits `debug`-level WebSocket diagnostics ([#6210](https://github.com/diegosouzapw/OmniRoute/issues/6210)) — the outbound WS URL (with the `access_token` **redacted** via `redactWsUrl()`), handshake success/failure, and each received SignalR frame's `type`/`target`. Previously the streaming path logged nothing, so an empty `content:null` response (the M365 Education / Starter tier symptom fixed in #6234) was undiagnosable even at `APP_LOG_LEVEL=debug`. The change is debug-level and side-effect-free — it does not alter streaming behavior or the frame parser, and the token never reaches the logs. Regression guard: `tests/unit/copilot-m365-web-logging-6210.test.ts` (thanks @qpeyba)
+
+- **fix(resilience):** a round-robin combo no longer returns `503 all upstream accounts are unavailable` when a compatibility-rejected target is actually healthy ([#6238](https://github.com/diegosouzapw/OmniRoute/issues/6238)). `filterTargetsByRequestCompatibility` drops request-incompatible targets (tool/vision/structured-output unsupported, or below the required context window) **before** any availability check runs, and its `compatible.length === 0` safety net only fired when _all_ targets were filtered — not when the kept targets later all turned out runtime-unavailable (circuit-open / cooldown / no credentials). So a combo could 503 while a compat-rejected-but-healthy provider sat unused. `handleRoundRobinCombo` now keeps the compat-rejected set and, when every compat-kept target was skipped without a single real attempt, probes those rejected targets as a **last-resort fallback tier** (via the new pure `open-sse/services/combo/comboCompatFallback.ts`) before crystallizing the 503. Regression guard: `tests/unit/combo-roundrobin-compat-fallback-6238.test.ts`. (thanks @ThongAccount)
+
+- **fix(startup):** best-effort self-heal for a corrupted Turbopack dev cache on Windows ([#6289](https://github.com/diegosouzapw/OmniRoute/issues/6289)). On Windows, `pnpm dev` can fail at startup when Turbopack `mmap`s a persistent-cache SST file and the OS refuses the mapping (`os error 1455` — "paging file too small"), which Turbopack surfaces as a misleading `Module not found: Can't resolve '@/shared/utils/machine'`. This is a **known upstream Turbopack cache-corruption bug — not our code**. The dev launcher (`scripts/dev/run-next.mjs`) now wraps `nextApp.prepare()` and, when it rejects with that signature (`isTurbopackCacheCorruption` in the new `scripts/dev/turbopackCacheHeal.mjs`), purges `.build/next/**/cache/turbopack` and retries **once** with a clear log. **Caveat — best-effort only:** the corruption often surfaces as a runtime overlay rather than a `prepare()` rejection, so this cannot always intercept it; the reliable remedy remains manually deleting the Turbopack cache dir. Regression guard: `tests/unit/turbopack-cache-heal-6289.test.ts`. (thanks @chirag127)
+
+- **fix(providers):** qodercli PAT auth no longer fails with `spawn qodercli ENOENT` on Windows ([#6263](https://github.com/diegosouzapw/OmniRoute/issues/6263)) — `spawnQoderCli` spawned the bare `qodercli` name with `shell:false` and an unenriched env, so the npm `.cmd` wrapper under `%APPDATA%\npm` (a user-PATH directory) was never resolved. It now resolves the absolute `.cmd`/`.exe` path through the existing `getCliRuntimeStatus("qoder")` resolver in `src/shared/services/cliRuntime.ts` (memoized), spawns with `shell` when the target is a `.cmd`/`.bat`, and uses the cliRuntime-enriched env (PATH + PATHEXT + APPDATA); the ENOENT error now lists the searched paths plus the `CLI_QODER_BIN` override. End-to-end spawn on a real Windows host is host-only (Hard Rule #18); the path-resolution logic is unit-tested. Regression guard: `tests/unit/qodercli-windows-resolve-6263.test.ts`. (thanks @chirag127)
+
+- **fix(sse):** the reasoning-token buffer no longer inflates **probe-sized `max_tokens`** ([#6274](https://github.com/diegosouzapw/OmniRoute/issues/6274)) — Claude Code's `/model` capability check sends `max_tokens: 1`, but for a thinking-capable model with a large output cap (e.g. `glm-5.2`) the #3587 headroom heuristic (`max(current + 1000, ceil(current * 1.5))`) rewrote it to `1001` and forwarded that upstream, wasting tokens on a request that was never a genuine reasoning budget. `resolveReasoningBufferedMaxTokens()` (`open-sse/services/reasoningTokenBuffer.ts`) now short-circuits and returns the caller's value verbatim when it is below the new `REASONING_BUFFER_MIN_TRIGGER` (256) threshold — a tiny explicit limit is a probe, not a reasoning request. Real budgets still receive the #3587 headroom unchanged, and the guard runs after the existing capability checks so unknown / non-reasoning models keep returning `null`. Regression guard: `tests/unit/reasoning-token-buffer-6274.test.ts`. (thanks @brightfiscalband)
+
+- **fix(cli):** `omniroute reset-password` now works as a real subcommand, and password resets over piped (non-TTY) stdin actually apply ([#6261](https://github.com/diegosouzapw/OmniRoute/issues/6261), [#6258](https://github.com/diegosouzapw/OmniRoute/issues/6258)). Two coupled defects: (1) **#6261** — `bin/omniroute.mjs` routed everything through Commander with only two pre-Commander bypasses (`--mcp`, `reset-encrypted-columns`), so `omniroute reset-password` was rejected as an unknown command; only the separate `omniroute-reset-password` bin worked, while the docs falsely advertised the subcommand (incl. a bogus "legacy alias still works"). A pre-Commander bypass mirroring `reset-encrypted-columns` now dynamically imports `bin/reset-password.mjs` (which self-executes) before Commander parses; the three doc lines were corrected. (2) **#6258** — `bin/reset-password.mjs` issued two sequential `rl.question` prompts; under piped stdin the second read never settled at EOF, so `main()` never reached `resetManagementPassword` and the reset was a silent no-op (both prompts printed, no success, password unchanged). The CLI now detects non-TTY stdin and reads it once (first line = password, second line = confirm if present, else reused), adds a `--password-stdin` flag (entire stdin is the password, no confirmation), and exits `0` explicitly so the success line always flushes; interactive TTY behavior is unchanged. Regression guard: `tests/unit/reset-password-cli-6261-6258.test.ts` (3). (thanks @chirag127)
+
+- **fix(db):** the mass-migration **safety abort** now tells the operator how to bypass it and stops flooding the log ([#6260](https://github.com/diegosouzapw/OmniRoute/issues/6260)) — after restoring a backup that wiped the migration tracking table, `runMigrations()` threw the abort on every downstream `ensureDbInitialized()`, re-logging the full banner 11+ times, and the message never mentioned the existing `OMNIROUTE_MAX_PENDING_MIGRATIONS` escape hatch. The abort text now appends a bypass hint (set `OMNIROUTE_MAX_PENDING_MIGRATIONS=0` in `server.env` / `DATA_DIR/.env`), and a new `MigrationSafetyAbortError` is memoized so repeated calls in the same process throw the same instance and emit a single concise line instead of the full cascade. Regression guard: `tests/unit/migration-safety-abort-6260.test.ts`. (thanks @chirag127)
+
+- **fix(auth):** importing a **distinct** Codex/ChatGPT OAuth `auth.json` is no longer falsely rejected as "already exists" when it belongs to a different user in the same workspace ([#6301](https://github.com/diegosouzapw/OmniRoute/issues/6301)). `findExistingCodexConnection` (in `src/lib/oauth/utils/codexAuthImport.ts`) deduped **only** on `providerSpecificData.workspaceId === accountId`, where `accountId` is the shared `chatgpt_account_id`/`tokens.account_id` — so two members of the same ChatGPT Team collapsed onto a single connection (409 `duplicate_account`). The id_token's `https://api.openai.com/auth` claim carries a per-user `chatgpt_user_id` alongside the workspace id (the device-flow path already persisted it as `chatgptUserId`, but the import path did not). Now `parseAndValidateCodexAuth` extracts `userId` (`chatgpt_user_id` → `user_id` → JWT `sub`) into `ParsedCodexAuth`, the create/update paths persist `chatgptUserId` in `providerSpecificData` (mirroring `codex.ts`), and dedup keys on `workspaceId` **AND** `chatgptUserId` — with a backward-compat fallback to legacy accountId-only matching when no stored connection for that workspace records a `chatgptUserId`, so genuinely-same accounts still dedup. Regression guard: `tests/unit/codex-auth-import-userid-dedup-6301.test.ts` (4). (thanks @anungma)
+
+- **fix(providers):** importing models for the **venice-web** provider no longer fails with a red "Provider venice-web does not support models listing" ([#6269](https://github.com/diegosouzapw/OmniRoute/issues/6269)). `venice-web` is a web-cookie provider with an executor but no upstream `/v1/models` endpoint and no registry `models`, so the models route fell through to the tail `400`. Mirroring the `jules`/`linkup-search`/`ollama-search` fix (#5569), it now ships a static local catalog entry in `src/lib/providers/staticModels.ts` — seeding the current Venice lineup (`venice-uncensored`, `llama-3.3-70b`, `qwen3-235b`, `qwen3-4b`, `deepseek-r1-671b`; Venice rotates its catalog, see docs.venice.ai/models/overview) — so the route returns `200` with `source:"local_catalog"`, `intentional:true`. Regression guard: `tests/unit/static-models-venice-web-6269.test.ts`. (thanks @chirag127)
+
+- **fix(api):** the specialty model catalogs (`/v1/embeddings`, `/v1/images`, `/v1/music`, `/v1/videos` model lists) are now derived from the **unified catalog filtered by a predicate** (`getSpecialtyModelsResponse`) instead of ad-hoc per-route logic, so they consistently respect active-credential visibility and stay in sync with the main catalog ([#6303](https://github.com/diegosouzapw/OmniRoute/pull/6303)). Regression guard: `tests/unit/specialty-model-catalog-routes.test.ts`. (thanks @makcimbx)
+- **fix(api):** the agent-bridge server route now resolves the MITM manager via a **dynamic `import("@/mitm/manager.runtime")`** so Turbopack does not statically pull the stub (or over-bundle the manager), and the agent-skills generator anchors its output base path with `path.join(process.cwd(), …)` so Turbopack's static analyzer stops tracing the whole project root ([#6329](https://github.com/diegosouzapw/OmniRoute/issues/6329), [#6366](https://github.com/diegosouzapw/OmniRoute/pull/6366)). Regression guard: `tests/unit/agent-bridge-server-route-dynamic-import.test.ts`. (thanks @Iammilansoni)
+- **fix(api):** internal probes (combo-test, cloud-sync verify) now pick a **management-scoped / allow-all API key** instead of naively grabbing `getApiKeys()[0]` — a restricted `self:usage` first row made the probe fail with "Model X is not allowed for this API key" even when the combo path was healthy (`pickApiKeyForInternalUse` in `src/lib/db/apiKeys.ts`). The API-manager model editor also falls back to `/api/models?all=true` when `/v1/models` is catalog-protected ([#6372](https://github.com/diegosouzapw/OmniRoute/pull/6372)). Regression guard: `tests/unit/pick-internal-api-key-6372.test.ts`. (thanks @jmengit)
+- **fix(live-ws):** the Live Dashboard WebSocket server now **rejects on bind failure** (e.g. `EADDRINUSE` when the API bridge already holds the port) instead of letting the error surface as an unhandled `error` event that crash-loops the process — the `error` listener is attached to `wss` (not `server`) and releases the EventBus subscription on a failed start ([#6324](https://github.com/diegosouzapw/OmniRoute/issues/6324)). Regression guard: `tests/unit/live-ws-eaddrinuse-6324.test.ts`. (thanks @vinayakkulkarni)
+- **fix(dashboard):** the Home provider-topology widget now trusts the live provider-metrics snapshot — it uses `topology.errorProvider` and live `activeRequests` directly instead of re-deriving state from a stale `lastErrorAt` or applying a frontend timeout filter, so the topology reflects real-time provider health ([#6322](https://github.com/diegosouzapw/OmniRoute/pull/6322)). Regression guard: `tests/unit/home-provider-topology-live-state.test.ts`. (thanks @xz-dev)
+- **fix(sse):** strip zero-width markers from streamed **tool-call arguments** — a follow-up to [#5857](https://github.com/diegosouzapw/OmniRoute/pull/5857). That PR removed injected zero-width joiners (U+200D) from streamed assistant text/reasoning but deliberately left tool-call argument JSON byte-exact. The request-side obfuscation (`open-sse/services/claudeCodeObfuscation.ts`) injects ZWJ into agent words — including the temp path inside the Bash tool description — and Claude models copy that verbatim into generated commands, which are delivered as tool-call arguments rather than assistant text. As a result the ZWJ survived and corrupted code blocks (e.g. a temp path rendered with an invisible joiner). Now `open-sse/handlers/responseSanitizer.ts` strips zero-width code points from tool-call argument strings at every emit site (OpenAI non-stream/stream chat `tool_calls` + legacy `function_call`, native Responses `function_call` items, the OpenAI→Responses conversion, and the native Responses streaming `response.function_call_arguments.delta/.done` events). Only zero-width code points are removed; JSON structure and all other bytes stay identical (no parse/restringify), so normal arguments remain byte-exact. Regression guard: 6 new cases in `tests/unit/response-sanitizer.test.ts` (suite 50/50).
+
+- **fix(nodejs):** the default app log path now resolves under `DATA_DIR` (`~/.omniroute/logs/application/app.log`) instead of `process.cwd()` ([#6197](https://github.com/diegosouzapw/OmniRoute/issues/6197)) — the globally-installed CLI runs from an arbitrary working directory, so anchoring the default to cwd made file logging silently write to (or no-op under) an unrelated directory, contradicting the documented `.env.example` default. `getAppLogFilePath()` now computes the default lazily via the pure `resolveDataDir()` resolver (honours a per-process `DATA_DIR`, no directory-creation side effect); an explicit `APP_LOG_FILE_PATH` still wins. Regression guard: `tests/unit/logenv-datadir-path-6197.test.ts` (3).
+
+- **fix(docker):** AgentBridge/`startMitm` no longer aborts in containers/headless when the Antigravity-default DNS step can't write `/etc/hosts` ([#6127](https://github.com/diegosouzapw/OmniRoute/issues/6127)), and the privileged command's stderr now reaches `app.log` instead of only a bare exit code hitting the toast ([#6198](https://github.com/diegosouzapw/OmniRoute/issues/6198)). The default DNS step (`addDNSEntry`) was called unguarded while cert install and the two sibling DNS steps were each best-effort — in the runtime Docker image (`USER node`, no `sudo`, read-only `/etc/hosts`) it threw `Command failed with code 1` out of `startMitmInternal` and killed the whole start, discarding the stderr. The three DNS steps are extracted into a best-effort `provisionDnsEntries()` where each failure is logged with the full `err` (stderr included, folded in by `systemCommands.ts`) and never aborts the start. Regression guard: `tests/unit/mitm-dns-graceful-degrade-6127.test.ts` (4).
+
+- **fix(providers):** copilot-m365-web now supports the M365 Education "Starter / OfficeWebIncludedCopilot" tier and no longer returns an empty `content:null` stream ([#6210](https://github.com/diegosouzapw/OmniRoute/issues/6210)). Two gaps: (1) `buildWsUrl()` hardcoded the individual-consumer scenario (`OfficeWebPaidConsumerCopilot`, `isEdu=false`) — the EDU tier is now opt-in via `providerSpecificData.tier="edu"`, emitting `scenario=OfficeWebIncludedCopilot`/`isEdu=true` (the individual path is unchanged); (2) the EDU/GPT-5.5 path streams deltas via `arguments[0].writeAtCursor` (incremental) instead of only `messages[].text` (accumulated snapshots), which the parser dropped — a new `accumulateBotContent()` folds both formats, with `type:2 item.result.message` as a last-resort fallback. Regression guard: `tests/unit/copilot-m365-edu-writeatcursor-6210.test.ts` (10). (thanks @qpeyba)
+
+- **fix(providers):** GitLab Duo executor now feeds tool results back into the prompt instead of looping ([#6220](https://github.com/diegosouzapw/OmniRoute/issues/6220)) — `buildPrompt()` branched only on `system`/`user` and took `userParts.at(-1)`, silently dropping the `assistant{tool_calls}` + `tool{result}` turns the client appended, so the reconstructed prompt was byte-identical to turn 1 and the model re-emitted the same `<tool>` call forever. When a tool exchange is present the full conversation is now serialized, folding each tool result back keyed by its `tool_call_id`; simple conversations keep the legacy shape. Complements the tool_call emission from [#6051](https://github.com/diegosouzapw/OmniRoute/issues/6051) (the `kilo-duplicate` label was a false positive — different, sequential defect). Regression guard: `tests/unit/gitlab-tool-result-feedback-6220.test.ts` (4).
+
+- **fix(providers):** opencode-go/opencode-zen can now synthesize the OpenCode CLI identity headers Cloudflare requires on VPS egress ([#5997](https://github.com/diegosouzapw/OmniRoute/issues/5997)) — on a datacenter VPS, `opencode.ai/zen/go/v1/chat/completions` 403s (HTML challenge) requests lacking CLI identity, while the reporter's control curl proved that `User-Agent: opencode-cli/1.0.0` + `x-opencode-client: cli` + `x-opencode-project: default` + fresh request/session UUIDs succeed. Opt-in via `OPENCODE_SYNTHESIZE_CLI_HEADERS=true` (values overridable via `OPENCODE_GO_USER_AGENT`/`OPENCODE_USER_AGENT`/`OPENCODE_CLIENT`/`OPENCODE_PROJECT`); it fills only headers the client did not already send. Kept **off by default** — the forward-only path is deliberate (fabricating a wrong value risks upstream rejection; a prior dedup regressed with `opencode/local`), so this replaces the fragile local header-injection shim without changing default behavior. Regression guard: `tests/unit/opencode-cli-headers-synthesis-5997.test.ts` (6). (thanks @aleksesipenko)
+- fix(resilience): sticky session affinity now evicts and fails over to another account when the pinned account is exhausted/unavailable (#6219)
+
+- fix(sse): Responses API passthrough now drops internal commentary-phase output before forwarding to clients (gated by RESPONSES_PASSTHROUGH_DROP_COMMENTARY, default on) (#6199)
+
+- **fix(sse):** tool-call function schemas with a root `type: null` are now coerced to `type: "object"` before dispatch ([#6359](https://github.com/diegosouzapw/OmniRoute/issues/6359)) — clients like the Codex app emit `parameters: { type: null, ... }` for some tools, which OpenAI-compatible upstreams reject with `400 Invalid schema for function '...': schema must be a JSON Schema of 'type: "object"', got 'type: null'`, failing the whole request. `toolSchemaSanitizer` already stripped the null; it now re-adds the mandatory root `"object"` type (and empty `properties`/open `additionalProperties` when absent). Combinator roots (`anyOf`/`oneOf`/`allOf`) and explicit root types are left untouched. Regression guard: 5 new cases in `tests/unit/tool-schema-sanitizer.test.mjs`.
+
+- **fix(docker):** AgentBridge no longer fails to start on npm/Electron/VPS installs with "MITM manager stub reached at runtime" ([#6344](https://github.com/diegosouzapw/OmniRoute/issues/6344)) — v3.8.45 flipped the production bundler default to Turbopack, but `next.config.mjs` aliased `@/mitm/manager` to its Docker-only degraded stub **unconditionally**. That was harmless while Docker (which sets the alias intentionally for #3390 graceful degradation) was the sole Turbopack consumer, but once every artifact built with Turbopack the stub shipped to all non-Docker users and `startMitm` threw on the first Agent-Bridge start. The alias is now opt-in via `OMNIROUTE_MITM_STUB=1` (set only by the Dockerfile) through the shared `scripts/build/mitm-stub-flag.mjs` helper; default builds bundle the real manager. Regression guard: `tests/unit/mitm-stub-alias-6344.test.mjs` (4).
+
+- **fix(proxy):** stop the v3.8.44 proxy regression that leaked the real IP and disabled healthy proxies ([#6246](https://github.com/diegosouzapw/OmniRoute/issues/6246)). Two coupled defects from the new health scheduler: (1) **IP leak** — when a proxy assigned to a connection was marked `inactive`, resolution fell through to a **direct** egress instead of blocking, exposing the operator's real IP; (2) **over-deactivation** — the sweep flipped a proxy to `inactive` on the **first** failed probe and counted our own 5s timeout / a probe-target `5xx` as the proxy's fault, so healthy paid proxies vanished from egress selection ("my proxies are not being used anymore"). Fix: the sweep decision is extracted into a pure, network-free `decideProxyHealthAction` (`src/lib/proxyHealth/decision.ts`) — by default the health check now **only counts/logs and never downgrades status** (a proxy is downgraded/removed only with `PROXY_AUTO_REMOVE=true`, after `PROXY_AUTO_REMOVE_AFTER` **consecutive** conclusive failures); probes are classified tri-state so an inconclusive result (our timeout, or a `5xx` from the probe target) never penalizes the proxy, and the probe timeout is raised 5s→15s. Separately, `safeResolveProxy` now **fails closed** via the existing policy: a connection whose assigned proxy is dead is blocked instead of leaking direct (`hasBlockingProxyAssignment`), honoring the explicit `proxy off` toggles and the `PROXY_FAIL_OPEN=true` opt-out. Existing proxies stuck `inactive` by the old behavior need a one-time manual re-activate (the operator owns proxy status). Regression guards: `tests/unit/proxy-health-decide-action-6246.test.ts`, `tests/unit/proxy-assigned-unavailable-6246.test.ts`.
+
+- **fix(proxy):** make "Test All" read-only and add bulk enable/disable ([#6246](https://github.com/diegosouzapw/OmniRoute/issues/6246)). Complements the core fail-closed / scheduler fix (#6296) with the two remaining reporter asks. (1) The **"Test All" button** (`POST /api/settings/proxies/auto-test`) used to flip a proxy to `inactive` on a failed reachability probe; since the egress selector excludes `inactive` proxies, a flaky probe (an unreachable `httpbin.org`, a proxy that blocks `HEAD`, or a slow paid proxy) silently disabled every proxy that failed — "Test All" is now **read-only by default** (only the operator sets a proxy active/inactive; opt back into the legacy test-and-set with `PROXY_HEALTH_AUTO_DEACTIVATE=true`). (2) Adds a **bulk enable/disable** proxies endpoint + toolbar action (`POST /api/settings/proxies/batch-activate`) so an operator can re-activate proxies in one click. Regression guard: `tests/unit/proxy-health-6246.test.ts`. (thanks @tenshiak)
+
+- **chatcore (tools): stop the default 128-tool cap from silently dropping opencode's `task`/MCP tools.** opencode (used as an MCP/agent host) sends a large tool list; when it exceeds the speculative `MAX_TOOLS_LIMIT` (128) default, `truncateToolList` did a blind `tools.slice(0, 128)`, dropping every tool past index 128 — including opencode's built-in `task` tool (subagent launch) and many MCP tools, so models routed through OmniRoute could no longer spawn subagents or reach part of their tools. The cap exists to avoid upstream `400`s for providers with real hard limits (e.g. grok-cli 200), so it is kept for those: detection of the opencode client (`isOpencodeClient` — any `x-opencode-*` header, or `opencode` in the user-agent) now only bypasses the **speculative 128 default**, never a known provider ceiling. Precedence is explicit — a proactive/detected provider limit always truncates (even for opencode); otherwise opencode forwards its full tool list; otherwise the unchanged 128 default applies to every other client. Refactors `getEffectiveToolLimit` into `getKnownToolLimit(provider) ?? DEFAULT_LIMIT` (byte-identical for existing callers) and fixes a cosmetic debug-log that reported the truncated count instead of the original. Regression guard: `tests/unit/tool-limit-detector.test.ts`.
+
+- **fix(mitm):** the macOS MITM-cert install check now matches the system keychain again. `security find-certificate -a -Z` prints the SHA-1 as a colon-less hex string, but the installed-check compared it against `getCertFingerprint()`'s colon-separated form, so the substring match never hit — the cert was reported as not-installed and re-prompted for the sudo install on every run. Fingerprints are now normalized (colons stripped, upper-cased) on both sides via the extracted `macCertOutputHasFingerprint` helper. Regression guard: `tests/unit/mitm-cert-mac-fingerprint.test.ts`. ([#6204](https://github.com/diegosouzapw/OmniRoute/pull/6204), closes [#6134](https://github.com/diegosouzapw/OmniRoute/issues/6134) — thanks @rianonehub)
+
+- **fix(api):** `/v1/messages/count_tokens` now counts `tool_use`, `tool_result` and `thinking` content blocks (and array-form `system` prompts) in the local-estimation path, instead of only `text`. Real agentic conversations keep ~95% of their tokens inside tool results; the previous estimate returned near-zero for them, which silently broke Claude Code's auto-compaction (context grew past the window with no compaction until the upstream API rejected the request). The real provider-side count path is unchanged. Regression guard: `tests/unit/messages-count-tokens-route.test.ts`. ([#6221](https://github.com/diegosouzapw/OmniRoute/pull/6221) — thanks @luweiCN)
+
+- **fix(antigravity):** strip a trailing assistant prefill turn for Vertex Claude models to avoid upstream 400s ([#6114](https://github.com/diegosouzapw/OmniRoute/pull/6114)). Regression guard: `tests/unit/antigravity-claude-prefill-strip.test.ts`. (thanks @anki1kr)
+
+- **fix(security):** the mutable cloud-agent routes (`/api/cloud/credentials/update`, `/api/cloud/models/alias`) now require management auth instead of being treated as public. They were classified as public API routes, so a request without management credentials could update stored cloud-agent credentials and model aliases. They are removed from the public-route set, classified as management routes in the authz pipeline, and gated by `requireManagementAuth`; cloud **read**/auth routes stay public. Regression guards: `tests/unit/cloud-write-auth.test.ts`, `tests/unit/authz/classify.test.ts`, `tests/unit/public-api-routes.test.ts`. ([#6233](https://github.com/diegosouzapw/OmniRoute/pull/6233) — thanks @vittoroliveira-dev)
+
+- **refactor(dashboard):** extract the onboarding-wizard "Open provider details" link target into a pure, unit-tested `buildProviderDetailsHref(connection)` helper. The wizard already routes by `connection.id` (the node UUID) rather than the provider category slug (#6144/#6145); this hardens that behavior behind a tested helper that guards a missing id/connection. Regression guard: `tests/unit/provider-onboarding-href.test.ts`. ([#6166](https://github.com/diegosouzapw/OmniRoute/pull/6166) — thanks @KooshaPari)
+
+- **fix(security):** the doubao synthetic device-id generator now derives its digits via an unbiased crypto-random draw (rejection sampling over `crypto.randomBytes()`) instead of a `% 10` reduction, closing a CodeQL `js/biased-cryptographic-random` finding.
+- **fix(agentSkills):** the GitHub-skills generator now resolves `outputDir` to an absolute path before writing, fixing a regression introduced by #6366 (relative-to-cwd base path) that could write generated skill files to the wrong directory.
+- **fix(security):** `/api/keys/{id}/devices` now checks the HTTP method before auth/validation, returning a `405` for non-GET/DELETE verbs instead of a misleading `401`/`500` (closes a `dast-smoke` QUERY-method finding).
+- **fix(quality):** clear the last 2 heavy quality-gate reds on the release tip (cycle pre-flight).
+- **fix(mitm):** the test suite and CI can never mutate the OS trust store — `OMNIROUTE_SKIP_SYSTEM_TRUST=1` is set globally for tests/CI so `installCert`/`uninstallCert`/`installTproxyCa` skip the privileged OS dispatch ([#6310](https://github.com/diegosouzapw/OmniRoute/pull/6310); full detail is under the [3.8.45] section below — this branch received it via the parallel-cycle sync-back).
+- **fix(api):** `POST /api/github-skills` now Zod-validates its request body; documented the new quality-gate env vars and pinned the merge-integrity GitHub Actions to a commit SHA.
+- **fix(skills):** generate the missing `omni-github-skills` registry entry and align the agent-skills catalog-count tests (follow-up to #6186).
+- **fix(quality):** clear the cycle's 11 net-new ESLint errors and make `validate-release-green` suppressions-aware.
+
+### 📝 Maintenance
+
+- **i18n(it):** add 118 missing Italian (`it`) translations (net-additive — no existing keys dropped, valid JSON), improving Italian UI coverage. ([#6212](https://github.com/diegosouzapw/OmniRoute/pull/6212) — thanks @serverless83)
+- **chore(providers):** remove deprecated **MiMo V2** model entries from the catalogs (xiaomi-mimo, opencode-go, zenmux-free, audio TTS) — the upstream V2 line is superseded by MiMo V2.5; drops `mimo-v2-tts`, `mimo-v2-pro`, `mimo-v2-omni`, `mimo-v2-flash`, `mimo-v2-flash-free` and realigns the provider-catalog tests. ([#6248](https://github.com/diegosouzapw/OmniRoute/pull/6248) — thanks @backryun)
+- **chore(release):** ~50 commits on this branch are v3.8.45 pre-flight/hardening fixes and CI-perf work that landed here via the parallel-cycle sync-back (`sync-next-cycle.mjs`, Hard Rule #21) **after** the `v3.8.45` git tag was cut, and are already fully documented under the **[3.8.45]** section below — listed here only so the per-cycle commit-coverage check (`npm run release:uncovered`) doesn't flag them as gaps. Provider/catalog/UX/backend fixes: #6041, #6078, #6108, #6135, #6148, #6149, #6154, #6158, #6161, #6162, #6163, #6164, #6165, #6170, #6177, #6178, #6181, #6186, #6187, #6191, #6193, #6194, #6195, #6200, #6205, #6208, #6209, #6211, #6213, #6223, #6224, #6225, #6226, #6227, #6228, #6229, #6230, #6235, #6291, #6292. CI/release-pipeline work: #6167, #6203, #6214, #6215, #6218, #6273, #6275, #6283, #6284, #6285, #6300, #6305.
+- **chore(release):** additional zero-ref release-cycle plumbing on this branch, kept out of `release:uncovered` on purpose (no `#N` in the commit subject to cite): opening the v3.8.46 cycle, opening/closing the v3.8.45 cycle, the finalized [3.8.45] CHANGELOG i18n sync-back to 42 mirrors, the v3.8.45 cognitive/cyclomatic and file-size drift rebaselines, ESLint stale-suppression pruning (4,273 → 4,233), and clearing test-masking/docs-all pre-flight reds for v3.8.45.
+
+### ⚡ Performance & Infrastructure
+
+- **perf(release-green):** the pre-flight validator (`scripts/quality/validate-release-green.mjs`) now runs its 4 slow suites (unit / vitest / integration / pack-artifact) **concurrently** via `Promise.all` — pre-flight wall time drops from ~the sum of the suites to ~the slowest one (~30min saved per round; Phase 0 was the nº1 bottleneck of the v3.8.45 release benchmark, 2h54 of 6h34 e2e). Guard: `tests/unit/validate-release-green.test.ts` ("runs the slow suites CONCURRENTLY"). ([#6319](https://github.com/diegosouzapw/OmniRoute/pull/6319))
+- **fix(ci):** `scripts/release/sync-next-cycle.mjs` — two defects found live in its first production run (v3.8.45 Phase 5): (1) the `git()` helper's default 1 MiB `maxBuffer` crashed with `ENOBUFS` on `git show origin/main:CHANGELOG.md` (the CHANGELOG alone is >1 MiB) — widened to 64 MiB; (2) the i18n resync only propagated the `[NEXT]` (TBD) section, leaving the just-shipped finalized section as "— TBD" in all 42 mirrors — it now also syncs `[prevVersion]` bounded by the heading below it (new exported pure helper `versionAfter`). Guards: +5 tests in `tests/unit/sync-next-cycle.test.ts` (8/8). ([#6327](https://github.com/diegosouzapw/OmniRoute/pull/6327))
+- **test(ci):** concurrency-sensitive flaky tests are **quarantined into a serial pass** (`tests/unit/serial/`, `--test-concurrency=1`, appended to every unit runner incl. sharded variants — the serial pass is sharded too so concurrent shard jobs never self-collide). Initial set: `glm-coding-plan-monthly-3580`, `quota-division-blocks`, `provider-health-autopilot`, `combo-health-autopilot` — the class behind the ~28min CI wedges/re-runs (two live 1h+ wedges cancelled during this PR's own validation). Discovery + TIA gates track the new glob; systemic root cause (async logger writing after teardown) tracked in [#6360](https://github.com/diegosouzapw/OmniRoute/issues/6360). Guard: `tests/unit/test-serial-quarantine.test.ts` (4). ([#6347](https://github.com/diegosouzapw/OmniRoute/pull/6347))
+
+### 🙌 Contributors
+
+Thanks to everyone whose work landed in v3.8.46:
+
+| Contributor                                                    | PRs / Issues           |
+| -------------------------------------------------------------- | ---------------------- |
+| [@2220258345](https://github.com/2220258345)                   | direct commit / report |
+| [@abdofallah](https://github.com/abdofallah)                   | direct commit / report |
+| [@adentdk](https://github.com/adentdk)                         | direct commit / report |
+| [@aleksesipenko](https://github.com/aleksesipenko)             | direct commit / report |
+| [@anki1kr](https://github.com/anki1kr)                         | direct commit / report |
+| [@anungma](https://github.com/anungma)                         | direct commit / report |
+| [@arpicato](https://github.com/arpicato)                       | direct commit / report |
+| [@backryun](https://github.com/backryun)                       | #6248                  |
+| [@binsarjr](https://github.com/binsarjr)                       | direct commit / report |
+| [@brightfiscalband](https://github.com/brightfiscalband)       | direct commit / report |
+| [@chirag127](https://github.com/chirag127)                     | #6501, #6506           |
+| [@developerjillur](https://github.com/developerjillur)         | direct commit / report |
+| [@dilneiss](https://github.com/dilneiss)                       | #6499                  |
+| [@dtybnrj](https://github.com/dtybnrj)                         | direct commit / report |
+| [@Forcerecon](https://github.com/Forcerecon)                   | direct commit / report |
+| [@hao3039032](https://github.com/hao3039032)                   | direct commit / report |
+| [@Iammilansoni](https://github.com/Iammilansoni)               | #6150, #6245           |
+| [@jmengit](https://github.com/jmengit)                         | direct commit / report |
+| [@jordansilly77-stack](https://github.com/jordansilly77-stack) | direct commit / report |
+| [@JxnLexn](https://github.com/JxnLexn)                         | direct commit / report |
+| [@KooshaPari](https://github.com/KooshaPari)                   | #6166                  |
+| [@loopyd](https://github.com/loopyd)                           | direct commit / report |
+| [@luweiCN](https://github.com/luweiCN)                         | #6221                  |
+| [@makcimbx](https://github.com/makcimbx)                       | direct commit / report |
+| [@muflifadla38](https://github.com/muflifadla38)               | direct commit / report |
+| [@newnol](https://github.com/newnol)                           | direct commit / report |
+| [@ofekbetzalel](https://github.com/ofekbetzalel)               | direct commit / report |
+| [@ohahe52-dot](https://github.com/ohahe52-dot)                 | direct commit / report |
+| [@phidinhmanh](https://github.com/phidinhmanh)                 | direct commit / report |
+| [@powellnorma](https://github.com/powellnorma)                 | direct commit / report |
+| [@qpeyba](https://github.com/qpeyba)                           | direct commit / report |
+| [@RaviTharuma](https://github.com/RaviTharuma)                 | direct commit / report |
+| [@RCrushMe](https://github.com/RCrushMe)                       | direct commit / report |
+| [@rianonehub](https://github.com/rianonehub)                   | #6134, #6204           |
+| [@serverless83](https://github.com/serverless83)               | #6212                  |
+| [@swingtempo](https://github.com/swingtempo)                   | direct commit / report |
+| [@tenshiak](https://github.com/tenshiak)                       | direct commit / report |
+| [@ThongAccount](https://github.com/ThongAccount)               | direct commit / report |
+| [@UnrealAryan](https://github.com/UnrealAryan)                 | direct commit / report |
+| [@vinayakkulkarni](https://github.com/vinayakkulkarni)         | direct commit / report |
+| [@vittoroliveira-dev](https://github.com/vittoroliveira-dev)   | #6233                  |
+| [@warelik](https://github.com/warelik)                         | direct commit / report |
+| [@xxy9468615](https://github.com/xxy9468615)                   | direct commit / report |
+| [@xz-dev](https://github.com/xz-dev)                           | direct commit / report |
+| [@yanpaing007](https://github.com/yanpaing007)                 | direct commit / report |
+| [@diegosouzapw](https://github.com/diegosouzapw)               | maintainer             |
+
+---
+
 ## [3.8.45] — 2026-07-06
 
 ### ✨ New Features

docs/i18n/mr/CHANGELOG.md

@@ -6,6 +6,213 @@
 
 ## [3.8.31] — 2026-06-20
 
+## [3.8.46] — TBD
+
+### ✨ New Features
+
+- **feat(sse):** **hide paid-only models from `auto/*` routing** when `hidePaidModels` is on ([#6512](https://github.com/diegosouzapw/OmniRoute/issues/6512)) — follow-up to #6328/#6495. PR #6495 hid paid-only models from the `GET /v1/models` listing, but `auto/*` combos (`auto/best-coding`, `auto/glm`, …) could still pick a paid-only backend into their candidate pool → a 402/403 at request time. `createVirtualAutoCombo` now filters the candidate pool through the new pure `open-sse/services/autoCombo/paidModelFilter.ts` (`filterPaidOnlyCandidates`), applying the same free-model predicate #6495 uses in `catalog.ts` (`providerHasFreeModels(provider) && isFreeModel(provider, {id})`) whenever `settings.hidePaidModels === true`. Applied before the category/tier/family narrowing, so it covers every `auto/*` combo; an all-paid pool degrades to the existing graceful empty-pool path. **Opt-in — default OFF leaves the pool unchanged** (identity). Regression guard: `tests/unit/autoCombo/paid-model-filter-6512.test.ts` (4, incl. the default-off identity guard).
+- **feat(sse):** **provider-family auto combos** — `auto/glm`, `auto/minimax`, `auto/mimo`, `auto/zai`, `auto/gemma`, `auto/llama`, `auto/gemini` ([#6453](https://github.com/diegosouzapw/OmniRoute/issues/6453)) — new routable ids that materialize an on-demand virtual combo spanning whatever installed backends currently expose that model family, degrading gracefully as backends rotate. A new pure `open-sse/services/autoCombo/modelFamily.ts` (`detectModelFamily`) classifies by model-id prefix for six families; `zai` is instead resolved by provider id (z.ai's hosted API serves the same `glm-*` model ids as every other GLM backend, so `auto/zai` means "route to my z.ai backend specifically" vs `auto/glm`'s "any connected GLM backend"). Reuses the existing `createVirtualAutoCombo` on-demand materialization path (no DB writes) and the `/v1/models` catalog advertising loop. Regression guard: `tests/unit/autoCombo/provider-family-combos.test.ts` (11).
+- **feat(proxy):** native **proxy-pool round-robin / egress IP rotation** ([#6365](https://github.com/diegosouzapw/OmniRoute/issues/6365)) — a scope (global / provider / account) can now hold **multiple** proxies as a pool with a rotation strategy, so outbound requests cycle their egress IP instead of pinning one proxy per scope. Migration `117_proxy_pool_rotation.sql` lifts the `UNIQUE(scope, scope_id)` constraint (rebuild via the canonical rename/copy/drop; existing single assignments become 1-element pools) and adds a `proxy_scope_rotation` companion table holding the per-scope strategy + a persisted monotonic round-robin cursor. Strategies: `round-robin` (default, monotonic cursor — never `Math.random`), `random`, and `sticky-per-N-min`. Resolution (`resolveProxyForScopeFromRegistry` / `resolveProxyForConnectionFromRegistry`) now fetches the alive, position-ordered candidate set (unchanged `PROXY_ALIVE_PREDICATE`) and applies the strategy; an empty / all-dead pool still returns `null` — the #6246 fail-closed guard is **untouched** (never falls through to direct egress). Backend + DB only; dashboard pool-builder UI is a follow-up. Regression guard: `tests/unit/proxy-pool-rotation-6365.test.ts` (8, incl. fail-closed + backward-compat).
+- **feat(providers):** end-to-end **tool/function calling on the native Gemini `/v1beta` endpoint** ([#6222](https://github.com/diegosouzapw/OmniRoute/issues/6222)) — both directions of the Gemini↔OpenAI conversion now preserve tool data (previously silently dropped). Request side: `convertGeminiToInternal` (extracted to its own testable module) maps `tools[].functionDeclarations` → OpenAI `tools`, prior `functionCall` parts → assistant `tool_calls`, and `functionResponse` parts → `tool`-role messages. Response side: `convertOpenAIResponseToGemini` emits `parts[].functionCall {name,args}` from `message.tool_calls`, and the streaming `openAIChunkToGeminiChunk` accumulates fragmented `tool_calls` deltas by index into complete `functionCall` parts. The non-Gemini client paths (Claude, OpenAI-Responses) already preserved tool calls — this closes the gap specific to the native Gemini surface. Regression guard: `tests/unit/v1beta-gemini-tool-calling-6222.test.ts` (6, incl. a streaming SSE round-trip).
+- **feat(providers):** copilot-m365-web **enterprise / work tier** support ([#6334](https://github.com/diegosouzapw/OmniRoute/issues/6334)) — mirrors the EDU-tier pattern (#6210): `M365ConnectionParams` gains an `agent` field, a new opt-in `M365_ENTERPRISE_OVERRIDES` preset (`agent=work`, `scenario=officeweb`, `licenseType=Premium`) applies via `providerSpecificData.tier="enterprise"` (alias `"work"`), and `agent` is also overridable directly via `providerSpecificData.agent`. `buildWsUrl` was hardcoding `agent="web"` (the one enterprise-distinguishing param with no override path), so a Premium work account handshook then returned an empty stream. The individual and EDU paths are untouched. Kilo's dup flag vs #6210 (EDU tier) was a false positive — different tier. Regression guard: `tests/unit/copilot-m365-enterprise-6334.test.ts` (7). End-to-end confirmation on a real Premium work account is a live-VPS validation follow-up (Hard Rule #18). (thanks @Forcerecon)
+- **feat(api):** standardized, provider-agnostic **`effort` + `thinking` request params** ([#6241](https://github.com/diegosouzapw/OmniRoute/issues/6241)) — a thin standardization layer over the existing mature per-provider reasoning plumbing (no provider mapper touched). `providerChatCompletionSchema` gains a canonical `effort` (reusing the shared `none/low/medium/high/xhigh` vocabulary — the UI tiers `extra`/`max` collapse onto `xhigh`) and a boolean `thinking`. A pure `normalizeReasoningRequest` (wired once in `src/sse/handlers/chat.ts`, before any reasoning field is read) folds them onto the fields the translators already consume (`reasoning_effort` / `reasoning.effort` / `thinking`), so they fan out to Anthropic / Gemini / xAI / Responses — an explicit client `reasoning_effort` / object-shaped `thinking` always wins (backward-compatible). `/models` additively exposes `supportsThinking` + `effort_tiers` so the frontend can render the toggles (UI component is a follow-up). Regression guard: `tests/unit/effort-thinking-standardization-6241.test.ts` (12). (thanks @Iammilansoni, @shabeer)
+- **feat(combo):** new **`pipeline` (sequential) combo strategy** ([#6297](https://github.com/diegosouzapw/OmniRoute/issues/6297)) — the 18th routing strategy runs targets **in order**, threading each step's output into the next step's input, with an optional per-step `prompt` (system instruction); only the final step's response is returned. Distinct from `fusion` (parallel fan-out + judge). Implemented as a self-contained `open-sse/services/pipeline.ts` (sibling to `fusion.ts`), dispatched from `combo.ts`; the step list reuses `combo.models` order and reads an optional `prompt` off each target (backward-compatible — ignored by every other strategy). Intermediate steps run non-streaming with tools stripped (complete prose to thread forward); the final step keeps the client's `stream` flag + tools. A failing/empty/unparseable intermediate step fails the whole pipeline explicitly via a sanitized error (never silently swallowed). Kilo's dup flag vs #563 was a false positive (that's model→chain selection; this is a sequential chain). Regression guard: `tests/unit/combo-pipeline-strategy.test.ts` (5). (thanks @ofekbetzalel)
+- **feat(ci):** `check:test-masking` now flags **inline-reimplemented prod conditions** ([#6348](https://github.com/diegosouzapw/OmniRoute/issues/6348)) — a new **report-only** subcheck (v2, 6A.10 family) catches the wrong-shape contract test: a test that recomputes the condition under test inline instead of importing/exercising the real function (the #6216 class, where `=== 500` → `>= 500` stayed green because the test re-implemented the branch). For each added/modified test file it warns when the file textually duplicates a ≥3-token conditional from a production file touched in the same PR **and** does not import the symbol/module owning it, via a pure, fixture-tested `findReimplementedConditions()` with an allowlist mirroring `assertReductionAllowlist`. Report-only for now (does not fail the gate) — to be promoted to blocking after a triage cycle. Regression guard: `tests/unit/check-test-masking.test.ts` (45).
+- **feat(sse):** per-connection routing override (native vs CLIProxyAPI) ([#6339](https://github.com/diegosouzapw/OmniRoute/issues/6339)) — the previously-dead `isCliproxyapiDeepModeEnabled` helper is now wired into `resolveExecutorWithProxy`: a single connection can opt itself into the CLIProxyAPI passthrough executor via `providerSpecificData.cliproxyapiMode="claude-native"`, with precedence **connection override > provider `upstream_proxy_config` mode > default**. `resolveExecutorWithProxy` now receives the resolved connection's `providerSpecificData` (threaded from `chatCore.ts`), so one connection can deep-route while the provider's other connections stay native — no DB schema change (the toggle rides in `providerSpecificData`). Also resolves the same-provider mixing ask in #6340. Regression guard: `tests/unit/chatcore-executor-proxy.test.ts` (9). (thanks @RaviTharuma)
+- **feat(dashboard):** "Add session cookie" modal now shows a prominent **"Open ‹host› →"** link to the provider's own site ([#6268](https://github.com/diegosouzapw/OmniRoute/issues/6268)) — every `-web` cookie-session provider (chatgpt-web, claude-web, gemini-web, kimi-web, lmarena, qwen-web, m365-copilot-web, …) renders a one-click external link (opening the provider's login/home page in a new tab) so operators no longer tab away to retype the URL mid-setup. The host resolves from a pure, unit-tested `resolveWebProviderHost()` (prefers `WEB_COOKIE_PROVIDERS[id].website`, falls back to the registry `baseUrl` origin); non-web providers render exactly as before. Kilo's dup flag vs #6265 (modal-too-small-on-1080p) was a false positive — distinct concern. Regression guard: `tests/unit/resolve-web-provider-host.test.ts` (5). (thanks @chirag127)
+- **feat(providers):** add **DigitalOcean** AI (serverless inference) as an OpenAI-compatible API-key provider ([#6373](https://github.com/diegosouzapw/OmniRoute/pull/6373)) — base `https://inference.do-ai.run/v1`, wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/digitalocean/`, `src/shared/constants/providers/apikey/inference-hosts.ts`). Regression guard: `tests/unit/digitalocean-provider.test.ts`. (thanks @newnol)
+- **feat(providers):** add **Huancheng Public API** (`hcnsec`) as an OpenAI-compatible regional provider ([#6410](https://github.com/diegosouzapw/OmniRoute/pull/6410)) — Xinjiang Huancheng Cybersecurity's public LLM platform (base `https://api.hcnsec.cn/v1`, free credits via daily check-ins), wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/hcnsec/`, `src/shared/constants/providers/apikey/regional.ts`). Regression guard: `tests/unit/hcnsec-provider.test.ts`. (thanks @UnrealAryan)
+- **feat(dashboard):** the web-session credential guide now shows an **"Open {host}" link** ([#6316](https://github.com/diegosouzapw/OmniRoute/pull/6316)) to the provider's sign-in site (derived from the provider `website` via `getProviderWebsiteHost`), so you can jump straight to the page where the cookie/session must be captured. Regression guard: `tests/unit/web-session-provider-link-6316.test.ts`. (thanks @jordansilly77-stack)
+- **feat(cerebras):** add the **Gemma 4 31B** model (`gemma-4-31b`) to the Cerebras registry + pricing table ([#6331](https://github.com/diegosouzapw/OmniRoute/pull/6331)). Regression guard extends `tests/unit/t28-model-catalog-updates.test.ts`. (thanks @backryun)
+- **feat(providers):** add **Yuanbao (web)** as a cookie-session provider ([#6196](https://github.com/diegosouzapw/OmniRoute/issues/6196)) — `yuanbao-web` (Tencent Yuanbao, `yuanbao.tencent.com`) with cookie-only auth (`hy_user`/`hy_token` + public agent id), SSE→OpenAI translation incl. `reasoning_content`, exposing DeepSeek V3/R1 + Hunyuan / Hunyuan-T1. Regression guard: `tests/unit/providers-yuanbao-web.test.ts`. `together-web` was **deferred** (no verifiable web-session endpoint — needs a captured request) and `huggingchat-web` **dropped** (the existing `huggingchat` already is a web-cookie provider). (thanks @chirag127)
+- **feat(providers):** route the built-in **agentrouter** through the dynamic Claude-Code wire image ([#6056](https://github.com/diegosouzapw/OmniRoute/issues/6056)) — a small static allow-set (`CC_WIRE_IMAGE_BUILTINS` in `open-sse/services/ccWireImageBuiltins.ts`), consulted by `isClaudeCodeCompatible` / `isClaudeCodeCompatibleProvider` / `applyFingerprint`, makes agentrouter adopt the CC wire-image headers + fingerprint **while guarding the CC baseUrl/auth branches** so it keeps its own registry `baseUrl` and `x-api-key` auth. Regression guard: `tests/unit/agentrouter-cc-wire-image.test.ts` (asserts the wire image is applied AND agentrouter's baseUrl/auth are preserved). Live WAF-acceptance against agentrouter.org is a VPS validation follow-up (Hard Rule #18).
+- **feat(providers):** **bulk-add API keys for Cloudflare Workers AI** ([#6174](https://github.com/diegosouzapw/OmniRoute/issues/6174)) — `cloudflare-ai` is removed from the bulk-add exclusion list and the bulk parser gains a 3-field `name|accountId|apiKey` mode; the bulk route now builds a **per-entry** `providerSpecificData` so each key carries its own `accountId` (fixing the previous shared-object reuse), and both the create + key-validation paths receive it. Regression guard: `tests/unit/bulk-api-key-parser-cloudflare.test.ts`. (thanks @muflifadla38)
+- **feat(dashboard):** routing/settings UX clarity ([#6147](https://github.com/diegosouzapw/OmniRoute/issues/6147)) — (1) weighted combos show the **effective routing share %** next to each weight when weights don't sum to 100 (`WeightTotalBar.tsx`); (2) the status widget's user-facing **"Cloud Sync" label is renamed** to "Remote Settings Sync" (`CloudSyncStatus.tsx`; internal ids/state untouched); (3) built-in providers gain an **opt-in advanced base-URL override** (`isBaseUrlOverrideEligibleProvider`, hidden behind an "Advanced" toggle, reusing the existing `providerSpecificData.baseUrl` persistence — not globally widened). Regression guard: `tests/unit/routing-settings-ux-6147.test.ts`.
+- **feat(combo):** add an option to **disable session stickiness**, per-combo or globally — round-robin / random combos can rotate to a different connection on every request instead of pinning a whole conversation to one connection by its first-message hash. Resolution precedence per-combo `config.disableSessionStickiness` → global `settings.disableSessionStickiness` → default `false` (preserves the #3825 prompt-cache/504 fix); gates **both** stickiness call sites in `open-sse/services/combo.ts`. Exposed as a global toggle (Combo Defaults) and a per-combo Inherit/on/off control. ([#6168](https://github.com/diegosouzapw/OmniRoute/issues/6168)) Regression guard: `tests/unit/combo-disable-session-stickiness.test.ts`. (thanks @RCrushMe)
+- **feat(docker):** add the `OMNIROUTE_NO_SUDO` env flag for root-less / user-namespaced deployments — the MITM cert-trust command path (`resolveSudoSpawn` in `src/mitm/systemCommands.ts`) now strips the leading `sudo` when the flag is truthy, in addition to the existing root / sudo-missing cases, so the Proxy Agent runs without `sudo` (the operator trusts the CA manually, e.g. via `NODE_EXTRA_CA_CERTS`). Argv-array `spawn` preserved — no shell interpolation (Hard Rule #13). ([#6122](https://github.com/diegosouzapw/OmniRoute/issues/6122)) Regression guard: `tests/unit/mitm-systemCommands-no-sudo.test.ts`. (thanks @powellnorma)
+- **feat(providers):** add **Requesty** as an OpenAI-compatible gateway provider (BYOK, base `https://router.requesty.ai/v1`, ~200 free requests/day) — wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/requesty/`, `src/shared/constants/providers/apikey/gateways.ts`). ([#6120](https://github.com/diegosouzapw/OmniRoute/issues/6120)) Regression guard: `tests/unit/requesty-provider.test.ts`. (thanks @chirag127)
+- **feat(dashboard):** add **configured-only / available-only filters** to the Free Provider Rankings page ([#6150](https://github.com/diegosouzapw/OmniRoute/issues/6150)) — hide providers you haven't configured, or whose connections are all rate-limited / out of quota, via server-side query params (`?configuredOnly` / `?availableOnly` on `GET /api/free-provider-rankings`) backed by a testable lib helper reusing the in-process connection state (no Redis). Both filters default off, so the default view is unchanged; this supersedes the earlier client-side "Configured Only" toggle (#6245) with an available-only dimension and unit-tested logic. Regression guard: `tests/unit/freeProviderRankings-filters.test.ts`.
+
+### 🔧 Bug Fixes
+
+- **fix(dashboard):** adding a second API key connection for the same provider no longer silently overwrites the first — the Add-API-key modal now derives a unique default connection name (`main`, then `main-2`, `main-3`, …) so the backend name-based upsert can't collide (#6499 — thanks @dilneiss).
+- **fix(compression):** the session-dedup engine now also deduplicates a large multi-line block repeated **within a single message** (intra-message dedup), not just across turns; the compression-preview API surfaces a `fallbackReason`, and the fusion panel reports how many models were rate-limited vs failed on total-panel failure (#6501 — thanks @chirag127).
+- **fix(compression):** a stacked-pipeline step naming an unregistered engine now surfaces a `validationErrors` entry instead of silently no-op'ing, so misconfigured pipelines are visible in the preview API (#6506 — thanks @chirag127).
+- **feat(usage):** add a **Codex reset-credit redemption** flow to the Provider Limits UI ([#6361](https://github.com/diegosouzapw/OmniRoute/pull/6361)) — a `useCodexResetCreditRedemption` hook + `/api/usage/codex-reset-credit` route + `codexResetCredits` lib let you redeem banked Codex reset credits from the quota card. Regression guard: `tests/unit/codex-reset-credits.test.ts`. (thanks @JxnLexn)
+- **feat(glm):** add **team-plan quota settings** for `glm-cn` connections ([#6351](https://github.com/diegosouzapw/OmniRoute/pull/6351)) — a dedicated `GlmTeamQuotaFields` form section (team quota id / limits) threaded through the Add/Edit connection modals, persisted via `providerSpecificData`, with the GLM usage service reading the team quota. Regression guards: `tests/unit/glm-team-quota.test.ts`, `provider-specific-data-schema.test.ts`. (thanks @hao3039032)
+- **feat(providers):** add **TinyFish** web-fetch/search support ([#6349](https://github.com/diegosouzapw/OmniRoute/pull/6349)) — a `tinyfish-fetch` executor + `/v1/web/fetch` route + MCP web-fetch tool, registered as a specialty-media provider with request-validation and a search-provider catalog entry. Regression guards: `tests/unit/executor-tinyfish-fetch.test.ts`, `web-fetch-handler.test.ts`, `mcp-web-fetch-tool.test.ts`, `provider-validation-tinyfish.test.ts`. (thanks @dtybnrj)
+- **fix(cli):** `omniroute launch-codex` now spawns `codex.cmd` through a shell on Windows (the npm `.cmd` shim is unresolvable by bare `spawn` → ENOENT), mirroring the qodercli Windows fix (#6263) ([#6312](https://github.com/diegosouzapw/OmniRoute/pull/6312)). Regression guard: `tests/unit/launch-codex-windows-spawn-6312.test.ts`. (thanks @swingtempo)
+- **fix(codex):** isolate the **Spark** quota from the shared Codex quota and stabilize the quota UI ordering / hydration so per-scope limits render consistently ([#6336](https://github.com/diegosouzapw/OmniRoute/pull/6336)). Regression guards: `tests/unit/codex-quota-selection-hydration.test.ts`, `provider-limits-ui.test.ts` + 3 more. (thanks @xz-dev)
+- **feat(api):** add a `hidePaidModels` setting that filters paid-only models out of the `/v1/models` catalog. Regression guard: `tests/unit/models-catalog-hide-paid.test.ts`. (thanks @chirag127)
+- **fix(api-manager):** the fallback model picker now preserves combos instead of dropping them when a primary model is unavailable ([#6443](https://github.com/diegosouzapw/OmniRoute/pull/6443)). Regression guard: `tests/unit/api-manager-page-static.test.ts`. (thanks @jmengit)
+- **fix(providers):** recoverable Antigravity / Cloud-Code (Gemini Code Assist) `403` responses ([#6452](https://github.com/diegosouzapw/OmniRoute/pull/6452)) are now classified as a retryable project-config error instead of a terminal account ban, so a fixable project/API-disabled 403 no longer forces a ~1-year cooldown / full OAuth reconnect. Regression guard: `tests/unit/errorclassifier-antigravity-403.test.ts`. (thanks @developerjillur)
+- **fix(mitm):** `sanitizeHeaders` now redacts `Set-Cookie` response headers so upstream session cookies never leak into logs / diagnostics ([#6451](https://github.com/diegosouzapw/OmniRoute/pull/6451)). Regression guard: `tests/unit/mitm-sanitize-headers.test.ts`. (thanks @developerjillur)
+- **fix(api):** `/api/compression/preview` now accepts `mode: "caveman"` and correctly handles stacked / zero-compression previews ([#6425](https://github.com/diegosouzapw/OmniRoute/issues/6425)). Regression guard: `tests/unit/api/compression-preview-caveman-and-stacked-6425.test.ts`. (thanks @chirag127)
+- **feat(providers):** add **Zed** hosted LLM aggregator as a native-app provider ([#6118](https://github.com/diegosouzapw/OmniRoute/pull/6118)) — OAuth sign-in via the Zed hosted flow, registered through the shared provider registry + executor. Regression guards: `tests/unit/zed-oauth-provider.test.ts`, `zed-import-utils.test.ts`, `zed-docker-detect.test.ts`, `mitm-handler-zed.test.ts`. VPS-validated via live operator login (Hard Rule #18).
+- **fix(oauth):** the Kiro SSO-cache auto-import now **preserves the IDC region** — cross-region Amazon Q / Kiro profiles imported from the SSO cache are no longer collapsed to the default region ([#6113](https://github.com/diegosouzapw/OmniRoute/pull/6113)). Regression guard: `tests/unit/kiro-auto-import-idc-2059.test.ts`. VPS-validated via live operator login (Hard Rule #18).
+- **fix(dashboard):** passthrough model aliases no longer collide when two namespaced model ids share a last segment (port from 9router#1850, [#6431](https://github.com/diegosouzapw/OmniRoute/pull/6431)). `enx/gpt-5.5` and `enx/codebuddy/gpt-5.5` both auto-generated the alias `gpt-5.5`, so the second model could never be added (the UI just alerted "alias already exists"). Aliases are now disambiguated deterministically — bare last segment when free, then parent-qualified (`codebuddy-gpt-5.5`), then a numeric suffix — while re-adding the exact same model id is still blocked. Regression guard: `tests/unit/passthrough-alias-1850.test.ts`. (thanks @arpicato)
+- **fix(translator):** preserve a Gemini `functionResponse` co-located with other parts (another `functionCall`, or trailing `text`) in the same content when translating **Gemini → OpenAI** ([#6376](https://github.com/diegosouzapw/OmniRoute/pull/6376)). `convertGeminiContent()` early-returned the tool message on the first `functionResponse` part, dropping any co-located parts; such contents are now pre-split (one tool message per `functionResponse`, emitted first, plus one message for the remaining parts). Regression guard: `tests/unit/gemini-to-openai-function-response.test.ts`. (thanks @warelik)
+- **fix(headroom):** detect a python interpreter managed by **mise / pyenv / asdf / conda** (port from 9router#2353, [#6382](https://github.com/diegosouzapw/OmniRoute/pull/6382)). Headroom's python probe (`src/lib/headroom/detect.ts`) searched a hardcoded `PATH`, but version managers expose their interpreters via shim dirs that only join `PATH` through interactive-shell activation — which the non-interactive server never runs, so a managed python (≥3.10) was invisible and Headroom reported it missing. The search path now prepends the well-known shim/bin dirs (`~/.local/share/mise/shims`, `~/.pyenv/shims`, `~/.asdf/shims`, `$CONDA_PREFIX/bin`, `~/.local/bin`, respecting `MISE_DATA_DIR`/`PYENV_ROOT`/`ASDF_DATA_DIR` when set), and a new `HEADROOM_PYTHON` env override lets operators point straight at their interpreter (mirroring `HEADROOM_URL`). Still shell-free (`execFileSync`). Regression guard: `tests/unit/headroom-detect.test.ts` (5). (thanks @loopyd)
+- **fix(executors):** strip the OpenAI-Codex/Claude-CLI `client_metadata` passthrough field for **NVIDIA** requests (port from 9router#1887, [#6411](https://github.com/diegosouzapw/OmniRoute/pull/6411)). NVIDIA's OpenAI-compatible wrapper rejects it with `400 Unsupported parameter`, the same class already handled for `cerebras`/`mistral`; `nvidia` (executor `default`) was missing from the strip allowlist so Codex/Claude-Code passthrough requests 400'd. Regression guard: `tests/unit/executor-default-strip-client-metadata.test.ts` (+nvidia case). (thanks @phidinhmanh)
+- **fix(translator):** strip the Claude-style `thinking` field for **NVIDIA `z-ai/glm-5.2`** (port from 9router#2023, [#6413](https://github.com/diegosouzapw/OmniRoute/pull/6413)). NVIDIA's OpenAI-compatible wrapper 400s on `thinking` (a Claude-format client routed here leaves a `thinking:{type:"adaptive"}`); the existing strip rule only dropped `reasoning`. Same class already handled for `minimax-m2.7`. Regression guard: `tests/unit/nvidia-minimax-thinking-strip.test.ts` (+glm-5.2 case). (thanks @phidinhmanh)
+- **fix(translator):** suppress the streamed `</think>` close marker for the **Antigravity IDE** client (port from 9router#1061, [#6415](https://github.com/diegosouzapw/OmniRoute/pull/6415)). On thinking-only turns Antigravity rendered a bare `</think>` as the sole visible content, tripping its loop-detection and wasting requests. Antigravity's UA (`vscode/<v> (Antigravity/<v>)`) is added to the marker-suppress allowlist (alongside OpenCode); Claude Code / Cursor still get the marker, and `x-omniroute-thinking-marker: on` force-restores it. Regression guard: `tests/unit/think-close-marker-suppress-5245.test.ts`. (thanks @abdofallah)
+- **fix(executors):** strip nested `reasoning_content` from messages for **Mistral** (port from 9router#1649, [#6417](https://github.com/diegosouzapw/OmniRoute/pull/6417)). Mistral's API returns `422 extra_forbidden` when an assistant message carries `reasoning_content` (replayed thinking from a prior turn, e.g. via the Codex `/responses` path); the generic top-level 400 field-downgrade retry never covered the nested per-message field. `DefaultExecutor` now strips it for provider `mistral` only, so DeepSeek (which requires replayed `reasoning_content`) is unaffected. Regression guard: `tests/unit/mistral-strip-reasoning-content-1649.test.ts`. (thanks @xxy9468615)
+- **fix(executors):** strip the `client_metadata` passthrough field on the **OpenCode** path (port from 9router#1442, [#6418](https://github.com/diegosouzapw/OmniRoute/pull/6418)). OpenCode upstreams (e.g. `kimi-k2.6` via opencode-go) reject it with `400 "Extra inputs are not permitted, field: 'client_metadata'"`; the DefaultExecutor strip only covered cerebras/mistral and `OpencodeExecutor` extends `BaseExecutor` directly, so nothing removed it there. Regression guard: `tests/unit/opencode-strip-client-metadata-1442.test.ts`. (thanks @yanpaing007)
+- **fix(executors):** inject the `reasoning_content` echo for the native **Moonshot Kimi** provider (port from 9router#1480, [#6419](https://github.com/diegosouzapw/OmniRoute/pull/6419)). Kimi (executor `default`) is a thinking-mode upstream that 400s with "reasoning_content must be passed back" when a prior assistant turn lacks it; the placeholder injection was only wired into the OpenCode meta-provider, so direct multi-turn Kimi conversations failed. Scoped to `kimi` (gateway-served models matching the thinking-model name pattern are unaffected). Regression guard: `tests/unit/kimi-native-reasoning-injected-1480.test.ts`. (thanks @2220258345)
+- **fix(executors):** recover from a strict gateway's `context_management: Extra inputs are not permitted` 400 (port from 9router#1468, [#6420](https://github.com/diegosouzapw/OmniRoute/pull/6420)). **Claude Code** always sends a top-level `context_management` field; strict anthropic-compatible gateways reject it. The dedicated context-editing 400-fallback only fired when OmniRoute's own `contextEditing` feature was enabled (default off), so a client-sent field passed through untouched and 400'd. `context_management` is now in the generic reactive field-strip list, so it's stripped-and-retried once regardless of the feature flag (with correct request re-signing for claude-compatible relays). Regression guard: `tests/unit/provider-field-strips.test.ts`. (thanks @ohahe52-dot)
+- **fix(network):** enable **RFC 8305 Happy Eyeballs** (`autoSelectFamily`) on the direct-egress undici dispatcher (port from 9router#1237, [#6423](https://github.com/diegosouzapw/OmniRoute/pull/6423)). When DNS returns both IPv6 (AAAA) and IPv4 (A) and the IPv6 route is broken (e.g. a NAT64 `64:ff9b::` prefix without routing), undici tried IPv6 first and hung until `ETIMEDOUT` (then a 502 + account lockout), even though `curl` reached the same host. The direct dispatcher now races both families and uses whichever connects first. Proxy paths pin family via `proxyTls` and are unaffected. Regression guard: `tests/unit/direct-dispatcher-pipelining-4580.test.ts`. (thanks @adentdk)
+- **fix(combo):** round-robin now advances the rotation pointer past the model that **actually served**, not the eagerly-scheduled one (port from 9router#948, [#6428](https://github.com/diegosouzapw/OmniRoute/pull/6428)). With `stickyLimit: 1` (true round-robin), when the scheduled model failed and a _different_ model served via fallback, the counter had already advanced +1 from the scheduled index — so the next request reused the fallback-served model, degrading round-robin into hot-spotting on whichever model was healthy. The pointer now advances to the served index + 1 (mirroring the sticky-limit>1 path). Session-stickiness (#3825) and distribution are preserved. Regression guard: `tests/unit/combo-rr-fallback-advance-948.test.ts`. (thanks @binsarjr)
+- **fix(sse):** a non-string `model` field is now rejected with a `400` before the resolver, instead of crashing downstream `.toLowerCase()`/`.split()` calls into an empty-body `500` that escapes the error sanitizer ([#6407](https://github.com/diegosouzapw/OmniRoute/issues/6407)). Regression guard: `tests/unit/chat-non-string-model-6407.test.ts`. (thanks @chirag127)
+- **fix(api):** unknown `/api/*` routes now return a JSON `404` (instead of the dashboard HTML shell) and scalar chat params (`model`/`temperature`/etc.) are validated **before** the provider lookup so malformed requests fail fast with a clear `400` ([#6424](https://github.com/diegosouzapw/OmniRoute/issues/6424), [#6412](https://github.com/diegosouzapw/OmniRoute/issues/6412)). Regression guards: `tests/unit/api/api-catchall-json-404.test.ts`, `tests/unit/chat-early-schema-validation-6412.test.ts`. (thanks @chirag127)
+- **fix(api):** `/v1/chat/completions` now rejects a non-JSON `Content-Type` with a `400` before parsing the body ([#6414](https://github.com/diegosouzapw/OmniRoute/issues/6414)). Regression guard: `tests/unit/v1-chat-completions-content-type-6414.test.ts`. (thanks @chirag127)
+- **fix(api):** the `X-OmniRoute-Compression` response header is now echoed on `/v1/chat/completions` and `/v1/completions` ([#6422](https://github.com/diegosouzapw/OmniRoute/issues/6422)). Regression guard: `tests/unit/compression-header-echo-6422.test.ts`. (thanks @chirag127)
+- **fix(api):** concurrent `GET /v1/models` requests are coalesced into a single catalog build ([#6408](https://github.com/diegosouzapw/OmniRoute/issues/6408)). Regression guard: `tests/unit/v1-models-concurrent-6408.test.ts`. (thanks @chirag127)
+- **fix(api):** `/v1/completions` now echoes the requested `body.model` in its JSON + streamed responses ([#6429](https://github.com/diegosouzapw/OmniRoute/pull/6429)). Regression guard: `tests/unit/completions-body-model-echo.test.ts`. (thanks @chirag127)
+- **fix(api):** env-var master keys now see the full `/v1/models` catalog ([#6406](https://github.com/diegosouzapw/OmniRoute/issues/6406)). Regression guard: `tests/unit/models-catalog-envkey-6406.test.ts`. (thanks @chirag127)
+- **fix(api):** non-streaming `/v1/completions` responses now echo `body.model` aligned with the `X-OmniRoute-Model` header ([#6426](https://github.com/diegosouzapw/OmniRoute/issues/6426)). Regression guard: `tests/unit/v1-completions-model-header-match-6426.test.ts`. (thanks @chirag127)
+- **fix(api):** unknown `/v1/*` routes now return a JSON `404 not_found` instead of the Next.js dashboard HTML shell ([#6405](https://github.com/diegosouzapw/OmniRoute/issues/6405)). Regression guard: `tests/unit/api/v1-catchall-json-404.test.ts`. (thanks @chirag127)
+- **fix(api):** the per-connection provider models route now degrades to the shipped catalog when a provider's `/models` endpoint answers with a redirect ([#6267](https://github.com/diegosouzapw/OmniRoute/issues/6267)) — a `qwen-web` import failed with a raw `Redirect blocked … (307)` 503. `safeOutboundFetch` throws `REDIRECT_BLOCKED` on the 307, `getSafeOutboundFetchErrorStatus` maps it to 503, and `buildDiscoveryErrorFallbackResponse` treated every 503 as a hard error — so the non-empty `getModelsByProviderId("qwen-web")` catalog was never surfaced. A models-endpoint redirect is not a fixable-config error (unlike `URL_GUARD_BLOCKED`/`INVALID_URL`, which stay hard errors), so it now falls back to the local/cached catalog before the 503 short-circuit. General fix — covers any config-driven provider that 307s. Regression guard: `tests/unit/provider-models-qwen-web-redirect-6267.test.ts`. (thanks @chirag127)
+- **fix(api):** the per-connection provider models route (MCP `list_models_catalog` + the dashboard import view) now merges USER-ADDED custom models into its response ([#6247](https://github.com/diegosouzapw/OmniRoute/issues/6247)) — custom models live in the `key_value` namespace `customModels`, which the live REST `/api/v1/models` already merges, but `src/app/api/providers/[id]/models/route.ts` never read `getCustomModels`, so custom models were dropped on both the discovery-success and local_catalog paths. They are now folded into the returned model list (deduped by id, stamped `owned_by: provider`), fixing MCP + the dashboard import view in one place. Regression guard: `tests/unit/provider-models-custom-merge-6247.test.ts`. (thanks @RCrushMe)
+
+- **fix(providers):** GitLab Duo tool-calling follow-up turns no longer fail upstream with `422 {"detail":"Validation error"}` (tokens 0/0, rejected pre-inference). The #6234 tool-result-feedback fix serialized the **entire** multi-turn conversation into GitLab's single-file `code_suggestions` (`small_file`) generation endpoint — folded history that turn-N sent as an oversized `current_file.content_above_cursor` **and** duplicated verbatim into `user_instruction`, tripping the AI-Gateway's `small_file` validation guard. The executor now **bounds** that prompt: it keeps system + latest user message + the most-recent tool round (dropping older turns), caps oversized tool results, and stops duplicating the full prompt into `user_instruction` (which now carries only the short latest user message) — while still feeding the most-recent tool result back so the agent continues (`open-sse/executors/gitlab.ts`, [#6220](https://github.com/diegosouzapw/OmniRoute/issues/6220)). The unit test covers the bounding logic; the upstream 422→200 clearing is VPS-only (Hard Rule #18). Regression guard: `tests/unit/gitlab-tool-exchange-bounded-6220.test.ts`.
+
+- **fix(i18n):** the provider-detail (`/dashboard/providers/[id]`) connection-status filter labels no longer render as `__MISSING__:All` / `__MISSING__:Active` / `__MISSING__:Error` / `__MISSING__:Banned` / `__MISSING__:CreditsExhausted` in non-English locales (notably pt-BR) ([#6290](https://github.com/diegosouzapw/OmniRoute/issues/6290)). Root cause was **not** the namespace mismatch the issue guessed — the `providers.filter*` keys resolve correctly in `en.json`; the debt lived in the locale mirrors (`src/i18n/messages/*.json`), where these five keys carried the `__MISSING__:` sync sentinel in ~15 locales and were absent entirely in ~26 others, so next-intl found the key and echoed the sentinel verbatim. All 40 non-English/-Chinese mirrors now ship real translations for the five `providers.filter*` labels. Regression guard: `tests/unit/i18n-provider-filter-keys-6290.test.ts`. (thanks @diegosouzapw)
+
+- **fix(providers):** the `copilot-m365-web` streaming executor now emits `debug`-level WebSocket diagnostics ([#6210](https://github.com/diegosouzapw/OmniRoute/issues/6210)) — the outbound WS URL (with the `access_token` **redacted** via `redactWsUrl()`), handshake success/failure, and each received SignalR frame's `type`/`target`. Previously the streaming path logged nothing, so an empty `content:null` response (the M365 Education / Starter tier symptom fixed in #6234) was undiagnosable even at `APP_LOG_LEVEL=debug`. The change is debug-level and side-effect-free — it does not alter streaming behavior or the frame parser, and the token never reaches the logs. Regression guard: `tests/unit/copilot-m365-web-logging-6210.test.ts` (thanks @qpeyba)
+
+- **fix(resilience):** a round-robin combo no longer returns `503 all upstream accounts are unavailable` when a compatibility-rejected target is actually healthy ([#6238](https://github.com/diegosouzapw/OmniRoute/issues/6238)). `filterTargetsByRequestCompatibility` drops request-incompatible targets (tool/vision/structured-output unsupported, or below the required context window) **before** any availability check runs, and its `compatible.length === 0` safety net only fired when _all_ targets were filtered — not when the kept targets later all turned out runtime-unavailable (circuit-open / cooldown / no credentials). So a combo could 503 while a compat-rejected-but-healthy provider sat unused. `handleRoundRobinCombo` now keeps the compat-rejected set and, when every compat-kept target was skipped without a single real attempt, probes those rejected targets as a **last-resort fallback tier** (via the new pure `open-sse/services/combo/comboCompatFallback.ts`) before crystallizing the 503. Regression guard: `tests/unit/combo-roundrobin-compat-fallback-6238.test.ts`. (thanks @ThongAccount)
+
+- **fix(startup):** best-effort self-heal for a corrupted Turbopack dev cache on Windows ([#6289](https://github.com/diegosouzapw/OmniRoute/issues/6289)). On Windows, `pnpm dev` can fail at startup when Turbopack `mmap`s a persistent-cache SST file and the OS refuses the mapping (`os error 1455` — "paging file too small"), which Turbopack surfaces as a misleading `Module not found: Can't resolve '@/shared/utils/machine'`. This is a **known upstream Turbopack cache-corruption bug — not our code**. The dev launcher (`scripts/dev/run-next.mjs`) now wraps `nextApp.prepare()` and, when it rejects with that signature (`isTurbopackCacheCorruption` in the new `scripts/dev/turbopackCacheHeal.mjs`), purges `.build/next/**/cache/turbopack` and retries **once** with a clear log. **Caveat — best-effort only:** the corruption often surfaces as a runtime overlay rather than a `prepare()` rejection, so this cannot always intercept it; the reliable remedy remains manually deleting the Turbopack cache dir. Regression guard: `tests/unit/turbopack-cache-heal-6289.test.ts`. (thanks @chirag127)
+
+- **fix(providers):** qodercli PAT auth no longer fails with `spawn qodercli ENOENT` on Windows ([#6263](https://github.com/diegosouzapw/OmniRoute/issues/6263)) — `spawnQoderCli` spawned the bare `qodercli` name with `shell:false` and an unenriched env, so the npm `.cmd` wrapper under `%APPDATA%\npm` (a user-PATH directory) was never resolved. It now resolves the absolute `.cmd`/`.exe` path through the existing `getCliRuntimeStatus("qoder")` resolver in `src/shared/services/cliRuntime.ts` (memoized), spawns with `shell` when the target is a `.cmd`/`.bat`, and uses the cliRuntime-enriched env (PATH + PATHEXT + APPDATA); the ENOENT error now lists the searched paths plus the `CLI_QODER_BIN` override. End-to-end spawn on a real Windows host is host-only (Hard Rule #18); the path-resolution logic is unit-tested. Regression guard: `tests/unit/qodercli-windows-resolve-6263.test.ts`. (thanks @chirag127)
+
+- **fix(sse):** the reasoning-token buffer no longer inflates **probe-sized `max_tokens`** ([#6274](https://github.com/diegosouzapw/OmniRoute/issues/6274)) — Claude Code's `/model` capability check sends `max_tokens: 1`, but for a thinking-capable model with a large output cap (e.g. `glm-5.2`) the #3587 headroom heuristic (`max(current + 1000, ceil(current * 1.5))`) rewrote it to `1001` and forwarded that upstream, wasting tokens on a request that was never a genuine reasoning budget. `resolveReasoningBufferedMaxTokens()` (`open-sse/services/reasoningTokenBuffer.ts`) now short-circuits and returns the caller's value verbatim when it is below the new `REASONING_BUFFER_MIN_TRIGGER` (256) threshold — a tiny explicit limit is a probe, not a reasoning request. Real budgets still receive the #3587 headroom unchanged, and the guard runs after the existing capability checks so unknown / non-reasoning models keep returning `null`. Regression guard: `tests/unit/reasoning-token-buffer-6274.test.ts`. (thanks @brightfiscalband)
+
+- **fix(cli):** `omniroute reset-password` now works as a real subcommand, and password resets over piped (non-TTY) stdin actually apply ([#6261](https://github.com/diegosouzapw/OmniRoute/issues/6261), [#6258](https://github.com/diegosouzapw/OmniRoute/issues/6258)). Two coupled defects: (1) **#6261** — `bin/omniroute.mjs` routed everything through Commander with only two pre-Commander bypasses (`--mcp`, `reset-encrypted-columns`), so `omniroute reset-password` was rejected as an unknown command; only the separate `omniroute-reset-password` bin worked, while the docs falsely advertised the subcommand (incl. a bogus "legacy alias still works"). A pre-Commander bypass mirroring `reset-encrypted-columns` now dynamically imports `bin/reset-password.mjs` (which self-executes) before Commander parses; the three doc lines were corrected. (2) **#6258** — `bin/reset-password.mjs` issued two sequential `rl.question` prompts; under piped stdin the second read never settled at EOF, so `main()` never reached `resetManagementPassword` and the reset was a silent no-op (both prompts printed, no success, password unchanged). The CLI now detects non-TTY stdin and reads it once (first line = password, second line = confirm if present, else reused), adds a `--password-stdin` flag (entire stdin is the password, no confirmation), and exits `0` explicitly so the success line always flushes; interactive TTY behavior is unchanged. Regression guard: `tests/unit/reset-password-cli-6261-6258.test.ts` (3). (thanks @chirag127)
+
+- **fix(db):** the mass-migration **safety abort** now tells the operator how to bypass it and stops flooding the log ([#6260](https://github.com/diegosouzapw/OmniRoute/issues/6260)) — after restoring a backup that wiped the migration tracking table, `runMigrations()` threw the abort on every downstream `ensureDbInitialized()`, re-logging the full banner 11+ times, and the message never mentioned the existing `OMNIROUTE_MAX_PENDING_MIGRATIONS` escape hatch. The abort text now appends a bypass hint (set `OMNIROUTE_MAX_PENDING_MIGRATIONS=0` in `server.env` / `DATA_DIR/.env`), and a new `MigrationSafetyAbortError` is memoized so repeated calls in the same process throw the same instance and emit a single concise line instead of the full cascade. Regression guard: `tests/unit/migration-safety-abort-6260.test.ts`. (thanks @chirag127)
+
+- **fix(auth):** importing a **distinct** Codex/ChatGPT OAuth `auth.json` is no longer falsely rejected as "already exists" when it belongs to a different user in the same workspace ([#6301](https://github.com/diegosouzapw/OmniRoute/issues/6301)). `findExistingCodexConnection` (in `src/lib/oauth/utils/codexAuthImport.ts`) deduped **only** on `providerSpecificData.workspaceId === accountId`, where `accountId` is the shared `chatgpt_account_id`/`tokens.account_id` — so two members of the same ChatGPT Team collapsed onto a single connection (409 `duplicate_account`). The id_token's `https://api.openai.com/auth` claim carries a per-user `chatgpt_user_id` alongside the workspace id (the device-flow path already persisted it as `chatgptUserId`, but the import path did not). Now `parseAndValidateCodexAuth` extracts `userId` (`chatgpt_user_id` → `user_id` → JWT `sub`) into `ParsedCodexAuth`, the create/update paths persist `chatgptUserId` in `providerSpecificData` (mirroring `codex.ts`), and dedup keys on `workspaceId` **AND** `chatgptUserId` — with a backward-compat fallback to legacy accountId-only matching when no stored connection for that workspace records a `chatgptUserId`, so genuinely-same accounts still dedup. Regression guard: `tests/unit/codex-auth-import-userid-dedup-6301.test.ts` (4). (thanks @anungma)
+
+- **fix(providers):** importing models for the **venice-web** provider no longer fails with a red "Provider venice-web does not support models listing" ([#6269](https://github.com/diegosouzapw/OmniRoute/issues/6269)). `venice-web` is a web-cookie provider with an executor but no upstream `/v1/models` endpoint and no registry `models`, so the models route fell through to the tail `400`. Mirroring the `jules`/`linkup-search`/`ollama-search` fix (#5569), it now ships a static local catalog entry in `src/lib/providers/staticModels.ts` — seeding the current Venice lineup (`venice-uncensored`, `llama-3.3-70b`, `qwen3-235b`, `qwen3-4b`, `deepseek-r1-671b`; Venice rotates its catalog, see docs.venice.ai/models/overview) — so the route returns `200` with `source:"local_catalog"`, `intentional:true`. Regression guard: `tests/unit/static-models-venice-web-6269.test.ts`. (thanks @chirag127)
+
+- **fix(api):** the specialty model catalogs (`/v1/embeddings`, `/v1/images`, `/v1/music`, `/v1/videos` model lists) are now derived from the **unified catalog filtered by a predicate** (`getSpecialtyModelsResponse`) instead of ad-hoc per-route logic, so they consistently respect active-credential visibility and stay in sync with the main catalog ([#6303](https://github.com/diegosouzapw/OmniRoute/pull/6303)). Regression guard: `tests/unit/specialty-model-catalog-routes.test.ts`. (thanks @makcimbx)
+- **fix(api):** the agent-bridge server route now resolves the MITM manager via a **dynamic `import("@/mitm/manager.runtime")`** so Turbopack does not statically pull the stub (or over-bundle the manager), and the agent-skills generator anchors its output base path with `path.join(process.cwd(), …)` so Turbopack's static analyzer stops tracing the whole project root ([#6329](https://github.com/diegosouzapw/OmniRoute/issues/6329), [#6366](https://github.com/diegosouzapw/OmniRoute/pull/6366)). Regression guard: `tests/unit/agent-bridge-server-route-dynamic-import.test.ts`. (thanks @Iammilansoni)
+- **fix(api):** internal probes (combo-test, cloud-sync verify) now pick a **management-scoped / allow-all API key** instead of naively grabbing `getApiKeys()[0]` — a restricted `self:usage` first row made the probe fail with "Model X is not allowed for this API key" even when the combo path was healthy (`pickApiKeyForInternalUse` in `src/lib/db/apiKeys.ts`). The API-manager model editor also falls back to `/api/models?all=true` when `/v1/models` is catalog-protected ([#6372](https://github.com/diegosouzapw/OmniRoute/pull/6372)). Regression guard: `tests/unit/pick-internal-api-key-6372.test.ts`. (thanks @jmengit)
+- **fix(live-ws):** the Live Dashboard WebSocket server now **rejects on bind failure** (e.g. `EADDRINUSE` when the API bridge already holds the port) instead of letting the error surface as an unhandled `error` event that crash-loops the process — the `error` listener is attached to `wss` (not `server`) and releases the EventBus subscription on a failed start ([#6324](https://github.com/diegosouzapw/OmniRoute/issues/6324)). Regression guard: `tests/unit/live-ws-eaddrinuse-6324.test.ts`. (thanks @vinayakkulkarni)
+- **fix(dashboard):** the Home provider-topology widget now trusts the live provider-metrics snapshot — it uses `topology.errorProvider` and live `activeRequests` directly instead of re-deriving state from a stale `lastErrorAt` or applying a frontend timeout filter, so the topology reflects real-time provider health ([#6322](https://github.com/diegosouzapw/OmniRoute/pull/6322)). Regression guard: `tests/unit/home-provider-topology-live-state.test.ts`. (thanks @xz-dev)
+- **fix(sse):** strip zero-width markers from streamed **tool-call arguments** — a follow-up to [#5857](https://github.com/diegosouzapw/OmniRoute/pull/5857). That PR removed injected zero-width joiners (U+200D) from streamed assistant text/reasoning but deliberately left tool-call argument JSON byte-exact. The request-side obfuscation (`open-sse/services/claudeCodeObfuscation.ts`) injects ZWJ into agent words — including the temp path inside the Bash tool description — and Claude models copy that verbatim into generated commands, which are delivered as tool-call arguments rather than assistant text. As a result the ZWJ survived and corrupted code blocks (e.g. a temp path rendered with an invisible joiner). Now `open-sse/handlers/responseSanitizer.ts` strips zero-width code points from tool-call argument strings at every emit site (OpenAI non-stream/stream chat `tool_calls` + legacy `function_call`, native Responses `function_call` items, the OpenAI→Responses conversion, and the native Responses streaming `response.function_call_arguments.delta/.done` events). Only zero-width code points are removed; JSON structure and all other bytes stay identical (no parse/restringify), so normal arguments remain byte-exact. Regression guard: 6 new cases in `tests/unit/response-sanitizer.test.ts` (suite 50/50).
+
+- **fix(nodejs):** the default app log path now resolves under `DATA_DIR` (`~/.omniroute/logs/application/app.log`) instead of `process.cwd()` ([#6197](https://github.com/diegosouzapw/OmniRoute/issues/6197)) — the globally-installed CLI runs from an arbitrary working directory, so anchoring the default to cwd made file logging silently write to (or no-op under) an unrelated directory, contradicting the documented `.env.example` default. `getAppLogFilePath()` now computes the default lazily via the pure `resolveDataDir()` resolver (honours a per-process `DATA_DIR`, no directory-creation side effect); an explicit `APP_LOG_FILE_PATH` still wins. Regression guard: `tests/unit/logenv-datadir-path-6197.test.ts` (3).
+
+- **fix(docker):** AgentBridge/`startMitm` no longer aborts in containers/headless when the Antigravity-default DNS step can't write `/etc/hosts` ([#6127](https://github.com/diegosouzapw/OmniRoute/issues/6127)), and the privileged command's stderr now reaches `app.log` instead of only a bare exit code hitting the toast ([#6198](https://github.com/diegosouzapw/OmniRoute/issues/6198)). The default DNS step (`addDNSEntry`) was called unguarded while cert install and the two sibling DNS steps were each best-effort — in the runtime Docker image (`USER node`, no `sudo`, read-only `/etc/hosts`) it threw `Command failed with code 1` out of `startMitmInternal` and killed the whole start, discarding the stderr. The three DNS steps are extracted into a best-effort `provisionDnsEntries()` where each failure is logged with the full `err` (stderr included, folded in by `systemCommands.ts`) and never aborts the start. Regression guard: `tests/unit/mitm-dns-graceful-degrade-6127.test.ts` (4).
+
+- **fix(providers):** copilot-m365-web now supports the M365 Education "Starter / OfficeWebIncludedCopilot" tier and no longer returns an empty `content:null` stream ([#6210](https://github.com/diegosouzapw/OmniRoute/issues/6210)). Two gaps: (1) `buildWsUrl()` hardcoded the individual-consumer scenario (`OfficeWebPaidConsumerCopilot`, `isEdu=false`) — the EDU tier is now opt-in via `providerSpecificData.tier="edu"`, emitting `scenario=OfficeWebIncludedCopilot`/`isEdu=true` (the individual path is unchanged); (2) the EDU/GPT-5.5 path streams deltas via `arguments[0].writeAtCursor` (incremental) instead of only `messages[].text` (accumulated snapshots), which the parser dropped — a new `accumulateBotContent()` folds both formats, with `type:2 item.result.message` as a last-resort fallback. Regression guard: `tests/unit/copilot-m365-edu-writeatcursor-6210.test.ts` (10). (thanks @qpeyba)
+
+- **fix(providers):** GitLab Duo executor now feeds tool results back into the prompt instead of looping ([#6220](https://github.com/diegosouzapw/OmniRoute/issues/6220)) — `buildPrompt()` branched only on `system`/`user` and took `userParts.at(-1)`, silently dropping the `assistant{tool_calls}` + `tool{result}` turns the client appended, so the reconstructed prompt was byte-identical to turn 1 and the model re-emitted the same `<tool>` call forever. When a tool exchange is present the full conversation is now serialized, folding each tool result back keyed by its `tool_call_id`; simple conversations keep the legacy shape. Complements the tool_call emission from [#6051](https://github.com/diegosouzapw/OmniRoute/issues/6051) (the `kilo-duplicate` label was a false positive — different, sequential defect). Regression guard: `tests/unit/gitlab-tool-result-feedback-6220.test.ts` (4).
+
+- **fix(providers):** opencode-go/opencode-zen can now synthesize the OpenCode CLI identity headers Cloudflare requires on VPS egress ([#5997](https://github.com/diegosouzapw/OmniRoute/issues/5997)) — on a datacenter VPS, `opencode.ai/zen/go/v1/chat/completions` 403s (HTML challenge) requests lacking CLI identity, while the reporter's control curl proved that `User-Agent: opencode-cli/1.0.0` + `x-opencode-client: cli` + `x-opencode-project: default` + fresh request/session UUIDs succeed. Opt-in via `OPENCODE_SYNTHESIZE_CLI_HEADERS=true` (values overridable via `OPENCODE_GO_USER_AGENT`/`OPENCODE_USER_AGENT`/`OPENCODE_CLIENT`/`OPENCODE_PROJECT`); it fills only headers the client did not already send. Kept **off by default** — the forward-only path is deliberate (fabricating a wrong value risks upstream rejection; a prior dedup regressed with `opencode/local`), so this replaces the fragile local header-injection shim without changing default behavior. Regression guard: `tests/unit/opencode-cli-headers-synthesis-5997.test.ts` (6). (thanks @aleksesipenko)
+- fix(resilience): sticky session affinity now evicts and fails over to another account when the pinned account is exhausted/unavailable (#6219)
+
+- fix(sse): Responses API passthrough now drops internal commentary-phase output before forwarding to clients (gated by RESPONSES_PASSTHROUGH_DROP_COMMENTARY, default on) (#6199)
+
+- **fix(sse):** tool-call function schemas with a root `type: null` are now coerced to `type: "object"` before dispatch ([#6359](https://github.com/diegosouzapw/OmniRoute/issues/6359)) — clients like the Codex app emit `parameters: { type: null, ... }` for some tools, which OpenAI-compatible upstreams reject with `400 Invalid schema for function '...': schema must be a JSON Schema of 'type: "object"', got 'type: null'`, failing the whole request. `toolSchemaSanitizer` already stripped the null; it now re-adds the mandatory root `"object"` type (and empty `properties`/open `additionalProperties` when absent). Combinator roots (`anyOf`/`oneOf`/`allOf`) and explicit root types are left untouched. Regression guard: 5 new cases in `tests/unit/tool-schema-sanitizer.test.mjs`.
+
+- **fix(docker):** AgentBridge no longer fails to start on npm/Electron/VPS installs with "MITM manager stub reached at runtime" ([#6344](https://github.com/diegosouzapw/OmniRoute/issues/6344)) — v3.8.45 flipped the production bundler default to Turbopack, but `next.config.mjs` aliased `@/mitm/manager` to its Docker-only degraded stub **unconditionally**. That was harmless while Docker (which sets the alias intentionally for #3390 graceful degradation) was the sole Turbopack consumer, but once every artifact built with Turbopack the stub shipped to all non-Docker users and `startMitm` threw on the first Agent-Bridge start. The alias is now opt-in via `OMNIROUTE_MITM_STUB=1` (set only by the Dockerfile) through the shared `scripts/build/mitm-stub-flag.mjs` helper; default builds bundle the real manager. Regression guard: `tests/unit/mitm-stub-alias-6344.test.mjs` (4).
+
+- **fix(proxy):** stop the v3.8.44 proxy regression that leaked the real IP and disabled healthy proxies ([#6246](https://github.com/diegosouzapw/OmniRoute/issues/6246)). Two coupled defects from the new health scheduler: (1) **IP leak** — when a proxy assigned to a connection was marked `inactive`, resolution fell through to a **direct** egress instead of blocking, exposing the operator's real IP; (2) **over-deactivation** — the sweep flipped a proxy to `inactive` on the **first** failed probe and counted our own 5s timeout / a probe-target `5xx` as the proxy's fault, so healthy paid proxies vanished from egress selection ("my proxies are not being used anymore"). Fix: the sweep decision is extracted into a pure, network-free `decideProxyHealthAction` (`src/lib/proxyHealth/decision.ts`) — by default the health check now **only counts/logs and never downgrades status** (a proxy is downgraded/removed only with `PROXY_AUTO_REMOVE=true`, after `PROXY_AUTO_REMOVE_AFTER` **consecutive** conclusive failures); probes are classified tri-state so an inconclusive result (our timeout, or a `5xx` from the probe target) never penalizes the proxy, and the probe timeout is raised 5s→15s. Separately, `safeResolveProxy` now **fails closed** via the existing policy: a connection whose assigned proxy is dead is blocked instead of leaking direct (`hasBlockingProxyAssignment`), honoring the explicit `proxy off` toggles and the `PROXY_FAIL_OPEN=true` opt-out. Existing proxies stuck `inactive` by the old behavior need a one-time manual re-activate (the operator owns proxy status). Regression guards: `tests/unit/proxy-health-decide-action-6246.test.ts`, `tests/unit/proxy-assigned-unavailable-6246.test.ts`.
+
+- **fix(proxy):** make "Test All" read-only and add bulk enable/disable ([#6246](https://github.com/diegosouzapw/OmniRoute/issues/6246)). Complements the core fail-closed / scheduler fix (#6296) with the two remaining reporter asks. (1) The **"Test All" button** (`POST /api/settings/proxies/auto-test`) used to flip a proxy to `inactive` on a failed reachability probe; since the egress selector excludes `inactive` proxies, a flaky probe (an unreachable `httpbin.org`, a proxy that blocks `HEAD`, or a slow paid proxy) silently disabled every proxy that failed — "Test All" is now **read-only by default** (only the operator sets a proxy active/inactive; opt back into the legacy test-and-set with `PROXY_HEALTH_AUTO_DEACTIVATE=true`). (2) Adds a **bulk enable/disable** proxies endpoint + toolbar action (`POST /api/settings/proxies/batch-activate`) so an operator can re-activate proxies in one click. Regression guard: `tests/unit/proxy-health-6246.test.ts`. (thanks @tenshiak)
+
+- **chatcore (tools): stop the default 128-tool cap from silently dropping opencode's `task`/MCP tools.** opencode (used as an MCP/agent host) sends a large tool list; when it exceeds the speculative `MAX_TOOLS_LIMIT` (128) default, `truncateToolList` did a blind `tools.slice(0, 128)`, dropping every tool past index 128 — including opencode's built-in `task` tool (subagent launch) and many MCP tools, so models routed through OmniRoute could no longer spawn subagents or reach part of their tools. The cap exists to avoid upstream `400`s for providers with real hard limits (e.g. grok-cli 200), so it is kept for those: detection of the opencode client (`isOpencodeClient` — any `x-opencode-*` header, or `opencode` in the user-agent) now only bypasses the **speculative 128 default**, never a known provider ceiling. Precedence is explicit — a proactive/detected provider limit always truncates (even for opencode); otherwise opencode forwards its full tool list; otherwise the unchanged 128 default applies to every other client. Refactors `getEffectiveToolLimit` into `getKnownToolLimit(provider) ?? DEFAULT_LIMIT` (byte-identical for existing callers) and fixes a cosmetic debug-log that reported the truncated count instead of the original. Regression guard: `tests/unit/tool-limit-detector.test.ts`.
+
+- **fix(mitm):** the macOS MITM-cert install check now matches the system keychain again. `security find-certificate -a -Z` prints the SHA-1 as a colon-less hex string, but the installed-check compared it against `getCertFingerprint()`'s colon-separated form, so the substring match never hit — the cert was reported as not-installed and re-prompted for the sudo install on every run. Fingerprints are now normalized (colons stripped, upper-cased) on both sides via the extracted `macCertOutputHasFingerprint` helper. Regression guard: `tests/unit/mitm-cert-mac-fingerprint.test.ts`. ([#6204](https://github.com/diegosouzapw/OmniRoute/pull/6204), closes [#6134](https://github.com/diegosouzapw/OmniRoute/issues/6134) — thanks @rianonehub)
+
+- **fix(api):** `/v1/messages/count_tokens` now counts `tool_use`, `tool_result` and `thinking` content blocks (and array-form `system` prompts) in the local-estimation path, instead of only `text`. Real agentic conversations keep ~95% of their tokens inside tool results; the previous estimate returned near-zero for them, which silently broke Claude Code's auto-compaction (context grew past the window with no compaction until the upstream API rejected the request). The real provider-side count path is unchanged. Regression guard: `tests/unit/messages-count-tokens-route.test.ts`. ([#6221](https://github.com/diegosouzapw/OmniRoute/pull/6221) — thanks @luweiCN)
+
+- **fix(antigravity):** strip a trailing assistant prefill turn for Vertex Claude models to avoid upstream 400s ([#6114](https://github.com/diegosouzapw/OmniRoute/pull/6114)). Regression guard: `tests/unit/antigravity-claude-prefill-strip.test.ts`. (thanks @anki1kr)
+
+- **fix(security):** the mutable cloud-agent routes (`/api/cloud/credentials/update`, `/api/cloud/models/alias`) now require management auth instead of being treated as public. They were classified as public API routes, so a request without management credentials could update stored cloud-agent credentials and model aliases. They are removed from the public-route set, classified as management routes in the authz pipeline, and gated by `requireManagementAuth`; cloud **read**/auth routes stay public. Regression guards: `tests/unit/cloud-write-auth.test.ts`, `tests/unit/authz/classify.test.ts`, `tests/unit/public-api-routes.test.ts`. ([#6233](https://github.com/diegosouzapw/OmniRoute/pull/6233) — thanks @vittoroliveira-dev)
+
+- **refactor(dashboard):** extract the onboarding-wizard "Open provider details" link target into a pure, unit-tested `buildProviderDetailsHref(connection)` helper. The wizard already routes by `connection.id` (the node UUID) rather than the provider category slug (#6144/#6145); this hardens that behavior behind a tested helper that guards a missing id/connection. Regression guard: `tests/unit/provider-onboarding-href.test.ts`. ([#6166](https://github.com/diegosouzapw/OmniRoute/pull/6166) — thanks @KooshaPari)
+
+- **fix(security):** the doubao synthetic device-id generator now derives its digits via an unbiased crypto-random draw (rejection sampling over `crypto.randomBytes()`) instead of a `% 10` reduction, closing a CodeQL `js/biased-cryptographic-random` finding.
+- **fix(agentSkills):** the GitHub-skills generator now resolves `outputDir` to an absolute path before writing, fixing a regression introduced by #6366 (relative-to-cwd base path) that could write generated skill files to the wrong directory.
+- **fix(security):** `/api/keys/{id}/devices` now checks the HTTP method before auth/validation, returning a `405` for non-GET/DELETE verbs instead of a misleading `401`/`500` (closes a `dast-smoke` QUERY-method finding).
+- **fix(quality):** clear the last 2 heavy quality-gate reds on the release tip (cycle pre-flight).
+- **fix(mitm):** the test suite and CI can never mutate the OS trust store — `OMNIROUTE_SKIP_SYSTEM_TRUST=1` is set globally for tests/CI so `installCert`/`uninstallCert`/`installTproxyCa` skip the privileged OS dispatch ([#6310](https://github.com/diegosouzapw/OmniRoute/pull/6310); full detail is under the [3.8.45] section below — this branch received it via the parallel-cycle sync-back).
+- **fix(api):** `POST /api/github-skills` now Zod-validates its request body; documented the new quality-gate env vars and pinned the merge-integrity GitHub Actions to a commit SHA.
+- **fix(skills):** generate the missing `omni-github-skills` registry entry and align the agent-skills catalog-count tests (follow-up to #6186).
+- **fix(quality):** clear the cycle's 11 net-new ESLint errors and make `validate-release-green` suppressions-aware.
+
+### 📝 Maintenance
+
+- **i18n(it):** add 118 missing Italian (`it`) translations (net-additive — no existing keys dropped, valid JSON), improving Italian UI coverage. ([#6212](https://github.com/diegosouzapw/OmniRoute/pull/6212) — thanks @serverless83)
+- **chore(providers):** remove deprecated **MiMo V2** model entries from the catalogs (xiaomi-mimo, opencode-go, zenmux-free, audio TTS) — the upstream V2 line is superseded by MiMo V2.5; drops `mimo-v2-tts`, `mimo-v2-pro`, `mimo-v2-omni`, `mimo-v2-flash`, `mimo-v2-flash-free` and realigns the provider-catalog tests. ([#6248](https://github.com/diegosouzapw/OmniRoute/pull/6248) — thanks @backryun)
+- **chore(release):** ~50 commits on this branch are v3.8.45 pre-flight/hardening fixes and CI-perf work that landed here via the parallel-cycle sync-back (`sync-next-cycle.mjs`, Hard Rule #21) **after** the `v3.8.45` git tag was cut, and are already fully documented under the **[3.8.45]** section below — listed here only so the per-cycle commit-coverage check (`npm run release:uncovered`) doesn't flag them as gaps. Provider/catalog/UX/backend fixes: #6041, #6078, #6108, #6135, #6148, #6149, #6154, #6158, #6161, #6162, #6163, #6164, #6165, #6170, #6177, #6178, #6181, #6186, #6187, #6191, #6193, #6194, #6195, #6200, #6205, #6208, #6209, #6211, #6213, #6223, #6224, #6225, #6226, #6227, #6228, #6229, #6230, #6235, #6291, #6292. CI/release-pipeline work: #6167, #6203, #6214, #6215, #6218, #6273, #6275, #6283, #6284, #6285, #6300, #6305.
+- **chore(release):** additional zero-ref release-cycle plumbing on this branch, kept out of `release:uncovered` on purpose (no `#N` in the commit subject to cite): opening the v3.8.46 cycle, opening/closing the v3.8.45 cycle, the finalized [3.8.45] CHANGELOG i18n sync-back to 42 mirrors, the v3.8.45 cognitive/cyclomatic and file-size drift rebaselines, ESLint stale-suppression pruning (4,273 → 4,233), and clearing test-masking/docs-all pre-flight reds for v3.8.45.
+
+### ⚡ Performance & Infrastructure
+
+- **perf(release-green):** the pre-flight validator (`scripts/quality/validate-release-green.mjs`) now runs its 4 slow suites (unit / vitest / integration / pack-artifact) **concurrently** via `Promise.all` — pre-flight wall time drops from ~the sum of the suites to ~the slowest one (~30min saved per round; Phase 0 was the nº1 bottleneck of the v3.8.45 release benchmark, 2h54 of 6h34 e2e). Guard: `tests/unit/validate-release-green.test.ts` ("runs the slow suites CONCURRENTLY"). ([#6319](https://github.com/diegosouzapw/OmniRoute/pull/6319))
+- **fix(ci):** `scripts/release/sync-next-cycle.mjs` — two defects found live in its first production run (v3.8.45 Phase 5): (1) the `git()` helper's default 1 MiB `maxBuffer` crashed with `ENOBUFS` on `git show origin/main:CHANGELOG.md` (the CHANGELOG alone is >1 MiB) — widened to 64 MiB; (2) the i18n resync only propagated the `[NEXT]` (TBD) section, leaving the just-shipped finalized section as "— TBD" in all 42 mirrors — it now also syncs `[prevVersion]` bounded by the heading below it (new exported pure helper `versionAfter`). Guards: +5 tests in `tests/unit/sync-next-cycle.test.ts` (8/8). ([#6327](https://github.com/diegosouzapw/OmniRoute/pull/6327))
+- **test(ci):** concurrency-sensitive flaky tests are **quarantined into a serial pass** (`tests/unit/serial/`, `--test-concurrency=1`, appended to every unit runner incl. sharded variants — the serial pass is sharded too so concurrent shard jobs never self-collide). Initial set: `glm-coding-plan-monthly-3580`, `quota-division-blocks`, `provider-health-autopilot`, `combo-health-autopilot` — the class behind the ~28min CI wedges/re-runs (two live 1h+ wedges cancelled during this PR's own validation). Discovery + TIA gates track the new glob; systemic root cause (async logger writing after teardown) tracked in [#6360](https://github.com/diegosouzapw/OmniRoute/issues/6360). Guard: `tests/unit/test-serial-quarantine.test.ts` (4). ([#6347](https://github.com/diegosouzapw/OmniRoute/pull/6347))
+
+### 🙌 Contributors
+
+Thanks to everyone whose work landed in v3.8.46:
+
+| Contributor                                                    | PRs / Issues           |
+| -------------------------------------------------------------- | ---------------------- |
+| [@2220258345](https://github.com/2220258345)                   | direct commit / report |
+| [@abdofallah](https://github.com/abdofallah)                   | direct commit / report |
+| [@adentdk](https://github.com/adentdk)                         | direct commit / report |
+| [@aleksesipenko](https://github.com/aleksesipenko)             | direct commit / report |
+| [@anki1kr](https://github.com/anki1kr)                         | direct commit / report |
+| [@anungma](https://github.com/anungma)                         | direct commit / report |
+| [@arpicato](https://github.com/arpicato)                       | direct commit / report |
+| [@backryun](https://github.com/backryun)                       | #6248                  |
+| [@binsarjr](https://github.com/binsarjr)                       | direct commit / report |
+| [@brightfiscalband](https://github.com/brightfiscalband)       | direct commit / report |
+| [@chirag127](https://github.com/chirag127)                     | #6501, #6506           |
+| [@developerjillur](https://github.com/developerjillur)         | direct commit / report |
+| [@dilneiss](https://github.com/dilneiss)                       | #6499                  |
+| [@dtybnrj](https://github.com/dtybnrj)                         | direct commit / report |
+| [@Forcerecon](https://github.com/Forcerecon)                   | direct commit / report |
+| [@hao3039032](https://github.com/hao3039032)                   | direct commit / report |
+| [@Iammilansoni](https://github.com/Iammilansoni)               | #6150, #6245           |
+| [@jmengit](https://github.com/jmengit)                         | direct commit / report |
+| [@jordansilly77-stack](https://github.com/jordansilly77-stack) | direct commit / report |
+| [@JxnLexn](https://github.com/JxnLexn)                         | direct commit / report |
+| [@KooshaPari](https://github.com/KooshaPari)                   | #6166                  |
+| [@loopyd](https://github.com/loopyd)                           | direct commit / report |
+| [@luweiCN](https://github.com/luweiCN)                         | #6221                  |
+| [@makcimbx](https://github.com/makcimbx)                       | direct commit / report |
+| [@muflifadla38](https://github.com/muflifadla38)               | direct commit / report |
+| [@newnol](https://github.com/newnol)                           | direct commit / report |
+| [@ofekbetzalel](https://github.com/ofekbetzalel)               | direct commit / report |
+| [@ohahe52-dot](https://github.com/ohahe52-dot)                 | direct commit / report |
+| [@phidinhmanh](https://github.com/phidinhmanh)                 | direct commit / report |
+| [@powellnorma](https://github.com/powellnorma)                 | direct commit / report |
+| [@qpeyba](https://github.com/qpeyba)                           | direct commit / report |
+| [@RaviTharuma](https://github.com/RaviTharuma)                 | direct commit / report |
+| [@RCrushMe](https://github.com/RCrushMe)                       | direct commit / report |
+| [@rianonehub](https://github.com/rianonehub)                   | #6134, #6204           |
+| [@serverless83](https://github.com/serverless83)               | #6212                  |
+| [@swingtempo](https://github.com/swingtempo)                   | direct commit / report |
+| [@tenshiak](https://github.com/tenshiak)                       | direct commit / report |
+| [@ThongAccount](https://github.com/ThongAccount)               | direct commit / report |
+| [@UnrealAryan](https://github.com/UnrealAryan)                 | direct commit / report |
+| [@vinayakkulkarni](https://github.com/vinayakkulkarni)         | direct commit / report |
+| [@vittoroliveira-dev](https://github.com/vittoroliveira-dev)   | #6233                  |
+| [@warelik](https://github.com/warelik)                         | direct commit / report |
+| [@xxy9468615](https://github.com/xxy9468615)                   | direct commit / report |
+| [@xz-dev](https://github.com/xz-dev)                           | direct commit / report |
+| [@yanpaing007](https://github.com/yanpaing007)                 | direct commit / report |
+| [@diegosouzapw](https://github.com/diegosouzapw)               | maintainer             |
+
+---
+
 ## [3.8.45] — 2026-07-06
 
 ### ✨ New Features

docs/i18n/ms/CHANGELOG.md

@@ -6,6 +6,213 @@
 
 ## [3.8.31] — 2026-06-20
 
+## [3.8.46] — TBD
+
+### ✨ New Features
+
+- **feat(sse):** **hide paid-only models from `auto/*` routing** when `hidePaidModels` is on ([#6512](https://github.com/diegosouzapw/OmniRoute/issues/6512)) — follow-up to #6328/#6495. PR #6495 hid paid-only models from the `GET /v1/models` listing, but `auto/*` combos (`auto/best-coding`, `auto/glm`, …) could still pick a paid-only backend into their candidate pool → a 402/403 at request time. `createVirtualAutoCombo` now filters the candidate pool through the new pure `open-sse/services/autoCombo/paidModelFilter.ts` (`filterPaidOnlyCandidates`), applying the same free-model predicate #6495 uses in `catalog.ts` (`providerHasFreeModels(provider) && isFreeModel(provider, {id})`) whenever `settings.hidePaidModels === true`. Applied before the category/tier/family narrowing, so it covers every `auto/*` combo; an all-paid pool degrades to the existing graceful empty-pool path. **Opt-in — default OFF leaves the pool unchanged** (identity). Regression guard: `tests/unit/autoCombo/paid-model-filter-6512.test.ts` (4, incl. the default-off identity guard).
+- **feat(sse):** **provider-family auto combos** — `auto/glm`, `auto/minimax`, `auto/mimo`, `auto/zai`, `auto/gemma`, `auto/llama`, `auto/gemini` ([#6453](https://github.com/diegosouzapw/OmniRoute/issues/6453)) — new routable ids that materialize an on-demand virtual combo spanning whatever installed backends currently expose that model family, degrading gracefully as backends rotate. A new pure `open-sse/services/autoCombo/modelFamily.ts` (`detectModelFamily`) classifies by model-id prefix for six families; `zai` is instead resolved by provider id (z.ai's hosted API serves the same `glm-*` model ids as every other GLM backend, so `auto/zai` means "route to my z.ai backend specifically" vs `auto/glm`'s "any connected GLM backend"). Reuses the existing `createVirtualAutoCombo` on-demand materialization path (no DB writes) and the `/v1/models` catalog advertising loop. Regression guard: `tests/unit/autoCombo/provider-family-combos.test.ts` (11).
+- **feat(proxy):** native **proxy-pool round-robin / egress IP rotation** ([#6365](https://github.com/diegosouzapw/OmniRoute/issues/6365)) — a scope (global / provider / account) can now hold **multiple** proxies as a pool with a rotation strategy, so outbound requests cycle their egress IP instead of pinning one proxy per scope. Migration `117_proxy_pool_rotation.sql` lifts the `UNIQUE(scope, scope_id)` constraint (rebuild via the canonical rename/copy/drop; existing single assignments become 1-element pools) and adds a `proxy_scope_rotation` companion table holding the per-scope strategy + a persisted monotonic round-robin cursor. Strategies: `round-robin` (default, monotonic cursor — never `Math.random`), `random`, and `sticky-per-N-min`. Resolution (`resolveProxyForScopeFromRegistry` / `resolveProxyForConnectionFromRegistry`) now fetches the alive, position-ordered candidate set (unchanged `PROXY_ALIVE_PREDICATE`) and applies the strategy; an empty / all-dead pool still returns `null` — the #6246 fail-closed guard is **untouched** (never falls through to direct egress). Backend + DB only; dashboard pool-builder UI is a follow-up. Regression guard: `tests/unit/proxy-pool-rotation-6365.test.ts` (8, incl. fail-closed + backward-compat).
+- **feat(providers):** end-to-end **tool/function calling on the native Gemini `/v1beta` endpoint** ([#6222](https://github.com/diegosouzapw/OmniRoute/issues/6222)) — both directions of the Gemini↔OpenAI conversion now preserve tool data (previously silently dropped). Request side: `convertGeminiToInternal` (extracted to its own testable module) maps `tools[].functionDeclarations` → OpenAI `tools`, prior `functionCall` parts → assistant `tool_calls`, and `functionResponse` parts → `tool`-role messages. Response side: `convertOpenAIResponseToGemini` emits `parts[].functionCall {name,args}` from `message.tool_calls`, and the streaming `openAIChunkToGeminiChunk` accumulates fragmented `tool_calls` deltas by index into complete `functionCall` parts. The non-Gemini client paths (Claude, OpenAI-Responses) already preserved tool calls — this closes the gap specific to the native Gemini surface. Regression guard: `tests/unit/v1beta-gemini-tool-calling-6222.test.ts` (6, incl. a streaming SSE round-trip).
+- **feat(providers):** copilot-m365-web **enterprise / work tier** support ([#6334](https://github.com/diegosouzapw/OmniRoute/issues/6334)) — mirrors the EDU-tier pattern (#6210): `M365ConnectionParams` gains an `agent` field, a new opt-in `M365_ENTERPRISE_OVERRIDES` preset (`agent=work`, `scenario=officeweb`, `licenseType=Premium`) applies via `providerSpecificData.tier="enterprise"` (alias `"work"`), and `agent` is also overridable directly via `providerSpecificData.agent`. `buildWsUrl` was hardcoding `agent="web"` (the one enterprise-distinguishing param with no override path), so a Premium work account handshook then returned an empty stream. The individual and EDU paths are untouched. Kilo's dup flag vs #6210 (EDU tier) was a false positive — different tier. Regression guard: `tests/unit/copilot-m365-enterprise-6334.test.ts` (7). End-to-end confirmation on a real Premium work account is a live-VPS validation follow-up (Hard Rule #18). (thanks @Forcerecon)
+- **feat(api):** standardized, provider-agnostic **`effort` + `thinking` request params** ([#6241](https://github.com/diegosouzapw/OmniRoute/issues/6241)) — a thin standardization layer over the existing mature per-provider reasoning plumbing (no provider mapper touched). `providerChatCompletionSchema` gains a canonical `effort` (reusing the shared `none/low/medium/high/xhigh` vocabulary — the UI tiers `extra`/`max` collapse onto `xhigh`) and a boolean `thinking`. A pure `normalizeReasoningRequest` (wired once in `src/sse/handlers/chat.ts`, before any reasoning field is read) folds them onto the fields the translators already consume (`reasoning_effort` / `reasoning.effort` / `thinking`), so they fan out to Anthropic / Gemini / xAI / Responses — an explicit client `reasoning_effort` / object-shaped `thinking` always wins (backward-compatible). `/models` additively exposes `supportsThinking` + `effort_tiers` so the frontend can render the toggles (UI component is a follow-up). Regression guard: `tests/unit/effort-thinking-standardization-6241.test.ts` (12). (thanks @Iammilansoni, @shabeer)
+- **feat(combo):** new **`pipeline` (sequential) combo strategy** ([#6297](https://github.com/diegosouzapw/OmniRoute/issues/6297)) — the 18th routing strategy runs targets **in order**, threading each step's output into the next step's input, with an optional per-step `prompt` (system instruction); only the final step's response is returned. Distinct from `fusion` (parallel fan-out + judge). Implemented as a self-contained `open-sse/services/pipeline.ts` (sibling to `fusion.ts`), dispatched from `combo.ts`; the step list reuses `combo.models` order and reads an optional `prompt` off each target (backward-compatible — ignored by every other strategy). Intermediate steps run non-streaming with tools stripped (complete prose to thread forward); the final step keeps the client's `stream` flag + tools. A failing/empty/unparseable intermediate step fails the whole pipeline explicitly via a sanitized error (never silently swallowed). Kilo's dup flag vs #563 was a false positive (that's model→chain selection; this is a sequential chain). Regression guard: `tests/unit/combo-pipeline-strategy.test.ts` (5). (thanks @ofekbetzalel)
+- **feat(ci):** `check:test-masking` now flags **inline-reimplemented prod conditions** ([#6348](https://github.com/diegosouzapw/OmniRoute/issues/6348)) — a new **report-only** subcheck (v2, 6A.10 family) catches the wrong-shape contract test: a test that recomputes the condition under test inline instead of importing/exercising the real function (the #6216 class, where `=== 500` → `>= 500` stayed green because the test re-implemented the branch). For each added/modified test file it warns when the file textually duplicates a ≥3-token conditional from a production file touched in the same PR **and** does not import the symbol/module owning it, via a pure, fixture-tested `findReimplementedConditions()` with an allowlist mirroring `assertReductionAllowlist`. Report-only for now (does not fail the gate) — to be promoted to blocking after a triage cycle. Regression guard: `tests/unit/check-test-masking.test.ts` (45).
+- **feat(sse):** per-connection routing override (native vs CLIProxyAPI) ([#6339](https://github.com/diegosouzapw/OmniRoute/issues/6339)) — the previously-dead `isCliproxyapiDeepModeEnabled` helper is now wired into `resolveExecutorWithProxy`: a single connection can opt itself into the CLIProxyAPI passthrough executor via `providerSpecificData.cliproxyapiMode="claude-native"`, with precedence **connection override > provider `upstream_proxy_config` mode > default**. `resolveExecutorWithProxy` now receives the resolved connection's `providerSpecificData` (threaded from `chatCore.ts`), so one connection can deep-route while the provider's other connections stay native — no DB schema change (the toggle rides in `providerSpecificData`). Also resolves the same-provider mixing ask in #6340. Regression guard: `tests/unit/chatcore-executor-proxy.test.ts` (9). (thanks @RaviTharuma)
+- **feat(dashboard):** "Add session cookie" modal now shows a prominent **"Open ‹host› →"** link to the provider's own site ([#6268](https://github.com/diegosouzapw/OmniRoute/issues/6268)) — every `-web` cookie-session provider (chatgpt-web, claude-web, gemini-web, kimi-web, lmarena, qwen-web, m365-copilot-web, …) renders a one-click external link (opening the provider's login/home page in a new tab) so operators no longer tab away to retype the URL mid-setup. The host resolves from a pure, unit-tested `resolveWebProviderHost()` (prefers `WEB_COOKIE_PROVIDERS[id].website`, falls back to the registry `baseUrl` origin); non-web providers render exactly as before. Kilo's dup flag vs #6265 (modal-too-small-on-1080p) was a false positive — distinct concern. Regression guard: `tests/unit/resolve-web-provider-host.test.ts` (5). (thanks @chirag127)
+- **feat(providers):** add **DigitalOcean** AI (serverless inference) as an OpenAI-compatible API-key provider ([#6373](https://github.com/diegosouzapw/OmniRoute/pull/6373)) — base `https://inference.do-ai.run/v1`, wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/digitalocean/`, `src/shared/constants/providers/apikey/inference-hosts.ts`). Regression guard: `tests/unit/digitalocean-provider.test.ts`. (thanks @newnol)
+- **feat(providers):** add **Huancheng Public API** (`hcnsec`) as an OpenAI-compatible regional provider ([#6410](https://github.com/diegosouzapw/OmniRoute/pull/6410)) — Xinjiang Huancheng Cybersecurity's public LLM platform (base `https://api.hcnsec.cn/v1`, free credits via daily check-ins), wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/hcnsec/`, `src/shared/constants/providers/apikey/regional.ts`). Regression guard: `tests/unit/hcnsec-provider.test.ts`. (thanks @UnrealAryan)
+- **feat(dashboard):** the web-session credential guide now shows an **"Open {host}" link** ([#6316](https://github.com/diegosouzapw/OmniRoute/pull/6316)) to the provider's sign-in site (derived from the provider `website` via `getProviderWebsiteHost`), so you can jump straight to the page where the cookie/session must be captured. Regression guard: `tests/unit/web-session-provider-link-6316.test.ts`. (thanks @jordansilly77-stack)
+- **feat(cerebras):** add the **Gemma 4 31B** model (`gemma-4-31b`) to the Cerebras registry + pricing table ([#6331](https://github.com/diegosouzapw/OmniRoute/pull/6331)). Regression guard extends `tests/unit/t28-model-catalog-updates.test.ts`. (thanks @backryun)
+- **feat(providers):** add **Yuanbao (web)** as a cookie-session provider ([#6196](https://github.com/diegosouzapw/OmniRoute/issues/6196)) — `yuanbao-web` (Tencent Yuanbao, `yuanbao.tencent.com`) with cookie-only auth (`hy_user`/`hy_token` + public agent id), SSE→OpenAI translation incl. `reasoning_content`, exposing DeepSeek V3/R1 + Hunyuan / Hunyuan-T1. Regression guard: `tests/unit/providers-yuanbao-web.test.ts`. `together-web` was **deferred** (no verifiable web-session endpoint — needs a captured request) and `huggingchat-web` **dropped** (the existing `huggingchat` already is a web-cookie provider). (thanks @chirag127)
+- **feat(providers):** route the built-in **agentrouter** through the dynamic Claude-Code wire image ([#6056](https://github.com/diegosouzapw/OmniRoute/issues/6056)) — a small static allow-set (`CC_WIRE_IMAGE_BUILTINS` in `open-sse/services/ccWireImageBuiltins.ts`), consulted by `isClaudeCodeCompatible` / `isClaudeCodeCompatibleProvider` / `applyFingerprint`, makes agentrouter adopt the CC wire-image headers + fingerprint **while guarding the CC baseUrl/auth branches** so it keeps its own registry `baseUrl` and `x-api-key` auth. Regression guard: `tests/unit/agentrouter-cc-wire-image.test.ts` (asserts the wire image is applied AND agentrouter's baseUrl/auth are preserved). Live WAF-acceptance against agentrouter.org is a VPS validation follow-up (Hard Rule #18).
+- **feat(providers):** **bulk-add API keys for Cloudflare Workers AI** ([#6174](https://github.com/diegosouzapw/OmniRoute/issues/6174)) — `cloudflare-ai` is removed from the bulk-add exclusion list and the bulk parser gains a 3-field `name|accountId|apiKey` mode; the bulk route now builds a **per-entry** `providerSpecificData` so each key carries its own `accountId` (fixing the previous shared-object reuse), and both the create + key-validation paths receive it. Regression guard: `tests/unit/bulk-api-key-parser-cloudflare.test.ts`. (thanks @muflifadla38)
+- **feat(dashboard):** routing/settings UX clarity ([#6147](https://github.com/diegosouzapw/OmniRoute/issues/6147)) — (1) weighted combos show the **effective routing share %** next to each weight when weights don't sum to 100 (`WeightTotalBar.tsx`); (2) the status widget's user-facing **"Cloud Sync" label is renamed** to "Remote Settings Sync" (`CloudSyncStatus.tsx`; internal ids/state untouched); (3) built-in providers gain an **opt-in advanced base-URL override** (`isBaseUrlOverrideEligibleProvider`, hidden behind an "Advanced" toggle, reusing the existing `providerSpecificData.baseUrl` persistence — not globally widened). Regression guard: `tests/unit/routing-settings-ux-6147.test.ts`.
+- **feat(combo):** add an option to **disable session stickiness**, per-combo or globally — round-robin / random combos can rotate to a different connection on every request instead of pinning a whole conversation to one connection by its first-message hash. Resolution precedence per-combo `config.disableSessionStickiness` → global `settings.disableSessionStickiness` → default `false` (preserves the #3825 prompt-cache/504 fix); gates **both** stickiness call sites in `open-sse/services/combo.ts`. Exposed as a global toggle (Combo Defaults) and a per-combo Inherit/on/off control. ([#6168](https://github.com/diegosouzapw/OmniRoute/issues/6168)) Regression guard: `tests/unit/combo-disable-session-stickiness.test.ts`. (thanks @RCrushMe)
+- **feat(docker):** add the `OMNIROUTE_NO_SUDO` env flag for root-less / user-namespaced deployments — the MITM cert-trust command path (`resolveSudoSpawn` in `src/mitm/systemCommands.ts`) now strips the leading `sudo` when the flag is truthy, in addition to the existing root / sudo-missing cases, so the Proxy Agent runs without `sudo` (the operator trusts the CA manually, e.g. via `NODE_EXTRA_CA_CERTS`). Argv-array `spawn` preserved — no shell interpolation (Hard Rule #13). ([#6122](https://github.com/diegosouzapw/OmniRoute/issues/6122)) Regression guard: `tests/unit/mitm-systemCommands-no-sudo.test.ts`. (thanks @powellnorma)
+- **feat(providers):** add **Requesty** as an OpenAI-compatible gateway provider (BYOK, base `https://router.requesty.ai/v1`, ~200 free requests/day) — wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/requesty/`, `src/shared/constants/providers/apikey/gateways.ts`). ([#6120](https://github.com/diegosouzapw/OmniRoute/issues/6120)) Regression guard: `tests/unit/requesty-provider.test.ts`. (thanks @chirag127)
+- **feat(dashboard):** add **configured-only / available-only filters** to the Free Provider Rankings page ([#6150](https://github.com/diegosouzapw/OmniRoute/issues/6150)) — hide providers you haven't configured, or whose connections are all rate-limited / out of quota, via server-side query params (`?configuredOnly` / `?availableOnly` on `GET /api/free-provider-rankings`) backed by a testable lib helper reusing the in-process connection state (no Redis). Both filters default off, so the default view is unchanged; this supersedes the earlier client-side "Configured Only" toggle (#6245) with an available-only dimension and unit-tested logic. Regression guard: `tests/unit/freeProviderRankings-filters.test.ts`.
+
+### 🔧 Bug Fixes
+
+- **fix(dashboard):** adding a second API key connection for the same provider no longer silently overwrites the first — the Add-API-key modal now derives a unique default connection name (`main`, then `main-2`, `main-3`, …) so the backend name-based upsert can't collide (#6499 — thanks @dilneiss).
+- **fix(compression):** the session-dedup engine now also deduplicates a large multi-line block repeated **within a single message** (intra-message dedup), not just across turns; the compression-preview API surfaces a `fallbackReason`, and the fusion panel reports how many models were rate-limited vs failed on total-panel failure (#6501 — thanks @chirag127).
+- **fix(compression):** a stacked-pipeline step naming an unregistered engine now surfaces a `validationErrors` entry instead of silently no-op'ing, so misconfigured pipelines are visible in the preview API (#6506 — thanks @chirag127).
+- **feat(usage):** add a **Codex reset-credit redemption** flow to the Provider Limits UI ([#6361](https://github.com/diegosouzapw/OmniRoute/pull/6361)) — a `useCodexResetCreditRedemption` hook + `/api/usage/codex-reset-credit` route + `codexResetCredits` lib let you redeem banked Codex reset credits from the quota card. Regression guard: `tests/unit/codex-reset-credits.test.ts`. (thanks @JxnLexn)
+- **feat(glm):** add **team-plan quota settings** for `glm-cn` connections ([#6351](https://github.com/diegosouzapw/OmniRoute/pull/6351)) — a dedicated `GlmTeamQuotaFields` form section (team quota id / limits) threaded through the Add/Edit connection modals, persisted via `providerSpecificData`, with the GLM usage service reading the team quota. Regression guards: `tests/unit/glm-team-quota.test.ts`, `provider-specific-data-schema.test.ts`. (thanks @hao3039032)
+- **feat(providers):** add **TinyFish** web-fetch/search support ([#6349](https://github.com/diegosouzapw/OmniRoute/pull/6349)) — a `tinyfish-fetch` executor + `/v1/web/fetch` route + MCP web-fetch tool, registered as a specialty-media provider with request-validation and a search-provider catalog entry. Regression guards: `tests/unit/executor-tinyfish-fetch.test.ts`, `web-fetch-handler.test.ts`, `mcp-web-fetch-tool.test.ts`, `provider-validation-tinyfish.test.ts`. (thanks @dtybnrj)
+- **fix(cli):** `omniroute launch-codex` now spawns `codex.cmd` through a shell on Windows (the npm `.cmd` shim is unresolvable by bare `spawn` → ENOENT), mirroring the qodercli Windows fix (#6263) ([#6312](https://github.com/diegosouzapw/OmniRoute/pull/6312)). Regression guard: `tests/unit/launch-codex-windows-spawn-6312.test.ts`. (thanks @swingtempo)
+- **fix(codex):** isolate the **Spark** quota from the shared Codex quota and stabilize the quota UI ordering / hydration so per-scope limits render consistently ([#6336](https://github.com/diegosouzapw/OmniRoute/pull/6336)). Regression guards: `tests/unit/codex-quota-selection-hydration.test.ts`, `provider-limits-ui.test.ts` + 3 more. (thanks @xz-dev)
+- **feat(api):** add a `hidePaidModels` setting that filters paid-only models out of the `/v1/models` catalog. Regression guard: `tests/unit/models-catalog-hide-paid.test.ts`. (thanks @chirag127)
+- **fix(api-manager):** the fallback model picker now preserves combos instead of dropping them when a primary model is unavailable ([#6443](https://github.com/diegosouzapw/OmniRoute/pull/6443)). Regression guard: `tests/unit/api-manager-page-static.test.ts`. (thanks @jmengit)
+- **fix(providers):** recoverable Antigravity / Cloud-Code (Gemini Code Assist) `403` responses ([#6452](https://github.com/diegosouzapw/OmniRoute/pull/6452)) are now classified as a retryable project-config error instead of a terminal account ban, so a fixable project/API-disabled 403 no longer forces a ~1-year cooldown / full OAuth reconnect. Regression guard: `tests/unit/errorclassifier-antigravity-403.test.ts`. (thanks @developerjillur)
+- **fix(mitm):** `sanitizeHeaders` now redacts `Set-Cookie` response headers so upstream session cookies never leak into logs / diagnostics ([#6451](https://github.com/diegosouzapw/OmniRoute/pull/6451)). Regression guard: `tests/unit/mitm-sanitize-headers.test.ts`. (thanks @developerjillur)
+- **fix(api):** `/api/compression/preview` now accepts `mode: "caveman"` and correctly handles stacked / zero-compression previews ([#6425](https://github.com/diegosouzapw/OmniRoute/issues/6425)). Regression guard: `tests/unit/api/compression-preview-caveman-and-stacked-6425.test.ts`. (thanks @chirag127)
+- **feat(providers):** add **Zed** hosted LLM aggregator as a native-app provider ([#6118](https://github.com/diegosouzapw/OmniRoute/pull/6118)) — OAuth sign-in via the Zed hosted flow, registered through the shared provider registry + executor. Regression guards: `tests/unit/zed-oauth-provider.test.ts`, `zed-import-utils.test.ts`, `zed-docker-detect.test.ts`, `mitm-handler-zed.test.ts`. VPS-validated via live operator login (Hard Rule #18).
+- **fix(oauth):** the Kiro SSO-cache auto-import now **preserves the IDC region** — cross-region Amazon Q / Kiro profiles imported from the SSO cache are no longer collapsed to the default region ([#6113](https://github.com/diegosouzapw/OmniRoute/pull/6113)). Regression guard: `tests/unit/kiro-auto-import-idc-2059.test.ts`. VPS-validated via live operator login (Hard Rule #18).
+- **fix(dashboard):** passthrough model aliases no longer collide when two namespaced model ids share a last segment (port from 9router#1850, [#6431](https://github.com/diegosouzapw/OmniRoute/pull/6431)). `enx/gpt-5.5` and `enx/codebuddy/gpt-5.5` both auto-generated the alias `gpt-5.5`, so the second model could never be added (the UI just alerted "alias already exists"). Aliases are now disambiguated deterministically — bare last segment when free, then parent-qualified (`codebuddy-gpt-5.5`), then a numeric suffix — while re-adding the exact same model id is still blocked. Regression guard: `tests/unit/passthrough-alias-1850.test.ts`. (thanks @arpicato)
+- **fix(translator):** preserve a Gemini `functionResponse` co-located with other parts (another `functionCall`, or trailing `text`) in the same content when translating **Gemini → OpenAI** ([#6376](https://github.com/diegosouzapw/OmniRoute/pull/6376)). `convertGeminiContent()` early-returned the tool message on the first `functionResponse` part, dropping any co-located parts; such contents are now pre-split (one tool message per `functionResponse`, emitted first, plus one message for the remaining parts). Regression guard: `tests/unit/gemini-to-openai-function-response.test.ts`. (thanks @warelik)
+- **fix(headroom):** detect a python interpreter managed by **mise / pyenv / asdf / conda** (port from 9router#2353, [#6382](https://github.com/diegosouzapw/OmniRoute/pull/6382)). Headroom's python probe (`src/lib/headroom/detect.ts`) searched a hardcoded `PATH`, but version managers expose their interpreters via shim dirs that only join `PATH` through interactive-shell activation — which the non-interactive server never runs, so a managed python (≥3.10) was invisible and Headroom reported it missing. The search path now prepends the well-known shim/bin dirs (`~/.local/share/mise/shims`, `~/.pyenv/shims`, `~/.asdf/shims`, `$CONDA_PREFIX/bin`, `~/.local/bin`, respecting `MISE_DATA_DIR`/`PYENV_ROOT`/`ASDF_DATA_DIR` when set), and a new `HEADROOM_PYTHON` env override lets operators point straight at their interpreter (mirroring `HEADROOM_URL`). Still shell-free (`execFileSync`). Regression guard: `tests/unit/headroom-detect.test.ts` (5). (thanks @loopyd)
+- **fix(executors):** strip the OpenAI-Codex/Claude-CLI `client_metadata` passthrough field for **NVIDIA** requests (port from 9router#1887, [#6411](https://github.com/diegosouzapw/OmniRoute/pull/6411)). NVIDIA's OpenAI-compatible wrapper rejects it with `400 Unsupported parameter`, the same class already handled for `cerebras`/`mistral`; `nvidia` (executor `default`) was missing from the strip allowlist so Codex/Claude-Code passthrough requests 400'd. Regression guard: `tests/unit/executor-default-strip-client-metadata.test.ts` (+nvidia case). (thanks @phidinhmanh)
+- **fix(translator):** strip the Claude-style `thinking` field for **NVIDIA `z-ai/glm-5.2`** (port from 9router#2023, [#6413](https://github.com/diegosouzapw/OmniRoute/pull/6413)). NVIDIA's OpenAI-compatible wrapper 400s on `thinking` (a Claude-format client routed here leaves a `thinking:{type:"adaptive"}`); the existing strip rule only dropped `reasoning`. Same class already handled for `minimax-m2.7`. Regression guard: `tests/unit/nvidia-minimax-thinking-strip.test.ts` (+glm-5.2 case). (thanks @phidinhmanh)
+- **fix(translator):** suppress the streamed `</think>` close marker for the **Antigravity IDE** client (port from 9router#1061, [#6415](https://github.com/diegosouzapw/OmniRoute/pull/6415)). On thinking-only turns Antigravity rendered a bare `</think>` as the sole visible content, tripping its loop-detection and wasting requests. Antigravity's UA (`vscode/<v> (Antigravity/<v>)`) is added to the marker-suppress allowlist (alongside OpenCode); Claude Code / Cursor still get the marker, and `x-omniroute-thinking-marker: on` force-restores it. Regression guard: `tests/unit/think-close-marker-suppress-5245.test.ts`. (thanks @abdofallah)
+- **fix(executors):** strip nested `reasoning_content` from messages for **Mistral** (port from 9router#1649, [#6417](https://github.com/diegosouzapw/OmniRoute/pull/6417)). Mistral's API returns `422 extra_forbidden` when an assistant message carries `reasoning_content` (replayed thinking from a prior turn, e.g. via the Codex `/responses` path); the generic top-level 400 field-downgrade retry never covered the nested per-message field. `DefaultExecutor` now strips it for provider `mistral` only, so DeepSeek (which requires replayed `reasoning_content`) is unaffected. Regression guard: `tests/unit/mistral-strip-reasoning-content-1649.test.ts`. (thanks @xxy9468615)
+- **fix(executors):** strip the `client_metadata` passthrough field on the **OpenCode** path (port from 9router#1442, [#6418](https://github.com/diegosouzapw/OmniRoute/pull/6418)). OpenCode upstreams (e.g. `kimi-k2.6` via opencode-go) reject it with `400 "Extra inputs are not permitted, field: 'client_metadata'"`; the DefaultExecutor strip only covered cerebras/mistral and `OpencodeExecutor` extends `BaseExecutor` directly, so nothing removed it there. Regression guard: `tests/unit/opencode-strip-client-metadata-1442.test.ts`. (thanks @yanpaing007)
+- **fix(executors):** inject the `reasoning_content` echo for the native **Moonshot Kimi** provider (port from 9router#1480, [#6419](https://github.com/diegosouzapw/OmniRoute/pull/6419)). Kimi (executor `default`) is a thinking-mode upstream that 400s with "reasoning_content must be passed back" when a prior assistant turn lacks it; the placeholder injection was only wired into the OpenCode meta-provider, so direct multi-turn Kimi conversations failed. Scoped to `kimi` (gateway-served models matching the thinking-model name pattern are unaffected). Regression guard: `tests/unit/kimi-native-reasoning-injected-1480.test.ts`. (thanks @2220258345)
+- **fix(executors):** recover from a strict gateway's `context_management: Extra inputs are not permitted` 400 (port from 9router#1468, [#6420](https://github.com/diegosouzapw/OmniRoute/pull/6420)). **Claude Code** always sends a top-level `context_management` field; strict anthropic-compatible gateways reject it. The dedicated context-editing 400-fallback only fired when OmniRoute's own `contextEditing` feature was enabled (default off), so a client-sent field passed through untouched and 400'd. `context_management` is now in the generic reactive field-strip list, so it's stripped-and-retried once regardless of the feature flag (with correct request re-signing for claude-compatible relays). Regression guard: `tests/unit/provider-field-strips.test.ts`. (thanks @ohahe52-dot)
+- **fix(network):** enable **RFC 8305 Happy Eyeballs** (`autoSelectFamily`) on the direct-egress undici dispatcher (port from 9router#1237, [#6423](https://github.com/diegosouzapw/OmniRoute/pull/6423)). When DNS returns both IPv6 (AAAA) and IPv4 (A) and the IPv6 route is broken (e.g. a NAT64 `64:ff9b::` prefix without routing), undici tried IPv6 first and hung until `ETIMEDOUT` (then a 502 + account lockout), even though `curl` reached the same host. The direct dispatcher now races both families and uses whichever connects first. Proxy paths pin family via `proxyTls` and are unaffected. Regression guard: `tests/unit/direct-dispatcher-pipelining-4580.test.ts`. (thanks @adentdk)
+- **fix(combo):** round-robin now advances the rotation pointer past the model that **actually served**, not the eagerly-scheduled one (port from 9router#948, [#6428](https://github.com/diegosouzapw/OmniRoute/pull/6428)). With `stickyLimit: 1` (true round-robin), when the scheduled model failed and a _different_ model served via fallback, the counter had already advanced +1 from the scheduled index — so the next request reused the fallback-served model, degrading round-robin into hot-spotting on whichever model was healthy. The pointer now advances to the served index + 1 (mirroring the sticky-limit>1 path). Session-stickiness (#3825) and distribution are preserved. Regression guard: `tests/unit/combo-rr-fallback-advance-948.test.ts`. (thanks @binsarjr)
+- **fix(sse):** a non-string `model` field is now rejected with a `400` before the resolver, instead of crashing downstream `.toLowerCase()`/`.split()` calls into an empty-body `500` that escapes the error sanitizer ([#6407](https://github.com/diegosouzapw/OmniRoute/issues/6407)). Regression guard: `tests/unit/chat-non-string-model-6407.test.ts`. (thanks @chirag127)
+- **fix(api):** unknown `/api/*` routes now return a JSON `404` (instead of the dashboard HTML shell) and scalar chat params (`model`/`temperature`/etc.) are validated **before** the provider lookup so malformed requests fail fast with a clear `400` ([#6424](https://github.com/diegosouzapw/OmniRoute/issues/6424), [#6412](https://github.com/diegosouzapw/OmniRoute/issues/6412)). Regression guards: `tests/unit/api/api-catchall-json-404.test.ts`, `tests/unit/chat-early-schema-validation-6412.test.ts`. (thanks @chirag127)
+- **fix(api):** `/v1/chat/completions` now rejects a non-JSON `Content-Type` with a `400` before parsing the body ([#6414](https://github.com/diegosouzapw/OmniRoute/issues/6414)). Regression guard: `tests/unit/v1-chat-completions-content-type-6414.test.ts`. (thanks @chirag127)
+- **fix(api):** the `X-OmniRoute-Compression` response header is now echoed on `/v1/chat/completions` and `/v1/completions` ([#6422](https://github.com/diegosouzapw/OmniRoute/issues/6422)). Regression guard: `tests/unit/compression-header-echo-6422.test.ts`. (thanks @chirag127)
+- **fix(api):** concurrent `GET /v1/models` requests are coalesced into a single catalog build ([#6408](https://github.com/diegosouzapw/OmniRoute/issues/6408)). Regression guard: `tests/unit/v1-models-concurrent-6408.test.ts`. (thanks @chirag127)
+- **fix(api):** `/v1/completions` now echoes the requested `body.model` in its JSON + streamed responses ([#6429](https://github.com/diegosouzapw/OmniRoute/pull/6429)). Regression guard: `tests/unit/completions-body-model-echo.test.ts`. (thanks @chirag127)
+- **fix(api):** env-var master keys now see the full `/v1/models` catalog ([#6406](https://github.com/diegosouzapw/OmniRoute/issues/6406)). Regression guard: `tests/unit/models-catalog-envkey-6406.test.ts`. (thanks @chirag127)
+- **fix(api):** non-streaming `/v1/completions` responses now echo `body.model` aligned with the `X-OmniRoute-Model` header ([#6426](https://github.com/diegosouzapw/OmniRoute/issues/6426)). Regression guard: `tests/unit/v1-completions-model-header-match-6426.test.ts`. (thanks @chirag127)
+- **fix(api):** unknown `/v1/*` routes now return a JSON `404 not_found` instead of the Next.js dashboard HTML shell ([#6405](https://github.com/diegosouzapw/OmniRoute/issues/6405)). Regression guard: `tests/unit/api/v1-catchall-json-404.test.ts`. (thanks @chirag127)
+- **fix(api):** the per-connection provider models route now degrades to the shipped catalog when a provider's `/models` endpoint answers with a redirect ([#6267](https://github.com/diegosouzapw/OmniRoute/issues/6267)) — a `qwen-web` import failed with a raw `Redirect blocked … (307)` 503. `safeOutboundFetch` throws `REDIRECT_BLOCKED` on the 307, `getSafeOutboundFetchErrorStatus` maps it to 503, and `buildDiscoveryErrorFallbackResponse` treated every 503 as a hard error — so the non-empty `getModelsByProviderId("qwen-web")` catalog was never surfaced. A models-endpoint redirect is not a fixable-config error (unlike `URL_GUARD_BLOCKED`/`INVALID_URL`, which stay hard errors), so it now falls back to the local/cached catalog before the 503 short-circuit. General fix — covers any config-driven provider that 307s. Regression guard: `tests/unit/provider-models-qwen-web-redirect-6267.test.ts`. (thanks @chirag127)
+- **fix(api):** the per-connection provider models route (MCP `list_models_catalog` + the dashboard import view) now merges USER-ADDED custom models into its response ([#6247](https://github.com/diegosouzapw/OmniRoute/issues/6247)) — custom models live in the `key_value` namespace `customModels`, which the live REST `/api/v1/models` already merges, but `src/app/api/providers/[id]/models/route.ts` never read `getCustomModels`, so custom models were dropped on both the discovery-success and local_catalog paths. They are now folded into the returned model list (deduped by id, stamped `owned_by: provider`), fixing MCP + the dashboard import view in one place. Regression guard: `tests/unit/provider-models-custom-merge-6247.test.ts`. (thanks @RCrushMe)
+
+- **fix(providers):** GitLab Duo tool-calling follow-up turns no longer fail upstream with `422 {"detail":"Validation error"}` (tokens 0/0, rejected pre-inference). The #6234 tool-result-feedback fix serialized the **entire** multi-turn conversation into GitLab's single-file `code_suggestions` (`small_file`) generation endpoint — folded history that turn-N sent as an oversized `current_file.content_above_cursor` **and** duplicated verbatim into `user_instruction`, tripping the AI-Gateway's `small_file` validation guard. The executor now **bounds** that prompt: it keeps system + latest user message + the most-recent tool round (dropping older turns), caps oversized tool results, and stops duplicating the full prompt into `user_instruction` (which now carries only the short latest user message) — while still feeding the most-recent tool result back so the agent continues (`open-sse/executors/gitlab.ts`, [#6220](https://github.com/diegosouzapw/OmniRoute/issues/6220)). The unit test covers the bounding logic; the upstream 422→200 clearing is VPS-only (Hard Rule #18). Regression guard: `tests/unit/gitlab-tool-exchange-bounded-6220.test.ts`.
+
+- **fix(i18n):** the provider-detail (`/dashboard/providers/[id]`) connection-status filter labels no longer render as `__MISSING__:All` / `__MISSING__:Active` / `__MISSING__:Error` / `__MISSING__:Banned` / `__MISSING__:CreditsExhausted` in non-English locales (notably pt-BR) ([#6290](https://github.com/diegosouzapw/OmniRoute/issues/6290)). Root cause was **not** the namespace mismatch the issue guessed — the `providers.filter*` keys resolve correctly in `en.json`; the debt lived in the locale mirrors (`src/i18n/messages/*.json`), where these five keys carried the `__MISSING__:` sync sentinel in ~15 locales and were absent entirely in ~26 others, so next-intl found the key and echoed the sentinel verbatim. All 40 non-English/-Chinese mirrors now ship real translations for the five `providers.filter*` labels. Regression guard: `tests/unit/i18n-provider-filter-keys-6290.test.ts`. (thanks @diegosouzapw)
+
+- **fix(providers):** the `copilot-m365-web` streaming executor now emits `debug`-level WebSocket diagnostics ([#6210](https://github.com/diegosouzapw/OmniRoute/issues/6210)) — the outbound WS URL (with the `access_token` **redacted** via `redactWsUrl()`), handshake success/failure, and each received SignalR frame's `type`/`target`. Previously the streaming path logged nothing, so an empty `content:null` response (the M365 Education / Starter tier symptom fixed in #6234) was undiagnosable even at `APP_LOG_LEVEL=debug`. The change is debug-level and side-effect-free — it does not alter streaming behavior or the frame parser, and the token never reaches the logs. Regression guard: `tests/unit/copilot-m365-web-logging-6210.test.ts` (thanks @qpeyba)
+
+- **fix(resilience):** a round-robin combo no longer returns `503 all upstream accounts are unavailable` when a compatibility-rejected target is actually healthy ([#6238](https://github.com/diegosouzapw/OmniRoute/issues/6238)). `filterTargetsByRequestCompatibility` drops request-incompatible targets (tool/vision/structured-output unsupported, or below the required context window) **before** any availability check runs, and its `compatible.length === 0` safety net only fired when _all_ targets were filtered — not when the kept targets later all turned out runtime-unavailable (circuit-open / cooldown / no credentials). So a combo could 503 while a compat-rejected-but-healthy provider sat unused. `handleRoundRobinCombo` now keeps the compat-rejected set and, when every compat-kept target was skipped without a single real attempt, probes those rejected targets as a **last-resort fallback tier** (via the new pure `open-sse/services/combo/comboCompatFallback.ts`) before crystallizing the 503. Regression guard: `tests/unit/combo-roundrobin-compat-fallback-6238.test.ts`. (thanks @ThongAccount)
+
+- **fix(startup):** best-effort self-heal for a corrupted Turbopack dev cache on Windows ([#6289](https://github.com/diegosouzapw/OmniRoute/issues/6289)). On Windows, `pnpm dev` can fail at startup when Turbopack `mmap`s a persistent-cache SST file and the OS refuses the mapping (`os error 1455` — "paging file too small"), which Turbopack surfaces as a misleading `Module not found: Can't resolve '@/shared/utils/machine'`. This is a **known upstream Turbopack cache-corruption bug — not our code**. The dev launcher (`scripts/dev/run-next.mjs`) now wraps `nextApp.prepare()` and, when it rejects with that signature (`isTurbopackCacheCorruption` in the new `scripts/dev/turbopackCacheHeal.mjs`), purges `.build/next/**/cache/turbopack` and retries **once** with a clear log. **Caveat — best-effort only:** the corruption often surfaces as a runtime overlay rather than a `prepare()` rejection, so this cannot always intercept it; the reliable remedy remains manually deleting the Turbopack cache dir. Regression guard: `tests/unit/turbopack-cache-heal-6289.test.ts`. (thanks @chirag127)
+
+- **fix(providers):** qodercli PAT auth no longer fails with `spawn qodercli ENOENT` on Windows ([#6263](https://github.com/diegosouzapw/OmniRoute/issues/6263)) — `spawnQoderCli` spawned the bare `qodercli` name with `shell:false` and an unenriched env, so the npm `.cmd` wrapper under `%APPDATA%\npm` (a user-PATH directory) was never resolved. It now resolves the absolute `.cmd`/`.exe` path through the existing `getCliRuntimeStatus("qoder")` resolver in `src/shared/services/cliRuntime.ts` (memoized), spawns with `shell` when the target is a `.cmd`/`.bat`, and uses the cliRuntime-enriched env (PATH + PATHEXT + APPDATA); the ENOENT error now lists the searched paths plus the `CLI_QODER_BIN` override. End-to-end spawn on a real Windows host is host-only (Hard Rule #18); the path-resolution logic is unit-tested. Regression guard: `tests/unit/qodercli-windows-resolve-6263.test.ts`. (thanks @chirag127)
+
+- **fix(sse):** the reasoning-token buffer no longer inflates **probe-sized `max_tokens`** ([#6274](https://github.com/diegosouzapw/OmniRoute/issues/6274)) — Claude Code's `/model` capability check sends `max_tokens: 1`, but for a thinking-capable model with a large output cap (e.g. `glm-5.2`) the #3587 headroom heuristic (`max(current + 1000, ceil(current * 1.5))`) rewrote it to `1001` and forwarded that upstream, wasting tokens on a request that was never a genuine reasoning budget. `resolveReasoningBufferedMaxTokens()` (`open-sse/services/reasoningTokenBuffer.ts`) now short-circuits and returns the caller's value verbatim when it is below the new `REASONING_BUFFER_MIN_TRIGGER` (256) threshold — a tiny explicit limit is a probe, not a reasoning request. Real budgets still receive the #3587 headroom unchanged, and the guard runs after the existing capability checks so unknown / non-reasoning models keep returning `null`. Regression guard: `tests/unit/reasoning-token-buffer-6274.test.ts`. (thanks @brightfiscalband)
+
+- **fix(cli):** `omniroute reset-password` now works as a real subcommand, and password resets over piped (non-TTY) stdin actually apply ([#6261](https://github.com/diegosouzapw/OmniRoute/issues/6261), [#6258](https://github.com/diegosouzapw/OmniRoute/issues/6258)). Two coupled defects: (1) **#6261** — `bin/omniroute.mjs` routed everything through Commander with only two pre-Commander bypasses (`--mcp`, `reset-encrypted-columns`), so `omniroute reset-password` was rejected as an unknown command; only the separate `omniroute-reset-password` bin worked, while the docs falsely advertised the subcommand (incl. a bogus "legacy alias still works"). A pre-Commander bypass mirroring `reset-encrypted-columns` now dynamically imports `bin/reset-password.mjs` (which self-executes) before Commander parses; the three doc lines were corrected. (2) **#6258** — `bin/reset-password.mjs` issued two sequential `rl.question` prompts; under piped stdin the second read never settled at EOF, so `main()` never reached `resetManagementPassword` and the reset was a silent no-op (both prompts printed, no success, password unchanged). The CLI now detects non-TTY stdin and reads it once (first line = password, second line = confirm if present, else reused), adds a `--password-stdin` flag (entire stdin is the password, no confirmation), and exits `0` explicitly so the success line always flushes; interactive TTY behavior is unchanged. Regression guard: `tests/unit/reset-password-cli-6261-6258.test.ts` (3). (thanks @chirag127)
+
+- **fix(db):** the mass-migration **safety abort** now tells the operator how to bypass it and stops flooding the log ([#6260](https://github.com/diegosouzapw/OmniRoute/issues/6260)) — after restoring a backup that wiped the migration tracking table, `runMigrations()` threw the abort on every downstream `ensureDbInitialized()`, re-logging the full banner 11+ times, and the message never mentioned the existing `OMNIROUTE_MAX_PENDING_MIGRATIONS` escape hatch. The abort text now appends a bypass hint (set `OMNIROUTE_MAX_PENDING_MIGRATIONS=0` in `server.env` / `DATA_DIR/.env`), and a new `MigrationSafetyAbortError` is memoized so repeated calls in the same process throw the same instance and emit a single concise line instead of the full cascade. Regression guard: `tests/unit/migration-safety-abort-6260.test.ts`. (thanks @chirag127)
+
+- **fix(auth):** importing a **distinct** Codex/ChatGPT OAuth `auth.json` is no longer falsely rejected as "already exists" when it belongs to a different user in the same workspace ([#6301](https://github.com/diegosouzapw/OmniRoute/issues/6301)). `findExistingCodexConnection` (in `src/lib/oauth/utils/codexAuthImport.ts`) deduped **only** on `providerSpecificData.workspaceId === accountId`, where `accountId` is the shared `chatgpt_account_id`/`tokens.account_id` — so two members of the same ChatGPT Team collapsed onto a single connection (409 `duplicate_account`). The id_token's `https://api.openai.com/auth` claim carries a per-user `chatgpt_user_id` alongside the workspace id (the device-flow path already persisted it as `chatgptUserId`, but the import path did not). Now `parseAndValidateCodexAuth` extracts `userId` (`chatgpt_user_id` → `user_id` → JWT `sub`) into `ParsedCodexAuth`, the create/update paths persist `chatgptUserId` in `providerSpecificData` (mirroring `codex.ts`), and dedup keys on `workspaceId` **AND** `chatgptUserId` — with a backward-compat fallback to legacy accountId-only matching when no stored connection for that workspace records a `chatgptUserId`, so genuinely-same accounts still dedup. Regression guard: `tests/unit/codex-auth-import-userid-dedup-6301.test.ts` (4). (thanks @anungma)
+
+- **fix(providers):** importing models for the **venice-web** provider no longer fails with a red "Provider venice-web does not support models listing" ([#6269](https://github.com/diegosouzapw/OmniRoute/issues/6269)). `venice-web` is a web-cookie provider with an executor but no upstream `/v1/models` endpoint and no registry `models`, so the models route fell through to the tail `400`. Mirroring the `jules`/`linkup-search`/`ollama-search` fix (#5569), it now ships a static local catalog entry in `src/lib/providers/staticModels.ts` — seeding the current Venice lineup (`venice-uncensored`, `llama-3.3-70b`, `qwen3-235b`, `qwen3-4b`, `deepseek-r1-671b`; Venice rotates its catalog, see docs.venice.ai/models/overview) — so the route returns `200` with `source:"local_catalog"`, `intentional:true`. Regression guard: `tests/unit/static-models-venice-web-6269.test.ts`. (thanks @chirag127)
+
+- **fix(api):** the specialty model catalogs (`/v1/embeddings`, `/v1/images`, `/v1/music`, `/v1/videos` model lists) are now derived from the **unified catalog filtered by a predicate** (`getSpecialtyModelsResponse`) instead of ad-hoc per-route logic, so they consistently respect active-credential visibility and stay in sync with the main catalog ([#6303](https://github.com/diegosouzapw/OmniRoute/pull/6303)). Regression guard: `tests/unit/specialty-model-catalog-routes.test.ts`. (thanks @makcimbx)
+- **fix(api):** the agent-bridge server route now resolves the MITM manager via a **dynamic `import("@/mitm/manager.runtime")`** so Turbopack does not statically pull the stub (or over-bundle the manager), and the agent-skills generator anchors its output base path with `path.join(process.cwd(), …)` so Turbopack's static analyzer stops tracing the whole project root ([#6329](https://github.com/diegosouzapw/OmniRoute/issues/6329), [#6366](https://github.com/diegosouzapw/OmniRoute/pull/6366)). Regression guard: `tests/unit/agent-bridge-server-route-dynamic-import.test.ts`. (thanks @Iammilansoni)
+- **fix(api):** internal probes (combo-test, cloud-sync verify) now pick a **management-scoped / allow-all API key** instead of naively grabbing `getApiKeys()[0]` — a restricted `self:usage` first row made the probe fail with "Model X is not allowed for this API key" even when the combo path was healthy (`pickApiKeyForInternalUse` in `src/lib/db/apiKeys.ts`). The API-manager model editor also falls back to `/api/models?all=true` when `/v1/models` is catalog-protected ([#6372](https://github.com/diegosouzapw/OmniRoute/pull/6372)). Regression guard: `tests/unit/pick-internal-api-key-6372.test.ts`. (thanks @jmengit)
+- **fix(live-ws):** the Live Dashboard WebSocket server now **rejects on bind failure** (e.g. `EADDRINUSE` when the API bridge already holds the port) instead of letting the error surface as an unhandled `error` event that crash-loops the process — the `error` listener is attached to `wss` (not `server`) and releases the EventBus subscription on a failed start ([#6324](https://github.com/diegosouzapw/OmniRoute/issues/6324)). Regression guard: `tests/unit/live-ws-eaddrinuse-6324.test.ts`. (thanks @vinayakkulkarni)
+- **fix(dashboard):** the Home provider-topology widget now trusts the live provider-metrics snapshot — it uses `topology.errorProvider` and live `activeRequests` directly instead of re-deriving state from a stale `lastErrorAt` or applying a frontend timeout filter, so the topology reflects real-time provider health ([#6322](https://github.com/diegosouzapw/OmniRoute/pull/6322)). Regression guard: `tests/unit/home-provider-topology-live-state.test.ts`. (thanks @xz-dev)
+- **fix(sse):** strip zero-width markers from streamed **tool-call arguments** — a follow-up to [#5857](https://github.com/diegosouzapw/OmniRoute/pull/5857). That PR removed injected zero-width joiners (U+200D) from streamed assistant text/reasoning but deliberately left tool-call argument JSON byte-exact. The request-side obfuscation (`open-sse/services/claudeCodeObfuscation.ts`) injects ZWJ into agent words — including the temp path inside the Bash tool description — and Claude models copy that verbatim into generated commands, which are delivered as tool-call arguments rather than assistant text. As a result the ZWJ survived and corrupted code blocks (e.g. a temp path rendered with an invisible joiner). Now `open-sse/handlers/responseSanitizer.ts` strips zero-width code points from tool-call argument strings at every emit site (OpenAI non-stream/stream chat `tool_calls` + legacy `function_call`, native Responses `function_call` items, the OpenAI→Responses conversion, and the native Responses streaming `response.function_call_arguments.delta/.done` events). Only zero-width code points are removed; JSON structure and all other bytes stay identical (no parse/restringify), so normal arguments remain byte-exact. Regression guard: 6 new cases in `tests/unit/response-sanitizer.test.ts` (suite 50/50).
+
+- **fix(nodejs):** the default app log path now resolves under `DATA_DIR` (`~/.omniroute/logs/application/app.log`) instead of `process.cwd()` ([#6197](https://github.com/diegosouzapw/OmniRoute/issues/6197)) — the globally-installed CLI runs from an arbitrary working directory, so anchoring the default to cwd made file logging silently write to (or no-op under) an unrelated directory, contradicting the documented `.env.example` default. `getAppLogFilePath()` now computes the default lazily via the pure `resolveDataDir()` resolver (honours a per-process `DATA_DIR`, no directory-creation side effect); an explicit `APP_LOG_FILE_PATH` still wins. Regression guard: `tests/unit/logenv-datadir-path-6197.test.ts` (3).
+
+- **fix(docker):** AgentBridge/`startMitm` no longer aborts in containers/headless when the Antigravity-default DNS step can't write `/etc/hosts` ([#6127](https://github.com/diegosouzapw/OmniRoute/issues/6127)), and the privileged command's stderr now reaches `app.log` instead of only a bare exit code hitting the toast ([#6198](https://github.com/diegosouzapw/OmniRoute/issues/6198)). The default DNS step (`addDNSEntry`) was called unguarded while cert install and the two sibling DNS steps were each best-effort — in the runtime Docker image (`USER node`, no `sudo`, read-only `/etc/hosts`) it threw `Command failed with code 1` out of `startMitmInternal` and killed the whole start, discarding the stderr. The three DNS steps are extracted into a best-effort `provisionDnsEntries()` where each failure is logged with the full `err` (stderr included, folded in by `systemCommands.ts`) and never aborts the start. Regression guard: `tests/unit/mitm-dns-graceful-degrade-6127.test.ts` (4).
+
+- **fix(providers):** copilot-m365-web now supports the M365 Education "Starter / OfficeWebIncludedCopilot" tier and no longer returns an empty `content:null` stream ([#6210](https://github.com/diegosouzapw/OmniRoute/issues/6210)). Two gaps: (1) `buildWsUrl()` hardcoded the individual-consumer scenario (`OfficeWebPaidConsumerCopilot`, `isEdu=false`) — the EDU tier is now opt-in via `providerSpecificData.tier="edu"`, emitting `scenario=OfficeWebIncludedCopilot`/`isEdu=true` (the individual path is unchanged); (2) the EDU/GPT-5.5 path streams deltas via `arguments[0].writeAtCursor` (incremental) instead of only `messages[].text` (accumulated snapshots), which the parser dropped — a new `accumulateBotContent()` folds both formats, with `type:2 item.result.message` as a last-resort fallback. Regression guard: `tests/unit/copilot-m365-edu-writeatcursor-6210.test.ts` (10). (thanks @qpeyba)
+
+- **fix(providers):** GitLab Duo executor now feeds tool results back into the prompt instead of looping ([#6220](https://github.com/diegosouzapw/OmniRoute/issues/6220)) — `buildPrompt()` branched only on `system`/`user` and took `userParts.at(-1)`, silently dropping the `assistant{tool_calls}` + `tool{result}` turns the client appended, so the reconstructed prompt was byte-identical to turn 1 and the model re-emitted the same `<tool>` call forever. When a tool exchange is present the full conversation is now serialized, folding each tool result back keyed by its `tool_call_id`; simple conversations keep the legacy shape. Complements the tool_call emission from [#6051](https://github.com/diegosouzapw/OmniRoute/issues/6051) (the `kilo-duplicate` label was a false positive — different, sequential defect). Regression guard: `tests/unit/gitlab-tool-result-feedback-6220.test.ts` (4).
+
+- **fix(providers):** opencode-go/opencode-zen can now synthesize the OpenCode CLI identity headers Cloudflare requires on VPS egress ([#5997](https://github.com/diegosouzapw/OmniRoute/issues/5997)) — on a datacenter VPS, `opencode.ai/zen/go/v1/chat/completions` 403s (HTML challenge) requests lacking CLI identity, while the reporter's control curl proved that `User-Agent: opencode-cli/1.0.0` + `x-opencode-client: cli` + `x-opencode-project: default` + fresh request/session UUIDs succeed. Opt-in via `OPENCODE_SYNTHESIZE_CLI_HEADERS=true` (values overridable via `OPENCODE_GO_USER_AGENT`/`OPENCODE_USER_AGENT`/`OPENCODE_CLIENT`/`OPENCODE_PROJECT`); it fills only headers the client did not already send. Kept **off by default** — the forward-only path is deliberate (fabricating a wrong value risks upstream rejection; a prior dedup regressed with `opencode/local`), so this replaces the fragile local header-injection shim without changing default behavior. Regression guard: `tests/unit/opencode-cli-headers-synthesis-5997.test.ts` (6). (thanks @aleksesipenko)
+- fix(resilience): sticky session affinity now evicts and fails over to another account when the pinned account is exhausted/unavailable (#6219)
+
+- fix(sse): Responses API passthrough now drops internal commentary-phase output before forwarding to clients (gated by RESPONSES_PASSTHROUGH_DROP_COMMENTARY, default on) (#6199)
+
+- **fix(sse):** tool-call function schemas with a root `type: null` are now coerced to `type: "object"` before dispatch ([#6359](https://github.com/diegosouzapw/OmniRoute/issues/6359)) — clients like the Codex app emit `parameters: { type: null, ... }` for some tools, which OpenAI-compatible upstreams reject with `400 Invalid schema for function '...': schema must be a JSON Schema of 'type: "object"', got 'type: null'`, failing the whole request. `toolSchemaSanitizer` already stripped the null; it now re-adds the mandatory root `"object"` type (and empty `properties`/open `additionalProperties` when absent). Combinator roots (`anyOf`/`oneOf`/`allOf`) and explicit root types are left untouched. Regression guard: 5 new cases in `tests/unit/tool-schema-sanitizer.test.mjs`.
+
+- **fix(docker):** AgentBridge no longer fails to start on npm/Electron/VPS installs with "MITM manager stub reached at runtime" ([#6344](https://github.com/diegosouzapw/OmniRoute/issues/6344)) — v3.8.45 flipped the production bundler default to Turbopack, but `next.config.mjs` aliased `@/mitm/manager` to its Docker-only degraded stub **unconditionally**. That was harmless while Docker (which sets the alias intentionally for #3390 graceful degradation) was the sole Turbopack consumer, but once every artifact built with Turbopack the stub shipped to all non-Docker users and `startMitm` threw on the first Agent-Bridge start. The alias is now opt-in via `OMNIROUTE_MITM_STUB=1` (set only by the Dockerfile) through the shared `scripts/build/mitm-stub-flag.mjs` helper; default builds bundle the real manager. Regression guard: `tests/unit/mitm-stub-alias-6344.test.mjs` (4).
+
+- **fix(proxy):** stop the v3.8.44 proxy regression that leaked the real IP and disabled healthy proxies ([#6246](https://github.com/diegosouzapw/OmniRoute/issues/6246)). Two coupled defects from the new health scheduler: (1) **IP leak** — when a proxy assigned to a connection was marked `inactive`, resolution fell through to a **direct** egress instead of blocking, exposing the operator's real IP; (2) **over-deactivation** — the sweep flipped a proxy to `inactive` on the **first** failed probe and counted our own 5s timeout / a probe-target `5xx` as the proxy's fault, so healthy paid proxies vanished from egress selection ("my proxies are not being used anymore"). Fix: the sweep decision is extracted into a pure, network-free `decideProxyHealthAction` (`src/lib/proxyHealth/decision.ts`) — by default the health check now **only counts/logs and never downgrades status** (a proxy is downgraded/removed only with `PROXY_AUTO_REMOVE=true`, after `PROXY_AUTO_REMOVE_AFTER` **consecutive** conclusive failures); probes are classified tri-state so an inconclusive result (our timeout, or a `5xx` from the probe target) never penalizes the proxy, and the probe timeout is raised 5s→15s. Separately, `safeResolveProxy` now **fails closed** via the existing policy: a connection whose assigned proxy is dead is blocked instead of leaking direct (`hasBlockingProxyAssignment`), honoring the explicit `proxy off` toggles and the `PROXY_FAIL_OPEN=true` opt-out. Existing proxies stuck `inactive` by the old behavior need a one-time manual re-activate (the operator owns proxy status). Regression guards: `tests/unit/proxy-health-decide-action-6246.test.ts`, `tests/unit/proxy-assigned-unavailable-6246.test.ts`.
+
+- **fix(proxy):** make "Test All" read-only and add bulk enable/disable ([#6246](https://github.com/diegosouzapw/OmniRoute/issues/6246)). Complements the core fail-closed / scheduler fix (#6296) with the two remaining reporter asks. (1) The **"Test All" button** (`POST /api/settings/proxies/auto-test`) used to flip a proxy to `inactive` on a failed reachability probe; since the egress selector excludes `inactive` proxies, a flaky probe (an unreachable `httpbin.org`, a proxy that blocks `HEAD`, or a slow paid proxy) silently disabled every proxy that failed — "Test All" is now **read-only by default** (only the operator sets a proxy active/inactive; opt back into the legacy test-and-set with `PROXY_HEALTH_AUTO_DEACTIVATE=true`). (2) Adds a **bulk enable/disable** proxies endpoint + toolbar action (`POST /api/settings/proxies/batch-activate`) so an operator can re-activate proxies in one click. Regression guard: `tests/unit/proxy-health-6246.test.ts`. (thanks @tenshiak)
+
+- **chatcore (tools): stop the default 128-tool cap from silently dropping opencode's `task`/MCP tools.** opencode (used as an MCP/agent host) sends a large tool list; when it exceeds the speculative `MAX_TOOLS_LIMIT` (128) default, `truncateToolList` did a blind `tools.slice(0, 128)`, dropping every tool past index 128 — including opencode's built-in `task` tool (subagent launch) and many MCP tools, so models routed through OmniRoute could no longer spawn subagents or reach part of their tools. The cap exists to avoid upstream `400`s for providers with real hard limits (e.g. grok-cli 200), so it is kept for those: detection of the opencode client (`isOpencodeClient` — any `x-opencode-*` header, or `opencode` in the user-agent) now only bypasses the **speculative 128 default**, never a known provider ceiling. Precedence is explicit — a proactive/detected provider limit always truncates (even for opencode); otherwise opencode forwards its full tool list; otherwise the unchanged 128 default applies to every other client. Refactors `getEffectiveToolLimit` into `getKnownToolLimit(provider) ?? DEFAULT_LIMIT` (byte-identical for existing callers) and fixes a cosmetic debug-log that reported the truncated count instead of the original. Regression guard: `tests/unit/tool-limit-detector.test.ts`.
+
+- **fix(mitm):** the macOS MITM-cert install check now matches the system keychain again. `security find-certificate -a -Z` prints the SHA-1 as a colon-less hex string, but the installed-check compared it against `getCertFingerprint()`'s colon-separated form, so the substring match never hit — the cert was reported as not-installed and re-prompted for the sudo install on every run. Fingerprints are now normalized (colons stripped, upper-cased) on both sides via the extracted `macCertOutputHasFingerprint` helper. Regression guard: `tests/unit/mitm-cert-mac-fingerprint.test.ts`. ([#6204](https://github.com/diegosouzapw/OmniRoute/pull/6204), closes [#6134](https://github.com/diegosouzapw/OmniRoute/issues/6134) — thanks @rianonehub)
+
+- **fix(api):** `/v1/messages/count_tokens` now counts `tool_use`, `tool_result` and `thinking` content blocks (and array-form `system` prompts) in the local-estimation path, instead of only `text`. Real agentic conversations keep ~95% of their tokens inside tool results; the previous estimate returned near-zero for them, which silently broke Claude Code's auto-compaction (context grew past the window with no compaction until the upstream API rejected the request). The real provider-side count path is unchanged. Regression guard: `tests/unit/messages-count-tokens-route.test.ts`. ([#6221](https://github.com/diegosouzapw/OmniRoute/pull/6221) — thanks @luweiCN)
+
+- **fix(antigravity):** strip a trailing assistant prefill turn for Vertex Claude models to avoid upstream 400s ([#6114](https://github.com/diegosouzapw/OmniRoute/pull/6114)). Regression guard: `tests/unit/antigravity-claude-prefill-strip.test.ts`. (thanks @anki1kr)
+
+- **fix(security):** the mutable cloud-agent routes (`/api/cloud/credentials/update`, `/api/cloud/models/alias`) now require management auth instead of being treated as public. They were classified as public API routes, so a request without management credentials could update stored cloud-agent credentials and model aliases. They are removed from the public-route set, classified as management routes in the authz pipeline, and gated by `requireManagementAuth`; cloud **read**/auth routes stay public. Regression guards: `tests/unit/cloud-write-auth.test.ts`, `tests/unit/authz/classify.test.ts`, `tests/unit/public-api-routes.test.ts`. ([#6233](https://github.com/diegosouzapw/OmniRoute/pull/6233) — thanks @vittoroliveira-dev)
+
+- **refactor(dashboard):** extract the onboarding-wizard "Open provider details" link target into a pure, unit-tested `buildProviderDetailsHref(connection)` helper. The wizard already routes by `connection.id` (the node UUID) rather than the provider category slug (#6144/#6145); this hardens that behavior behind a tested helper that guards a missing id/connection. Regression guard: `tests/unit/provider-onboarding-href.test.ts`. ([#6166](https://github.com/diegosouzapw/OmniRoute/pull/6166) — thanks @KooshaPari)
+
+- **fix(security):** the doubao synthetic device-id generator now derives its digits via an unbiased crypto-random draw (rejection sampling over `crypto.randomBytes()`) instead of a `% 10` reduction, closing a CodeQL `js/biased-cryptographic-random` finding.
+- **fix(agentSkills):** the GitHub-skills generator now resolves `outputDir` to an absolute path before writing, fixing a regression introduced by #6366 (relative-to-cwd base path) that could write generated skill files to the wrong directory.
+- **fix(security):** `/api/keys/{id}/devices` now checks the HTTP method before auth/validation, returning a `405` for non-GET/DELETE verbs instead of a misleading `401`/`500` (closes a `dast-smoke` QUERY-method finding).
+- **fix(quality):** clear the last 2 heavy quality-gate reds on the release tip (cycle pre-flight).
+- **fix(mitm):** the test suite and CI can never mutate the OS trust store — `OMNIROUTE_SKIP_SYSTEM_TRUST=1` is set globally for tests/CI so `installCert`/`uninstallCert`/`installTproxyCa` skip the privileged OS dispatch ([#6310](https://github.com/diegosouzapw/OmniRoute/pull/6310); full detail is under the [3.8.45] section below — this branch received it via the parallel-cycle sync-back).
+- **fix(api):** `POST /api/github-skills` now Zod-validates its request body; documented the new quality-gate env vars and pinned the merge-integrity GitHub Actions to a commit SHA.
+- **fix(skills):** generate the missing `omni-github-skills` registry entry and align the agent-skills catalog-count tests (follow-up to #6186).
+- **fix(quality):** clear the cycle's 11 net-new ESLint errors and make `validate-release-green` suppressions-aware.
+
+### 📝 Maintenance
+
+- **i18n(it):** add 118 missing Italian (`it`) translations (net-additive — no existing keys dropped, valid JSON), improving Italian UI coverage. ([#6212](https://github.com/diegosouzapw/OmniRoute/pull/6212) — thanks @serverless83)
+- **chore(providers):** remove deprecated **MiMo V2** model entries from the catalogs (xiaomi-mimo, opencode-go, zenmux-free, audio TTS) — the upstream V2 line is superseded by MiMo V2.5; drops `mimo-v2-tts`, `mimo-v2-pro`, `mimo-v2-omni`, `mimo-v2-flash`, `mimo-v2-flash-free` and realigns the provider-catalog tests. ([#6248](https://github.com/diegosouzapw/OmniRoute/pull/6248) — thanks @backryun)
+- **chore(release):** ~50 commits on this branch are v3.8.45 pre-flight/hardening fixes and CI-perf work that landed here via the parallel-cycle sync-back (`sync-next-cycle.mjs`, Hard Rule #21) **after** the `v3.8.45` git tag was cut, and are already fully documented under the **[3.8.45]** section below — listed here only so the per-cycle commit-coverage check (`npm run release:uncovered`) doesn't flag them as gaps. Provider/catalog/UX/backend fixes: #6041, #6078, #6108, #6135, #6148, #6149, #6154, #6158, #6161, #6162, #6163, #6164, #6165, #6170, #6177, #6178, #6181, #6186, #6187, #6191, #6193, #6194, #6195, #6200, #6205, #6208, #6209, #6211, #6213, #6223, #6224, #6225, #6226, #6227, #6228, #6229, #6230, #6235, #6291, #6292. CI/release-pipeline work: #6167, #6203, #6214, #6215, #6218, #6273, #6275, #6283, #6284, #6285, #6300, #6305.
+- **chore(release):** additional zero-ref release-cycle plumbing on this branch, kept out of `release:uncovered` on purpose (no `#N` in the commit subject to cite): opening the v3.8.46 cycle, opening/closing the v3.8.45 cycle, the finalized [3.8.45] CHANGELOG i18n sync-back to 42 mirrors, the v3.8.45 cognitive/cyclomatic and file-size drift rebaselines, ESLint stale-suppression pruning (4,273 → 4,233), and clearing test-masking/docs-all pre-flight reds for v3.8.45.
+
+### ⚡ Performance & Infrastructure
+
+- **perf(release-green):** the pre-flight validator (`scripts/quality/validate-release-green.mjs`) now runs its 4 slow suites (unit / vitest / integration / pack-artifact) **concurrently** via `Promise.all` — pre-flight wall time drops from ~the sum of the suites to ~the slowest one (~30min saved per round; Phase 0 was the nº1 bottleneck of the v3.8.45 release benchmark, 2h54 of 6h34 e2e). Guard: `tests/unit/validate-release-green.test.ts` ("runs the slow suites CONCURRENTLY"). ([#6319](https://github.com/diegosouzapw/OmniRoute/pull/6319))
+- **fix(ci):** `scripts/release/sync-next-cycle.mjs` — two defects found live in its first production run (v3.8.45 Phase 5): (1) the `git()` helper's default 1 MiB `maxBuffer` crashed with `ENOBUFS` on `git show origin/main:CHANGELOG.md` (the CHANGELOG alone is >1 MiB) — widened to 64 MiB; (2) the i18n resync only propagated the `[NEXT]` (TBD) section, leaving the just-shipped finalized section as "— TBD" in all 42 mirrors — it now also syncs `[prevVersion]` bounded by the heading below it (new exported pure helper `versionAfter`). Guards: +5 tests in `tests/unit/sync-next-cycle.test.ts` (8/8). ([#6327](https://github.com/diegosouzapw/OmniRoute/pull/6327))
+- **test(ci):** concurrency-sensitive flaky tests are **quarantined into a serial pass** (`tests/unit/serial/`, `--test-concurrency=1`, appended to every unit runner incl. sharded variants — the serial pass is sharded too so concurrent shard jobs never self-collide). Initial set: `glm-coding-plan-monthly-3580`, `quota-division-blocks`, `provider-health-autopilot`, `combo-health-autopilot` — the class behind the ~28min CI wedges/re-runs (two live 1h+ wedges cancelled during this PR's own validation). Discovery + TIA gates track the new glob; systemic root cause (async logger writing after teardown) tracked in [#6360](https://github.com/diegosouzapw/OmniRoute/issues/6360). Guard: `tests/unit/test-serial-quarantine.test.ts` (4). ([#6347](https://github.com/diegosouzapw/OmniRoute/pull/6347))
+
+### 🙌 Contributors
+
+Thanks to everyone whose work landed in v3.8.46:
+
+| Contributor                                                    | PRs / Issues           |
+| -------------------------------------------------------------- | ---------------------- |
+| [@2220258345](https://github.com/2220258345)                   | direct commit / report |
+| [@abdofallah](https://github.com/abdofallah)                   | direct commit / report |
+| [@adentdk](https://github.com/adentdk)                         | direct commit / report |
+| [@aleksesipenko](https://github.com/aleksesipenko)             | direct commit / report |
+| [@anki1kr](https://github.com/anki1kr)                         | direct commit / report |
+| [@anungma](https://github.com/anungma)                         | direct commit / report |
+| [@arpicato](https://github.com/arpicato)                       | direct commit / report |
+| [@backryun](https://github.com/backryun)                       | #6248                  |
+| [@binsarjr](https://github.com/binsarjr)                       | direct commit / report |
+| [@brightfiscalband](https://github.com/brightfiscalband)       | direct commit / report |
+| [@chirag127](https://github.com/chirag127)                     | #6501, #6506           |
+| [@developerjillur](https://github.com/developerjillur)         | direct commit / report |
+| [@dilneiss](https://github.com/dilneiss)                       | #6499                  |
+| [@dtybnrj](https://github.com/dtybnrj)                         | direct commit / report |
+| [@Forcerecon](https://github.com/Forcerecon)                   | direct commit / report |
+| [@hao3039032](https://github.com/hao3039032)                   | direct commit / report |
+| [@Iammilansoni](https://github.com/Iammilansoni)               | #6150, #6245           |
+| [@jmengit](https://github.com/jmengit)                         | direct commit / report |
+| [@jordansilly77-stack](https://github.com/jordansilly77-stack) | direct commit / report |
+| [@JxnLexn](https://github.com/JxnLexn)                         | direct commit / report |
+| [@KooshaPari](https://github.com/KooshaPari)                   | #6166                  |
+| [@loopyd](https://github.com/loopyd)                           | direct commit / report |
+| [@luweiCN](https://github.com/luweiCN)                         | #6221                  |
+| [@makcimbx](https://github.com/makcimbx)                       | direct commit / report |
+| [@muflifadla38](https://github.com/muflifadla38)               | direct commit / report |
+| [@newnol](https://github.com/newnol)                           | direct commit / report |
+| [@ofekbetzalel](https://github.com/ofekbetzalel)               | direct commit / report |
+| [@ohahe52-dot](https://github.com/ohahe52-dot)                 | direct commit / report |
+| [@phidinhmanh](https://github.com/phidinhmanh)                 | direct commit / report |
+| [@powellnorma](https://github.com/powellnorma)                 | direct commit / report |
+| [@qpeyba](https://github.com/qpeyba)                           | direct commit / report |
+| [@RaviTharuma](https://github.com/RaviTharuma)                 | direct commit / report |
+| [@RCrushMe](https://github.com/RCrushMe)                       | direct commit / report |
+| [@rianonehub](https://github.com/rianonehub)                   | #6134, #6204           |
+| [@serverless83](https://github.com/serverless83)               | #6212                  |
+| [@swingtempo](https://github.com/swingtempo)                   | direct commit / report |
+| [@tenshiak](https://github.com/tenshiak)                       | direct commit / report |
+| [@ThongAccount](https://github.com/ThongAccount)               | direct commit / report |
+| [@UnrealAryan](https://github.com/UnrealAryan)                 | direct commit / report |
+| [@vinayakkulkarni](https://github.com/vinayakkulkarni)         | direct commit / report |
+| [@vittoroliveira-dev](https://github.com/vittoroliveira-dev)   | #6233                  |
+| [@warelik](https://github.com/warelik)                         | direct commit / report |
+| [@xxy9468615](https://github.com/xxy9468615)                   | direct commit / report |
+| [@xz-dev](https://github.com/xz-dev)                           | direct commit / report |
+| [@yanpaing007](https://github.com/yanpaing007)                 | direct commit / report |
+| [@diegosouzapw](https://github.com/diegosouzapw)               | maintainer             |
+
+---
+
 ## [3.8.45] — 2026-07-06
 
 ### ✨ New Features

docs/i18n/nl/CHANGELOG.md

@@ -6,6 +6,213 @@
 
 ## [3.8.31] — 2026-06-20
 
+## [3.8.46] — TBD
+
+### ✨ New Features
+
+- **feat(sse):** **hide paid-only models from `auto/*` routing** when `hidePaidModels` is on ([#6512](https://github.com/diegosouzapw/OmniRoute/issues/6512)) — follow-up to #6328/#6495. PR #6495 hid paid-only models from the `GET /v1/models` listing, but `auto/*` combos (`auto/best-coding`, `auto/glm`, …) could still pick a paid-only backend into their candidate pool → a 402/403 at request time. `createVirtualAutoCombo` now filters the candidate pool through the new pure `open-sse/services/autoCombo/paidModelFilter.ts` (`filterPaidOnlyCandidates`), applying the same free-model predicate #6495 uses in `catalog.ts` (`providerHasFreeModels(provider) && isFreeModel(provider, {id})`) whenever `settings.hidePaidModels === true`. Applied before the category/tier/family narrowing, so it covers every `auto/*` combo; an all-paid pool degrades to the existing graceful empty-pool path. **Opt-in — default OFF leaves the pool unchanged** (identity). Regression guard: `tests/unit/autoCombo/paid-model-filter-6512.test.ts` (4, incl. the default-off identity guard).
+- **feat(sse):** **provider-family auto combos** — `auto/glm`, `auto/minimax`, `auto/mimo`, `auto/zai`, `auto/gemma`, `auto/llama`, `auto/gemini` ([#6453](https://github.com/diegosouzapw/OmniRoute/issues/6453)) — new routable ids that materialize an on-demand virtual combo spanning whatever installed backends currently expose that model family, degrading gracefully as backends rotate. A new pure `open-sse/services/autoCombo/modelFamily.ts` (`detectModelFamily`) classifies by model-id prefix for six families; `zai` is instead resolved by provider id (z.ai's hosted API serves the same `glm-*` model ids as every other GLM backend, so `auto/zai` means "route to my z.ai backend specifically" vs `auto/glm`'s "any connected GLM backend"). Reuses the existing `createVirtualAutoCombo` on-demand materialization path (no DB writes) and the `/v1/models` catalog advertising loop. Regression guard: `tests/unit/autoCombo/provider-family-combos.test.ts` (11).
+- **feat(proxy):** native **proxy-pool round-robin / egress IP rotation** ([#6365](https://github.com/diegosouzapw/OmniRoute/issues/6365)) — a scope (global / provider / account) can now hold **multiple** proxies as a pool with a rotation strategy, so outbound requests cycle their egress IP instead of pinning one proxy per scope. Migration `117_proxy_pool_rotation.sql` lifts the `UNIQUE(scope, scope_id)` constraint (rebuild via the canonical rename/copy/drop; existing single assignments become 1-element pools) and adds a `proxy_scope_rotation` companion table holding the per-scope strategy + a persisted monotonic round-robin cursor. Strategies: `round-robin` (default, monotonic cursor — never `Math.random`), `random`, and `sticky-per-N-min`. Resolution (`resolveProxyForScopeFromRegistry` / `resolveProxyForConnectionFromRegistry`) now fetches the alive, position-ordered candidate set (unchanged `PROXY_ALIVE_PREDICATE`) and applies the strategy; an empty / all-dead pool still returns `null` — the #6246 fail-closed guard is **untouched** (never falls through to direct egress). Backend + DB only; dashboard pool-builder UI is a follow-up. Regression guard: `tests/unit/proxy-pool-rotation-6365.test.ts` (8, incl. fail-closed + backward-compat).
+- **feat(providers):** end-to-end **tool/function calling on the native Gemini `/v1beta` endpoint** ([#6222](https://github.com/diegosouzapw/OmniRoute/issues/6222)) — both directions of the Gemini↔OpenAI conversion now preserve tool data (previously silently dropped). Request side: `convertGeminiToInternal` (extracted to its own testable module) maps `tools[].functionDeclarations` → OpenAI `tools`, prior `functionCall` parts → assistant `tool_calls`, and `functionResponse` parts → `tool`-role messages. Response side: `convertOpenAIResponseToGemini` emits `parts[].functionCall {name,args}` from `message.tool_calls`, and the streaming `openAIChunkToGeminiChunk` accumulates fragmented `tool_calls` deltas by index into complete `functionCall` parts. The non-Gemini client paths (Claude, OpenAI-Responses) already preserved tool calls — this closes the gap specific to the native Gemini surface. Regression guard: `tests/unit/v1beta-gemini-tool-calling-6222.test.ts` (6, incl. a streaming SSE round-trip).
+- **feat(providers):** copilot-m365-web **enterprise / work tier** support ([#6334](https://github.com/diegosouzapw/OmniRoute/issues/6334)) — mirrors the EDU-tier pattern (#6210): `M365ConnectionParams` gains an `agent` field, a new opt-in `M365_ENTERPRISE_OVERRIDES` preset (`agent=work`, `scenario=officeweb`, `licenseType=Premium`) applies via `providerSpecificData.tier="enterprise"` (alias `"work"`), and `agent` is also overridable directly via `providerSpecificData.agent`. `buildWsUrl` was hardcoding `agent="web"` (the one enterprise-distinguishing param with no override path), so a Premium work account handshook then returned an empty stream. The individual and EDU paths are untouched. Kilo's dup flag vs #6210 (EDU tier) was a false positive — different tier. Regression guard: `tests/unit/copilot-m365-enterprise-6334.test.ts` (7). End-to-end confirmation on a real Premium work account is a live-VPS validation follow-up (Hard Rule #18). (thanks @Forcerecon)
+- **feat(api):** standardized, provider-agnostic **`effort` + `thinking` request params** ([#6241](https://github.com/diegosouzapw/OmniRoute/issues/6241)) — a thin standardization layer over the existing mature per-provider reasoning plumbing (no provider mapper touched). `providerChatCompletionSchema` gains a canonical `effort` (reusing the shared `none/low/medium/high/xhigh` vocabulary — the UI tiers `extra`/`max` collapse onto `xhigh`) and a boolean `thinking`. A pure `normalizeReasoningRequest` (wired once in `src/sse/handlers/chat.ts`, before any reasoning field is read) folds them onto the fields the translators already consume (`reasoning_effort` / `reasoning.effort` / `thinking`), so they fan out to Anthropic / Gemini / xAI / Responses — an explicit client `reasoning_effort` / object-shaped `thinking` always wins (backward-compatible). `/models` additively exposes `supportsThinking` + `effort_tiers` so the frontend can render the toggles (UI component is a follow-up). Regression guard: `tests/unit/effort-thinking-standardization-6241.test.ts` (12). (thanks @Iammilansoni, @shabeer)
+- **feat(combo):** new **`pipeline` (sequential) combo strategy** ([#6297](https://github.com/diegosouzapw/OmniRoute/issues/6297)) — the 18th routing strategy runs targets **in order**, threading each step's output into the next step's input, with an optional per-step `prompt` (system instruction); only the final step's response is returned. Distinct from `fusion` (parallel fan-out + judge). Implemented as a self-contained `open-sse/services/pipeline.ts` (sibling to `fusion.ts`), dispatched from `combo.ts`; the step list reuses `combo.models` order and reads an optional `prompt` off each target (backward-compatible — ignored by every other strategy). Intermediate steps run non-streaming with tools stripped (complete prose to thread forward); the final step keeps the client's `stream` flag + tools. A failing/empty/unparseable intermediate step fails the whole pipeline explicitly via a sanitized error (never silently swallowed). Kilo's dup flag vs #563 was a false positive (that's model→chain selection; this is a sequential chain). Regression guard: `tests/unit/combo-pipeline-strategy.test.ts` (5). (thanks @ofekbetzalel)
+- **feat(ci):** `check:test-masking` now flags **inline-reimplemented prod conditions** ([#6348](https://github.com/diegosouzapw/OmniRoute/issues/6348)) — a new **report-only** subcheck (v2, 6A.10 family) catches the wrong-shape contract test: a test that recomputes the condition under test inline instead of importing/exercising the real function (the #6216 class, where `=== 500` → `>= 500` stayed green because the test re-implemented the branch). For each added/modified test file it warns when the file textually duplicates a ≥3-token conditional from a production file touched in the same PR **and** does not import the symbol/module owning it, via a pure, fixture-tested `findReimplementedConditions()` with an allowlist mirroring `assertReductionAllowlist`. Report-only for now (does not fail the gate) — to be promoted to blocking after a triage cycle. Regression guard: `tests/unit/check-test-masking.test.ts` (45).
+- **feat(sse):** per-connection routing override (native vs CLIProxyAPI) ([#6339](https://github.com/diegosouzapw/OmniRoute/issues/6339)) — the previously-dead `isCliproxyapiDeepModeEnabled` helper is now wired into `resolveExecutorWithProxy`: a single connection can opt itself into the CLIProxyAPI passthrough executor via `providerSpecificData.cliproxyapiMode="claude-native"`, with precedence **connection override > provider `upstream_proxy_config` mode > default**. `resolveExecutorWithProxy` now receives the resolved connection's `providerSpecificData` (threaded from `chatCore.ts`), so one connection can deep-route while the provider's other connections stay native — no DB schema change (the toggle rides in `providerSpecificData`). Also resolves the same-provider mixing ask in #6340. Regression guard: `tests/unit/chatcore-executor-proxy.test.ts` (9). (thanks @RaviTharuma)
+- **feat(dashboard):** "Add session cookie" modal now shows a prominent **"Open ‹host› →"** link to the provider's own site ([#6268](https://github.com/diegosouzapw/OmniRoute/issues/6268)) — every `-web` cookie-session provider (chatgpt-web, claude-web, gemini-web, kimi-web, lmarena, qwen-web, m365-copilot-web, …) renders a one-click external link (opening the provider's login/home page in a new tab) so operators no longer tab away to retype the URL mid-setup. The host resolves from a pure, unit-tested `resolveWebProviderHost()` (prefers `WEB_COOKIE_PROVIDERS[id].website`, falls back to the registry `baseUrl` origin); non-web providers render exactly as before. Kilo's dup flag vs #6265 (modal-too-small-on-1080p) was a false positive — distinct concern. Regression guard: `tests/unit/resolve-web-provider-host.test.ts` (5). (thanks @chirag127)
+- **feat(providers):** add **DigitalOcean** AI (serverless inference) as an OpenAI-compatible API-key provider ([#6373](https://github.com/diegosouzapw/OmniRoute/pull/6373)) — base `https://inference.do-ai.run/v1`, wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/digitalocean/`, `src/shared/constants/providers/apikey/inference-hosts.ts`). Regression guard: `tests/unit/digitalocean-provider.test.ts`. (thanks @newnol)
+- **feat(providers):** add **Huancheng Public API** (`hcnsec`) as an OpenAI-compatible regional provider ([#6410](https://github.com/diegosouzapw/OmniRoute/pull/6410)) — Xinjiang Huancheng Cybersecurity's public LLM platform (base `https://api.hcnsec.cn/v1`, free credits via daily check-ins), wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/hcnsec/`, `src/shared/constants/providers/apikey/regional.ts`). Regression guard: `tests/unit/hcnsec-provider.test.ts`. (thanks @UnrealAryan)
+- **feat(dashboard):** the web-session credential guide now shows an **"Open {host}" link** ([#6316](https://github.com/diegosouzapw/OmniRoute/pull/6316)) to the provider's sign-in site (derived from the provider `website` via `getProviderWebsiteHost`), so you can jump straight to the page where the cookie/session must be captured. Regression guard: `tests/unit/web-session-provider-link-6316.test.ts`. (thanks @jordansilly77-stack)
+- **feat(cerebras):** add the **Gemma 4 31B** model (`gemma-4-31b`) to the Cerebras registry + pricing table ([#6331](https://github.com/diegosouzapw/OmniRoute/pull/6331)). Regression guard extends `tests/unit/t28-model-catalog-updates.test.ts`. (thanks @backryun)
+- **feat(providers):** add **Yuanbao (web)** as a cookie-session provider ([#6196](https://github.com/diegosouzapw/OmniRoute/issues/6196)) — `yuanbao-web` (Tencent Yuanbao, `yuanbao.tencent.com`) with cookie-only auth (`hy_user`/`hy_token` + public agent id), SSE→OpenAI translation incl. `reasoning_content`, exposing DeepSeek V3/R1 + Hunyuan / Hunyuan-T1. Regression guard: `tests/unit/providers-yuanbao-web.test.ts`. `together-web` was **deferred** (no verifiable web-session endpoint — needs a captured request) and `huggingchat-web` **dropped** (the existing `huggingchat` already is a web-cookie provider). (thanks @chirag127)
+- **feat(providers):** route the built-in **agentrouter** through the dynamic Claude-Code wire image ([#6056](https://github.com/diegosouzapw/OmniRoute/issues/6056)) — a small static allow-set (`CC_WIRE_IMAGE_BUILTINS` in `open-sse/services/ccWireImageBuiltins.ts`), consulted by `isClaudeCodeCompatible` / `isClaudeCodeCompatibleProvider` / `applyFingerprint`, makes agentrouter adopt the CC wire-image headers + fingerprint **while guarding the CC baseUrl/auth branches** so it keeps its own registry `baseUrl` and `x-api-key` auth. Regression guard: `tests/unit/agentrouter-cc-wire-image.test.ts` (asserts the wire image is applied AND agentrouter's baseUrl/auth are preserved). Live WAF-acceptance against agentrouter.org is a VPS validation follow-up (Hard Rule #18).
+- **feat(providers):** **bulk-add API keys for Cloudflare Workers AI** ([#6174](https://github.com/diegosouzapw/OmniRoute/issues/6174)) — `cloudflare-ai` is removed from the bulk-add exclusion list and the bulk parser gains a 3-field `name|accountId|apiKey` mode; the bulk route now builds a **per-entry** `providerSpecificData` so each key carries its own `accountId` (fixing the previous shared-object reuse), and both the create + key-validation paths receive it. Regression guard: `tests/unit/bulk-api-key-parser-cloudflare.test.ts`. (thanks @muflifadla38)
+- **feat(dashboard):** routing/settings UX clarity ([#6147](https://github.com/diegosouzapw/OmniRoute/issues/6147)) — (1) weighted combos show the **effective routing share %** next to each weight when weights don't sum to 100 (`WeightTotalBar.tsx`); (2) the status widget's user-facing **"Cloud Sync" label is renamed** to "Remote Settings Sync" (`CloudSyncStatus.tsx`; internal ids/state untouched); (3) built-in providers gain an **opt-in advanced base-URL override** (`isBaseUrlOverrideEligibleProvider`, hidden behind an "Advanced" toggle, reusing the existing `providerSpecificData.baseUrl` persistence — not globally widened). Regression guard: `tests/unit/routing-settings-ux-6147.test.ts`.
+- **feat(combo):** add an option to **disable session stickiness**, per-combo or globally — round-robin / random combos can rotate to a different connection on every request instead of pinning a whole conversation to one connection by its first-message hash. Resolution precedence per-combo `config.disableSessionStickiness` → global `settings.disableSessionStickiness` → default `false` (preserves the #3825 prompt-cache/504 fix); gates **both** stickiness call sites in `open-sse/services/combo.ts`. Exposed as a global toggle (Combo Defaults) and a per-combo Inherit/on/off control. ([#6168](https://github.com/diegosouzapw/OmniRoute/issues/6168)) Regression guard: `tests/unit/combo-disable-session-stickiness.test.ts`. (thanks @RCrushMe)
+- **feat(docker):** add the `OMNIROUTE_NO_SUDO` env flag for root-less / user-namespaced deployments — the MITM cert-trust command path (`resolveSudoSpawn` in `src/mitm/systemCommands.ts`) now strips the leading `sudo` when the flag is truthy, in addition to the existing root / sudo-missing cases, so the Proxy Agent runs without `sudo` (the operator trusts the CA manually, e.g. via `NODE_EXTRA_CA_CERTS`). Argv-array `spawn` preserved — no shell interpolation (Hard Rule #13). ([#6122](https://github.com/diegosouzapw/OmniRoute/issues/6122)) Regression guard: `tests/unit/mitm-systemCommands-no-sudo.test.ts`. (thanks @powellnorma)
+- **feat(providers):** add **Requesty** as an OpenAI-compatible gateway provider (BYOK, base `https://router.requesty.ai/v1`, ~200 free requests/day) — wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/requesty/`, `src/shared/constants/providers/apikey/gateways.ts`). ([#6120](https://github.com/diegosouzapw/OmniRoute/issues/6120)) Regression guard: `tests/unit/requesty-provider.test.ts`. (thanks @chirag127)
+- **feat(dashboard):** add **configured-only / available-only filters** to the Free Provider Rankings page ([#6150](https://github.com/diegosouzapw/OmniRoute/issues/6150)) — hide providers you haven't configured, or whose connections are all rate-limited / out of quota, via server-side query params (`?configuredOnly` / `?availableOnly` on `GET /api/free-provider-rankings`) backed by a testable lib helper reusing the in-process connection state (no Redis). Both filters default off, so the default view is unchanged; this supersedes the earlier client-side "Configured Only" toggle (#6245) with an available-only dimension and unit-tested logic. Regression guard: `tests/unit/freeProviderRankings-filters.test.ts`.
+
+### 🔧 Bug Fixes
+
+- **fix(dashboard):** adding a second API key connection for the same provider no longer silently overwrites the first — the Add-API-key modal now derives a unique default connection name (`main`, then `main-2`, `main-3`, …) so the backend name-based upsert can't collide (#6499 — thanks @dilneiss).
+- **fix(compression):** the session-dedup engine now also deduplicates a large multi-line block repeated **within a single message** (intra-message dedup), not just across turns; the compression-preview API surfaces a `fallbackReason`, and the fusion panel reports how many models were rate-limited vs failed on total-panel failure (#6501 — thanks @chirag127).
+- **fix(compression):** a stacked-pipeline step naming an unregistered engine now surfaces a `validationErrors` entry instead of silently no-op'ing, so misconfigured pipelines are visible in the preview API (#6506 — thanks @chirag127).
+- **feat(usage):** add a **Codex reset-credit redemption** flow to the Provider Limits UI ([#6361](https://github.com/diegosouzapw/OmniRoute/pull/6361)) — a `useCodexResetCreditRedemption` hook + `/api/usage/codex-reset-credit` route + `codexResetCredits` lib let you redeem banked Codex reset credits from the quota card. Regression guard: `tests/unit/codex-reset-credits.test.ts`. (thanks @JxnLexn)
+- **feat(glm):** add **team-plan quota settings** for `glm-cn` connections ([#6351](https://github.com/diegosouzapw/OmniRoute/pull/6351)) — a dedicated `GlmTeamQuotaFields` form section (team quota id / limits) threaded through the Add/Edit connection modals, persisted via `providerSpecificData`, with the GLM usage service reading the team quota. Regression guards: `tests/unit/glm-team-quota.test.ts`, `provider-specific-data-schema.test.ts`. (thanks @hao3039032)
+- **feat(providers):** add **TinyFish** web-fetch/search support ([#6349](https://github.com/diegosouzapw/OmniRoute/pull/6349)) — a `tinyfish-fetch` executor + `/v1/web/fetch` route + MCP web-fetch tool, registered as a specialty-media provider with request-validation and a search-provider catalog entry. Regression guards: `tests/unit/executor-tinyfish-fetch.test.ts`, `web-fetch-handler.test.ts`, `mcp-web-fetch-tool.test.ts`, `provider-validation-tinyfish.test.ts`. (thanks @dtybnrj)
+- **fix(cli):** `omniroute launch-codex` now spawns `codex.cmd` through a shell on Windows (the npm `.cmd` shim is unresolvable by bare `spawn` → ENOENT), mirroring the qodercli Windows fix (#6263) ([#6312](https://github.com/diegosouzapw/OmniRoute/pull/6312)). Regression guard: `tests/unit/launch-codex-windows-spawn-6312.test.ts`. (thanks @swingtempo)
+- **fix(codex):** isolate the **Spark** quota from the shared Codex quota and stabilize the quota UI ordering / hydration so per-scope limits render consistently ([#6336](https://github.com/diegosouzapw/OmniRoute/pull/6336)). Regression guards: `tests/unit/codex-quota-selection-hydration.test.ts`, `provider-limits-ui.test.ts` + 3 more. (thanks @xz-dev)
+- **feat(api):** add a `hidePaidModels` setting that filters paid-only models out of the `/v1/models` catalog. Regression guard: `tests/unit/models-catalog-hide-paid.test.ts`. (thanks @chirag127)
+- **fix(api-manager):** the fallback model picker now preserves combos instead of dropping them when a primary model is unavailable ([#6443](https://github.com/diegosouzapw/OmniRoute/pull/6443)). Regression guard: `tests/unit/api-manager-page-static.test.ts`. (thanks @jmengit)
+- **fix(providers):** recoverable Antigravity / Cloud-Code (Gemini Code Assist) `403` responses ([#6452](https://github.com/diegosouzapw/OmniRoute/pull/6452)) are now classified as a retryable project-config error instead of a terminal account ban, so a fixable project/API-disabled 403 no longer forces a ~1-year cooldown / full OAuth reconnect. Regression guard: `tests/unit/errorclassifier-antigravity-403.test.ts`. (thanks @developerjillur)
+- **fix(mitm):** `sanitizeHeaders` now redacts `Set-Cookie` response headers so upstream session cookies never leak into logs / diagnostics ([#6451](https://github.com/diegosouzapw/OmniRoute/pull/6451)). Regression guard: `tests/unit/mitm-sanitize-headers.test.ts`. (thanks @developerjillur)
+- **fix(api):** `/api/compression/preview` now accepts `mode: "caveman"` and correctly handles stacked / zero-compression previews ([#6425](https://github.com/diegosouzapw/OmniRoute/issues/6425)). Regression guard: `tests/unit/api/compression-preview-caveman-and-stacked-6425.test.ts`. (thanks @chirag127)
+- **feat(providers):** add **Zed** hosted LLM aggregator as a native-app provider ([#6118](https://github.com/diegosouzapw/OmniRoute/pull/6118)) — OAuth sign-in via the Zed hosted flow, registered through the shared provider registry + executor. Regression guards: `tests/unit/zed-oauth-provider.test.ts`, `zed-import-utils.test.ts`, `zed-docker-detect.test.ts`, `mitm-handler-zed.test.ts`. VPS-validated via live operator login (Hard Rule #18).
+- **fix(oauth):** the Kiro SSO-cache auto-import now **preserves the IDC region** — cross-region Amazon Q / Kiro profiles imported from the SSO cache are no longer collapsed to the default region ([#6113](https://github.com/diegosouzapw/OmniRoute/pull/6113)). Regression guard: `tests/unit/kiro-auto-import-idc-2059.test.ts`. VPS-validated via live operator login (Hard Rule #18).
+- **fix(dashboard):** passthrough model aliases no longer collide when two namespaced model ids share a last segment (port from 9router#1850, [#6431](https://github.com/diegosouzapw/OmniRoute/pull/6431)). `enx/gpt-5.5` and `enx/codebuddy/gpt-5.5` both auto-generated the alias `gpt-5.5`, so the second model could never be added (the UI just alerted "alias already exists"). Aliases are now disambiguated deterministically — bare last segment when free, then parent-qualified (`codebuddy-gpt-5.5`), then a numeric suffix — while re-adding the exact same model id is still blocked. Regression guard: `tests/unit/passthrough-alias-1850.test.ts`. (thanks @arpicato)
+- **fix(translator):** preserve a Gemini `functionResponse` co-located with other parts (another `functionCall`, or trailing `text`) in the same content when translating **Gemini → OpenAI** ([#6376](https://github.com/diegosouzapw/OmniRoute/pull/6376)). `convertGeminiContent()` early-returned the tool message on the first `functionResponse` part, dropping any co-located parts; such contents are now pre-split (one tool message per `functionResponse`, emitted first, plus one message for the remaining parts). Regression guard: `tests/unit/gemini-to-openai-function-response.test.ts`. (thanks @warelik)
+- **fix(headroom):** detect a python interpreter managed by **mise / pyenv / asdf / conda** (port from 9router#2353, [#6382](https://github.com/diegosouzapw/OmniRoute/pull/6382)). Headroom's python probe (`src/lib/headroom/detect.ts`) searched a hardcoded `PATH`, but version managers expose their interpreters via shim dirs that only join `PATH` through interactive-shell activation — which the non-interactive server never runs, so a managed python (≥3.10) was invisible and Headroom reported it missing. The search path now prepends the well-known shim/bin dirs (`~/.local/share/mise/shims`, `~/.pyenv/shims`, `~/.asdf/shims`, `$CONDA_PREFIX/bin`, `~/.local/bin`, respecting `MISE_DATA_DIR`/`PYENV_ROOT`/`ASDF_DATA_DIR` when set), and a new `HEADROOM_PYTHON` env override lets operators point straight at their interpreter (mirroring `HEADROOM_URL`). Still shell-free (`execFileSync`). Regression guard: `tests/unit/headroom-detect.test.ts` (5). (thanks @loopyd)
+- **fix(executors):** strip the OpenAI-Codex/Claude-CLI `client_metadata` passthrough field for **NVIDIA** requests (port from 9router#1887, [#6411](https://github.com/diegosouzapw/OmniRoute/pull/6411)). NVIDIA's OpenAI-compatible wrapper rejects it with `400 Unsupported parameter`, the same class already handled for `cerebras`/`mistral`; `nvidia` (executor `default`) was missing from the strip allowlist so Codex/Claude-Code passthrough requests 400'd. Regression guard: `tests/unit/executor-default-strip-client-metadata.test.ts` (+nvidia case). (thanks @phidinhmanh)
+- **fix(translator):** strip the Claude-style `thinking` field for **NVIDIA `z-ai/glm-5.2`** (port from 9router#2023, [#6413](https://github.com/diegosouzapw/OmniRoute/pull/6413)). NVIDIA's OpenAI-compatible wrapper 400s on `thinking` (a Claude-format client routed here leaves a `thinking:{type:"adaptive"}`); the existing strip rule only dropped `reasoning`. Same class already handled for `minimax-m2.7`. Regression guard: `tests/unit/nvidia-minimax-thinking-strip.test.ts` (+glm-5.2 case). (thanks @phidinhmanh)
+- **fix(translator):** suppress the streamed `</think>` close marker for the **Antigravity IDE** client (port from 9router#1061, [#6415](https://github.com/diegosouzapw/OmniRoute/pull/6415)). On thinking-only turns Antigravity rendered a bare `</think>` as the sole visible content, tripping its loop-detection and wasting requests. Antigravity's UA (`vscode/<v> (Antigravity/<v>)`) is added to the marker-suppress allowlist (alongside OpenCode); Claude Code / Cursor still get the marker, and `x-omniroute-thinking-marker: on` force-restores it. Regression guard: `tests/unit/think-close-marker-suppress-5245.test.ts`. (thanks @abdofallah)
+- **fix(executors):** strip nested `reasoning_content` from messages for **Mistral** (port from 9router#1649, [#6417](https://github.com/diegosouzapw/OmniRoute/pull/6417)). Mistral's API returns `422 extra_forbidden` when an assistant message carries `reasoning_content` (replayed thinking from a prior turn, e.g. via the Codex `/responses` path); the generic top-level 400 field-downgrade retry never covered the nested per-message field. `DefaultExecutor` now strips it for provider `mistral` only, so DeepSeek (which requires replayed `reasoning_content`) is unaffected. Regression guard: `tests/unit/mistral-strip-reasoning-content-1649.test.ts`. (thanks @xxy9468615)
+- **fix(executors):** strip the `client_metadata` passthrough field on the **OpenCode** path (port from 9router#1442, [#6418](https://github.com/diegosouzapw/OmniRoute/pull/6418)). OpenCode upstreams (e.g. `kimi-k2.6` via opencode-go) reject it with `400 "Extra inputs are not permitted, field: 'client_metadata'"`; the DefaultExecutor strip only covered cerebras/mistral and `OpencodeExecutor` extends `BaseExecutor` directly, so nothing removed it there. Regression guard: `tests/unit/opencode-strip-client-metadata-1442.test.ts`. (thanks @yanpaing007)
+- **fix(executors):** inject the `reasoning_content` echo for the native **Moonshot Kimi** provider (port from 9router#1480, [#6419](https://github.com/diegosouzapw/OmniRoute/pull/6419)). Kimi (executor `default`) is a thinking-mode upstream that 400s with "reasoning_content must be passed back" when a prior assistant turn lacks it; the placeholder injection was only wired into the OpenCode meta-provider, so direct multi-turn Kimi conversations failed. Scoped to `kimi` (gateway-served models matching the thinking-model name pattern are unaffected). Regression guard: `tests/unit/kimi-native-reasoning-injected-1480.test.ts`. (thanks @2220258345)
+- **fix(executors):** recover from a strict gateway's `context_management: Extra inputs are not permitted` 400 (port from 9router#1468, [#6420](https://github.com/diegosouzapw/OmniRoute/pull/6420)). **Claude Code** always sends a top-level `context_management` field; strict anthropic-compatible gateways reject it. The dedicated context-editing 400-fallback only fired when OmniRoute's own `contextEditing` feature was enabled (default off), so a client-sent field passed through untouched and 400'd. `context_management` is now in the generic reactive field-strip list, so it's stripped-and-retried once regardless of the feature flag (with correct request re-signing for claude-compatible relays). Regression guard: `tests/unit/provider-field-strips.test.ts`. (thanks @ohahe52-dot)
+- **fix(network):** enable **RFC 8305 Happy Eyeballs** (`autoSelectFamily`) on the direct-egress undici dispatcher (port from 9router#1237, [#6423](https://github.com/diegosouzapw/OmniRoute/pull/6423)). When DNS returns both IPv6 (AAAA) and IPv4 (A) and the IPv6 route is broken (e.g. a NAT64 `64:ff9b::` prefix without routing), undici tried IPv6 first and hung until `ETIMEDOUT` (then a 502 + account lockout), even though `curl` reached the same host. The direct dispatcher now races both families and uses whichever connects first. Proxy paths pin family via `proxyTls` and are unaffected. Regression guard: `tests/unit/direct-dispatcher-pipelining-4580.test.ts`. (thanks @adentdk)
+- **fix(combo):** round-robin now advances the rotation pointer past the model that **actually served**, not the eagerly-scheduled one (port from 9router#948, [#6428](https://github.com/diegosouzapw/OmniRoute/pull/6428)). With `stickyLimit: 1` (true round-robin), when the scheduled model failed and a _different_ model served via fallback, the counter had already advanced +1 from the scheduled index — so the next request reused the fallback-served model, degrading round-robin into hot-spotting on whichever model was healthy. The pointer now advances to the served index + 1 (mirroring the sticky-limit>1 path). Session-stickiness (#3825) and distribution are preserved. Regression guard: `tests/unit/combo-rr-fallback-advance-948.test.ts`. (thanks @binsarjr)
+- **fix(sse):** a non-string `model` field is now rejected with a `400` before the resolver, instead of crashing downstream `.toLowerCase()`/`.split()` calls into an empty-body `500` that escapes the error sanitizer ([#6407](https://github.com/diegosouzapw/OmniRoute/issues/6407)). Regression guard: `tests/unit/chat-non-string-model-6407.test.ts`. (thanks @chirag127)
+- **fix(api):** unknown `/api/*` routes now return a JSON `404` (instead of the dashboard HTML shell) and scalar chat params (`model`/`temperature`/etc.) are validated **before** the provider lookup so malformed requests fail fast with a clear `400` ([#6424](https://github.com/diegosouzapw/OmniRoute/issues/6424), [#6412](https://github.com/diegosouzapw/OmniRoute/issues/6412)). Regression guards: `tests/unit/api/api-catchall-json-404.test.ts`, `tests/unit/chat-early-schema-validation-6412.test.ts`. (thanks @chirag127)
+- **fix(api):** `/v1/chat/completions` now rejects a non-JSON `Content-Type` with a `400` before parsing the body ([#6414](https://github.com/diegosouzapw/OmniRoute/issues/6414)). Regression guard: `tests/unit/v1-chat-completions-content-type-6414.test.ts`. (thanks @chirag127)
+- **fix(api):** the `X-OmniRoute-Compression` response header is now echoed on `/v1/chat/completions` and `/v1/completions` ([#6422](https://github.com/diegosouzapw/OmniRoute/issues/6422)). Regression guard: `tests/unit/compression-header-echo-6422.test.ts`. (thanks @chirag127)
+- **fix(api):** concurrent `GET /v1/models` requests are coalesced into a single catalog build ([#6408](https://github.com/diegosouzapw/OmniRoute/issues/6408)). Regression guard: `tests/unit/v1-models-concurrent-6408.test.ts`. (thanks @chirag127)
+- **fix(api):** `/v1/completions` now echoes the requested `body.model` in its JSON + streamed responses ([#6429](https://github.com/diegosouzapw/OmniRoute/pull/6429)). Regression guard: `tests/unit/completions-body-model-echo.test.ts`. (thanks @chirag127)
+- **fix(api):** env-var master keys now see the full `/v1/models` catalog ([#6406](https://github.com/diegosouzapw/OmniRoute/issues/6406)). Regression guard: `tests/unit/models-catalog-envkey-6406.test.ts`. (thanks @chirag127)
+- **fix(api):** non-streaming `/v1/completions` responses now echo `body.model` aligned with the `X-OmniRoute-Model` header ([#6426](https://github.com/diegosouzapw/OmniRoute/issues/6426)). Regression guard: `tests/unit/v1-completions-model-header-match-6426.test.ts`. (thanks @chirag127)
+- **fix(api):** unknown `/v1/*` routes now return a JSON `404 not_found` instead of the Next.js dashboard HTML shell ([#6405](https://github.com/diegosouzapw/OmniRoute/issues/6405)). Regression guard: `tests/unit/api/v1-catchall-json-404.test.ts`. (thanks @chirag127)
+- **fix(api):** the per-connection provider models route now degrades to the shipped catalog when a provider's `/models` endpoint answers with a redirect ([#6267](https://github.com/diegosouzapw/OmniRoute/issues/6267)) — a `qwen-web` import failed with a raw `Redirect blocked … (307)` 503. `safeOutboundFetch` throws `REDIRECT_BLOCKED` on the 307, `getSafeOutboundFetchErrorStatus` maps it to 503, and `buildDiscoveryErrorFallbackResponse` treated every 503 as a hard error — so the non-empty `getModelsByProviderId("qwen-web")` catalog was never surfaced. A models-endpoint redirect is not a fixable-config error (unlike `URL_GUARD_BLOCKED`/`INVALID_URL`, which stay hard errors), so it now falls back to the local/cached catalog before the 503 short-circuit. General fix — covers any config-driven provider that 307s. Regression guard: `tests/unit/provider-models-qwen-web-redirect-6267.test.ts`. (thanks @chirag127)
+- **fix(api):** the per-connection provider models route (MCP `list_models_catalog` + the dashboard import view) now merges USER-ADDED custom models into its response ([#6247](https://github.com/diegosouzapw/OmniRoute/issues/6247)) — custom models live in the `key_value` namespace `customModels`, which the live REST `/api/v1/models` already merges, but `src/app/api/providers/[id]/models/route.ts` never read `getCustomModels`, so custom models were dropped on both the discovery-success and local_catalog paths. They are now folded into the returned model list (deduped by id, stamped `owned_by: provider`), fixing MCP + the dashboard import view in one place. Regression guard: `tests/unit/provider-models-custom-merge-6247.test.ts`. (thanks @RCrushMe)
+
+- **fix(providers):** GitLab Duo tool-calling follow-up turns no longer fail upstream with `422 {"detail":"Validation error"}` (tokens 0/0, rejected pre-inference). The #6234 tool-result-feedback fix serialized the **entire** multi-turn conversation into GitLab's single-file `code_suggestions` (`small_file`) generation endpoint — folded history that turn-N sent as an oversized `current_file.content_above_cursor` **and** duplicated verbatim into `user_instruction`, tripping the AI-Gateway's `small_file` validation guard. The executor now **bounds** that prompt: it keeps system + latest user message + the most-recent tool round (dropping older turns), caps oversized tool results, and stops duplicating the full prompt into `user_instruction` (which now carries only the short latest user message) — while still feeding the most-recent tool result back so the agent continues (`open-sse/executors/gitlab.ts`, [#6220](https://github.com/diegosouzapw/OmniRoute/issues/6220)). The unit test covers the bounding logic; the upstream 422→200 clearing is VPS-only (Hard Rule #18). Regression guard: `tests/unit/gitlab-tool-exchange-bounded-6220.test.ts`.
+
+- **fix(i18n):** the provider-detail (`/dashboard/providers/[id]`) connection-status filter labels no longer render as `__MISSING__:All` / `__MISSING__:Active` / `__MISSING__:Error` / `__MISSING__:Banned` / `__MISSING__:CreditsExhausted` in non-English locales (notably pt-BR) ([#6290](https://github.com/diegosouzapw/OmniRoute/issues/6290)). Root cause was **not** the namespace mismatch the issue guessed — the `providers.filter*` keys resolve correctly in `en.json`; the debt lived in the locale mirrors (`src/i18n/messages/*.json`), where these five keys carried the `__MISSING__:` sync sentinel in ~15 locales and were absent entirely in ~26 others, so next-intl found the key and echoed the sentinel verbatim. All 40 non-English/-Chinese mirrors now ship real translations for the five `providers.filter*` labels. Regression guard: `tests/unit/i18n-provider-filter-keys-6290.test.ts`. (thanks @diegosouzapw)
+
+- **fix(providers):** the `copilot-m365-web` streaming executor now emits `debug`-level WebSocket diagnostics ([#6210](https://github.com/diegosouzapw/OmniRoute/issues/6210)) — the outbound WS URL (with the `access_token` **redacted** via `redactWsUrl()`), handshake success/failure, and each received SignalR frame's `type`/`target`. Previously the streaming path logged nothing, so an empty `content:null` response (the M365 Education / Starter tier symptom fixed in #6234) was undiagnosable even at `APP_LOG_LEVEL=debug`. The change is debug-level and side-effect-free — it does not alter streaming behavior or the frame parser, and the token never reaches the logs. Regression guard: `tests/unit/copilot-m365-web-logging-6210.test.ts` (thanks @qpeyba)
+
+- **fix(resilience):** a round-robin combo no longer returns `503 all upstream accounts are unavailable` when a compatibility-rejected target is actually healthy ([#6238](https://github.com/diegosouzapw/OmniRoute/issues/6238)). `filterTargetsByRequestCompatibility` drops request-incompatible targets (tool/vision/structured-output unsupported, or below the required context window) **before** any availability check runs, and its `compatible.length === 0` safety net only fired when _all_ targets were filtered — not when the kept targets later all turned out runtime-unavailable (circuit-open / cooldown / no credentials). So a combo could 503 while a compat-rejected-but-healthy provider sat unused. `handleRoundRobinCombo` now keeps the compat-rejected set and, when every compat-kept target was skipped without a single real attempt, probes those rejected targets as a **last-resort fallback tier** (via the new pure `open-sse/services/combo/comboCompatFallback.ts`) before crystallizing the 503. Regression guard: `tests/unit/combo-roundrobin-compat-fallback-6238.test.ts`. (thanks @ThongAccount)
+
+- **fix(startup):** best-effort self-heal for a corrupted Turbopack dev cache on Windows ([#6289](https://github.com/diegosouzapw/OmniRoute/issues/6289)). On Windows, `pnpm dev` can fail at startup when Turbopack `mmap`s a persistent-cache SST file and the OS refuses the mapping (`os error 1455` — "paging file too small"), which Turbopack surfaces as a misleading `Module not found: Can't resolve '@/shared/utils/machine'`. This is a **known upstream Turbopack cache-corruption bug — not our code**. The dev launcher (`scripts/dev/run-next.mjs`) now wraps `nextApp.prepare()` and, when it rejects with that signature (`isTurbopackCacheCorruption` in the new `scripts/dev/turbopackCacheHeal.mjs`), purges `.build/next/**/cache/turbopack` and retries **once** with a clear log. **Caveat — best-effort only:** the corruption often surfaces as a runtime overlay rather than a `prepare()` rejection, so this cannot always intercept it; the reliable remedy remains manually deleting the Turbopack cache dir. Regression guard: `tests/unit/turbopack-cache-heal-6289.test.ts`. (thanks @chirag127)
+
+- **fix(providers):** qodercli PAT auth no longer fails with `spawn qodercli ENOENT` on Windows ([#6263](https://github.com/diegosouzapw/OmniRoute/issues/6263)) — `spawnQoderCli` spawned the bare `qodercli` name with `shell:false` and an unenriched env, so the npm `.cmd` wrapper under `%APPDATA%\npm` (a user-PATH directory) was never resolved. It now resolves the absolute `.cmd`/`.exe` path through the existing `getCliRuntimeStatus("qoder")` resolver in `src/shared/services/cliRuntime.ts` (memoized), spawns with `shell` when the target is a `.cmd`/`.bat`, and uses the cliRuntime-enriched env (PATH + PATHEXT + APPDATA); the ENOENT error now lists the searched paths plus the `CLI_QODER_BIN` override. End-to-end spawn on a real Windows host is host-only (Hard Rule #18); the path-resolution logic is unit-tested. Regression guard: `tests/unit/qodercli-windows-resolve-6263.test.ts`. (thanks @chirag127)
+
+- **fix(sse):** the reasoning-token buffer no longer inflates **probe-sized `max_tokens`** ([#6274](https://github.com/diegosouzapw/OmniRoute/issues/6274)) — Claude Code's `/model` capability check sends `max_tokens: 1`, but for a thinking-capable model with a large output cap (e.g. `glm-5.2`) the #3587 headroom heuristic (`max(current + 1000, ceil(current * 1.5))`) rewrote it to `1001` and forwarded that upstream, wasting tokens on a request that was never a genuine reasoning budget. `resolveReasoningBufferedMaxTokens()` (`open-sse/services/reasoningTokenBuffer.ts`) now short-circuits and returns the caller's value verbatim when it is below the new `REASONING_BUFFER_MIN_TRIGGER` (256) threshold — a tiny explicit limit is a probe, not a reasoning request. Real budgets still receive the #3587 headroom unchanged, and the guard runs after the existing capability checks so unknown / non-reasoning models keep returning `null`. Regression guard: `tests/unit/reasoning-token-buffer-6274.test.ts`. (thanks @brightfiscalband)
+
+- **fix(cli):** `omniroute reset-password` now works as a real subcommand, and password resets over piped (non-TTY) stdin actually apply ([#6261](https://github.com/diegosouzapw/OmniRoute/issues/6261), [#6258](https://github.com/diegosouzapw/OmniRoute/issues/6258)). Two coupled defects: (1) **#6261** — `bin/omniroute.mjs` routed everything through Commander with only two pre-Commander bypasses (`--mcp`, `reset-encrypted-columns`), so `omniroute reset-password` was rejected as an unknown command; only the separate `omniroute-reset-password` bin worked, while the docs falsely advertised the subcommand (incl. a bogus "legacy alias still works"). A pre-Commander bypass mirroring `reset-encrypted-columns` now dynamically imports `bin/reset-password.mjs` (which self-executes) before Commander parses; the three doc lines were corrected. (2) **#6258** — `bin/reset-password.mjs` issued two sequential `rl.question` prompts; under piped stdin the second read never settled at EOF, so `main()` never reached `resetManagementPassword` and the reset was a silent no-op (both prompts printed, no success, password unchanged). The CLI now detects non-TTY stdin and reads it once (first line = password, second line = confirm if present, else reused), adds a `--password-stdin` flag (entire stdin is the password, no confirmation), and exits `0` explicitly so the success line always flushes; interactive TTY behavior is unchanged. Regression guard: `tests/unit/reset-password-cli-6261-6258.test.ts` (3). (thanks @chirag127)
+
+- **fix(db):** the mass-migration **safety abort** now tells the operator how to bypass it and stops flooding the log ([#6260](https://github.com/diegosouzapw/OmniRoute/issues/6260)) — after restoring a backup that wiped the migration tracking table, `runMigrations()` threw the abort on every downstream `ensureDbInitialized()`, re-logging the full banner 11+ times, and the message never mentioned the existing `OMNIROUTE_MAX_PENDING_MIGRATIONS` escape hatch. The abort text now appends a bypass hint (set `OMNIROUTE_MAX_PENDING_MIGRATIONS=0` in `server.env` / `DATA_DIR/.env`), and a new `MigrationSafetyAbortError` is memoized so repeated calls in the same process throw the same instance and emit a single concise line instead of the full cascade. Regression guard: `tests/unit/migration-safety-abort-6260.test.ts`. (thanks @chirag127)
+
+- **fix(auth):** importing a **distinct** Codex/ChatGPT OAuth `auth.json` is no longer falsely rejected as "already exists" when it belongs to a different user in the same workspace ([#6301](https://github.com/diegosouzapw/OmniRoute/issues/6301)). `findExistingCodexConnection` (in `src/lib/oauth/utils/codexAuthImport.ts`) deduped **only** on `providerSpecificData.workspaceId === accountId`, where `accountId` is the shared `chatgpt_account_id`/`tokens.account_id` — so two members of the same ChatGPT Team collapsed onto a single connection (409 `duplicate_account`). The id_token's `https://api.openai.com/auth` claim carries a per-user `chatgpt_user_id` alongside the workspace id (the device-flow path already persisted it as `chatgptUserId`, but the import path did not). Now `parseAndValidateCodexAuth` extracts `userId` (`chatgpt_user_id` → `user_id` → JWT `sub`) into `ParsedCodexAuth`, the create/update paths persist `chatgptUserId` in `providerSpecificData` (mirroring `codex.ts`), and dedup keys on `workspaceId` **AND** `chatgptUserId` — with a backward-compat fallback to legacy accountId-only matching when no stored connection for that workspace records a `chatgptUserId`, so genuinely-same accounts still dedup. Regression guard: `tests/unit/codex-auth-import-userid-dedup-6301.test.ts` (4). (thanks @anungma)
+
+- **fix(providers):** importing models for the **venice-web** provider no longer fails with a red "Provider venice-web does not support models listing" ([#6269](https://github.com/diegosouzapw/OmniRoute/issues/6269)). `venice-web` is a web-cookie provider with an executor but no upstream `/v1/models` endpoint and no registry `models`, so the models route fell through to the tail `400`. Mirroring the `jules`/`linkup-search`/`ollama-search` fix (#5569), it now ships a static local catalog entry in `src/lib/providers/staticModels.ts` — seeding the current Venice lineup (`venice-uncensored`, `llama-3.3-70b`, `qwen3-235b`, `qwen3-4b`, `deepseek-r1-671b`; Venice rotates its catalog, see docs.venice.ai/models/overview) — so the route returns `200` with `source:"local_catalog"`, `intentional:true`. Regression guard: `tests/unit/static-models-venice-web-6269.test.ts`. (thanks @chirag127)
+
+- **fix(api):** the specialty model catalogs (`/v1/embeddings`, `/v1/images`, `/v1/music`, `/v1/videos` model lists) are now derived from the **unified catalog filtered by a predicate** (`getSpecialtyModelsResponse`) instead of ad-hoc per-route logic, so they consistently respect active-credential visibility and stay in sync with the main catalog ([#6303](https://github.com/diegosouzapw/OmniRoute/pull/6303)). Regression guard: `tests/unit/specialty-model-catalog-routes.test.ts`. (thanks @makcimbx)
+- **fix(api):** the agent-bridge server route now resolves the MITM manager via a **dynamic `import("@/mitm/manager.runtime")`** so Turbopack does not statically pull the stub (or over-bundle the manager), and the agent-skills generator anchors its output base path with `path.join(process.cwd(), …)` so Turbopack's static analyzer stops tracing the whole project root ([#6329](https://github.com/diegosouzapw/OmniRoute/issues/6329), [#6366](https://github.com/diegosouzapw/OmniRoute/pull/6366)). Regression guard: `tests/unit/agent-bridge-server-route-dynamic-import.test.ts`. (thanks @Iammilansoni)
+- **fix(api):** internal probes (combo-test, cloud-sync verify) now pick a **management-scoped / allow-all API key** instead of naively grabbing `getApiKeys()[0]` — a restricted `self:usage` first row made the probe fail with "Model X is not allowed for this API key" even when the combo path was healthy (`pickApiKeyForInternalUse` in `src/lib/db/apiKeys.ts`). The API-manager model editor also falls back to `/api/models?all=true` when `/v1/models` is catalog-protected ([#6372](https://github.com/diegosouzapw/OmniRoute/pull/6372)). Regression guard: `tests/unit/pick-internal-api-key-6372.test.ts`. (thanks @jmengit)
+- **fix(live-ws):** the Live Dashboard WebSocket server now **rejects on bind failure** (e.g. `EADDRINUSE` when the API bridge already holds the port) instead of letting the error surface as an unhandled `error` event that crash-loops the process — the `error` listener is attached to `wss` (not `server`) and releases the EventBus subscription on a failed start ([#6324](https://github.com/diegosouzapw/OmniRoute/issues/6324)). Regression guard: `tests/unit/live-ws-eaddrinuse-6324.test.ts`. (thanks @vinayakkulkarni)
+- **fix(dashboard):** the Home provider-topology widget now trusts the live provider-metrics snapshot — it uses `topology.errorProvider` and live `activeRequests` directly instead of re-deriving state from a stale `lastErrorAt` or applying a frontend timeout filter, so the topology reflects real-time provider health ([#6322](https://github.com/diegosouzapw/OmniRoute/pull/6322)). Regression guard: `tests/unit/home-provider-topology-live-state.test.ts`. (thanks @xz-dev)
+- **fix(sse):** strip zero-width markers from streamed **tool-call arguments** — a follow-up to [#5857](https://github.com/diegosouzapw/OmniRoute/pull/5857). That PR removed injected zero-width joiners (U+200D) from streamed assistant text/reasoning but deliberately left tool-call argument JSON byte-exact. The request-side obfuscation (`open-sse/services/claudeCodeObfuscation.ts`) injects ZWJ into agent words — including the temp path inside the Bash tool description — and Claude models copy that verbatim into generated commands, which are delivered as tool-call arguments rather than assistant text. As a result the ZWJ survived and corrupted code blocks (e.g. a temp path rendered with an invisible joiner). Now `open-sse/handlers/responseSanitizer.ts` strips zero-width code points from tool-call argument strings at every emit site (OpenAI non-stream/stream chat `tool_calls` + legacy `function_call`, native Responses `function_call` items, the OpenAI→Responses conversion, and the native Responses streaming `response.function_call_arguments.delta/.done` events). Only zero-width code points are removed; JSON structure and all other bytes stay identical (no parse/restringify), so normal arguments remain byte-exact. Regression guard: 6 new cases in `tests/unit/response-sanitizer.test.ts` (suite 50/50).
+
+- **fix(nodejs):** the default app log path now resolves under `DATA_DIR` (`~/.omniroute/logs/application/app.log`) instead of `process.cwd()` ([#6197](https://github.com/diegosouzapw/OmniRoute/issues/6197)) — the globally-installed CLI runs from an arbitrary working directory, so anchoring the default to cwd made file logging silently write to (or no-op under) an unrelated directory, contradicting the documented `.env.example` default. `getAppLogFilePath()` now computes the default lazily via the pure `resolveDataDir()` resolver (honours a per-process `DATA_DIR`, no directory-creation side effect); an explicit `APP_LOG_FILE_PATH` still wins. Regression guard: `tests/unit/logenv-datadir-path-6197.test.ts` (3).
+
+- **fix(docker):** AgentBridge/`startMitm` no longer aborts in containers/headless when the Antigravity-default DNS step can't write `/etc/hosts` ([#6127](https://github.com/diegosouzapw/OmniRoute/issues/6127)), and the privileged command's stderr now reaches `app.log` instead of only a bare exit code hitting the toast ([#6198](https://github.com/diegosouzapw/OmniRoute/issues/6198)). The default DNS step (`addDNSEntry`) was called unguarded while cert install and the two sibling DNS steps were each best-effort — in the runtime Docker image (`USER node`, no `sudo`, read-only `/etc/hosts`) it threw `Command failed with code 1` out of `startMitmInternal` and killed the whole start, discarding the stderr. The three DNS steps are extracted into a best-effort `provisionDnsEntries()` where each failure is logged with the full `err` (stderr included, folded in by `systemCommands.ts`) and never aborts the start. Regression guard: `tests/unit/mitm-dns-graceful-degrade-6127.test.ts` (4).
+
+- **fix(providers):** copilot-m365-web now supports the M365 Education "Starter / OfficeWebIncludedCopilot" tier and no longer returns an empty `content:null` stream ([#6210](https://github.com/diegosouzapw/OmniRoute/issues/6210)). Two gaps: (1) `buildWsUrl()` hardcoded the individual-consumer scenario (`OfficeWebPaidConsumerCopilot`, `isEdu=false`) — the EDU tier is now opt-in via `providerSpecificData.tier="edu"`, emitting `scenario=OfficeWebIncludedCopilot`/`isEdu=true` (the individual path is unchanged); (2) the EDU/GPT-5.5 path streams deltas via `arguments[0].writeAtCursor` (incremental) instead of only `messages[].text` (accumulated snapshots), which the parser dropped — a new `accumulateBotContent()` folds both formats, with `type:2 item.result.message` as a last-resort fallback. Regression guard: `tests/unit/copilot-m365-edu-writeatcursor-6210.test.ts` (10). (thanks @qpeyba)
+
+- **fix(providers):** GitLab Duo executor now feeds tool results back into the prompt instead of looping ([#6220](https://github.com/diegosouzapw/OmniRoute/issues/6220)) — `buildPrompt()` branched only on `system`/`user` and took `userParts.at(-1)`, silently dropping the `assistant{tool_calls}` + `tool{result}` turns the client appended, so the reconstructed prompt was byte-identical to turn 1 and the model re-emitted the same `<tool>` call forever. When a tool exchange is present the full conversation is now serialized, folding each tool result back keyed by its `tool_call_id`; simple conversations keep the legacy shape. Complements the tool_call emission from [#6051](https://github.com/diegosouzapw/OmniRoute/issues/6051) (the `kilo-duplicate` label was a false positive — different, sequential defect). Regression guard: `tests/unit/gitlab-tool-result-feedback-6220.test.ts` (4).
+
+- **fix(providers):** opencode-go/opencode-zen can now synthesize the OpenCode CLI identity headers Cloudflare requires on VPS egress ([#5997](https://github.com/diegosouzapw/OmniRoute/issues/5997)) — on a datacenter VPS, `opencode.ai/zen/go/v1/chat/completions` 403s (HTML challenge) requests lacking CLI identity, while the reporter's control curl proved that `User-Agent: opencode-cli/1.0.0` + `x-opencode-client: cli` + `x-opencode-project: default` + fresh request/session UUIDs succeed. Opt-in via `OPENCODE_SYNTHESIZE_CLI_HEADERS=true` (values overridable via `OPENCODE_GO_USER_AGENT`/`OPENCODE_USER_AGENT`/`OPENCODE_CLIENT`/`OPENCODE_PROJECT`); it fills only headers the client did not already send. Kept **off by default** — the forward-only path is deliberate (fabricating a wrong value risks upstream rejection; a prior dedup regressed with `opencode/local`), so this replaces the fragile local header-injection shim without changing default behavior. Regression guard: `tests/unit/opencode-cli-headers-synthesis-5997.test.ts` (6). (thanks @aleksesipenko)
+- fix(resilience): sticky session affinity now evicts and fails over to another account when the pinned account is exhausted/unavailable (#6219)
+
+- fix(sse): Responses API passthrough now drops internal commentary-phase output before forwarding to clients (gated by RESPONSES_PASSTHROUGH_DROP_COMMENTARY, default on) (#6199)
+
+- **fix(sse):** tool-call function schemas with a root `type: null` are now coerced to `type: "object"` before dispatch ([#6359](https://github.com/diegosouzapw/OmniRoute/issues/6359)) — clients like the Codex app emit `parameters: { type: null, ... }` for some tools, which OpenAI-compatible upstreams reject with `400 Invalid schema for function '...': schema must be a JSON Schema of 'type: "object"', got 'type: null'`, failing the whole request. `toolSchemaSanitizer` already stripped the null; it now re-adds the mandatory root `"object"` type (and empty `properties`/open `additionalProperties` when absent). Combinator roots (`anyOf`/`oneOf`/`allOf`) and explicit root types are left untouched. Regression guard: 5 new cases in `tests/unit/tool-schema-sanitizer.test.mjs`.
+
+- **fix(docker):** AgentBridge no longer fails to start on npm/Electron/VPS installs with "MITM manager stub reached at runtime" ([#6344](https://github.com/diegosouzapw/OmniRoute/issues/6344)) — v3.8.45 flipped the production bundler default to Turbopack, but `next.config.mjs` aliased `@/mitm/manager` to its Docker-only degraded stub **unconditionally**. That was harmless while Docker (which sets the alias intentionally for #3390 graceful degradation) was the sole Turbopack consumer, but once every artifact built with Turbopack the stub shipped to all non-Docker users and `startMitm` threw on the first Agent-Bridge start. The alias is now opt-in via `OMNIROUTE_MITM_STUB=1` (set only by the Dockerfile) through the shared `scripts/build/mitm-stub-flag.mjs` helper; default builds bundle the real manager. Regression guard: `tests/unit/mitm-stub-alias-6344.test.mjs` (4).
+
+- **fix(proxy):** stop the v3.8.44 proxy regression that leaked the real IP and disabled healthy proxies ([#6246](https://github.com/diegosouzapw/OmniRoute/issues/6246)). Two coupled defects from the new health scheduler: (1) **IP leak** — when a proxy assigned to a connection was marked `inactive`, resolution fell through to a **direct** egress instead of blocking, exposing the operator's real IP; (2) **over-deactivation** — the sweep flipped a proxy to `inactive` on the **first** failed probe and counted our own 5s timeout / a probe-target `5xx` as the proxy's fault, so healthy paid proxies vanished from egress selection ("my proxies are not being used anymore"). Fix: the sweep decision is extracted into a pure, network-free `decideProxyHealthAction` (`src/lib/proxyHealth/decision.ts`) — by default the health check now **only counts/logs and never downgrades status** (a proxy is downgraded/removed only with `PROXY_AUTO_REMOVE=true`, after `PROXY_AUTO_REMOVE_AFTER` **consecutive** conclusive failures); probes are classified tri-state so an inconclusive result (our timeout, or a `5xx` from the probe target) never penalizes the proxy, and the probe timeout is raised 5s→15s. Separately, `safeResolveProxy` now **fails closed** via the existing policy: a connection whose assigned proxy is dead is blocked instead of leaking direct (`hasBlockingProxyAssignment`), honoring the explicit `proxy off` toggles and the `PROXY_FAIL_OPEN=true` opt-out. Existing proxies stuck `inactive` by the old behavior need a one-time manual re-activate (the operator owns proxy status). Regression guards: `tests/unit/proxy-health-decide-action-6246.test.ts`, `tests/unit/proxy-assigned-unavailable-6246.test.ts`.
+
+- **fix(proxy):** make "Test All" read-only and add bulk enable/disable ([#6246](https://github.com/diegosouzapw/OmniRoute/issues/6246)). Complements the core fail-closed / scheduler fix (#6296) with the two remaining reporter asks. (1) The **"Test All" button** (`POST /api/settings/proxies/auto-test`) used to flip a proxy to `inactive` on a failed reachability probe; since the egress selector excludes `inactive` proxies, a flaky probe (an unreachable `httpbin.org`, a proxy that blocks `HEAD`, or a slow paid proxy) silently disabled every proxy that failed — "Test All" is now **read-only by default** (only the operator sets a proxy active/inactive; opt back into the legacy test-and-set with `PROXY_HEALTH_AUTO_DEACTIVATE=true`). (2) Adds a **bulk enable/disable** proxies endpoint + toolbar action (`POST /api/settings/proxies/batch-activate`) so an operator can re-activate proxies in one click. Regression guard: `tests/unit/proxy-health-6246.test.ts`. (thanks @tenshiak)
+
+- **chatcore (tools): stop the default 128-tool cap from silently dropping opencode's `task`/MCP tools.** opencode (used as an MCP/agent host) sends a large tool list; when it exceeds the speculative `MAX_TOOLS_LIMIT` (128) default, `truncateToolList` did a blind `tools.slice(0, 128)`, dropping every tool past index 128 — including opencode's built-in `task` tool (subagent launch) and many MCP tools, so models routed through OmniRoute could no longer spawn subagents or reach part of their tools. The cap exists to avoid upstream `400`s for providers with real hard limits (e.g. grok-cli 200), so it is kept for those: detection of the opencode client (`isOpencodeClient` — any `x-opencode-*` header, or `opencode` in the user-agent) now only bypasses the **speculative 128 default**, never a known provider ceiling. Precedence is explicit — a proactive/detected provider limit always truncates (even for opencode); otherwise opencode forwards its full tool list; otherwise the unchanged 128 default applies to every other client. Refactors `getEffectiveToolLimit` into `getKnownToolLimit(provider) ?? DEFAULT_LIMIT` (byte-identical for existing callers) and fixes a cosmetic debug-log that reported the truncated count instead of the original. Regression guard: `tests/unit/tool-limit-detector.test.ts`.
+
+- **fix(mitm):** the macOS MITM-cert install check now matches the system keychain again. `security find-certificate -a -Z` prints the SHA-1 as a colon-less hex string, but the installed-check compared it against `getCertFingerprint()`'s colon-separated form, so the substring match never hit — the cert was reported as not-installed and re-prompted for the sudo install on every run. Fingerprints are now normalized (colons stripped, upper-cased) on both sides via the extracted `macCertOutputHasFingerprint` helper. Regression guard: `tests/unit/mitm-cert-mac-fingerprint.test.ts`. ([#6204](https://github.com/diegosouzapw/OmniRoute/pull/6204), closes [#6134](https://github.com/diegosouzapw/OmniRoute/issues/6134) — thanks @rianonehub)
+
+- **fix(api):** `/v1/messages/count_tokens` now counts `tool_use`, `tool_result` and `thinking` content blocks (and array-form `system` prompts) in the local-estimation path, instead of only `text`. Real agentic conversations keep ~95% of their tokens inside tool results; the previous estimate returned near-zero for them, which silently broke Claude Code's auto-compaction (context grew past the window with no compaction until the upstream API rejected the request). The real provider-side count path is unchanged. Regression guard: `tests/unit/messages-count-tokens-route.test.ts`. ([#6221](https://github.com/diegosouzapw/OmniRoute/pull/6221) — thanks @luweiCN)
+
+- **fix(antigravity):** strip a trailing assistant prefill turn for Vertex Claude models to avoid upstream 400s ([#6114](https://github.com/diegosouzapw/OmniRoute/pull/6114)). Regression guard: `tests/unit/antigravity-claude-prefill-strip.test.ts`. (thanks @anki1kr)
+
+- **fix(security):** the mutable cloud-agent routes (`/api/cloud/credentials/update`, `/api/cloud/models/alias`) now require management auth instead of being treated as public. They were classified as public API routes, so a request without management credentials could update stored cloud-agent credentials and model aliases. They are removed from the public-route set, classified as management routes in the authz pipeline, and gated by `requireManagementAuth`; cloud **read**/auth routes stay public. Regression guards: `tests/unit/cloud-write-auth.test.ts`, `tests/unit/authz/classify.test.ts`, `tests/unit/public-api-routes.test.ts`. ([#6233](https://github.com/diegosouzapw/OmniRoute/pull/6233) — thanks @vittoroliveira-dev)
+
+- **refactor(dashboard):** extract the onboarding-wizard "Open provider details" link target into a pure, unit-tested `buildProviderDetailsHref(connection)` helper. The wizard already routes by `connection.id` (the node UUID) rather than the provider category slug (#6144/#6145); this hardens that behavior behind a tested helper that guards a missing id/connection. Regression guard: `tests/unit/provider-onboarding-href.test.ts`. ([#6166](https://github.com/diegosouzapw/OmniRoute/pull/6166) — thanks @KooshaPari)
+
+- **fix(security):** the doubao synthetic device-id generator now derives its digits via an unbiased crypto-random draw (rejection sampling over `crypto.randomBytes()`) instead of a `% 10` reduction, closing a CodeQL `js/biased-cryptographic-random` finding.
+- **fix(agentSkills):** the GitHub-skills generator now resolves `outputDir` to an absolute path before writing, fixing a regression introduced by #6366 (relative-to-cwd base path) that could write generated skill files to the wrong directory.
+- **fix(security):** `/api/keys/{id}/devices` now checks the HTTP method before auth/validation, returning a `405` for non-GET/DELETE verbs instead of a misleading `401`/`500` (closes a `dast-smoke` QUERY-method finding).
+- **fix(quality):** clear the last 2 heavy quality-gate reds on the release tip (cycle pre-flight).
+- **fix(mitm):** the test suite and CI can never mutate the OS trust store — `OMNIROUTE_SKIP_SYSTEM_TRUST=1` is set globally for tests/CI so `installCert`/`uninstallCert`/`installTproxyCa` skip the privileged OS dispatch ([#6310](https://github.com/diegosouzapw/OmniRoute/pull/6310); full detail is under the [3.8.45] section below — this branch received it via the parallel-cycle sync-back).
+- **fix(api):** `POST /api/github-skills` now Zod-validates its request body; documented the new quality-gate env vars and pinned the merge-integrity GitHub Actions to a commit SHA.
+- **fix(skills):** generate the missing `omni-github-skills` registry entry and align the agent-skills catalog-count tests (follow-up to #6186).
+- **fix(quality):** clear the cycle's 11 net-new ESLint errors and make `validate-release-green` suppressions-aware.
+
+### 📝 Maintenance
+
+- **i18n(it):** add 118 missing Italian (`it`) translations (net-additive — no existing keys dropped, valid JSON), improving Italian UI coverage. ([#6212](https://github.com/diegosouzapw/OmniRoute/pull/6212) — thanks @serverless83)
+- **chore(providers):** remove deprecated **MiMo V2** model entries from the catalogs (xiaomi-mimo, opencode-go, zenmux-free, audio TTS) — the upstream V2 line is superseded by MiMo V2.5; drops `mimo-v2-tts`, `mimo-v2-pro`, `mimo-v2-omni`, `mimo-v2-flash`, `mimo-v2-flash-free` and realigns the provider-catalog tests. ([#6248](https://github.com/diegosouzapw/OmniRoute/pull/6248) — thanks @backryun)
+- **chore(release):** ~50 commits on this branch are v3.8.45 pre-flight/hardening fixes and CI-perf work that landed here via the parallel-cycle sync-back (`sync-next-cycle.mjs`, Hard Rule #21) **after** the `v3.8.45` git tag was cut, and are already fully documented under the **[3.8.45]** section below — listed here only so the per-cycle commit-coverage check (`npm run release:uncovered`) doesn't flag them as gaps. Provider/catalog/UX/backend fixes: #6041, #6078, #6108, #6135, #6148, #6149, #6154, #6158, #6161, #6162, #6163, #6164, #6165, #6170, #6177, #6178, #6181, #6186, #6187, #6191, #6193, #6194, #6195, #6200, #6205, #6208, #6209, #6211, #6213, #6223, #6224, #6225, #6226, #6227, #6228, #6229, #6230, #6235, #6291, #6292. CI/release-pipeline work: #6167, #6203, #6214, #6215, #6218, #6273, #6275, #6283, #6284, #6285, #6300, #6305.
+- **chore(release):** additional zero-ref release-cycle plumbing on this branch, kept out of `release:uncovered` on purpose (no `#N` in the commit subject to cite): opening the v3.8.46 cycle, opening/closing the v3.8.45 cycle, the finalized [3.8.45] CHANGELOG i18n sync-back to 42 mirrors, the v3.8.45 cognitive/cyclomatic and file-size drift rebaselines, ESLint stale-suppression pruning (4,273 → 4,233), and clearing test-masking/docs-all pre-flight reds for v3.8.45.
+
+### ⚡ Performance & Infrastructure
+
+- **perf(release-green):** the pre-flight validator (`scripts/quality/validate-release-green.mjs`) now runs its 4 slow suites (unit / vitest / integration / pack-artifact) **concurrently** via `Promise.all` — pre-flight wall time drops from ~the sum of the suites to ~the slowest one (~30min saved per round; Phase 0 was the nº1 bottleneck of the v3.8.45 release benchmark, 2h54 of 6h34 e2e). Guard: `tests/unit/validate-release-green.test.ts` ("runs the slow suites CONCURRENTLY"). ([#6319](https://github.com/diegosouzapw/OmniRoute/pull/6319))
+- **fix(ci):** `scripts/release/sync-next-cycle.mjs` — two defects found live in its first production run (v3.8.45 Phase 5): (1) the `git()` helper's default 1 MiB `maxBuffer` crashed with `ENOBUFS` on `git show origin/main:CHANGELOG.md` (the CHANGELOG alone is >1 MiB) — widened to 64 MiB; (2) the i18n resync only propagated the `[NEXT]` (TBD) section, leaving the just-shipped finalized section as "— TBD" in all 42 mirrors — it now also syncs `[prevVersion]` bounded by the heading below it (new exported pure helper `versionAfter`). Guards: +5 tests in `tests/unit/sync-next-cycle.test.ts` (8/8). ([#6327](https://github.com/diegosouzapw/OmniRoute/pull/6327))
+- **test(ci):** concurrency-sensitive flaky tests are **quarantined into a serial pass** (`tests/unit/serial/`, `--test-concurrency=1`, appended to every unit runner incl. sharded variants — the serial pass is sharded too so concurrent shard jobs never self-collide). Initial set: `glm-coding-plan-monthly-3580`, `quota-division-blocks`, `provider-health-autopilot`, `combo-health-autopilot` — the class behind the ~28min CI wedges/re-runs (two live 1h+ wedges cancelled during this PR's own validation). Discovery + TIA gates track the new glob; systemic root cause (async logger writing after teardown) tracked in [#6360](https://github.com/diegosouzapw/OmniRoute/issues/6360). Guard: `tests/unit/test-serial-quarantine.test.ts` (4). ([#6347](https://github.com/diegosouzapw/OmniRoute/pull/6347))
+
+### 🙌 Contributors
+
+Thanks to everyone whose work landed in v3.8.46:
+
+| Contributor                                                    | PRs / Issues           |
+| -------------------------------------------------------------- | ---------------------- |
+| [@2220258345](https://github.com/2220258345)                   | direct commit / report |
+| [@abdofallah](https://github.com/abdofallah)                   | direct commit / report |
+| [@adentdk](https://github.com/adentdk)                         | direct commit / report |
+| [@aleksesipenko](https://github.com/aleksesipenko)             | direct commit / report |
+| [@anki1kr](https://github.com/anki1kr)                         | direct commit / report |
+| [@anungma](https://github.com/anungma)                         | direct commit / report |
+| [@arpicato](https://github.com/arpicato)                       | direct commit / report |
+| [@backryun](https://github.com/backryun)                       | #6248                  |
+| [@binsarjr](https://github.com/binsarjr)                       | direct commit / report |
+| [@brightfiscalband](https://github.com/brightfiscalband)       | direct commit / report |
+| [@chirag127](https://github.com/chirag127)                     | #6501, #6506           |
+| [@developerjillur](https://github.com/developerjillur)         | direct commit / report |
+| [@dilneiss](https://github.com/dilneiss)                       | #6499                  |
+| [@dtybnrj](https://github.com/dtybnrj)                         | direct commit / report |
+| [@Forcerecon](https://github.com/Forcerecon)                   | direct commit / report |
+| [@hao3039032](https://github.com/hao3039032)                   | direct commit / report |
+| [@Iammilansoni](https://github.com/Iammilansoni)               | #6150, #6245           |
+| [@jmengit](https://github.com/jmengit)                         | direct commit / report |
+| [@jordansilly77-stack](https://github.com/jordansilly77-stack) | direct commit / report |
+| [@JxnLexn](https://github.com/JxnLexn)                         | direct commit / report |
+| [@KooshaPari](https://github.com/KooshaPari)                   | #6166                  |
+| [@loopyd](https://github.com/loopyd)                           | direct commit / report |
+| [@luweiCN](https://github.com/luweiCN)                         | #6221                  |
+| [@makcimbx](https://github.com/makcimbx)                       | direct commit / report |
+| [@muflifadla38](https://github.com/muflifadla38)               | direct commit / report |
+| [@newnol](https://github.com/newnol)                           | direct commit / report |
+| [@ofekbetzalel](https://github.com/ofekbetzalel)               | direct commit / report |
+| [@ohahe52-dot](https://github.com/ohahe52-dot)                 | direct commit / report |
+| [@phidinhmanh](https://github.com/phidinhmanh)                 | direct commit / report |
+| [@powellnorma](https://github.com/powellnorma)                 | direct commit / report |
+| [@qpeyba](https://github.com/qpeyba)                           | direct commit / report |
+| [@RaviTharuma](https://github.com/RaviTharuma)                 | direct commit / report |
+| [@RCrushMe](https://github.com/RCrushMe)                       | direct commit / report |
+| [@rianonehub](https://github.com/rianonehub)                   | #6134, #6204           |
+| [@serverless83](https://github.com/serverless83)               | #6212                  |
+| [@swingtempo](https://github.com/swingtempo)                   | direct commit / report |
+| [@tenshiak](https://github.com/tenshiak)                       | direct commit / report |
+| [@ThongAccount](https://github.com/ThongAccount)               | direct commit / report |
+| [@UnrealAryan](https://github.com/UnrealAryan)                 | direct commit / report |
+| [@vinayakkulkarni](https://github.com/vinayakkulkarni)         | direct commit / report |
+| [@vittoroliveira-dev](https://github.com/vittoroliveira-dev)   | #6233                  |
+| [@warelik](https://github.com/warelik)                         | direct commit / report |
+| [@xxy9468615](https://github.com/xxy9468615)                   | direct commit / report |
+| [@xz-dev](https://github.com/xz-dev)                           | direct commit / report |
+| [@yanpaing007](https://github.com/yanpaing007)                 | direct commit / report |
+| [@diegosouzapw](https://github.com/diegosouzapw)               | maintainer             |
+
+---
+
 ## [3.8.45] — 2026-07-06
 
 ### ✨ New Features

docs/i18n/no/CHANGELOG.md

@@ -6,6 +6,213 @@
 
 ## [3.8.31] — 2026-06-20
 
+## [3.8.46] — TBD
+
+### ✨ New Features
+
+- **feat(sse):** **hide paid-only models from `auto/*` routing** when `hidePaidModels` is on ([#6512](https://github.com/diegosouzapw/OmniRoute/issues/6512)) — follow-up to #6328/#6495. PR #6495 hid paid-only models from the `GET /v1/models` listing, but `auto/*` combos (`auto/best-coding`, `auto/glm`, …) could still pick a paid-only backend into their candidate pool → a 402/403 at request time. `createVirtualAutoCombo` now filters the candidate pool through the new pure `open-sse/services/autoCombo/paidModelFilter.ts` (`filterPaidOnlyCandidates`), applying the same free-model predicate #6495 uses in `catalog.ts` (`providerHasFreeModels(provider) && isFreeModel(provider, {id})`) whenever `settings.hidePaidModels === true`. Applied before the category/tier/family narrowing, so it covers every `auto/*` combo; an all-paid pool degrades to the existing graceful empty-pool path. **Opt-in — default OFF leaves the pool unchanged** (identity). Regression guard: `tests/unit/autoCombo/paid-model-filter-6512.test.ts` (4, incl. the default-off identity guard).
+- **feat(sse):** **provider-family auto combos** — `auto/glm`, `auto/minimax`, `auto/mimo`, `auto/zai`, `auto/gemma`, `auto/llama`, `auto/gemini` ([#6453](https://github.com/diegosouzapw/OmniRoute/issues/6453)) — new routable ids that materialize an on-demand virtual combo spanning whatever installed backends currently expose that model family, degrading gracefully as backends rotate. A new pure `open-sse/services/autoCombo/modelFamily.ts` (`detectModelFamily`) classifies by model-id prefix for six families; `zai` is instead resolved by provider id (z.ai's hosted API serves the same `glm-*` model ids as every other GLM backend, so `auto/zai` means "route to my z.ai backend specifically" vs `auto/glm`'s "any connected GLM backend"). Reuses the existing `createVirtualAutoCombo` on-demand materialization path (no DB writes) and the `/v1/models` catalog advertising loop. Regression guard: `tests/unit/autoCombo/provider-family-combos.test.ts` (11).
+- **feat(proxy):** native **proxy-pool round-robin / egress IP rotation** ([#6365](https://github.com/diegosouzapw/OmniRoute/issues/6365)) — a scope (global / provider / account) can now hold **multiple** proxies as a pool with a rotation strategy, so outbound requests cycle their egress IP instead of pinning one proxy per scope. Migration `117_proxy_pool_rotation.sql` lifts the `UNIQUE(scope, scope_id)` constraint (rebuild via the canonical rename/copy/drop; existing single assignments become 1-element pools) and adds a `proxy_scope_rotation` companion table holding the per-scope strategy + a persisted monotonic round-robin cursor. Strategies: `round-robin` (default, monotonic cursor — never `Math.random`), `random`, and `sticky-per-N-min`. Resolution (`resolveProxyForScopeFromRegistry` / `resolveProxyForConnectionFromRegistry`) now fetches the alive, position-ordered candidate set (unchanged `PROXY_ALIVE_PREDICATE`) and applies the strategy; an empty / all-dead pool still returns `null` — the #6246 fail-closed guard is **untouched** (never falls through to direct egress). Backend + DB only; dashboard pool-builder UI is a follow-up. Regression guard: `tests/unit/proxy-pool-rotation-6365.test.ts` (8, incl. fail-closed + backward-compat).
+- **feat(providers):** end-to-end **tool/function calling on the native Gemini `/v1beta` endpoint** ([#6222](https://github.com/diegosouzapw/OmniRoute/issues/6222)) — both directions of the Gemini↔OpenAI conversion now preserve tool data (previously silently dropped). Request side: `convertGeminiToInternal` (extracted to its own testable module) maps `tools[].functionDeclarations` → OpenAI `tools`, prior `functionCall` parts → assistant `tool_calls`, and `functionResponse` parts → `tool`-role messages. Response side: `convertOpenAIResponseToGemini` emits `parts[].functionCall {name,args}` from `message.tool_calls`, and the streaming `openAIChunkToGeminiChunk` accumulates fragmented `tool_calls` deltas by index into complete `functionCall` parts. The non-Gemini client paths (Claude, OpenAI-Responses) already preserved tool calls — this closes the gap specific to the native Gemini surface. Regression guard: `tests/unit/v1beta-gemini-tool-calling-6222.test.ts` (6, incl. a streaming SSE round-trip).
+- **feat(providers):** copilot-m365-web **enterprise / work tier** support ([#6334](https://github.com/diegosouzapw/OmniRoute/issues/6334)) — mirrors the EDU-tier pattern (#6210): `M365ConnectionParams` gains an `agent` field, a new opt-in `M365_ENTERPRISE_OVERRIDES` preset (`agent=work`, `scenario=officeweb`, `licenseType=Premium`) applies via `providerSpecificData.tier="enterprise"` (alias `"work"`), and `agent` is also overridable directly via `providerSpecificData.agent`. `buildWsUrl` was hardcoding `agent="web"` (the one enterprise-distinguishing param with no override path), so a Premium work account handshook then returned an empty stream. The individual and EDU paths are untouched. Kilo's dup flag vs #6210 (EDU tier) was a false positive — different tier. Regression guard: `tests/unit/copilot-m365-enterprise-6334.test.ts` (7). End-to-end confirmation on a real Premium work account is a live-VPS validation follow-up (Hard Rule #18). (thanks @Forcerecon)
+- **feat(api):** standardized, provider-agnostic **`effort` + `thinking` request params** ([#6241](https://github.com/diegosouzapw/OmniRoute/issues/6241)) — a thin standardization layer over the existing mature per-provider reasoning plumbing (no provider mapper touched). `providerChatCompletionSchema` gains a canonical `effort` (reusing the shared `none/low/medium/high/xhigh` vocabulary — the UI tiers `extra`/`max` collapse onto `xhigh`) and a boolean `thinking`. A pure `normalizeReasoningRequest` (wired once in `src/sse/handlers/chat.ts`, before any reasoning field is read) folds them onto the fields the translators already consume (`reasoning_effort` / `reasoning.effort` / `thinking`), so they fan out to Anthropic / Gemini / xAI / Responses — an explicit client `reasoning_effort` / object-shaped `thinking` always wins (backward-compatible). `/models` additively exposes `supportsThinking` + `effort_tiers` so the frontend can render the toggles (UI component is a follow-up). Regression guard: `tests/unit/effort-thinking-standardization-6241.test.ts` (12). (thanks @Iammilansoni, @shabeer)
+- **feat(combo):** new **`pipeline` (sequential) combo strategy** ([#6297](https://github.com/diegosouzapw/OmniRoute/issues/6297)) — the 18th routing strategy runs targets **in order**, threading each step's output into the next step's input, with an optional per-step `prompt` (system instruction); only the final step's response is returned. Distinct from `fusion` (parallel fan-out + judge). Implemented as a self-contained `open-sse/services/pipeline.ts` (sibling to `fusion.ts`), dispatched from `combo.ts`; the step list reuses `combo.models` order and reads an optional `prompt` off each target (backward-compatible — ignored by every other strategy). Intermediate steps run non-streaming with tools stripped (complete prose to thread forward); the final step keeps the client's `stream` flag + tools. A failing/empty/unparseable intermediate step fails the whole pipeline explicitly via a sanitized error (never silently swallowed). Kilo's dup flag vs #563 was a false positive (that's model→chain selection; this is a sequential chain). Regression guard: `tests/unit/combo-pipeline-strategy.test.ts` (5). (thanks @ofekbetzalel)
+- **feat(ci):** `check:test-masking` now flags **inline-reimplemented prod conditions** ([#6348](https://github.com/diegosouzapw/OmniRoute/issues/6348)) — a new **report-only** subcheck (v2, 6A.10 family) catches the wrong-shape contract test: a test that recomputes the condition under test inline instead of importing/exercising the real function (the #6216 class, where `=== 500` → `>= 500` stayed green because the test re-implemented the branch). For each added/modified test file it warns when the file textually duplicates a ≥3-token conditional from a production file touched in the same PR **and** does not import the symbol/module owning it, via a pure, fixture-tested `findReimplementedConditions()` with an allowlist mirroring `assertReductionAllowlist`. Report-only for now (does not fail the gate) — to be promoted to blocking after a triage cycle. Regression guard: `tests/unit/check-test-masking.test.ts` (45).
+- **feat(sse):** per-connection routing override (native vs CLIProxyAPI) ([#6339](https://github.com/diegosouzapw/OmniRoute/issues/6339)) — the previously-dead `isCliproxyapiDeepModeEnabled` helper is now wired into `resolveExecutorWithProxy`: a single connection can opt itself into the CLIProxyAPI passthrough executor via `providerSpecificData.cliproxyapiMode="claude-native"`, with precedence **connection override > provider `upstream_proxy_config` mode > default**. `resolveExecutorWithProxy` now receives the resolved connection's `providerSpecificData` (threaded from `chatCore.ts`), so one connection can deep-route while the provider's other connections stay native — no DB schema change (the toggle rides in `providerSpecificData`). Also resolves the same-provider mixing ask in #6340. Regression guard: `tests/unit/chatcore-executor-proxy.test.ts` (9). (thanks @RaviTharuma)
+- **feat(dashboard):** "Add session cookie" modal now shows a prominent **"Open ‹host› →"** link to the provider's own site ([#6268](https://github.com/diegosouzapw/OmniRoute/issues/6268)) — every `-web` cookie-session provider (chatgpt-web, claude-web, gemini-web, kimi-web, lmarena, qwen-web, m365-copilot-web, …) renders a one-click external link (opening the provider's login/home page in a new tab) so operators no longer tab away to retype the URL mid-setup. The host resolves from a pure, unit-tested `resolveWebProviderHost()` (prefers `WEB_COOKIE_PROVIDERS[id].website`, falls back to the registry `baseUrl` origin); non-web providers render exactly as before. Kilo's dup flag vs #6265 (modal-too-small-on-1080p) was a false positive — distinct concern. Regression guard: `tests/unit/resolve-web-provider-host.test.ts` (5). (thanks @chirag127)
+- **feat(providers):** add **DigitalOcean** AI (serverless inference) as an OpenAI-compatible API-key provider ([#6373](https://github.com/diegosouzapw/OmniRoute/pull/6373)) — base `https://inference.do-ai.run/v1`, wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/digitalocean/`, `src/shared/constants/providers/apikey/inference-hosts.ts`). Regression guard: `tests/unit/digitalocean-provider.test.ts`. (thanks @newnol)
+- **feat(providers):** add **Huancheng Public API** (`hcnsec`) as an OpenAI-compatible regional provider ([#6410](https://github.com/diegosouzapw/OmniRoute/pull/6410)) — Xinjiang Huancheng Cybersecurity's public LLM platform (base `https://api.hcnsec.cn/v1`, free credits via daily check-ins), wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/hcnsec/`, `src/shared/constants/providers/apikey/regional.ts`). Regression guard: `tests/unit/hcnsec-provider.test.ts`. (thanks @UnrealAryan)
+- **feat(dashboard):** the web-session credential guide now shows an **"Open {host}" link** ([#6316](https://github.com/diegosouzapw/OmniRoute/pull/6316)) to the provider's sign-in site (derived from the provider `website` via `getProviderWebsiteHost`), so you can jump straight to the page where the cookie/session must be captured. Regression guard: `tests/unit/web-session-provider-link-6316.test.ts`. (thanks @jordansilly77-stack)
+- **feat(cerebras):** add the **Gemma 4 31B** model (`gemma-4-31b`) to the Cerebras registry + pricing table ([#6331](https://github.com/diegosouzapw/OmniRoute/pull/6331)). Regression guard extends `tests/unit/t28-model-catalog-updates.test.ts`. (thanks @backryun)
+- **feat(providers):** add **Yuanbao (web)** as a cookie-session provider ([#6196](https://github.com/diegosouzapw/OmniRoute/issues/6196)) — `yuanbao-web` (Tencent Yuanbao, `yuanbao.tencent.com`) with cookie-only auth (`hy_user`/`hy_token` + public agent id), SSE→OpenAI translation incl. `reasoning_content`, exposing DeepSeek V3/R1 + Hunyuan / Hunyuan-T1. Regression guard: `tests/unit/providers-yuanbao-web.test.ts`. `together-web` was **deferred** (no verifiable web-session endpoint — needs a captured request) and `huggingchat-web` **dropped** (the existing `huggingchat` already is a web-cookie provider). (thanks @chirag127)
+- **feat(providers):** route the built-in **agentrouter** through the dynamic Claude-Code wire image ([#6056](https://github.com/diegosouzapw/OmniRoute/issues/6056)) — a small static allow-set (`CC_WIRE_IMAGE_BUILTINS` in `open-sse/services/ccWireImageBuiltins.ts`), consulted by `isClaudeCodeCompatible` / `isClaudeCodeCompatibleProvider` / `applyFingerprint`, makes agentrouter adopt the CC wire-image headers + fingerprint **while guarding the CC baseUrl/auth branches** so it keeps its own registry `baseUrl` and `x-api-key` auth. Regression guard: `tests/unit/agentrouter-cc-wire-image.test.ts` (asserts the wire image is applied AND agentrouter's baseUrl/auth are preserved). Live WAF-acceptance against agentrouter.org is a VPS validation follow-up (Hard Rule #18).
+- **feat(providers):** **bulk-add API keys for Cloudflare Workers AI** ([#6174](https://github.com/diegosouzapw/OmniRoute/issues/6174)) — `cloudflare-ai` is removed from the bulk-add exclusion list and the bulk parser gains a 3-field `name|accountId|apiKey` mode; the bulk route now builds a **per-entry** `providerSpecificData` so each key carries its own `accountId` (fixing the previous shared-object reuse), and both the create + key-validation paths receive it. Regression guard: `tests/unit/bulk-api-key-parser-cloudflare.test.ts`. (thanks @muflifadla38)
+- **feat(dashboard):** routing/settings UX clarity ([#6147](https://github.com/diegosouzapw/OmniRoute/issues/6147)) — (1) weighted combos show the **effective routing share %** next to each weight when weights don't sum to 100 (`WeightTotalBar.tsx`); (2) the status widget's user-facing **"Cloud Sync" label is renamed** to "Remote Settings Sync" (`CloudSyncStatus.tsx`; internal ids/state untouched); (3) built-in providers gain an **opt-in advanced base-URL override** (`isBaseUrlOverrideEligibleProvider`, hidden behind an "Advanced" toggle, reusing the existing `providerSpecificData.baseUrl` persistence — not globally widened). Regression guard: `tests/unit/routing-settings-ux-6147.test.ts`.
+- **feat(combo):** add an option to **disable session stickiness**, per-combo or globally — round-robin / random combos can rotate to a different connection on every request instead of pinning a whole conversation to one connection by its first-message hash. Resolution precedence per-combo `config.disableSessionStickiness` → global `settings.disableSessionStickiness` → default `false` (preserves the #3825 prompt-cache/504 fix); gates **both** stickiness call sites in `open-sse/services/combo.ts`. Exposed as a global toggle (Combo Defaults) and a per-combo Inherit/on/off control. ([#6168](https://github.com/diegosouzapw/OmniRoute/issues/6168)) Regression guard: `tests/unit/combo-disable-session-stickiness.test.ts`. (thanks @RCrushMe)
+- **feat(docker):** add the `OMNIROUTE_NO_SUDO` env flag for root-less / user-namespaced deployments — the MITM cert-trust command path (`resolveSudoSpawn` in `src/mitm/systemCommands.ts`) now strips the leading `sudo` when the flag is truthy, in addition to the existing root / sudo-missing cases, so the Proxy Agent runs without `sudo` (the operator trusts the CA manually, e.g. via `NODE_EXTRA_CA_CERTS`). Argv-array `spawn` preserved — no shell interpolation (Hard Rule #13). ([#6122](https://github.com/diegosouzapw/OmniRoute/issues/6122)) Regression guard: `tests/unit/mitm-systemCommands-no-sudo.test.ts`. (thanks @powellnorma)
+- **feat(providers):** add **Requesty** as an OpenAI-compatible gateway provider (BYOK, base `https://router.requesty.ai/v1`, ~200 free requests/day) — wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/requesty/`, `src/shared/constants/providers/apikey/gateways.ts`). ([#6120](https://github.com/diegosouzapw/OmniRoute/issues/6120)) Regression guard: `tests/unit/requesty-provider.test.ts`. (thanks @chirag127)
+- **feat(dashboard):** add **configured-only / available-only filters** to the Free Provider Rankings page ([#6150](https://github.com/diegosouzapw/OmniRoute/issues/6150)) — hide providers you haven't configured, or whose connections are all rate-limited / out of quota, via server-side query params (`?configuredOnly` / `?availableOnly` on `GET /api/free-provider-rankings`) backed by a testable lib helper reusing the in-process connection state (no Redis). Both filters default off, so the default view is unchanged; this supersedes the earlier client-side "Configured Only" toggle (#6245) with an available-only dimension and unit-tested logic. Regression guard: `tests/unit/freeProviderRankings-filters.test.ts`.
+
+### 🔧 Bug Fixes
+
+- **fix(dashboard):** adding a second API key connection for the same provider no longer silently overwrites the first — the Add-API-key modal now derives a unique default connection name (`main`, then `main-2`, `main-3`, …) so the backend name-based upsert can't collide (#6499 — thanks @dilneiss).
+- **fix(compression):** the session-dedup engine now also deduplicates a large multi-line block repeated **within a single message** (intra-message dedup), not just across turns; the compression-preview API surfaces a `fallbackReason`, and the fusion panel reports how many models were rate-limited vs failed on total-panel failure (#6501 — thanks @chirag127).
+- **fix(compression):** a stacked-pipeline step naming an unregistered engine now surfaces a `validationErrors` entry instead of silently no-op'ing, so misconfigured pipelines are visible in the preview API (#6506 — thanks @chirag127).
+- **feat(usage):** add a **Codex reset-credit redemption** flow to the Provider Limits UI ([#6361](https://github.com/diegosouzapw/OmniRoute/pull/6361)) — a `useCodexResetCreditRedemption` hook + `/api/usage/codex-reset-credit` route + `codexResetCredits` lib let you redeem banked Codex reset credits from the quota card. Regression guard: `tests/unit/codex-reset-credits.test.ts`. (thanks @JxnLexn)
+- **feat(glm):** add **team-plan quota settings** for `glm-cn` connections ([#6351](https://github.com/diegosouzapw/OmniRoute/pull/6351)) — a dedicated `GlmTeamQuotaFields` form section (team quota id / limits) threaded through the Add/Edit connection modals, persisted via `providerSpecificData`, with the GLM usage service reading the team quota. Regression guards: `tests/unit/glm-team-quota.test.ts`, `provider-specific-data-schema.test.ts`. (thanks @hao3039032)
+- **feat(providers):** add **TinyFish** web-fetch/search support ([#6349](https://github.com/diegosouzapw/OmniRoute/pull/6349)) — a `tinyfish-fetch` executor + `/v1/web/fetch` route + MCP web-fetch tool, registered as a specialty-media provider with request-validation and a search-provider catalog entry. Regression guards: `tests/unit/executor-tinyfish-fetch.test.ts`, `web-fetch-handler.test.ts`, `mcp-web-fetch-tool.test.ts`, `provider-validation-tinyfish.test.ts`. (thanks @dtybnrj)
+- **fix(cli):** `omniroute launch-codex` now spawns `codex.cmd` through a shell on Windows (the npm `.cmd` shim is unresolvable by bare `spawn` → ENOENT), mirroring the qodercli Windows fix (#6263) ([#6312](https://github.com/diegosouzapw/OmniRoute/pull/6312)). Regression guard: `tests/unit/launch-codex-windows-spawn-6312.test.ts`. (thanks @swingtempo)
+- **fix(codex):** isolate the **Spark** quota from the shared Codex quota and stabilize the quota UI ordering / hydration so per-scope limits render consistently ([#6336](https://github.com/diegosouzapw/OmniRoute/pull/6336)). Regression guards: `tests/unit/codex-quota-selection-hydration.test.ts`, `provider-limits-ui.test.ts` + 3 more. (thanks @xz-dev)
+- **feat(api):** add a `hidePaidModels` setting that filters paid-only models out of the `/v1/models` catalog. Regression guard: `tests/unit/models-catalog-hide-paid.test.ts`. (thanks @chirag127)
+- **fix(api-manager):** the fallback model picker now preserves combos instead of dropping them when a primary model is unavailable ([#6443](https://github.com/diegosouzapw/OmniRoute/pull/6443)). Regression guard: `tests/unit/api-manager-page-static.test.ts`. (thanks @jmengit)
+- **fix(providers):** recoverable Antigravity / Cloud-Code (Gemini Code Assist) `403` responses ([#6452](https://github.com/diegosouzapw/OmniRoute/pull/6452)) are now classified as a retryable project-config error instead of a terminal account ban, so a fixable project/API-disabled 403 no longer forces a ~1-year cooldown / full OAuth reconnect. Regression guard: `tests/unit/errorclassifier-antigravity-403.test.ts`. (thanks @developerjillur)
+- **fix(mitm):** `sanitizeHeaders` now redacts `Set-Cookie` response headers so upstream session cookies never leak into logs / diagnostics ([#6451](https://github.com/diegosouzapw/OmniRoute/pull/6451)). Regression guard: `tests/unit/mitm-sanitize-headers.test.ts`. (thanks @developerjillur)
+- **fix(api):** `/api/compression/preview` now accepts `mode: "caveman"` and correctly handles stacked / zero-compression previews ([#6425](https://github.com/diegosouzapw/OmniRoute/issues/6425)). Regression guard: `tests/unit/api/compression-preview-caveman-and-stacked-6425.test.ts`. (thanks @chirag127)
+- **feat(providers):** add **Zed** hosted LLM aggregator as a native-app provider ([#6118](https://github.com/diegosouzapw/OmniRoute/pull/6118)) — OAuth sign-in via the Zed hosted flow, registered through the shared provider registry + executor. Regression guards: `tests/unit/zed-oauth-provider.test.ts`, `zed-import-utils.test.ts`, `zed-docker-detect.test.ts`, `mitm-handler-zed.test.ts`. VPS-validated via live operator login (Hard Rule #18).
+- **fix(oauth):** the Kiro SSO-cache auto-import now **preserves the IDC region** — cross-region Amazon Q / Kiro profiles imported from the SSO cache are no longer collapsed to the default region ([#6113](https://github.com/diegosouzapw/OmniRoute/pull/6113)). Regression guard: `tests/unit/kiro-auto-import-idc-2059.test.ts`. VPS-validated via live operator login (Hard Rule #18).
+- **fix(dashboard):** passthrough model aliases no longer collide when two namespaced model ids share a last segment (port from 9router#1850, [#6431](https://github.com/diegosouzapw/OmniRoute/pull/6431)). `enx/gpt-5.5` and `enx/codebuddy/gpt-5.5` both auto-generated the alias `gpt-5.5`, so the second model could never be added (the UI just alerted "alias already exists"). Aliases are now disambiguated deterministically — bare last segment when free, then parent-qualified (`codebuddy-gpt-5.5`), then a numeric suffix — while re-adding the exact same model id is still blocked. Regression guard: `tests/unit/passthrough-alias-1850.test.ts`. (thanks @arpicato)
+- **fix(translator):** preserve a Gemini `functionResponse` co-located with other parts (another `functionCall`, or trailing `text`) in the same content when translating **Gemini → OpenAI** ([#6376](https://github.com/diegosouzapw/OmniRoute/pull/6376)). `convertGeminiContent()` early-returned the tool message on the first `functionResponse` part, dropping any co-located parts; such contents are now pre-split (one tool message per `functionResponse`, emitted first, plus one message for the remaining parts). Regression guard: `tests/unit/gemini-to-openai-function-response.test.ts`. (thanks @warelik)
+- **fix(headroom):** detect a python interpreter managed by **mise / pyenv / asdf / conda** (port from 9router#2353, [#6382](https://github.com/diegosouzapw/OmniRoute/pull/6382)). Headroom's python probe (`src/lib/headroom/detect.ts`) searched a hardcoded `PATH`, but version managers expose their interpreters via shim dirs that only join `PATH` through interactive-shell activation — which the non-interactive server never runs, so a managed python (≥3.10) was invisible and Headroom reported it missing. The search path now prepends the well-known shim/bin dirs (`~/.local/share/mise/shims`, `~/.pyenv/shims`, `~/.asdf/shims`, `$CONDA_PREFIX/bin`, `~/.local/bin`, respecting `MISE_DATA_DIR`/`PYENV_ROOT`/`ASDF_DATA_DIR` when set), and a new `HEADROOM_PYTHON` env override lets operators point straight at their interpreter (mirroring `HEADROOM_URL`). Still shell-free (`execFileSync`). Regression guard: `tests/unit/headroom-detect.test.ts` (5). (thanks @loopyd)
+- **fix(executors):** strip the OpenAI-Codex/Claude-CLI `client_metadata` passthrough field for **NVIDIA** requests (port from 9router#1887, [#6411](https://github.com/diegosouzapw/OmniRoute/pull/6411)). NVIDIA's OpenAI-compatible wrapper rejects it with `400 Unsupported parameter`, the same class already handled for `cerebras`/`mistral`; `nvidia` (executor `default`) was missing from the strip allowlist so Codex/Claude-Code passthrough requests 400'd. Regression guard: `tests/unit/executor-default-strip-client-metadata.test.ts` (+nvidia case). (thanks @phidinhmanh)
+- **fix(translator):** strip the Claude-style `thinking` field for **NVIDIA `z-ai/glm-5.2`** (port from 9router#2023, [#6413](https://github.com/diegosouzapw/OmniRoute/pull/6413)). NVIDIA's OpenAI-compatible wrapper 400s on `thinking` (a Claude-format client routed here leaves a `thinking:{type:"adaptive"}`); the existing strip rule only dropped `reasoning`. Same class already handled for `minimax-m2.7`. Regression guard: `tests/unit/nvidia-minimax-thinking-strip.test.ts` (+glm-5.2 case). (thanks @phidinhmanh)
+- **fix(translator):** suppress the streamed `</think>` close marker for the **Antigravity IDE** client (port from 9router#1061, [#6415](https://github.com/diegosouzapw/OmniRoute/pull/6415)). On thinking-only turns Antigravity rendered a bare `</think>` as the sole visible content, tripping its loop-detection and wasting requests. Antigravity's UA (`vscode/<v> (Antigravity/<v>)`) is added to the marker-suppress allowlist (alongside OpenCode); Claude Code / Cursor still get the marker, and `x-omniroute-thinking-marker: on` force-restores it. Regression guard: `tests/unit/think-close-marker-suppress-5245.test.ts`. (thanks @abdofallah)
+- **fix(executors):** strip nested `reasoning_content` from messages for **Mistral** (port from 9router#1649, [#6417](https://github.com/diegosouzapw/OmniRoute/pull/6417)). Mistral's API returns `422 extra_forbidden` when an assistant message carries `reasoning_content` (replayed thinking from a prior turn, e.g. via the Codex `/responses` path); the generic top-level 400 field-downgrade retry never covered the nested per-message field. `DefaultExecutor` now strips it for provider `mistral` only, so DeepSeek (which requires replayed `reasoning_content`) is unaffected. Regression guard: `tests/unit/mistral-strip-reasoning-content-1649.test.ts`. (thanks @xxy9468615)
+- **fix(executors):** strip the `client_metadata` passthrough field on the **OpenCode** path (port from 9router#1442, [#6418](https://github.com/diegosouzapw/OmniRoute/pull/6418)). OpenCode upstreams (e.g. `kimi-k2.6` via opencode-go) reject it with `400 "Extra inputs are not permitted, field: 'client_metadata'"`; the DefaultExecutor strip only covered cerebras/mistral and `OpencodeExecutor` extends `BaseExecutor` directly, so nothing removed it there. Regression guard: `tests/unit/opencode-strip-client-metadata-1442.test.ts`. (thanks @yanpaing007)
+- **fix(executors):** inject the `reasoning_content` echo for the native **Moonshot Kimi** provider (port from 9router#1480, [#6419](https://github.com/diegosouzapw/OmniRoute/pull/6419)). Kimi (executor `default`) is a thinking-mode upstream that 400s with "reasoning_content must be passed back" when a prior assistant turn lacks it; the placeholder injection was only wired into the OpenCode meta-provider, so direct multi-turn Kimi conversations failed. Scoped to `kimi` (gateway-served models matching the thinking-model name pattern are unaffected). Regression guard: `tests/unit/kimi-native-reasoning-injected-1480.test.ts`. (thanks @2220258345)
+- **fix(executors):** recover from a strict gateway's `context_management: Extra inputs are not permitted` 400 (port from 9router#1468, [#6420](https://github.com/diegosouzapw/OmniRoute/pull/6420)). **Claude Code** always sends a top-level `context_management` field; strict anthropic-compatible gateways reject it. The dedicated context-editing 400-fallback only fired when OmniRoute's own `contextEditing` feature was enabled (default off), so a client-sent field passed through untouched and 400'd. `context_management` is now in the generic reactive field-strip list, so it's stripped-and-retried once regardless of the feature flag (with correct request re-signing for claude-compatible relays). Regression guard: `tests/unit/provider-field-strips.test.ts`. (thanks @ohahe52-dot)
+- **fix(network):** enable **RFC 8305 Happy Eyeballs** (`autoSelectFamily`) on the direct-egress undici dispatcher (port from 9router#1237, [#6423](https://github.com/diegosouzapw/OmniRoute/pull/6423)). When DNS returns both IPv6 (AAAA) and IPv4 (A) and the IPv6 route is broken (e.g. a NAT64 `64:ff9b::` prefix without routing), undici tried IPv6 first and hung until `ETIMEDOUT` (then a 502 + account lockout), even though `curl` reached the same host. The direct dispatcher now races both families and uses whichever connects first. Proxy paths pin family via `proxyTls` and are unaffected. Regression guard: `tests/unit/direct-dispatcher-pipelining-4580.test.ts`. (thanks @adentdk)
+- **fix(combo):** round-robin now advances the rotation pointer past the model that **actually served**, not the eagerly-scheduled one (port from 9router#948, [#6428](https://github.com/diegosouzapw/OmniRoute/pull/6428)). With `stickyLimit: 1` (true round-robin), when the scheduled model failed and a _different_ model served via fallback, the counter had already advanced +1 from the scheduled index — so the next request reused the fallback-served model, degrading round-robin into hot-spotting on whichever model was healthy. The pointer now advances to the served index + 1 (mirroring the sticky-limit>1 path). Session-stickiness (#3825) and distribution are preserved. Regression guard: `tests/unit/combo-rr-fallback-advance-948.test.ts`. (thanks @binsarjr)
+- **fix(sse):** a non-string `model` field is now rejected with a `400` before the resolver, instead of crashing downstream `.toLowerCase()`/`.split()` calls into an empty-body `500` that escapes the error sanitizer ([#6407](https://github.com/diegosouzapw/OmniRoute/issues/6407)). Regression guard: `tests/unit/chat-non-string-model-6407.test.ts`. (thanks @chirag127)
+- **fix(api):** unknown `/api/*` routes now return a JSON `404` (instead of the dashboard HTML shell) and scalar chat params (`model`/`temperature`/etc.) are validated **before** the provider lookup so malformed requests fail fast with a clear `400` ([#6424](https://github.com/diegosouzapw/OmniRoute/issues/6424), [#6412](https://github.com/diegosouzapw/OmniRoute/issues/6412)). Regression guards: `tests/unit/api/api-catchall-json-404.test.ts`, `tests/unit/chat-early-schema-validation-6412.test.ts`. (thanks @chirag127)
+- **fix(api):** `/v1/chat/completions` now rejects a non-JSON `Content-Type` with a `400` before parsing the body ([#6414](https://github.com/diegosouzapw/OmniRoute/issues/6414)). Regression guard: `tests/unit/v1-chat-completions-content-type-6414.test.ts`. (thanks @chirag127)
+- **fix(api):** the `X-OmniRoute-Compression` response header is now echoed on `/v1/chat/completions` and `/v1/completions` ([#6422](https://github.com/diegosouzapw/OmniRoute/issues/6422)). Regression guard: `tests/unit/compression-header-echo-6422.test.ts`. (thanks @chirag127)
+- **fix(api):** concurrent `GET /v1/models` requests are coalesced into a single catalog build ([#6408](https://github.com/diegosouzapw/OmniRoute/issues/6408)). Regression guard: `tests/unit/v1-models-concurrent-6408.test.ts`. (thanks @chirag127)
+- **fix(api):** `/v1/completions` now echoes the requested `body.model` in its JSON + streamed responses ([#6429](https://github.com/diegosouzapw/OmniRoute/pull/6429)). Regression guard: `tests/unit/completions-body-model-echo.test.ts`. (thanks @chirag127)
+- **fix(api):** env-var master keys now see the full `/v1/models` catalog ([#6406](https://github.com/diegosouzapw/OmniRoute/issues/6406)). Regression guard: `tests/unit/models-catalog-envkey-6406.test.ts`. (thanks @chirag127)
+- **fix(api):** non-streaming `/v1/completions` responses now echo `body.model` aligned with the `X-OmniRoute-Model` header ([#6426](https://github.com/diegosouzapw/OmniRoute/issues/6426)). Regression guard: `tests/unit/v1-completions-model-header-match-6426.test.ts`. (thanks @chirag127)
+- **fix(api):** unknown `/v1/*` routes now return a JSON `404 not_found` instead of the Next.js dashboard HTML shell ([#6405](https://github.com/diegosouzapw/OmniRoute/issues/6405)). Regression guard: `tests/unit/api/v1-catchall-json-404.test.ts`. (thanks @chirag127)
+- **fix(api):** the per-connection provider models route now degrades to the shipped catalog when a provider's `/models` endpoint answers with a redirect ([#6267](https://github.com/diegosouzapw/OmniRoute/issues/6267)) — a `qwen-web` import failed with a raw `Redirect blocked … (307)` 503. `safeOutboundFetch` throws `REDIRECT_BLOCKED` on the 307, `getSafeOutboundFetchErrorStatus` maps it to 503, and `buildDiscoveryErrorFallbackResponse` treated every 503 as a hard error — so the non-empty `getModelsByProviderId("qwen-web")` catalog was never surfaced. A models-endpoint redirect is not a fixable-config error (unlike `URL_GUARD_BLOCKED`/`INVALID_URL`, which stay hard errors), so it now falls back to the local/cached catalog before the 503 short-circuit. General fix — covers any config-driven provider that 307s. Regression guard: `tests/unit/provider-models-qwen-web-redirect-6267.test.ts`. (thanks @chirag127)
+- **fix(api):** the per-connection provider models route (MCP `list_models_catalog` + the dashboard import view) now merges USER-ADDED custom models into its response ([#6247](https://github.com/diegosouzapw/OmniRoute/issues/6247)) — custom models live in the `key_value` namespace `customModels`, which the live REST `/api/v1/models` already merges, but `src/app/api/providers/[id]/models/route.ts` never read `getCustomModels`, so custom models were dropped on both the discovery-success and local_catalog paths. They are now folded into the returned model list (deduped by id, stamped `owned_by: provider`), fixing MCP + the dashboard import view in one place. Regression guard: `tests/unit/provider-models-custom-merge-6247.test.ts`. (thanks @RCrushMe)
+
+- **fix(providers):** GitLab Duo tool-calling follow-up turns no longer fail upstream with `422 {"detail":"Validation error"}` (tokens 0/0, rejected pre-inference). The #6234 tool-result-feedback fix serialized the **entire** multi-turn conversation into GitLab's single-file `code_suggestions` (`small_file`) generation endpoint — folded history that turn-N sent as an oversized `current_file.content_above_cursor` **and** duplicated verbatim into `user_instruction`, tripping the AI-Gateway's `small_file` validation guard. The executor now **bounds** that prompt: it keeps system + latest user message + the most-recent tool round (dropping older turns), caps oversized tool results, and stops duplicating the full prompt into `user_instruction` (which now carries only the short latest user message) — while still feeding the most-recent tool result back so the agent continues (`open-sse/executors/gitlab.ts`, [#6220](https://github.com/diegosouzapw/OmniRoute/issues/6220)). The unit test covers the bounding logic; the upstream 422→200 clearing is VPS-only (Hard Rule #18). Regression guard: `tests/unit/gitlab-tool-exchange-bounded-6220.test.ts`.
+
+- **fix(i18n):** the provider-detail (`/dashboard/providers/[id]`) connection-status filter labels no longer render as `__MISSING__:All` / `__MISSING__:Active` / `__MISSING__:Error` / `__MISSING__:Banned` / `__MISSING__:CreditsExhausted` in non-English locales (notably pt-BR) ([#6290](https://github.com/diegosouzapw/OmniRoute/issues/6290)). Root cause was **not** the namespace mismatch the issue guessed — the `providers.filter*` keys resolve correctly in `en.json`; the debt lived in the locale mirrors (`src/i18n/messages/*.json`), where these five keys carried the `__MISSING__:` sync sentinel in ~15 locales and were absent entirely in ~26 others, so next-intl found the key and echoed the sentinel verbatim. All 40 non-English/-Chinese mirrors now ship real translations for the five `providers.filter*` labels. Regression guard: `tests/unit/i18n-provider-filter-keys-6290.test.ts`. (thanks @diegosouzapw)
+
+- **fix(providers):** the `copilot-m365-web` streaming executor now emits `debug`-level WebSocket diagnostics ([#6210](https://github.com/diegosouzapw/OmniRoute/issues/6210)) — the outbound WS URL (with the `access_token` **redacted** via `redactWsUrl()`), handshake success/failure, and each received SignalR frame's `type`/`target`. Previously the streaming path logged nothing, so an empty `content:null` response (the M365 Education / Starter tier symptom fixed in #6234) was undiagnosable even at `APP_LOG_LEVEL=debug`. The change is debug-level and side-effect-free — it does not alter streaming behavior or the frame parser, and the token never reaches the logs. Regression guard: `tests/unit/copilot-m365-web-logging-6210.test.ts` (thanks @qpeyba)
+
+- **fix(resilience):** a round-robin combo no longer returns `503 all upstream accounts are unavailable` when a compatibility-rejected target is actually healthy ([#6238](https://github.com/diegosouzapw/OmniRoute/issues/6238)). `filterTargetsByRequestCompatibility` drops request-incompatible targets (tool/vision/structured-output unsupported, or below the required context window) **before** any availability check runs, and its `compatible.length === 0` safety net only fired when _all_ targets were filtered — not when the kept targets later all turned out runtime-unavailable (circuit-open / cooldown / no credentials). So a combo could 503 while a compat-rejected-but-healthy provider sat unused. `handleRoundRobinCombo` now keeps the compat-rejected set and, when every compat-kept target was skipped without a single real attempt, probes those rejected targets as a **last-resort fallback tier** (via the new pure `open-sse/services/combo/comboCompatFallback.ts`) before crystallizing the 503. Regression guard: `tests/unit/combo-roundrobin-compat-fallback-6238.test.ts`. (thanks @ThongAccount)
+
+- **fix(startup):** best-effort self-heal for a corrupted Turbopack dev cache on Windows ([#6289](https://github.com/diegosouzapw/OmniRoute/issues/6289)). On Windows, `pnpm dev` can fail at startup when Turbopack `mmap`s a persistent-cache SST file and the OS refuses the mapping (`os error 1455` — "paging file too small"), which Turbopack surfaces as a misleading `Module not found: Can't resolve '@/shared/utils/machine'`. This is a **known upstream Turbopack cache-corruption bug — not our code**. The dev launcher (`scripts/dev/run-next.mjs`) now wraps `nextApp.prepare()` and, when it rejects with that signature (`isTurbopackCacheCorruption` in the new `scripts/dev/turbopackCacheHeal.mjs`), purges `.build/next/**/cache/turbopack` and retries **once** with a clear log. **Caveat — best-effort only:** the corruption often surfaces as a runtime overlay rather than a `prepare()` rejection, so this cannot always intercept it; the reliable remedy remains manually deleting the Turbopack cache dir. Regression guard: `tests/unit/turbopack-cache-heal-6289.test.ts`. (thanks @chirag127)
+
+- **fix(providers):** qodercli PAT auth no longer fails with `spawn qodercli ENOENT` on Windows ([#6263](https://github.com/diegosouzapw/OmniRoute/issues/6263)) — `spawnQoderCli` spawned the bare `qodercli` name with `shell:false` and an unenriched env, so the npm `.cmd` wrapper under `%APPDATA%\npm` (a user-PATH directory) was never resolved. It now resolves the absolute `.cmd`/`.exe` path through the existing `getCliRuntimeStatus("qoder")` resolver in `src/shared/services/cliRuntime.ts` (memoized), spawns with `shell` when the target is a `.cmd`/`.bat`, and uses the cliRuntime-enriched env (PATH + PATHEXT + APPDATA); the ENOENT error now lists the searched paths plus the `CLI_QODER_BIN` override. End-to-end spawn on a real Windows host is host-only (Hard Rule #18); the path-resolution logic is unit-tested. Regression guard: `tests/unit/qodercli-windows-resolve-6263.test.ts`. (thanks @chirag127)
+
+- **fix(sse):** the reasoning-token buffer no longer inflates **probe-sized `max_tokens`** ([#6274](https://github.com/diegosouzapw/OmniRoute/issues/6274)) — Claude Code's `/model` capability check sends `max_tokens: 1`, but for a thinking-capable model with a large output cap (e.g. `glm-5.2`) the #3587 headroom heuristic (`max(current + 1000, ceil(current * 1.5))`) rewrote it to `1001` and forwarded that upstream, wasting tokens on a request that was never a genuine reasoning budget. `resolveReasoningBufferedMaxTokens()` (`open-sse/services/reasoningTokenBuffer.ts`) now short-circuits and returns the caller's value verbatim when it is below the new `REASONING_BUFFER_MIN_TRIGGER` (256) threshold — a tiny explicit limit is a probe, not a reasoning request. Real budgets still receive the #3587 headroom unchanged, and the guard runs after the existing capability checks so unknown / non-reasoning models keep returning `null`. Regression guard: `tests/unit/reasoning-token-buffer-6274.test.ts`. (thanks @brightfiscalband)
+
+- **fix(cli):** `omniroute reset-password` now works as a real subcommand, and password resets over piped (non-TTY) stdin actually apply ([#6261](https://github.com/diegosouzapw/OmniRoute/issues/6261), [#6258](https://github.com/diegosouzapw/OmniRoute/issues/6258)). Two coupled defects: (1) **#6261** — `bin/omniroute.mjs` routed everything through Commander with only two pre-Commander bypasses (`--mcp`, `reset-encrypted-columns`), so `omniroute reset-password` was rejected as an unknown command; only the separate `omniroute-reset-password` bin worked, while the docs falsely advertised the subcommand (incl. a bogus "legacy alias still works"). A pre-Commander bypass mirroring `reset-encrypted-columns` now dynamically imports `bin/reset-password.mjs` (which self-executes) before Commander parses; the three doc lines were corrected. (2) **#6258** — `bin/reset-password.mjs` issued two sequential `rl.question` prompts; under piped stdin the second read never settled at EOF, so `main()` never reached `resetManagementPassword` and the reset was a silent no-op (both prompts printed, no success, password unchanged). The CLI now detects non-TTY stdin and reads it once (first line = password, second line = confirm if present, else reused), adds a `--password-stdin` flag (entire stdin is the password, no confirmation), and exits `0` explicitly so the success line always flushes; interactive TTY behavior is unchanged. Regression guard: `tests/unit/reset-password-cli-6261-6258.test.ts` (3). (thanks @chirag127)
+
+- **fix(db):** the mass-migration **safety abort** now tells the operator how to bypass it and stops flooding the log ([#6260](https://github.com/diegosouzapw/OmniRoute/issues/6260)) — after restoring a backup that wiped the migration tracking table, `runMigrations()` threw the abort on every downstream `ensureDbInitialized()`, re-logging the full banner 11+ times, and the message never mentioned the existing `OMNIROUTE_MAX_PENDING_MIGRATIONS` escape hatch. The abort text now appends a bypass hint (set `OMNIROUTE_MAX_PENDING_MIGRATIONS=0` in `server.env` / `DATA_DIR/.env`), and a new `MigrationSafetyAbortError` is memoized so repeated calls in the same process throw the same instance and emit a single concise line instead of the full cascade. Regression guard: `tests/unit/migration-safety-abort-6260.test.ts`. (thanks @chirag127)
+
+- **fix(auth):** importing a **distinct** Codex/ChatGPT OAuth `auth.json` is no longer falsely rejected as "already exists" when it belongs to a different user in the same workspace ([#6301](https://github.com/diegosouzapw/OmniRoute/issues/6301)). `findExistingCodexConnection` (in `src/lib/oauth/utils/codexAuthImport.ts`) deduped **only** on `providerSpecificData.workspaceId === accountId`, where `accountId` is the shared `chatgpt_account_id`/`tokens.account_id` — so two members of the same ChatGPT Team collapsed onto a single connection (409 `duplicate_account`). The id_token's `https://api.openai.com/auth` claim carries a per-user `chatgpt_user_id` alongside the workspace id (the device-flow path already persisted it as `chatgptUserId`, but the import path did not). Now `parseAndValidateCodexAuth` extracts `userId` (`chatgpt_user_id` → `user_id` → JWT `sub`) into `ParsedCodexAuth`, the create/update paths persist `chatgptUserId` in `providerSpecificData` (mirroring `codex.ts`), and dedup keys on `workspaceId` **AND** `chatgptUserId` — with a backward-compat fallback to legacy accountId-only matching when no stored connection for that workspace records a `chatgptUserId`, so genuinely-same accounts still dedup. Regression guard: `tests/unit/codex-auth-import-userid-dedup-6301.test.ts` (4). (thanks @anungma)
+
+- **fix(providers):** importing models for the **venice-web** provider no longer fails with a red "Provider venice-web does not support models listing" ([#6269](https://github.com/diegosouzapw/OmniRoute/issues/6269)). `venice-web` is a web-cookie provider with an executor but no upstream `/v1/models` endpoint and no registry `models`, so the models route fell through to the tail `400`. Mirroring the `jules`/`linkup-search`/`ollama-search` fix (#5569), it now ships a static local catalog entry in `src/lib/providers/staticModels.ts` — seeding the current Venice lineup (`venice-uncensored`, `llama-3.3-70b`, `qwen3-235b`, `qwen3-4b`, `deepseek-r1-671b`; Venice rotates its catalog, see docs.venice.ai/models/overview) — so the route returns `200` with `source:"local_catalog"`, `intentional:true`. Regression guard: `tests/unit/static-models-venice-web-6269.test.ts`. (thanks @chirag127)
+
+- **fix(api):** the specialty model catalogs (`/v1/embeddings`, `/v1/images`, `/v1/music`, `/v1/videos` model lists) are now derived from the **unified catalog filtered by a predicate** (`getSpecialtyModelsResponse`) instead of ad-hoc per-route logic, so they consistently respect active-credential visibility and stay in sync with the main catalog ([#6303](https://github.com/diegosouzapw/OmniRoute/pull/6303)). Regression guard: `tests/unit/specialty-model-catalog-routes.test.ts`. (thanks @makcimbx)
+- **fix(api):** the agent-bridge server route now resolves the MITM manager via a **dynamic `import("@/mitm/manager.runtime")`** so Turbopack does not statically pull the stub (or over-bundle the manager), and the agent-skills generator anchors its output base path with `path.join(process.cwd(), …)` so Turbopack's static analyzer stops tracing the whole project root ([#6329](https://github.com/diegosouzapw/OmniRoute/issues/6329), [#6366](https://github.com/diegosouzapw/OmniRoute/pull/6366)). Regression guard: `tests/unit/agent-bridge-server-route-dynamic-import.test.ts`. (thanks @Iammilansoni)
+- **fix(api):** internal probes (combo-test, cloud-sync verify) now pick a **management-scoped / allow-all API key** instead of naively grabbing `getApiKeys()[0]` — a restricted `self:usage` first row made the probe fail with "Model X is not allowed for this API key" even when the combo path was healthy (`pickApiKeyForInternalUse` in `src/lib/db/apiKeys.ts`). The API-manager model editor also falls back to `/api/models?all=true` when `/v1/models` is catalog-protected ([#6372](https://github.com/diegosouzapw/OmniRoute/pull/6372)). Regression guard: `tests/unit/pick-internal-api-key-6372.test.ts`. (thanks @jmengit)
+- **fix(live-ws):** the Live Dashboard WebSocket server now **rejects on bind failure** (e.g. `EADDRINUSE` when the API bridge already holds the port) instead of letting the error surface as an unhandled `error` event that crash-loops the process — the `error` listener is attached to `wss` (not `server`) and releases the EventBus subscription on a failed start ([#6324](https://github.com/diegosouzapw/OmniRoute/issues/6324)). Regression guard: `tests/unit/live-ws-eaddrinuse-6324.test.ts`. (thanks @vinayakkulkarni)
+- **fix(dashboard):** the Home provider-topology widget now trusts the live provider-metrics snapshot — it uses `topology.errorProvider` and live `activeRequests` directly instead of re-deriving state from a stale `lastErrorAt` or applying a frontend timeout filter, so the topology reflects real-time provider health ([#6322](https://github.com/diegosouzapw/OmniRoute/pull/6322)). Regression guard: `tests/unit/home-provider-topology-live-state.test.ts`. (thanks @xz-dev)
+- **fix(sse):** strip zero-width markers from streamed **tool-call arguments** — a follow-up to [#5857](https://github.com/diegosouzapw/OmniRoute/pull/5857). That PR removed injected zero-width joiners (U+200D) from streamed assistant text/reasoning but deliberately left tool-call argument JSON byte-exact. The request-side obfuscation (`open-sse/services/claudeCodeObfuscation.ts`) injects ZWJ into agent words — including the temp path inside the Bash tool description — and Claude models copy that verbatim into generated commands, which are delivered as tool-call arguments rather than assistant text. As a result the ZWJ survived and corrupted code blocks (e.g. a temp path rendered with an invisible joiner). Now `open-sse/handlers/responseSanitizer.ts` strips zero-width code points from tool-call argument strings at every emit site (OpenAI non-stream/stream chat `tool_calls` + legacy `function_call`, native Responses `function_call` items, the OpenAI→Responses conversion, and the native Responses streaming `response.function_call_arguments.delta/.done` events). Only zero-width code points are removed; JSON structure and all other bytes stay identical (no parse/restringify), so normal arguments remain byte-exact. Regression guard: 6 new cases in `tests/unit/response-sanitizer.test.ts` (suite 50/50).
+
+- **fix(nodejs):** the default app log path now resolves under `DATA_DIR` (`~/.omniroute/logs/application/app.log`) instead of `process.cwd()` ([#6197](https://github.com/diegosouzapw/OmniRoute/issues/6197)) — the globally-installed CLI runs from an arbitrary working directory, so anchoring the default to cwd made file logging silently write to (or no-op under) an unrelated directory, contradicting the documented `.env.example` default. `getAppLogFilePath()` now computes the default lazily via the pure `resolveDataDir()` resolver (honours a per-process `DATA_DIR`, no directory-creation side effect); an explicit `APP_LOG_FILE_PATH` still wins. Regression guard: `tests/unit/logenv-datadir-path-6197.test.ts` (3).
+
+- **fix(docker):** AgentBridge/`startMitm` no longer aborts in containers/headless when the Antigravity-default DNS step can't write `/etc/hosts` ([#6127](https://github.com/diegosouzapw/OmniRoute/issues/6127)), and the privileged command's stderr now reaches `app.log` instead of only a bare exit code hitting the toast ([#6198](https://github.com/diegosouzapw/OmniRoute/issues/6198)). The default DNS step (`addDNSEntry`) was called unguarded while cert install and the two sibling DNS steps were each best-effort — in the runtime Docker image (`USER node`, no `sudo`, read-only `/etc/hosts`) it threw `Command failed with code 1` out of `startMitmInternal` and killed the whole start, discarding the stderr. The three DNS steps are extracted into a best-effort `provisionDnsEntries()` where each failure is logged with the full `err` (stderr included, folded in by `systemCommands.ts`) and never aborts the start. Regression guard: `tests/unit/mitm-dns-graceful-degrade-6127.test.ts` (4).
+
+- **fix(providers):** copilot-m365-web now supports the M365 Education "Starter / OfficeWebIncludedCopilot" tier and no longer returns an empty `content:null` stream ([#6210](https://github.com/diegosouzapw/OmniRoute/issues/6210)). Two gaps: (1) `buildWsUrl()` hardcoded the individual-consumer scenario (`OfficeWebPaidConsumerCopilot`, `isEdu=false`) — the EDU tier is now opt-in via `providerSpecificData.tier="edu"`, emitting `scenario=OfficeWebIncludedCopilot`/`isEdu=true` (the individual path is unchanged); (2) the EDU/GPT-5.5 path streams deltas via `arguments[0].writeAtCursor` (incremental) instead of only `messages[].text` (accumulated snapshots), which the parser dropped — a new `accumulateBotContent()` folds both formats, with `type:2 item.result.message` as a last-resort fallback. Regression guard: `tests/unit/copilot-m365-edu-writeatcursor-6210.test.ts` (10). (thanks @qpeyba)
+
+- **fix(providers):** GitLab Duo executor now feeds tool results back into the prompt instead of looping ([#6220](https://github.com/diegosouzapw/OmniRoute/issues/6220)) — `buildPrompt()` branched only on `system`/`user` and took `userParts.at(-1)`, silently dropping the `assistant{tool_calls}` + `tool{result}` turns the client appended, so the reconstructed prompt was byte-identical to turn 1 and the model re-emitted the same `<tool>` call forever. When a tool exchange is present the full conversation is now serialized, folding each tool result back keyed by its `tool_call_id`; simple conversations keep the legacy shape. Complements the tool_call emission from [#6051](https://github.com/diegosouzapw/OmniRoute/issues/6051) (the `kilo-duplicate` label was a false positive — different, sequential defect). Regression guard: `tests/unit/gitlab-tool-result-feedback-6220.test.ts` (4).
+
+- **fix(providers):** opencode-go/opencode-zen can now synthesize the OpenCode CLI identity headers Cloudflare requires on VPS egress ([#5997](https://github.com/diegosouzapw/OmniRoute/issues/5997)) — on a datacenter VPS, `opencode.ai/zen/go/v1/chat/completions` 403s (HTML challenge) requests lacking CLI identity, while the reporter's control curl proved that `User-Agent: opencode-cli/1.0.0` + `x-opencode-client: cli` + `x-opencode-project: default` + fresh request/session UUIDs succeed. Opt-in via `OPENCODE_SYNTHESIZE_CLI_HEADERS=true` (values overridable via `OPENCODE_GO_USER_AGENT`/`OPENCODE_USER_AGENT`/`OPENCODE_CLIENT`/`OPENCODE_PROJECT`); it fills only headers the client did not already send. Kept **off by default** — the forward-only path is deliberate (fabricating a wrong value risks upstream rejection; a prior dedup regressed with `opencode/local`), so this replaces the fragile local header-injection shim without changing default behavior. Regression guard: `tests/unit/opencode-cli-headers-synthesis-5997.test.ts` (6). (thanks @aleksesipenko)
+- fix(resilience): sticky session affinity now evicts and fails over to another account when the pinned account is exhausted/unavailable (#6219)
+
+- fix(sse): Responses API passthrough now drops internal commentary-phase output before forwarding to clients (gated by RESPONSES_PASSTHROUGH_DROP_COMMENTARY, default on) (#6199)
+
+- **fix(sse):** tool-call function schemas with a root `type: null` are now coerced to `type: "object"` before dispatch ([#6359](https://github.com/diegosouzapw/OmniRoute/issues/6359)) — clients like the Codex app emit `parameters: { type: null, ... }` for some tools, which OpenAI-compatible upstreams reject with `400 Invalid schema for function '...': schema must be a JSON Schema of 'type: "object"', got 'type: null'`, failing the whole request. `toolSchemaSanitizer` already stripped the null; it now re-adds the mandatory root `"object"` type (and empty `properties`/open `additionalProperties` when absent). Combinator roots (`anyOf`/`oneOf`/`allOf`) and explicit root types are left untouched. Regression guard: 5 new cases in `tests/unit/tool-schema-sanitizer.test.mjs`.
+
+- **fix(docker):** AgentBridge no longer fails to start on npm/Electron/VPS installs with "MITM manager stub reached at runtime" ([#6344](https://github.com/diegosouzapw/OmniRoute/issues/6344)) — v3.8.45 flipped the production bundler default to Turbopack, but `next.config.mjs` aliased `@/mitm/manager` to its Docker-only degraded stub **unconditionally**. That was harmless while Docker (which sets the alias intentionally for #3390 graceful degradation) was the sole Turbopack consumer, but once every artifact built with Turbopack the stub shipped to all non-Docker users and `startMitm` threw on the first Agent-Bridge start. The alias is now opt-in via `OMNIROUTE_MITM_STUB=1` (set only by the Dockerfile) through the shared `scripts/build/mitm-stub-flag.mjs` helper; default builds bundle the real manager. Regression guard: `tests/unit/mitm-stub-alias-6344.test.mjs` (4).
+
+- **fix(proxy):** stop the v3.8.44 proxy regression that leaked the real IP and disabled healthy proxies ([#6246](https://github.com/diegosouzapw/OmniRoute/issues/6246)). Two coupled defects from the new health scheduler: (1) **IP leak** — when a proxy assigned to a connection was marked `inactive`, resolution fell through to a **direct** egress instead of blocking, exposing the operator's real IP; (2) **over-deactivation** — the sweep flipped a proxy to `inactive` on the **first** failed probe and counted our own 5s timeout / a probe-target `5xx` as the proxy's fault, so healthy paid proxies vanished from egress selection ("my proxies are not being used anymore"). Fix: the sweep decision is extracted into a pure, network-free `decideProxyHealthAction` (`src/lib/proxyHealth/decision.ts`) — by default the health check now **only counts/logs and never downgrades status** (a proxy is downgraded/removed only with `PROXY_AUTO_REMOVE=true`, after `PROXY_AUTO_REMOVE_AFTER` **consecutive** conclusive failures); probes are classified tri-state so an inconclusive result (our timeout, or a `5xx` from the probe target) never penalizes the proxy, and the probe timeout is raised 5s→15s. Separately, `safeResolveProxy` now **fails closed** via the existing policy: a connection whose assigned proxy is dead is blocked instead of leaking direct (`hasBlockingProxyAssignment`), honoring the explicit `proxy off` toggles and the `PROXY_FAIL_OPEN=true` opt-out. Existing proxies stuck `inactive` by the old behavior need a one-time manual re-activate (the operator owns proxy status). Regression guards: `tests/unit/proxy-health-decide-action-6246.test.ts`, `tests/unit/proxy-assigned-unavailable-6246.test.ts`.
+
+- **fix(proxy):** make "Test All" read-only and add bulk enable/disable ([#6246](https://github.com/diegosouzapw/OmniRoute/issues/6246)). Complements the core fail-closed / scheduler fix (#6296) with the two remaining reporter asks. (1) The **"Test All" button** (`POST /api/settings/proxies/auto-test`) used to flip a proxy to `inactive` on a failed reachability probe; since the egress selector excludes `inactive` proxies, a flaky probe (an unreachable `httpbin.org`, a proxy that blocks `HEAD`, or a slow paid proxy) silently disabled every proxy that failed — "Test All" is now **read-only by default** (only the operator sets a proxy active/inactive; opt back into the legacy test-and-set with `PROXY_HEALTH_AUTO_DEACTIVATE=true`). (2) Adds a **bulk enable/disable** proxies endpoint + toolbar action (`POST /api/settings/proxies/batch-activate`) so an operator can re-activate proxies in one click. Regression guard: `tests/unit/proxy-health-6246.test.ts`. (thanks @tenshiak)
+
+- **chatcore (tools): stop the default 128-tool cap from silently dropping opencode's `task`/MCP tools.** opencode (used as an MCP/agent host) sends a large tool list; when it exceeds the speculative `MAX_TOOLS_LIMIT` (128) default, `truncateToolList` did a blind `tools.slice(0, 128)`, dropping every tool past index 128 — including opencode's built-in `task` tool (subagent launch) and many MCP tools, so models routed through OmniRoute could no longer spawn subagents or reach part of their tools. The cap exists to avoid upstream `400`s for providers with real hard limits (e.g. grok-cli 200), so it is kept for those: detection of the opencode client (`isOpencodeClient` — any `x-opencode-*` header, or `opencode` in the user-agent) now only bypasses the **speculative 128 default**, never a known provider ceiling. Precedence is explicit — a proactive/detected provider limit always truncates (even for opencode); otherwise opencode forwards its full tool list; otherwise the unchanged 128 default applies to every other client. Refactors `getEffectiveToolLimit` into `getKnownToolLimit(provider) ?? DEFAULT_LIMIT` (byte-identical for existing callers) and fixes a cosmetic debug-log that reported the truncated count instead of the original. Regression guard: `tests/unit/tool-limit-detector.test.ts`.
+
+- **fix(mitm):** the macOS MITM-cert install check now matches the system keychain again. `security find-certificate -a -Z` prints the SHA-1 as a colon-less hex string, but the installed-check compared it against `getCertFingerprint()`'s colon-separated form, so the substring match never hit — the cert was reported as not-installed and re-prompted for the sudo install on every run. Fingerprints are now normalized (colons stripped, upper-cased) on both sides via the extracted `macCertOutputHasFingerprint` helper. Regression guard: `tests/unit/mitm-cert-mac-fingerprint.test.ts`. ([#6204](https://github.com/diegosouzapw/OmniRoute/pull/6204), closes [#6134](https://github.com/diegosouzapw/OmniRoute/issues/6134) — thanks @rianonehub)
+
+- **fix(api):** `/v1/messages/count_tokens` now counts `tool_use`, `tool_result` and `thinking` content blocks (and array-form `system` prompts) in the local-estimation path, instead of only `text`. Real agentic conversations keep ~95% of their tokens inside tool results; the previous estimate returned near-zero for them, which silently broke Claude Code's auto-compaction (context grew past the window with no compaction until the upstream API rejected the request). The real provider-side count path is unchanged. Regression guard: `tests/unit/messages-count-tokens-route.test.ts`. ([#6221](https://github.com/diegosouzapw/OmniRoute/pull/6221) — thanks @luweiCN)
+
+- **fix(antigravity):** strip a trailing assistant prefill turn for Vertex Claude models to avoid upstream 400s ([#6114](https://github.com/diegosouzapw/OmniRoute/pull/6114)). Regression guard: `tests/unit/antigravity-claude-prefill-strip.test.ts`. (thanks @anki1kr)
+
+- **fix(security):** the mutable cloud-agent routes (`/api/cloud/credentials/update`, `/api/cloud/models/alias`) now require management auth instead of being treated as public. They were classified as public API routes, so a request without management credentials could update stored cloud-agent credentials and model aliases. They are removed from the public-route set, classified as management routes in the authz pipeline, and gated by `requireManagementAuth`; cloud **read**/auth routes stay public. Regression guards: `tests/unit/cloud-write-auth.test.ts`, `tests/unit/authz/classify.test.ts`, `tests/unit/public-api-routes.test.ts`. ([#6233](https://github.com/diegosouzapw/OmniRoute/pull/6233) — thanks @vittoroliveira-dev)
+
+- **refactor(dashboard):** extract the onboarding-wizard "Open provider details" link target into a pure, unit-tested `buildProviderDetailsHref(connection)` helper. The wizard already routes by `connection.id` (the node UUID) rather than the provider category slug (#6144/#6145); this hardens that behavior behind a tested helper that guards a missing id/connection. Regression guard: `tests/unit/provider-onboarding-href.test.ts`. ([#6166](https://github.com/diegosouzapw/OmniRoute/pull/6166) — thanks @KooshaPari)
+
+- **fix(security):** the doubao synthetic device-id generator now derives its digits via an unbiased crypto-random draw (rejection sampling over `crypto.randomBytes()`) instead of a `% 10` reduction, closing a CodeQL `js/biased-cryptographic-random` finding.
+- **fix(agentSkills):** the GitHub-skills generator now resolves `outputDir` to an absolute path before writing, fixing a regression introduced by #6366 (relative-to-cwd base path) that could write generated skill files to the wrong directory.
+- **fix(security):** `/api/keys/{id}/devices` now checks the HTTP method before auth/validation, returning a `405` for non-GET/DELETE verbs instead of a misleading `401`/`500` (closes a `dast-smoke` QUERY-method finding).
+- **fix(quality):** clear the last 2 heavy quality-gate reds on the release tip (cycle pre-flight).
+- **fix(mitm):** the test suite and CI can never mutate the OS trust store — `OMNIROUTE_SKIP_SYSTEM_TRUST=1` is set globally for tests/CI so `installCert`/`uninstallCert`/`installTproxyCa` skip the privileged OS dispatch ([#6310](https://github.com/diegosouzapw/OmniRoute/pull/6310); full detail is under the [3.8.45] section below — this branch received it via the parallel-cycle sync-back).
+- **fix(api):** `POST /api/github-skills` now Zod-validates its request body; documented the new quality-gate env vars and pinned the merge-integrity GitHub Actions to a commit SHA.
+- **fix(skills):** generate the missing `omni-github-skills` registry entry and align the agent-skills catalog-count tests (follow-up to #6186).
+- **fix(quality):** clear the cycle's 11 net-new ESLint errors and make `validate-release-green` suppressions-aware.
+
+### 📝 Maintenance
+
+- **i18n(it):** add 118 missing Italian (`it`) translations (net-additive — no existing keys dropped, valid JSON), improving Italian UI coverage. ([#6212](https://github.com/diegosouzapw/OmniRoute/pull/6212) — thanks @serverless83)
+- **chore(providers):** remove deprecated **MiMo V2** model entries from the catalogs (xiaomi-mimo, opencode-go, zenmux-free, audio TTS) — the upstream V2 line is superseded by MiMo V2.5; drops `mimo-v2-tts`, `mimo-v2-pro`, `mimo-v2-omni`, `mimo-v2-flash`, `mimo-v2-flash-free` and realigns the provider-catalog tests. ([#6248](https://github.com/diegosouzapw/OmniRoute/pull/6248) — thanks @backryun)
+- **chore(release):** ~50 commits on this branch are v3.8.45 pre-flight/hardening fixes and CI-perf work that landed here via the parallel-cycle sync-back (`sync-next-cycle.mjs`, Hard Rule #21) **after** the `v3.8.45` git tag was cut, and are already fully documented under the **[3.8.45]** section below — listed here only so the per-cycle commit-coverage check (`npm run release:uncovered`) doesn't flag them as gaps. Provider/catalog/UX/backend fixes: #6041, #6078, #6108, #6135, #6148, #6149, #6154, #6158, #6161, #6162, #6163, #6164, #6165, #6170, #6177, #6178, #6181, #6186, #6187, #6191, #6193, #6194, #6195, #6200, #6205, #6208, #6209, #6211, #6213, #6223, #6224, #6225, #6226, #6227, #6228, #6229, #6230, #6235, #6291, #6292. CI/release-pipeline work: #6167, #6203, #6214, #6215, #6218, #6273, #6275, #6283, #6284, #6285, #6300, #6305.
+- **chore(release):** additional zero-ref release-cycle plumbing on this branch, kept out of `release:uncovered` on purpose (no `#N` in the commit subject to cite): opening the v3.8.46 cycle, opening/closing the v3.8.45 cycle, the finalized [3.8.45] CHANGELOG i18n sync-back to 42 mirrors, the v3.8.45 cognitive/cyclomatic and file-size drift rebaselines, ESLint stale-suppression pruning (4,273 → 4,233), and clearing test-masking/docs-all pre-flight reds for v3.8.45.
+
+### ⚡ Performance & Infrastructure
+
+- **perf(release-green):** the pre-flight validator (`scripts/quality/validate-release-green.mjs`) now runs its 4 slow suites (unit / vitest / integration / pack-artifact) **concurrently** via `Promise.all` — pre-flight wall time drops from ~the sum of the suites to ~the slowest one (~30min saved per round; Phase 0 was the nº1 bottleneck of the v3.8.45 release benchmark, 2h54 of 6h34 e2e). Guard: `tests/unit/validate-release-green.test.ts` ("runs the slow suites CONCURRENTLY"). ([#6319](https://github.com/diegosouzapw/OmniRoute/pull/6319))
+- **fix(ci):** `scripts/release/sync-next-cycle.mjs` — two defects found live in its first production run (v3.8.45 Phase 5): (1) the `git()` helper's default 1 MiB `maxBuffer` crashed with `ENOBUFS` on `git show origin/main:CHANGELOG.md` (the CHANGELOG alone is >1 MiB) — widened to 64 MiB; (2) the i18n resync only propagated the `[NEXT]` (TBD) section, leaving the just-shipped finalized section as "— TBD" in all 42 mirrors — it now also syncs `[prevVersion]` bounded by the heading below it (new exported pure helper `versionAfter`). Guards: +5 tests in `tests/unit/sync-next-cycle.test.ts` (8/8). ([#6327](https://github.com/diegosouzapw/OmniRoute/pull/6327))
+- **test(ci):** concurrency-sensitive flaky tests are **quarantined into a serial pass** (`tests/unit/serial/`, `--test-concurrency=1`, appended to every unit runner incl. sharded variants — the serial pass is sharded too so concurrent shard jobs never self-collide). Initial set: `glm-coding-plan-monthly-3580`, `quota-division-blocks`, `provider-health-autopilot`, `combo-health-autopilot` — the class behind the ~28min CI wedges/re-runs (two live 1h+ wedges cancelled during this PR's own validation). Discovery + TIA gates track the new glob; systemic root cause (async logger writing after teardown) tracked in [#6360](https://github.com/diegosouzapw/OmniRoute/issues/6360). Guard: `tests/unit/test-serial-quarantine.test.ts` (4). ([#6347](https://github.com/diegosouzapw/OmniRoute/pull/6347))
+
+### 🙌 Contributors
+
+Thanks to everyone whose work landed in v3.8.46:
+
+| Contributor                                                    | PRs / Issues           |
+| -------------------------------------------------------------- | ---------------------- |
+| [@2220258345](https://github.com/2220258345)                   | direct commit / report |
+| [@abdofallah](https://github.com/abdofallah)                   | direct commit / report |
+| [@adentdk](https://github.com/adentdk)                         | direct commit / report |
+| [@aleksesipenko](https://github.com/aleksesipenko)             | direct commit / report |
+| [@anki1kr](https://github.com/anki1kr)                         | direct commit / report |
+| [@anungma](https://github.com/anungma)                         | direct commit / report |
+| [@arpicato](https://github.com/arpicato)                       | direct commit / report |
+| [@backryun](https://github.com/backryun)                       | #6248                  |
+| [@binsarjr](https://github.com/binsarjr)                       | direct commit / report |
+| [@brightfiscalband](https://github.com/brightfiscalband)       | direct commit / report |
+| [@chirag127](https://github.com/chirag127)                     | #6501, #6506           |
+| [@developerjillur](https://github.com/developerjillur)         | direct commit / report |
+| [@dilneiss](https://github.com/dilneiss)                       | #6499                  |
+| [@dtybnrj](https://github.com/dtybnrj)                         | direct commit / report |
+| [@Forcerecon](https://github.com/Forcerecon)                   | direct commit / report |
+| [@hao3039032](https://github.com/hao3039032)                   | direct commit / report |
+| [@Iammilansoni](https://github.com/Iammilansoni)               | #6150, #6245           |
+| [@jmengit](https://github.com/jmengit)                         | direct commit / report |
+| [@jordansilly77-stack](https://github.com/jordansilly77-stack) | direct commit / report |
+| [@JxnLexn](https://github.com/JxnLexn)                         | direct commit / report |
+| [@KooshaPari](https://github.com/KooshaPari)                   | #6166                  |
+| [@loopyd](https://github.com/loopyd)                           | direct commit / report |
+| [@luweiCN](https://github.com/luweiCN)                         | #6221                  |
+| [@makcimbx](https://github.com/makcimbx)                       | direct commit / report |
+| [@muflifadla38](https://github.com/muflifadla38)               | direct commit / report |
+| [@newnol](https://github.com/newnol)                           | direct commit / report |
+| [@ofekbetzalel](https://github.com/ofekbetzalel)               | direct commit / report |
+| [@ohahe52-dot](https://github.com/ohahe52-dot)                 | direct commit / report |
+| [@phidinhmanh](https://github.com/phidinhmanh)                 | direct commit / report |
+| [@powellnorma](https://github.com/powellnorma)                 | direct commit / report |
+| [@qpeyba](https://github.com/qpeyba)                           | direct commit / report |
+| [@RaviTharuma](https://github.com/RaviTharuma)                 | direct commit / report |
+| [@RCrushMe](https://github.com/RCrushMe)                       | direct commit / report |
+| [@rianonehub](https://github.com/rianonehub)                   | #6134, #6204           |
+| [@serverless83](https://github.com/serverless83)               | #6212                  |
+| [@swingtempo](https://github.com/swingtempo)                   | direct commit / report |
+| [@tenshiak](https://github.com/tenshiak)                       | direct commit / report |
+| [@ThongAccount](https://github.com/ThongAccount)               | direct commit / report |
+| [@UnrealAryan](https://github.com/UnrealAryan)                 | direct commit / report |
+| [@vinayakkulkarni](https://github.com/vinayakkulkarni)         | direct commit / report |
+| [@vittoroliveira-dev](https://github.com/vittoroliveira-dev)   | #6233                  |
+| [@warelik](https://github.com/warelik)                         | direct commit / report |
+| [@xxy9468615](https://github.com/xxy9468615)                   | direct commit / report |
+| [@xz-dev](https://github.com/xz-dev)                           | direct commit / report |
+| [@yanpaing007](https://github.com/yanpaing007)                 | direct commit / report |
+| [@diegosouzapw](https://github.com/diegosouzapw)               | maintainer             |
+
+---
+
 ## [3.8.45] — 2026-07-06
 
 ### ✨ New Features

docs/i18n/phi/CHANGELOG.md

@@ -6,6 +6,213 @@
 
 ## [3.8.31] — 2026-06-20
 
+## [3.8.46] — TBD
+
+### ✨ New Features
+
+- **feat(sse):** **hide paid-only models from `auto/*` routing** when `hidePaidModels` is on ([#6512](https://github.com/diegosouzapw/OmniRoute/issues/6512)) — follow-up to #6328/#6495. PR #6495 hid paid-only models from the `GET /v1/models` listing, but `auto/*` combos (`auto/best-coding`, `auto/glm`, …) could still pick a paid-only backend into their candidate pool → a 402/403 at request time. `createVirtualAutoCombo` now filters the candidate pool through the new pure `open-sse/services/autoCombo/paidModelFilter.ts` (`filterPaidOnlyCandidates`), applying the same free-model predicate #6495 uses in `catalog.ts` (`providerHasFreeModels(provider) && isFreeModel(provider, {id})`) whenever `settings.hidePaidModels === true`. Applied before the category/tier/family narrowing, so it covers every `auto/*` combo; an all-paid pool degrades to the existing graceful empty-pool path. **Opt-in — default OFF leaves the pool unchanged** (identity). Regression guard: `tests/unit/autoCombo/paid-model-filter-6512.test.ts` (4, incl. the default-off identity guard).
+- **feat(sse):** **provider-family auto combos** — `auto/glm`, `auto/minimax`, `auto/mimo`, `auto/zai`, `auto/gemma`, `auto/llama`, `auto/gemini` ([#6453](https://github.com/diegosouzapw/OmniRoute/issues/6453)) — new routable ids that materialize an on-demand virtual combo spanning whatever installed backends currently expose that model family, degrading gracefully as backends rotate. A new pure `open-sse/services/autoCombo/modelFamily.ts` (`detectModelFamily`) classifies by model-id prefix for six families; `zai` is instead resolved by provider id (z.ai's hosted API serves the same `glm-*` model ids as every other GLM backend, so `auto/zai` means "route to my z.ai backend specifically" vs `auto/glm`'s "any connected GLM backend"). Reuses the existing `createVirtualAutoCombo` on-demand materialization path (no DB writes) and the `/v1/models` catalog advertising loop. Regression guard: `tests/unit/autoCombo/provider-family-combos.test.ts` (11).
+- **feat(proxy):** native **proxy-pool round-robin / egress IP rotation** ([#6365](https://github.com/diegosouzapw/OmniRoute/issues/6365)) — a scope (global / provider / account) can now hold **multiple** proxies as a pool with a rotation strategy, so outbound requests cycle their egress IP instead of pinning one proxy per scope. Migration `117_proxy_pool_rotation.sql` lifts the `UNIQUE(scope, scope_id)` constraint (rebuild via the canonical rename/copy/drop; existing single assignments become 1-element pools) and adds a `proxy_scope_rotation` companion table holding the per-scope strategy + a persisted monotonic round-robin cursor. Strategies: `round-robin` (default, monotonic cursor — never `Math.random`), `random`, and `sticky-per-N-min`. Resolution (`resolveProxyForScopeFromRegistry` / `resolveProxyForConnectionFromRegistry`) now fetches the alive, position-ordered candidate set (unchanged `PROXY_ALIVE_PREDICATE`) and applies the strategy; an empty / all-dead pool still returns `null` — the #6246 fail-closed guard is **untouched** (never falls through to direct egress). Backend + DB only; dashboard pool-builder UI is a follow-up. Regression guard: `tests/unit/proxy-pool-rotation-6365.test.ts` (8, incl. fail-closed + backward-compat).
+- **feat(providers):** end-to-end **tool/function calling on the native Gemini `/v1beta` endpoint** ([#6222](https://github.com/diegosouzapw/OmniRoute/issues/6222)) — both directions of the Gemini↔OpenAI conversion now preserve tool data (previously silently dropped). Request side: `convertGeminiToInternal` (extracted to its own testable module) maps `tools[].functionDeclarations` → OpenAI `tools`, prior `functionCall` parts → assistant `tool_calls`, and `functionResponse` parts → `tool`-role messages. Response side: `convertOpenAIResponseToGemini` emits `parts[].functionCall {name,args}` from `message.tool_calls`, and the streaming `openAIChunkToGeminiChunk` accumulates fragmented `tool_calls` deltas by index into complete `functionCall` parts. The non-Gemini client paths (Claude, OpenAI-Responses) already preserved tool calls — this closes the gap specific to the native Gemini surface. Regression guard: `tests/unit/v1beta-gemini-tool-calling-6222.test.ts` (6, incl. a streaming SSE round-trip).
+- **feat(providers):** copilot-m365-web **enterprise / work tier** support ([#6334](https://github.com/diegosouzapw/OmniRoute/issues/6334)) — mirrors the EDU-tier pattern (#6210): `M365ConnectionParams` gains an `agent` field, a new opt-in `M365_ENTERPRISE_OVERRIDES` preset (`agent=work`, `scenario=officeweb`, `licenseType=Premium`) applies via `providerSpecificData.tier="enterprise"` (alias `"work"`), and `agent` is also overridable directly via `providerSpecificData.agent`. `buildWsUrl` was hardcoding `agent="web"` (the one enterprise-distinguishing param with no override path), so a Premium work account handshook then returned an empty stream. The individual and EDU paths are untouched. Kilo's dup flag vs #6210 (EDU tier) was a false positive — different tier. Regression guard: `tests/unit/copilot-m365-enterprise-6334.test.ts` (7). End-to-end confirmation on a real Premium work account is a live-VPS validation follow-up (Hard Rule #18). (thanks @Forcerecon)
+- **feat(api):** standardized, provider-agnostic **`effort` + `thinking` request params** ([#6241](https://github.com/diegosouzapw/OmniRoute/issues/6241)) — a thin standardization layer over the existing mature per-provider reasoning plumbing (no provider mapper touched). `providerChatCompletionSchema` gains a canonical `effort` (reusing the shared `none/low/medium/high/xhigh` vocabulary — the UI tiers `extra`/`max` collapse onto `xhigh`) and a boolean `thinking`. A pure `normalizeReasoningRequest` (wired once in `src/sse/handlers/chat.ts`, before any reasoning field is read) folds them onto the fields the translators already consume (`reasoning_effort` / `reasoning.effort` / `thinking`), so they fan out to Anthropic / Gemini / xAI / Responses — an explicit client `reasoning_effort` / object-shaped `thinking` always wins (backward-compatible). `/models` additively exposes `supportsThinking` + `effort_tiers` so the frontend can render the toggles (UI component is a follow-up). Regression guard: `tests/unit/effort-thinking-standardization-6241.test.ts` (12). (thanks @Iammilansoni, @shabeer)
+- **feat(combo):** new **`pipeline` (sequential) combo strategy** ([#6297](https://github.com/diegosouzapw/OmniRoute/issues/6297)) — the 18th routing strategy runs targets **in order**, threading each step's output into the next step's input, with an optional per-step `prompt` (system instruction); only the final step's response is returned. Distinct from `fusion` (parallel fan-out + judge). Implemented as a self-contained `open-sse/services/pipeline.ts` (sibling to `fusion.ts`), dispatched from `combo.ts`; the step list reuses `combo.models` order and reads an optional `prompt` off each target (backward-compatible — ignored by every other strategy). Intermediate steps run non-streaming with tools stripped (complete prose to thread forward); the final step keeps the client's `stream` flag + tools. A failing/empty/unparseable intermediate step fails the whole pipeline explicitly via a sanitized error (never silently swallowed). Kilo's dup flag vs #563 was a false positive (that's model→chain selection; this is a sequential chain). Regression guard: `tests/unit/combo-pipeline-strategy.test.ts` (5). (thanks @ofekbetzalel)
+- **feat(ci):** `check:test-masking` now flags **inline-reimplemented prod conditions** ([#6348](https://github.com/diegosouzapw/OmniRoute/issues/6348)) — a new **report-only** subcheck (v2, 6A.10 family) catches the wrong-shape contract test: a test that recomputes the condition under test inline instead of importing/exercising the real function (the #6216 class, where `=== 500` → `>= 500` stayed green because the test re-implemented the branch). For each added/modified test file it warns when the file textually duplicates a ≥3-token conditional from a production file touched in the same PR **and** does not import the symbol/module owning it, via a pure, fixture-tested `findReimplementedConditions()` with an allowlist mirroring `assertReductionAllowlist`. Report-only for now (does not fail the gate) — to be promoted to blocking after a triage cycle. Regression guard: `tests/unit/check-test-masking.test.ts` (45).
+- **feat(sse):** per-connection routing override (native vs CLIProxyAPI) ([#6339](https://github.com/diegosouzapw/OmniRoute/issues/6339)) — the previously-dead `isCliproxyapiDeepModeEnabled` helper is now wired into `resolveExecutorWithProxy`: a single connection can opt itself into the CLIProxyAPI passthrough executor via `providerSpecificData.cliproxyapiMode="claude-native"`, with precedence **connection override > provider `upstream_proxy_config` mode > default**. `resolveExecutorWithProxy` now receives the resolved connection's `providerSpecificData` (threaded from `chatCore.ts`), so one connection can deep-route while the provider's other connections stay native — no DB schema change (the toggle rides in `providerSpecificData`). Also resolves the same-provider mixing ask in #6340. Regression guard: `tests/unit/chatcore-executor-proxy.test.ts` (9). (thanks @RaviTharuma)
+- **feat(dashboard):** "Add session cookie" modal now shows a prominent **"Open ‹host› →"** link to the provider's own site ([#6268](https://github.com/diegosouzapw/OmniRoute/issues/6268)) — every `-web` cookie-session provider (chatgpt-web, claude-web, gemini-web, kimi-web, lmarena, qwen-web, m365-copilot-web, …) renders a one-click external link (opening the provider's login/home page in a new tab) so operators no longer tab away to retype the URL mid-setup. The host resolves from a pure, unit-tested `resolveWebProviderHost()` (prefers `WEB_COOKIE_PROVIDERS[id].website`, falls back to the registry `baseUrl` origin); non-web providers render exactly as before. Kilo's dup flag vs #6265 (modal-too-small-on-1080p) was a false positive — distinct concern. Regression guard: `tests/unit/resolve-web-provider-host.test.ts` (5). (thanks @chirag127)
+- **feat(providers):** add **DigitalOcean** AI (serverless inference) as an OpenAI-compatible API-key provider ([#6373](https://github.com/diegosouzapw/OmniRoute/pull/6373)) — base `https://inference.do-ai.run/v1`, wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/digitalocean/`, `src/shared/constants/providers/apikey/inference-hosts.ts`). Regression guard: `tests/unit/digitalocean-provider.test.ts`. (thanks @newnol)
+- **feat(providers):** add **Huancheng Public API** (`hcnsec`) as an OpenAI-compatible regional provider ([#6410](https://github.com/diegosouzapw/OmniRoute/pull/6410)) — Xinjiang Huancheng Cybersecurity's public LLM platform (base `https://api.hcnsec.cn/v1`, free credits via daily check-ins), wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/hcnsec/`, `src/shared/constants/providers/apikey/regional.ts`). Regression guard: `tests/unit/hcnsec-provider.test.ts`. (thanks @UnrealAryan)
+- **feat(dashboard):** the web-session credential guide now shows an **"Open {host}" link** ([#6316](https://github.com/diegosouzapw/OmniRoute/pull/6316)) to the provider's sign-in site (derived from the provider `website` via `getProviderWebsiteHost`), so you can jump straight to the page where the cookie/session must be captured. Regression guard: `tests/unit/web-session-provider-link-6316.test.ts`. (thanks @jordansilly77-stack)
+- **feat(cerebras):** add the **Gemma 4 31B** model (`gemma-4-31b`) to the Cerebras registry + pricing table ([#6331](https://github.com/diegosouzapw/OmniRoute/pull/6331)). Regression guard extends `tests/unit/t28-model-catalog-updates.test.ts`. (thanks @backryun)
+- **feat(providers):** add **Yuanbao (web)** as a cookie-session provider ([#6196](https://github.com/diegosouzapw/OmniRoute/issues/6196)) — `yuanbao-web` (Tencent Yuanbao, `yuanbao.tencent.com`) with cookie-only auth (`hy_user`/`hy_token` + public agent id), SSE→OpenAI translation incl. `reasoning_content`, exposing DeepSeek V3/R1 + Hunyuan / Hunyuan-T1. Regression guard: `tests/unit/providers-yuanbao-web.test.ts`. `together-web` was **deferred** (no verifiable web-session endpoint — needs a captured request) and `huggingchat-web` **dropped** (the existing `huggingchat` already is a web-cookie provider). (thanks @chirag127)
+- **feat(providers):** route the built-in **agentrouter** through the dynamic Claude-Code wire image ([#6056](https://github.com/diegosouzapw/OmniRoute/issues/6056)) — a small static allow-set (`CC_WIRE_IMAGE_BUILTINS` in `open-sse/services/ccWireImageBuiltins.ts`), consulted by `isClaudeCodeCompatible` / `isClaudeCodeCompatibleProvider` / `applyFingerprint`, makes agentrouter adopt the CC wire-image headers + fingerprint **while guarding the CC baseUrl/auth branches** so it keeps its own registry `baseUrl` and `x-api-key` auth. Regression guard: `tests/unit/agentrouter-cc-wire-image.test.ts` (asserts the wire image is applied AND agentrouter's baseUrl/auth are preserved). Live WAF-acceptance against agentrouter.org is a VPS validation follow-up (Hard Rule #18).
+- **feat(providers):** **bulk-add API keys for Cloudflare Workers AI** ([#6174](https://github.com/diegosouzapw/OmniRoute/issues/6174)) — `cloudflare-ai` is removed from the bulk-add exclusion list and the bulk parser gains a 3-field `name|accountId|apiKey` mode; the bulk route now builds a **per-entry** `providerSpecificData` so each key carries its own `accountId` (fixing the previous shared-object reuse), and both the create + key-validation paths receive it. Regression guard: `tests/unit/bulk-api-key-parser-cloudflare.test.ts`. (thanks @muflifadla38)
+- **feat(dashboard):** routing/settings UX clarity ([#6147](https://github.com/diegosouzapw/OmniRoute/issues/6147)) — (1) weighted combos show the **effective routing share %** next to each weight when weights don't sum to 100 (`WeightTotalBar.tsx`); (2) the status widget's user-facing **"Cloud Sync" label is renamed** to "Remote Settings Sync" (`CloudSyncStatus.tsx`; internal ids/state untouched); (3) built-in providers gain an **opt-in advanced base-URL override** (`isBaseUrlOverrideEligibleProvider`, hidden behind an "Advanced" toggle, reusing the existing `providerSpecificData.baseUrl` persistence — not globally widened). Regression guard: `tests/unit/routing-settings-ux-6147.test.ts`.
+- **feat(combo):** add an option to **disable session stickiness**, per-combo or globally — round-robin / random combos can rotate to a different connection on every request instead of pinning a whole conversation to one connection by its first-message hash. Resolution precedence per-combo `config.disableSessionStickiness` → global `settings.disableSessionStickiness` → default `false` (preserves the #3825 prompt-cache/504 fix); gates **both** stickiness call sites in `open-sse/services/combo.ts`. Exposed as a global toggle (Combo Defaults) and a per-combo Inherit/on/off control. ([#6168](https://github.com/diegosouzapw/OmniRoute/issues/6168)) Regression guard: `tests/unit/combo-disable-session-stickiness.test.ts`. (thanks @RCrushMe)
+- **feat(docker):** add the `OMNIROUTE_NO_SUDO` env flag for root-less / user-namespaced deployments — the MITM cert-trust command path (`resolveSudoSpawn` in `src/mitm/systemCommands.ts`) now strips the leading `sudo` when the flag is truthy, in addition to the existing root / sudo-missing cases, so the Proxy Agent runs without `sudo` (the operator trusts the CA manually, e.g. via `NODE_EXTRA_CA_CERTS`). Argv-array `spawn` preserved — no shell interpolation (Hard Rule #13). ([#6122](https://github.com/diegosouzapw/OmniRoute/issues/6122)) Regression guard: `tests/unit/mitm-systemCommands-no-sudo.test.ts`. (thanks @powellnorma)
+- **feat(providers):** add **Requesty** as an OpenAI-compatible gateway provider (BYOK, base `https://router.requesty.ai/v1`, ~200 free requests/day) — wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/requesty/`, `src/shared/constants/providers/apikey/gateways.ts`). ([#6120](https://github.com/diegosouzapw/OmniRoute/issues/6120)) Regression guard: `tests/unit/requesty-provider.test.ts`. (thanks @chirag127)
+- **feat(dashboard):** add **configured-only / available-only filters** to the Free Provider Rankings page ([#6150](https://github.com/diegosouzapw/OmniRoute/issues/6150)) — hide providers you haven't configured, or whose connections are all rate-limited / out of quota, via server-side query params (`?configuredOnly` / `?availableOnly` on `GET /api/free-provider-rankings`) backed by a testable lib helper reusing the in-process connection state (no Redis). Both filters default off, so the default view is unchanged; this supersedes the earlier client-side "Configured Only" toggle (#6245) with an available-only dimension and unit-tested logic. Regression guard: `tests/unit/freeProviderRankings-filters.test.ts`.
+
+### 🔧 Bug Fixes
+
+- **fix(dashboard):** adding a second API key connection for the same provider no longer silently overwrites the first — the Add-API-key modal now derives a unique default connection name (`main`, then `main-2`, `main-3`, …) so the backend name-based upsert can't collide (#6499 — thanks @dilneiss).
+- **fix(compression):** the session-dedup engine now also deduplicates a large multi-line block repeated **within a single message** (intra-message dedup), not just across turns; the compression-preview API surfaces a `fallbackReason`, and the fusion panel reports how many models were rate-limited vs failed on total-panel failure (#6501 — thanks @chirag127).
+- **fix(compression):** a stacked-pipeline step naming an unregistered engine now surfaces a `validationErrors` entry instead of silently no-op'ing, so misconfigured pipelines are visible in the preview API (#6506 — thanks @chirag127).
+- **feat(usage):** add a **Codex reset-credit redemption** flow to the Provider Limits UI ([#6361](https://github.com/diegosouzapw/OmniRoute/pull/6361)) — a `useCodexResetCreditRedemption` hook + `/api/usage/codex-reset-credit` route + `codexResetCredits` lib let you redeem banked Codex reset credits from the quota card. Regression guard: `tests/unit/codex-reset-credits.test.ts`. (thanks @JxnLexn)
+- **feat(glm):** add **team-plan quota settings** for `glm-cn` connections ([#6351](https://github.com/diegosouzapw/OmniRoute/pull/6351)) — a dedicated `GlmTeamQuotaFields` form section (team quota id / limits) threaded through the Add/Edit connection modals, persisted via `providerSpecificData`, with the GLM usage service reading the team quota. Regression guards: `tests/unit/glm-team-quota.test.ts`, `provider-specific-data-schema.test.ts`. (thanks @hao3039032)
+- **feat(providers):** add **TinyFish** web-fetch/search support ([#6349](https://github.com/diegosouzapw/OmniRoute/pull/6349)) — a `tinyfish-fetch` executor + `/v1/web/fetch` route + MCP web-fetch tool, registered as a specialty-media provider with request-validation and a search-provider catalog entry. Regression guards: `tests/unit/executor-tinyfish-fetch.test.ts`, `web-fetch-handler.test.ts`, `mcp-web-fetch-tool.test.ts`, `provider-validation-tinyfish.test.ts`. (thanks @dtybnrj)
+- **fix(cli):** `omniroute launch-codex` now spawns `codex.cmd` through a shell on Windows (the npm `.cmd` shim is unresolvable by bare `spawn` → ENOENT), mirroring the qodercli Windows fix (#6263) ([#6312](https://github.com/diegosouzapw/OmniRoute/pull/6312)). Regression guard: `tests/unit/launch-codex-windows-spawn-6312.test.ts`. (thanks @swingtempo)
+- **fix(codex):** isolate the **Spark** quota from the shared Codex quota and stabilize the quota UI ordering / hydration so per-scope limits render consistently ([#6336](https://github.com/diegosouzapw/OmniRoute/pull/6336)). Regression guards: `tests/unit/codex-quota-selection-hydration.test.ts`, `provider-limits-ui.test.ts` + 3 more. (thanks @xz-dev)
+- **feat(api):** add a `hidePaidModels` setting that filters paid-only models out of the `/v1/models` catalog. Regression guard: `tests/unit/models-catalog-hide-paid.test.ts`. (thanks @chirag127)
+- **fix(api-manager):** the fallback model picker now preserves combos instead of dropping them when a primary model is unavailable ([#6443](https://github.com/diegosouzapw/OmniRoute/pull/6443)). Regression guard: `tests/unit/api-manager-page-static.test.ts`. (thanks @jmengit)
+- **fix(providers):** recoverable Antigravity / Cloud-Code (Gemini Code Assist) `403` responses ([#6452](https://github.com/diegosouzapw/OmniRoute/pull/6452)) are now classified as a retryable project-config error instead of a terminal account ban, so a fixable project/API-disabled 403 no longer forces a ~1-year cooldown / full OAuth reconnect. Regression guard: `tests/unit/errorclassifier-antigravity-403.test.ts`. (thanks @developerjillur)
+- **fix(mitm):** `sanitizeHeaders` now redacts `Set-Cookie` response headers so upstream session cookies never leak into logs / diagnostics ([#6451](https://github.com/diegosouzapw/OmniRoute/pull/6451)). Regression guard: `tests/unit/mitm-sanitize-headers.test.ts`. (thanks @developerjillur)
+- **fix(api):** `/api/compression/preview` now accepts `mode: "caveman"` and correctly handles stacked / zero-compression previews ([#6425](https://github.com/diegosouzapw/OmniRoute/issues/6425)). Regression guard: `tests/unit/api/compression-preview-caveman-and-stacked-6425.test.ts`. (thanks @chirag127)
+- **feat(providers):** add **Zed** hosted LLM aggregator as a native-app provider ([#6118](https://github.com/diegosouzapw/OmniRoute/pull/6118)) — OAuth sign-in via the Zed hosted flow, registered through the shared provider registry + executor. Regression guards: `tests/unit/zed-oauth-provider.test.ts`, `zed-import-utils.test.ts`, `zed-docker-detect.test.ts`, `mitm-handler-zed.test.ts`. VPS-validated via live operator login (Hard Rule #18).
+- **fix(oauth):** the Kiro SSO-cache auto-import now **preserves the IDC region** — cross-region Amazon Q / Kiro profiles imported from the SSO cache are no longer collapsed to the default region ([#6113](https://github.com/diegosouzapw/OmniRoute/pull/6113)). Regression guard: `tests/unit/kiro-auto-import-idc-2059.test.ts`. VPS-validated via live operator login (Hard Rule #18).
+- **fix(dashboard):** passthrough model aliases no longer collide when two namespaced model ids share a last segment (port from 9router#1850, [#6431](https://github.com/diegosouzapw/OmniRoute/pull/6431)). `enx/gpt-5.5` and `enx/codebuddy/gpt-5.5` both auto-generated the alias `gpt-5.5`, so the second model could never be added (the UI just alerted "alias already exists"). Aliases are now disambiguated deterministically — bare last segment when free, then parent-qualified (`codebuddy-gpt-5.5`), then a numeric suffix — while re-adding the exact same model id is still blocked. Regression guard: `tests/unit/passthrough-alias-1850.test.ts`. (thanks @arpicato)
+- **fix(translator):** preserve a Gemini `functionResponse` co-located with other parts (another `functionCall`, or trailing `text`) in the same content when translating **Gemini → OpenAI** ([#6376](https://github.com/diegosouzapw/OmniRoute/pull/6376)). `convertGeminiContent()` early-returned the tool message on the first `functionResponse` part, dropping any co-located parts; such contents are now pre-split (one tool message per `functionResponse`, emitted first, plus one message for the remaining parts). Regression guard: `tests/unit/gemini-to-openai-function-response.test.ts`. (thanks @warelik)
+- **fix(headroom):** detect a python interpreter managed by **mise / pyenv / asdf / conda** (port from 9router#2353, [#6382](https://github.com/diegosouzapw/OmniRoute/pull/6382)). Headroom's python probe (`src/lib/headroom/detect.ts`) searched a hardcoded `PATH`, but version managers expose their interpreters via shim dirs that only join `PATH` through interactive-shell activation — which the non-interactive server never runs, so a managed python (≥3.10) was invisible and Headroom reported it missing. The search path now prepends the well-known shim/bin dirs (`~/.local/share/mise/shims`, `~/.pyenv/shims`, `~/.asdf/shims`, `$CONDA_PREFIX/bin`, `~/.local/bin`, respecting `MISE_DATA_DIR`/`PYENV_ROOT`/`ASDF_DATA_DIR` when set), and a new `HEADROOM_PYTHON` env override lets operators point straight at their interpreter (mirroring `HEADROOM_URL`). Still shell-free (`execFileSync`). Regression guard: `tests/unit/headroom-detect.test.ts` (5). (thanks @loopyd)
+- **fix(executors):** strip the OpenAI-Codex/Claude-CLI `client_metadata` passthrough field for **NVIDIA** requests (port from 9router#1887, [#6411](https://github.com/diegosouzapw/OmniRoute/pull/6411)). NVIDIA's OpenAI-compatible wrapper rejects it with `400 Unsupported parameter`, the same class already handled for `cerebras`/`mistral`; `nvidia` (executor `default`) was missing from the strip allowlist so Codex/Claude-Code passthrough requests 400'd. Regression guard: `tests/unit/executor-default-strip-client-metadata.test.ts` (+nvidia case). (thanks @phidinhmanh)
+- **fix(translator):** strip the Claude-style `thinking` field for **NVIDIA `z-ai/glm-5.2`** (port from 9router#2023, [#6413](https://github.com/diegosouzapw/OmniRoute/pull/6413)). NVIDIA's OpenAI-compatible wrapper 400s on `thinking` (a Claude-format client routed here leaves a `thinking:{type:"adaptive"}`); the existing strip rule only dropped `reasoning`. Same class already handled for `minimax-m2.7`. Regression guard: `tests/unit/nvidia-minimax-thinking-strip.test.ts` (+glm-5.2 case). (thanks @phidinhmanh)
+- **fix(translator):** suppress the streamed `</think>` close marker for the **Antigravity IDE** client (port from 9router#1061, [#6415](https://github.com/diegosouzapw/OmniRoute/pull/6415)). On thinking-only turns Antigravity rendered a bare `</think>` as the sole visible content, tripping its loop-detection and wasting requests. Antigravity's UA (`vscode/<v> (Antigravity/<v>)`) is added to the marker-suppress allowlist (alongside OpenCode); Claude Code / Cursor still get the marker, and `x-omniroute-thinking-marker: on` force-restores it. Regression guard: `tests/unit/think-close-marker-suppress-5245.test.ts`. (thanks @abdofallah)
+- **fix(executors):** strip nested `reasoning_content` from messages for **Mistral** (port from 9router#1649, [#6417](https://github.com/diegosouzapw/OmniRoute/pull/6417)). Mistral's API returns `422 extra_forbidden` when an assistant message carries `reasoning_content` (replayed thinking from a prior turn, e.g. via the Codex `/responses` path); the generic top-level 400 field-downgrade retry never covered the nested per-message field. `DefaultExecutor` now strips it for provider `mistral` only, so DeepSeek (which requires replayed `reasoning_content`) is unaffected. Regression guard: `tests/unit/mistral-strip-reasoning-content-1649.test.ts`. (thanks @xxy9468615)
+- **fix(executors):** strip the `client_metadata` passthrough field on the **OpenCode** path (port from 9router#1442, [#6418](https://github.com/diegosouzapw/OmniRoute/pull/6418)). OpenCode upstreams (e.g. `kimi-k2.6` via opencode-go) reject it with `400 "Extra inputs are not permitted, field: 'client_metadata'"`; the DefaultExecutor strip only covered cerebras/mistral and `OpencodeExecutor` extends `BaseExecutor` directly, so nothing removed it there. Regression guard: `tests/unit/opencode-strip-client-metadata-1442.test.ts`. (thanks @yanpaing007)
+- **fix(executors):** inject the `reasoning_content` echo for the native **Moonshot Kimi** provider (port from 9router#1480, [#6419](https://github.com/diegosouzapw/OmniRoute/pull/6419)). Kimi (executor `default`) is a thinking-mode upstream that 400s with "reasoning_content must be passed back" when a prior assistant turn lacks it; the placeholder injection was only wired into the OpenCode meta-provider, so direct multi-turn Kimi conversations failed. Scoped to `kimi` (gateway-served models matching the thinking-model name pattern are unaffected). Regression guard: `tests/unit/kimi-native-reasoning-injected-1480.test.ts`. (thanks @2220258345)
+- **fix(executors):** recover from a strict gateway's `context_management: Extra inputs are not permitted` 400 (port from 9router#1468, [#6420](https://github.com/diegosouzapw/OmniRoute/pull/6420)). **Claude Code** always sends a top-level `context_management` field; strict anthropic-compatible gateways reject it. The dedicated context-editing 400-fallback only fired when OmniRoute's own `contextEditing` feature was enabled (default off), so a client-sent field passed through untouched and 400'd. `context_management` is now in the generic reactive field-strip list, so it's stripped-and-retried once regardless of the feature flag (with correct request re-signing for claude-compatible relays). Regression guard: `tests/unit/provider-field-strips.test.ts`. (thanks @ohahe52-dot)
+- **fix(network):** enable **RFC 8305 Happy Eyeballs** (`autoSelectFamily`) on the direct-egress undici dispatcher (port from 9router#1237, [#6423](https://github.com/diegosouzapw/OmniRoute/pull/6423)). When DNS returns both IPv6 (AAAA) and IPv4 (A) and the IPv6 route is broken (e.g. a NAT64 `64:ff9b::` prefix without routing), undici tried IPv6 first and hung until `ETIMEDOUT` (then a 502 + account lockout), even though `curl` reached the same host. The direct dispatcher now races both families and uses whichever connects first. Proxy paths pin family via `proxyTls` and are unaffected. Regression guard: `tests/unit/direct-dispatcher-pipelining-4580.test.ts`. (thanks @adentdk)
+- **fix(combo):** round-robin now advances the rotation pointer past the model that **actually served**, not the eagerly-scheduled one (port from 9router#948, [#6428](https://github.com/diegosouzapw/OmniRoute/pull/6428)). With `stickyLimit: 1` (true round-robin), when the scheduled model failed and a _different_ model served via fallback, the counter had already advanced +1 from the scheduled index — so the next request reused the fallback-served model, degrading round-robin into hot-spotting on whichever model was healthy. The pointer now advances to the served index + 1 (mirroring the sticky-limit>1 path). Session-stickiness (#3825) and distribution are preserved. Regression guard: `tests/unit/combo-rr-fallback-advance-948.test.ts`. (thanks @binsarjr)
+- **fix(sse):** a non-string `model` field is now rejected with a `400` before the resolver, instead of crashing downstream `.toLowerCase()`/`.split()` calls into an empty-body `500` that escapes the error sanitizer ([#6407](https://github.com/diegosouzapw/OmniRoute/issues/6407)). Regression guard: `tests/unit/chat-non-string-model-6407.test.ts`. (thanks @chirag127)
+- **fix(api):** unknown `/api/*` routes now return a JSON `404` (instead of the dashboard HTML shell) and scalar chat params (`model`/`temperature`/etc.) are validated **before** the provider lookup so malformed requests fail fast with a clear `400` ([#6424](https://github.com/diegosouzapw/OmniRoute/issues/6424), [#6412](https://github.com/diegosouzapw/OmniRoute/issues/6412)). Regression guards: `tests/unit/api/api-catchall-json-404.test.ts`, `tests/unit/chat-early-schema-validation-6412.test.ts`. (thanks @chirag127)
+- **fix(api):** `/v1/chat/completions` now rejects a non-JSON `Content-Type` with a `400` before parsing the body ([#6414](https://github.com/diegosouzapw/OmniRoute/issues/6414)). Regression guard: `tests/unit/v1-chat-completions-content-type-6414.test.ts`. (thanks @chirag127)
+- **fix(api):** the `X-OmniRoute-Compression` response header is now echoed on `/v1/chat/completions` and `/v1/completions` ([#6422](https://github.com/diegosouzapw/OmniRoute/issues/6422)). Regression guard: `tests/unit/compression-header-echo-6422.test.ts`. (thanks @chirag127)
+- **fix(api):** concurrent `GET /v1/models` requests are coalesced into a single catalog build ([#6408](https://github.com/diegosouzapw/OmniRoute/issues/6408)). Regression guard: `tests/unit/v1-models-concurrent-6408.test.ts`. (thanks @chirag127)
+- **fix(api):** `/v1/completions` now echoes the requested `body.model` in its JSON + streamed responses ([#6429](https://github.com/diegosouzapw/OmniRoute/pull/6429)). Regression guard: `tests/unit/completions-body-model-echo.test.ts`. (thanks @chirag127)
+- **fix(api):** env-var master keys now see the full `/v1/models` catalog ([#6406](https://github.com/diegosouzapw/OmniRoute/issues/6406)). Regression guard: `tests/unit/models-catalog-envkey-6406.test.ts`. (thanks @chirag127)
+- **fix(api):** non-streaming `/v1/completions` responses now echo `body.model` aligned with the `X-OmniRoute-Model` header ([#6426](https://github.com/diegosouzapw/OmniRoute/issues/6426)). Regression guard: `tests/unit/v1-completions-model-header-match-6426.test.ts`. (thanks @chirag127)
+- **fix(api):** unknown `/v1/*` routes now return a JSON `404 not_found` instead of the Next.js dashboard HTML shell ([#6405](https://github.com/diegosouzapw/OmniRoute/issues/6405)). Regression guard: `tests/unit/api/v1-catchall-json-404.test.ts`. (thanks @chirag127)
+- **fix(api):** the per-connection provider models route now degrades to the shipped catalog when a provider's `/models` endpoint answers with a redirect ([#6267](https://github.com/diegosouzapw/OmniRoute/issues/6267)) — a `qwen-web` import failed with a raw `Redirect blocked … (307)` 503. `safeOutboundFetch` throws `REDIRECT_BLOCKED` on the 307, `getSafeOutboundFetchErrorStatus` maps it to 503, and `buildDiscoveryErrorFallbackResponse` treated every 503 as a hard error — so the non-empty `getModelsByProviderId("qwen-web")` catalog was never surfaced. A models-endpoint redirect is not a fixable-config error (unlike `URL_GUARD_BLOCKED`/`INVALID_URL`, which stay hard errors), so it now falls back to the local/cached catalog before the 503 short-circuit. General fix — covers any config-driven provider that 307s. Regression guard: `tests/unit/provider-models-qwen-web-redirect-6267.test.ts`. (thanks @chirag127)
+- **fix(api):** the per-connection provider models route (MCP `list_models_catalog` + the dashboard import view) now merges USER-ADDED custom models into its response ([#6247](https://github.com/diegosouzapw/OmniRoute/issues/6247)) — custom models live in the `key_value` namespace `customModels`, which the live REST `/api/v1/models` already merges, but `src/app/api/providers/[id]/models/route.ts` never read `getCustomModels`, so custom models were dropped on both the discovery-success and local_catalog paths. They are now folded into the returned model list (deduped by id, stamped `owned_by: provider`), fixing MCP + the dashboard import view in one place. Regression guard: `tests/unit/provider-models-custom-merge-6247.test.ts`. (thanks @RCrushMe)
+
+- **fix(providers):** GitLab Duo tool-calling follow-up turns no longer fail upstream with `422 {"detail":"Validation error"}` (tokens 0/0, rejected pre-inference). The #6234 tool-result-feedback fix serialized the **entire** multi-turn conversation into GitLab's single-file `code_suggestions` (`small_file`) generation endpoint — folded history that turn-N sent as an oversized `current_file.content_above_cursor` **and** duplicated verbatim into `user_instruction`, tripping the AI-Gateway's `small_file` validation guard. The executor now **bounds** that prompt: it keeps system + latest user message + the most-recent tool round (dropping older turns), caps oversized tool results, and stops duplicating the full prompt into `user_instruction` (which now carries only the short latest user message) — while still feeding the most-recent tool result back so the agent continues (`open-sse/executors/gitlab.ts`, [#6220](https://github.com/diegosouzapw/OmniRoute/issues/6220)). The unit test covers the bounding logic; the upstream 422→200 clearing is VPS-only (Hard Rule #18). Regression guard: `tests/unit/gitlab-tool-exchange-bounded-6220.test.ts`.
+
+- **fix(i18n):** the provider-detail (`/dashboard/providers/[id]`) connection-status filter labels no longer render as `__MISSING__:All` / `__MISSING__:Active` / `__MISSING__:Error` / `__MISSING__:Banned` / `__MISSING__:CreditsExhausted` in non-English locales (notably pt-BR) ([#6290](https://github.com/diegosouzapw/OmniRoute/issues/6290)). Root cause was **not** the namespace mismatch the issue guessed — the `providers.filter*` keys resolve correctly in `en.json`; the debt lived in the locale mirrors (`src/i18n/messages/*.json`), where these five keys carried the `__MISSING__:` sync sentinel in ~15 locales and were absent entirely in ~26 others, so next-intl found the key and echoed the sentinel verbatim. All 40 non-English/-Chinese mirrors now ship real translations for the five `providers.filter*` labels. Regression guard: `tests/unit/i18n-provider-filter-keys-6290.test.ts`. (thanks @diegosouzapw)
+
+- **fix(providers):** the `copilot-m365-web` streaming executor now emits `debug`-level WebSocket diagnostics ([#6210](https://github.com/diegosouzapw/OmniRoute/issues/6210)) — the outbound WS URL (with the `access_token` **redacted** via `redactWsUrl()`), handshake success/failure, and each received SignalR frame's `type`/`target`. Previously the streaming path logged nothing, so an empty `content:null` response (the M365 Education / Starter tier symptom fixed in #6234) was undiagnosable even at `APP_LOG_LEVEL=debug`. The change is debug-level and side-effect-free — it does not alter streaming behavior or the frame parser, and the token never reaches the logs. Regression guard: `tests/unit/copilot-m365-web-logging-6210.test.ts` (thanks @qpeyba)
+
+- **fix(resilience):** a round-robin combo no longer returns `503 all upstream accounts are unavailable` when a compatibility-rejected target is actually healthy ([#6238](https://github.com/diegosouzapw/OmniRoute/issues/6238)). `filterTargetsByRequestCompatibility` drops request-incompatible targets (tool/vision/structured-output unsupported, or below the required context window) **before** any availability check runs, and its `compatible.length === 0` safety net only fired when _all_ targets were filtered — not when the kept targets later all turned out runtime-unavailable (circuit-open / cooldown / no credentials). So a combo could 503 while a compat-rejected-but-healthy provider sat unused. `handleRoundRobinCombo` now keeps the compat-rejected set and, when every compat-kept target was skipped without a single real attempt, probes those rejected targets as a **last-resort fallback tier** (via the new pure `open-sse/services/combo/comboCompatFallback.ts`) before crystallizing the 503. Regression guard: `tests/unit/combo-roundrobin-compat-fallback-6238.test.ts`. (thanks @ThongAccount)
+
+- **fix(startup):** best-effort self-heal for a corrupted Turbopack dev cache on Windows ([#6289](https://github.com/diegosouzapw/OmniRoute/issues/6289)). On Windows, `pnpm dev` can fail at startup when Turbopack `mmap`s a persistent-cache SST file and the OS refuses the mapping (`os error 1455` — "paging file too small"), which Turbopack surfaces as a misleading `Module not found: Can't resolve '@/shared/utils/machine'`. This is a **known upstream Turbopack cache-corruption bug — not our code**. The dev launcher (`scripts/dev/run-next.mjs`) now wraps `nextApp.prepare()` and, when it rejects with that signature (`isTurbopackCacheCorruption` in the new `scripts/dev/turbopackCacheHeal.mjs`), purges `.build/next/**/cache/turbopack` and retries **once** with a clear log. **Caveat — best-effort only:** the corruption often surfaces as a runtime overlay rather than a `prepare()` rejection, so this cannot always intercept it; the reliable remedy remains manually deleting the Turbopack cache dir. Regression guard: `tests/unit/turbopack-cache-heal-6289.test.ts`. (thanks @chirag127)
+
+- **fix(providers):** qodercli PAT auth no longer fails with `spawn qodercli ENOENT` on Windows ([#6263](https://github.com/diegosouzapw/OmniRoute/issues/6263)) — `spawnQoderCli` spawned the bare `qodercli` name with `shell:false` and an unenriched env, so the npm `.cmd` wrapper under `%APPDATA%\npm` (a user-PATH directory) was never resolved. It now resolves the absolute `.cmd`/`.exe` path through the existing `getCliRuntimeStatus("qoder")` resolver in `src/shared/services/cliRuntime.ts` (memoized), spawns with `shell` when the target is a `.cmd`/`.bat`, and uses the cliRuntime-enriched env (PATH + PATHEXT + APPDATA); the ENOENT error now lists the searched paths plus the `CLI_QODER_BIN` override. End-to-end spawn on a real Windows host is host-only (Hard Rule #18); the path-resolution logic is unit-tested. Regression guard: `tests/unit/qodercli-windows-resolve-6263.test.ts`. (thanks @chirag127)
+
+- **fix(sse):** the reasoning-token buffer no longer inflates **probe-sized `max_tokens`** ([#6274](https://github.com/diegosouzapw/OmniRoute/issues/6274)) — Claude Code's `/model` capability check sends `max_tokens: 1`, but for a thinking-capable model with a large output cap (e.g. `glm-5.2`) the #3587 headroom heuristic (`max(current + 1000, ceil(current * 1.5))`) rewrote it to `1001` and forwarded that upstream, wasting tokens on a request that was never a genuine reasoning budget. `resolveReasoningBufferedMaxTokens()` (`open-sse/services/reasoningTokenBuffer.ts`) now short-circuits and returns the caller's value verbatim when it is below the new `REASONING_BUFFER_MIN_TRIGGER` (256) threshold — a tiny explicit limit is a probe, not a reasoning request. Real budgets still receive the #3587 headroom unchanged, and the guard runs after the existing capability checks so unknown / non-reasoning models keep returning `null`. Regression guard: `tests/unit/reasoning-token-buffer-6274.test.ts`. (thanks @brightfiscalband)
+
+- **fix(cli):** `omniroute reset-password` now works as a real subcommand, and password resets over piped (non-TTY) stdin actually apply ([#6261](https://github.com/diegosouzapw/OmniRoute/issues/6261), [#6258](https://github.com/diegosouzapw/OmniRoute/issues/6258)). Two coupled defects: (1) **#6261** — `bin/omniroute.mjs` routed everything through Commander with only two pre-Commander bypasses (`--mcp`, `reset-encrypted-columns`), so `omniroute reset-password` was rejected as an unknown command; only the separate `omniroute-reset-password` bin worked, while the docs falsely advertised the subcommand (incl. a bogus "legacy alias still works"). A pre-Commander bypass mirroring `reset-encrypted-columns` now dynamically imports `bin/reset-password.mjs` (which self-executes) before Commander parses; the three doc lines were corrected. (2) **#6258** — `bin/reset-password.mjs` issued two sequential `rl.question` prompts; under piped stdin the second read never settled at EOF, so `main()` never reached `resetManagementPassword` and the reset was a silent no-op (both prompts printed, no success, password unchanged). The CLI now detects non-TTY stdin and reads it once (first line = password, second line = confirm if present, else reused), adds a `--password-stdin` flag (entire stdin is the password, no confirmation), and exits `0` explicitly so the success line always flushes; interactive TTY behavior is unchanged. Regression guard: `tests/unit/reset-password-cli-6261-6258.test.ts` (3). (thanks @chirag127)
+
+- **fix(db):** the mass-migration **safety abort** now tells the operator how to bypass it and stops flooding the log ([#6260](https://github.com/diegosouzapw/OmniRoute/issues/6260)) — after restoring a backup that wiped the migration tracking table, `runMigrations()` threw the abort on every downstream `ensureDbInitialized()`, re-logging the full banner 11+ times, and the message never mentioned the existing `OMNIROUTE_MAX_PENDING_MIGRATIONS` escape hatch. The abort text now appends a bypass hint (set `OMNIROUTE_MAX_PENDING_MIGRATIONS=0` in `server.env` / `DATA_DIR/.env`), and a new `MigrationSafetyAbortError` is memoized so repeated calls in the same process throw the same instance and emit a single concise line instead of the full cascade. Regression guard: `tests/unit/migration-safety-abort-6260.test.ts`. (thanks @chirag127)
+
+- **fix(auth):** importing a **distinct** Codex/ChatGPT OAuth `auth.json` is no longer falsely rejected as "already exists" when it belongs to a different user in the same workspace ([#6301](https://github.com/diegosouzapw/OmniRoute/issues/6301)). `findExistingCodexConnection` (in `src/lib/oauth/utils/codexAuthImport.ts`) deduped **only** on `providerSpecificData.workspaceId === accountId`, where `accountId` is the shared `chatgpt_account_id`/`tokens.account_id` — so two members of the same ChatGPT Team collapsed onto a single connection (409 `duplicate_account`). The id_token's `https://api.openai.com/auth` claim carries a per-user `chatgpt_user_id` alongside the workspace id (the device-flow path already persisted it as `chatgptUserId`, but the import path did not). Now `parseAndValidateCodexAuth` extracts `userId` (`chatgpt_user_id` → `user_id` → JWT `sub`) into `ParsedCodexAuth`, the create/update paths persist `chatgptUserId` in `providerSpecificData` (mirroring `codex.ts`), and dedup keys on `workspaceId` **AND** `chatgptUserId` — with a backward-compat fallback to legacy accountId-only matching when no stored connection for that workspace records a `chatgptUserId`, so genuinely-same accounts still dedup. Regression guard: `tests/unit/codex-auth-import-userid-dedup-6301.test.ts` (4). (thanks @anungma)
+
+- **fix(providers):** importing models for the **venice-web** provider no longer fails with a red "Provider venice-web does not support models listing" ([#6269](https://github.com/diegosouzapw/OmniRoute/issues/6269)). `venice-web` is a web-cookie provider with an executor but no upstream `/v1/models` endpoint and no registry `models`, so the models route fell through to the tail `400`. Mirroring the `jules`/`linkup-search`/`ollama-search` fix (#5569), it now ships a static local catalog entry in `src/lib/providers/staticModels.ts` — seeding the current Venice lineup (`venice-uncensored`, `llama-3.3-70b`, `qwen3-235b`, `qwen3-4b`, `deepseek-r1-671b`; Venice rotates its catalog, see docs.venice.ai/models/overview) — so the route returns `200` with `source:"local_catalog"`, `intentional:true`. Regression guard: `tests/unit/static-models-venice-web-6269.test.ts`. (thanks @chirag127)
+
+- **fix(api):** the specialty model catalogs (`/v1/embeddings`, `/v1/images`, `/v1/music`, `/v1/videos` model lists) are now derived from the **unified catalog filtered by a predicate** (`getSpecialtyModelsResponse`) instead of ad-hoc per-route logic, so they consistently respect active-credential visibility and stay in sync with the main catalog ([#6303](https://github.com/diegosouzapw/OmniRoute/pull/6303)). Regression guard: `tests/unit/specialty-model-catalog-routes.test.ts`. (thanks @makcimbx)
+- **fix(api):** the agent-bridge server route now resolves the MITM manager via a **dynamic `import("@/mitm/manager.runtime")`** so Turbopack does not statically pull the stub (or over-bundle the manager), and the agent-skills generator anchors its output base path with `path.join(process.cwd(), …)` so Turbopack's static analyzer stops tracing the whole project root ([#6329](https://github.com/diegosouzapw/OmniRoute/issues/6329), [#6366](https://github.com/diegosouzapw/OmniRoute/pull/6366)). Regression guard: `tests/unit/agent-bridge-server-route-dynamic-import.test.ts`. (thanks @Iammilansoni)
+- **fix(api):** internal probes (combo-test, cloud-sync verify) now pick a **management-scoped / allow-all API key** instead of naively grabbing `getApiKeys()[0]` — a restricted `self:usage` first row made the probe fail with "Model X is not allowed for this API key" even when the combo path was healthy (`pickApiKeyForInternalUse` in `src/lib/db/apiKeys.ts`). The API-manager model editor also falls back to `/api/models?all=true` when `/v1/models` is catalog-protected ([#6372](https://github.com/diegosouzapw/OmniRoute/pull/6372)). Regression guard: `tests/unit/pick-internal-api-key-6372.test.ts`. (thanks @jmengit)
+- **fix(live-ws):** the Live Dashboard WebSocket server now **rejects on bind failure** (e.g. `EADDRINUSE` when the API bridge already holds the port) instead of letting the error surface as an unhandled `error` event that crash-loops the process — the `error` listener is attached to `wss` (not `server`) and releases the EventBus subscription on a failed start ([#6324](https://github.com/diegosouzapw/OmniRoute/issues/6324)). Regression guard: `tests/unit/live-ws-eaddrinuse-6324.test.ts`. (thanks @vinayakkulkarni)
+- **fix(dashboard):** the Home provider-topology widget now trusts the live provider-metrics snapshot — it uses `topology.errorProvider` and live `activeRequests` directly instead of re-deriving state from a stale `lastErrorAt` or applying a frontend timeout filter, so the topology reflects real-time provider health ([#6322](https://github.com/diegosouzapw/OmniRoute/pull/6322)). Regression guard: `tests/unit/home-provider-topology-live-state.test.ts`. (thanks @xz-dev)
+- **fix(sse):** strip zero-width markers from streamed **tool-call arguments** — a follow-up to [#5857](https://github.com/diegosouzapw/OmniRoute/pull/5857). That PR removed injected zero-width joiners (U+200D) from streamed assistant text/reasoning but deliberately left tool-call argument JSON byte-exact. The request-side obfuscation (`open-sse/services/claudeCodeObfuscation.ts`) injects ZWJ into agent words — including the temp path inside the Bash tool description — and Claude models copy that verbatim into generated commands, which are delivered as tool-call arguments rather than assistant text. As a result the ZWJ survived and corrupted code blocks (e.g. a temp path rendered with an invisible joiner). Now `open-sse/handlers/responseSanitizer.ts` strips zero-width code points from tool-call argument strings at every emit site (OpenAI non-stream/stream chat `tool_calls` + legacy `function_call`, native Responses `function_call` items, the OpenAI→Responses conversion, and the native Responses streaming `response.function_call_arguments.delta/.done` events). Only zero-width code points are removed; JSON structure and all other bytes stay identical (no parse/restringify), so normal arguments remain byte-exact. Regression guard: 6 new cases in `tests/unit/response-sanitizer.test.ts` (suite 50/50).
+
+- **fix(nodejs):** the default app log path now resolves under `DATA_DIR` (`~/.omniroute/logs/application/app.log`) instead of `process.cwd()` ([#6197](https://github.com/diegosouzapw/OmniRoute/issues/6197)) — the globally-installed CLI runs from an arbitrary working directory, so anchoring the default to cwd made file logging silently write to (or no-op under) an unrelated directory, contradicting the documented `.env.example` default. `getAppLogFilePath()` now computes the default lazily via the pure `resolveDataDir()` resolver (honours a per-process `DATA_DIR`, no directory-creation side effect); an explicit `APP_LOG_FILE_PATH` still wins. Regression guard: `tests/unit/logenv-datadir-path-6197.test.ts` (3).
+
+- **fix(docker):** AgentBridge/`startMitm` no longer aborts in containers/headless when the Antigravity-default DNS step can't write `/etc/hosts` ([#6127](https://github.com/diegosouzapw/OmniRoute/issues/6127)), and the privileged command's stderr now reaches `app.log` instead of only a bare exit code hitting the toast ([#6198](https://github.com/diegosouzapw/OmniRoute/issues/6198)). The default DNS step (`addDNSEntry`) was called unguarded while cert install and the two sibling DNS steps were each best-effort — in the runtime Docker image (`USER node`, no `sudo`, read-only `/etc/hosts`) it threw `Command failed with code 1` out of `startMitmInternal` and killed the whole start, discarding the stderr. The three DNS steps are extracted into a best-effort `provisionDnsEntries()` where each failure is logged with the full `err` (stderr included, folded in by `systemCommands.ts`) and never aborts the start. Regression guard: `tests/unit/mitm-dns-graceful-degrade-6127.test.ts` (4).
+
+- **fix(providers):** copilot-m365-web now supports the M365 Education "Starter / OfficeWebIncludedCopilot" tier and no longer returns an empty `content:null` stream ([#6210](https://github.com/diegosouzapw/OmniRoute/issues/6210)). Two gaps: (1) `buildWsUrl()` hardcoded the individual-consumer scenario (`OfficeWebPaidConsumerCopilot`, `isEdu=false`) — the EDU tier is now opt-in via `providerSpecificData.tier="edu"`, emitting `scenario=OfficeWebIncludedCopilot`/`isEdu=true` (the individual path is unchanged); (2) the EDU/GPT-5.5 path streams deltas via `arguments[0].writeAtCursor` (incremental) instead of only `messages[].text` (accumulated snapshots), which the parser dropped — a new `accumulateBotContent()` folds both formats, with `type:2 item.result.message` as a last-resort fallback. Regression guard: `tests/unit/copilot-m365-edu-writeatcursor-6210.test.ts` (10). (thanks @qpeyba)
+
+- **fix(providers):** GitLab Duo executor now feeds tool results back into the prompt instead of looping ([#6220](https://github.com/diegosouzapw/OmniRoute/issues/6220)) — `buildPrompt()` branched only on `system`/`user` and took `userParts.at(-1)`, silently dropping the `assistant{tool_calls}` + `tool{result}` turns the client appended, so the reconstructed prompt was byte-identical to turn 1 and the model re-emitted the same `<tool>` call forever. When a tool exchange is present the full conversation is now serialized, folding each tool result back keyed by its `tool_call_id`; simple conversations keep the legacy shape. Complements the tool_call emission from [#6051](https://github.com/diegosouzapw/OmniRoute/issues/6051) (the `kilo-duplicate` label was a false positive — different, sequential defect). Regression guard: `tests/unit/gitlab-tool-result-feedback-6220.test.ts` (4).
+
+- **fix(providers):** opencode-go/opencode-zen can now synthesize the OpenCode CLI identity headers Cloudflare requires on VPS egress ([#5997](https://github.com/diegosouzapw/OmniRoute/issues/5997)) — on a datacenter VPS, `opencode.ai/zen/go/v1/chat/completions` 403s (HTML challenge) requests lacking CLI identity, while the reporter's control curl proved that `User-Agent: opencode-cli/1.0.0` + `x-opencode-client: cli` + `x-opencode-project: default` + fresh request/session UUIDs succeed. Opt-in via `OPENCODE_SYNTHESIZE_CLI_HEADERS=true` (values overridable via `OPENCODE_GO_USER_AGENT`/`OPENCODE_USER_AGENT`/`OPENCODE_CLIENT`/`OPENCODE_PROJECT`); it fills only headers the client did not already send. Kept **off by default** — the forward-only path is deliberate (fabricating a wrong value risks upstream rejection; a prior dedup regressed with `opencode/local`), so this replaces the fragile local header-injection shim without changing default behavior. Regression guard: `tests/unit/opencode-cli-headers-synthesis-5997.test.ts` (6). (thanks @aleksesipenko)
+- fix(resilience): sticky session affinity now evicts and fails over to another account when the pinned account is exhausted/unavailable (#6219)
+
+- fix(sse): Responses API passthrough now drops internal commentary-phase output before forwarding to clients (gated by RESPONSES_PASSTHROUGH_DROP_COMMENTARY, default on) (#6199)
+
+- **fix(sse):** tool-call function schemas with a root `type: null` are now coerced to `type: "object"` before dispatch ([#6359](https://github.com/diegosouzapw/OmniRoute/issues/6359)) — clients like the Codex app emit `parameters: { type: null, ... }` for some tools, which OpenAI-compatible upstreams reject with `400 Invalid schema for function '...': schema must be a JSON Schema of 'type: "object"', got 'type: null'`, failing the whole request. `toolSchemaSanitizer` already stripped the null; it now re-adds the mandatory root `"object"` type (and empty `properties`/open `additionalProperties` when absent). Combinator roots (`anyOf`/`oneOf`/`allOf`) and explicit root types are left untouched. Regression guard: 5 new cases in `tests/unit/tool-schema-sanitizer.test.mjs`.
+
+- **fix(docker):** AgentBridge no longer fails to start on npm/Electron/VPS installs with "MITM manager stub reached at runtime" ([#6344](https://github.com/diegosouzapw/OmniRoute/issues/6344)) — v3.8.45 flipped the production bundler default to Turbopack, but `next.config.mjs` aliased `@/mitm/manager` to its Docker-only degraded stub **unconditionally**. That was harmless while Docker (which sets the alias intentionally for #3390 graceful degradation) was the sole Turbopack consumer, but once every artifact built with Turbopack the stub shipped to all non-Docker users and `startMitm` threw on the first Agent-Bridge start. The alias is now opt-in via `OMNIROUTE_MITM_STUB=1` (set only by the Dockerfile) through the shared `scripts/build/mitm-stub-flag.mjs` helper; default builds bundle the real manager. Regression guard: `tests/unit/mitm-stub-alias-6344.test.mjs` (4).
+
+- **fix(proxy):** stop the v3.8.44 proxy regression that leaked the real IP and disabled healthy proxies ([#6246](https://github.com/diegosouzapw/OmniRoute/issues/6246)). Two coupled defects from the new health scheduler: (1) **IP leak** — when a proxy assigned to a connection was marked `inactive`, resolution fell through to a **direct** egress instead of blocking, exposing the operator's real IP; (2) **over-deactivation** — the sweep flipped a proxy to `inactive` on the **first** failed probe and counted our own 5s timeout / a probe-target `5xx` as the proxy's fault, so healthy paid proxies vanished from egress selection ("my proxies are not being used anymore"). Fix: the sweep decision is extracted into a pure, network-free `decideProxyHealthAction` (`src/lib/proxyHealth/decision.ts`) — by default the health check now **only counts/logs and never downgrades status** (a proxy is downgraded/removed only with `PROXY_AUTO_REMOVE=true`, after `PROXY_AUTO_REMOVE_AFTER` **consecutive** conclusive failures); probes are classified tri-state so an inconclusive result (our timeout, or a `5xx` from the probe target) never penalizes the proxy, and the probe timeout is raised 5s→15s. Separately, `safeResolveProxy` now **fails closed** via the existing policy: a connection whose assigned proxy is dead is blocked instead of leaking direct (`hasBlockingProxyAssignment`), honoring the explicit `proxy off` toggles and the `PROXY_FAIL_OPEN=true` opt-out. Existing proxies stuck `inactive` by the old behavior need a one-time manual re-activate (the operator owns proxy status). Regression guards: `tests/unit/proxy-health-decide-action-6246.test.ts`, `tests/unit/proxy-assigned-unavailable-6246.test.ts`.
+
+- **fix(proxy):** make "Test All" read-only and add bulk enable/disable ([#6246](https://github.com/diegosouzapw/OmniRoute/issues/6246)). Complements the core fail-closed / scheduler fix (#6296) with the two remaining reporter asks. (1) The **"Test All" button** (`POST /api/settings/proxies/auto-test`) used to flip a proxy to `inactive` on a failed reachability probe; since the egress selector excludes `inactive` proxies, a flaky probe (an unreachable `httpbin.org`, a proxy that blocks `HEAD`, or a slow paid proxy) silently disabled every proxy that failed — "Test All" is now **read-only by default** (only the operator sets a proxy active/inactive; opt back into the legacy test-and-set with `PROXY_HEALTH_AUTO_DEACTIVATE=true`). (2) Adds a **bulk enable/disable** proxies endpoint + toolbar action (`POST /api/settings/proxies/batch-activate`) so an operator can re-activate proxies in one click. Regression guard: `tests/unit/proxy-health-6246.test.ts`. (thanks @tenshiak)
+
+- **chatcore (tools): stop the default 128-tool cap from silently dropping opencode's `task`/MCP tools.** opencode (used as an MCP/agent host) sends a large tool list; when it exceeds the speculative `MAX_TOOLS_LIMIT` (128) default, `truncateToolList` did a blind `tools.slice(0, 128)`, dropping every tool past index 128 — including opencode's built-in `task` tool (subagent launch) and many MCP tools, so models routed through OmniRoute could no longer spawn subagents or reach part of their tools. The cap exists to avoid upstream `400`s for providers with real hard limits (e.g. grok-cli 200), so it is kept for those: detection of the opencode client (`isOpencodeClient` — any `x-opencode-*` header, or `opencode` in the user-agent) now only bypasses the **speculative 128 default**, never a known provider ceiling. Precedence is explicit — a proactive/detected provider limit always truncates (even for opencode); otherwise opencode forwards its full tool list; otherwise the unchanged 128 default applies to every other client. Refactors `getEffectiveToolLimit` into `getKnownToolLimit(provider) ?? DEFAULT_LIMIT` (byte-identical for existing callers) and fixes a cosmetic debug-log that reported the truncated count instead of the original. Regression guard: `tests/unit/tool-limit-detector.test.ts`.
+
+- **fix(mitm):** the macOS MITM-cert install check now matches the system keychain again. `security find-certificate -a -Z` prints the SHA-1 as a colon-less hex string, but the installed-check compared it against `getCertFingerprint()`'s colon-separated form, so the substring match never hit — the cert was reported as not-installed and re-prompted for the sudo install on every run. Fingerprints are now normalized (colons stripped, upper-cased) on both sides via the extracted `macCertOutputHasFingerprint` helper. Regression guard: `tests/unit/mitm-cert-mac-fingerprint.test.ts`. ([#6204](https://github.com/diegosouzapw/OmniRoute/pull/6204), closes [#6134](https://github.com/diegosouzapw/OmniRoute/issues/6134) — thanks @rianonehub)
+
+- **fix(api):** `/v1/messages/count_tokens` now counts `tool_use`, `tool_result` and `thinking` content blocks (and array-form `system` prompts) in the local-estimation path, instead of only `text`. Real agentic conversations keep ~95% of their tokens inside tool results; the previous estimate returned near-zero for them, which silently broke Claude Code's auto-compaction (context grew past the window with no compaction until the upstream API rejected the request). The real provider-side count path is unchanged. Regression guard: `tests/unit/messages-count-tokens-route.test.ts`. ([#6221](https://github.com/diegosouzapw/OmniRoute/pull/6221) — thanks @luweiCN)
+
+- **fix(antigravity):** strip a trailing assistant prefill turn for Vertex Claude models to avoid upstream 400s ([#6114](https://github.com/diegosouzapw/OmniRoute/pull/6114)). Regression guard: `tests/unit/antigravity-claude-prefill-strip.test.ts`. (thanks @anki1kr)
+
+- **fix(security):** the mutable cloud-agent routes (`/api/cloud/credentials/update`, `/api/cloud/models/alias`) now require management auth instead of being treated as public. They were classified as public API routes, so a request without management credentials could update stored cloud-agent credentials and model aliases. They are removed from the public-route set, classified as management routes in the authz pipeline, and gated by `requireManagementAuth`; cloud **read**/auth routes stay public. Regression guards: `tests/unit/cloud-write-auth.test.ts`, `tests/unit/authz/classify.test.ts`, `tests/unit/public-api-routes.test.ts`. ([#6233](https://github.com/diegosouzapw/OmniRoute/pull/6233) — thanks @vittoroliveira-dev)
+
+- **refactor(dashboard):** extract the onboarding-wizard "Open provider details" link target into a pure, unit-tested `buildProviderDetailsHref(connection)` helper. The wizard already routes by `connection.id` (the node UUID) rather than the provider category slug (#6144/#6145); this hardens that behavior behind a tested helper that guards a missing id/connection. Regression guard: `tests/unit/provider-onboarding-href.test.ts`. ([#6166](https://github.com/diegosouzapw/OmniRoute/pull/6166) — thanks @KooshaPari)
+
+- **fix(security):** the doubao synthetic device-id generator now derives its digits via an unbiased crypto-random draw (rejection sampling over `crypto.randomBytes()`) instead of a `% 10` reduction, closing a CodeQL `js/biased-cryptographic-random` finding.
+- **fix(agentSkills):** the GitHub-skills generator now resolves `outputDir` to an absolute path before writing, fixing a regression introduced by #6366 (relative-to-cwd base path) that could write generated skill files to the wrong directory.
+- **fix(security):** `/api/keys/{id}/devices` now checks the HTTP method before auth/validation, returning a `405` for non-GET/DELETE verbs instead of a misleading `401`/`500` (closes a `dast-smoke` QUERY-method finding).
+- **fix(quality):** clear the last 2 heavy quality-gate reds on the release tip (cycle pre-flight).
+- **fix(mitm):** the test suite and CI can never mutate the OS trust store — `OMNIROUTE_SKIP_SYSTEM_TRUST=1` is set globally for tests/CI so `installCert`/`uninstallCert`/`installTproxyCa` skip the privileged OS dispatch ([#6310](https://github.com/diegosouzapw/OmniRoute/pull/6310); full detail is under the [3.8.45] section below — this branch received it via the parallel-cycle sync-back).
+- **fix(api):** `POST /api/github-skills` now Zod-validates its request body; documented the new quality-gate env vars and pinned the merge-integrity GitHub Actions to a commit SHA.
+- **fix(skills):** generate the missing `omni-github-skills` registry entry and align the agent-skills catalog-count tests (follow-up to #6186).
+- **fix(quality):** clear the cycle's 11 net-new ESLint errors and make `validate-release-green` suppressions-aware.
+
+### 📝 Maintenance
+
+- **i18n(it):** add 118 missing Italian (`it`) translations (net-additive — no existing keys dropped, valid JSON), improving Italian UI coverage. ([#6212](https://github.com/diegosouzapw/OmniRoute/pull/6212) — thanks @serverless83)
+- **chore(providers):** remove deprecated **MiMo V2** model entries from the catalogs (xiaomi-mimo, opencode-go, zenmux-free, audio TTS) — the upstream V2 line is superseded by MiMo V2.5; drops `mimo-v2-tts`, `mimo-v2-pro`, `mimo-v2-omni`, `mimo-v2-flash`, `mimo-v2-flash-free` and realigns the provider-catalog tests. ([#6248](https://github.com/diegosouzapw/OmniRoute/pull/6248) — thanks @backryun)
+- **chore(release):** ~50 commits on this branch are v3.8.45 pre-flight/hardening fixes and CI-perf work that landed here via the parallel-cycle sync-back (`sync-next-cycle.mjs`, Hard Rule #21) **after** the `v3.8.45` git tag was cut, and are already fully documented under the **[3.8.45]** section below — listed here only so the per-cycle commit-coverage check (`npm run release:uncovered`) doesn't flag them as gaps. Provider/catalog/UX/backend fixes: #6041, #6078, #6108, #6135, #6148, #6149, #6154, #6158, #6161, #6162, #6163, #6164, #6165, #6170, #6177, #6178, #6181, #6186, #6187, #6191, #6193, #6194, #6195, #6200, #6205, #6208, #6209, #6211, #6213, #6223, #6224, #6225, #6226, #6227, #6228, #6229, #6230, #6235, #6291, #6292. CI/release-pipeline work: #6167, #6203, #6214, #6215, #6218, #6273, #6275, #6283, #6284, #6285, #6300, #6305.
+- **chore(release):** additional zero-ref release-cycle plumbing on this branch, kept out of `release:uncovered` on purpose (no `#N` in the commit subject to cite): opening the v3.8.46 cycle, opening/closing the v3.8.45 cycle, the finalized [3.8.45] CHANGELOG i18n sync-back to 42 mirrors, the v3.8.45 cognitive/cyclomatic and file-size drift rebaselines, ESLint stale-suppression pruning (4,273 → 4,233), and clearing test-masking/docs-all pre-flight reds for v3.8.45.
+
+### ⚡ Performance & Infrastructure
+
+- **perf(release-green):** the pre-flight validator (`scripts/quality/validate-release-green.mjs`) now runs its 4 slow suites (unit / vitest / integration / pack-artifact) **concurrently** via `Promise.all` — pre-flight wall time drops from ~the sum of the suites to ~the slowest one (~30min saved per round; Phase 0 was the nº1 bottleneck of the v3.8.45 release benchmark, 2h54 of 6h34 e2e). Guard: `tests/unit/validate-release-green.test.ts` ("runs the slow suites CONCURRENTLY"). ([#6319](https://github.com/diegosouzapw/OmniRoute/pull/6319))
+- **fix(ci):** `scripts/release/sync-next-cycle.mjs` — two defects found live in its first production run (v3.8.45 Phase 5): (1) the `git()` helper's default 1 MiB `maxBuffer` crashed with `ENOBUFS` on `git show origin/main:CHANGELOG.md` (the CHANGELOG alone is >1 MiB) — widened to 64 MiB; (2) the i18n resync only propagated the `[NEXT]` (TBD) section, leaving the just-shipped finalized section as "— TBD" in all 42 mirrors — it now also syncs `[prevVersion]` bounded by the heading below it (new exported pure helper `versionAfter`). Guards: +5 tests in `tests/unit/sync-next-cycle.test.ts` (8/8). ([#6327](https://github.com/diegosouzapw/OmniRoute/pull/6327))
+- **test(ci):** concurrency-sensitive flaky tests are **quarantined into a serial pass** (`tests/unit/serial/`, `--test-concurrency=1`, appended to every unit runner incl. sharded variants — the serial pass is sharded too so concurrent shard jobs never self-collide). Initial set: `glm-coding-plan-monthly-3580`, `quota-division-blocks`, `provider-health-autopilot`, `combo-health-autopilot` — the class behind the ~28min CI wedges/re-runs (two live 1h+ wedges cancelled during this PR's own validation). Discovery + TIA gates track the new glob; systemic root cause (async logger writing after teardown) tracked in [#6360](https://github.com/diegosouzapw/OmniRoute/issues/6360). Guard: `tests/unit/test-serial-quarantine.test.ts` (4). ([#6347](https://github.com/diegosouzapw/OmniRoute/pull/6347))
+
+### 🙌 Contributors
+
+Thanks to everyone whose work landed in v3.8.46:
+
+| Contributor                                                    | PRs / Issues           |
+| -------------------------------------------------------------- | ---------------------- |
+| [@2220258345](https://github.com/2220258345)                   | direct commit / report |
+| [@abdofallah](https://github.com/abdofallah)                   | direct commit / report |
+| [@adentdk](https://github.com/adentdk)                         | direct commit / report |
+| [@aleksesipenko](https://github.com/aleksesipenko)             | direct commit / report |
+| [@anki1kr](https://github.com/anki1kr)                         | direct commit / report |
+| [@anungma](https://github.com/anungma)                         | direct commit / report |
+| [@arpicato](https://github.com/arpicato)                       | direct commit / report |
+| [@backryun](https://github.com/backryun)                       | #6248                  |
+| [@binsarjr](https://github.com/binsarjr)                       | direct commit / report |
+| [@brightfiscalband](https://github.com/brightfiscalband)       | direct commit / report |
+| [@chirag127](https://github.com/chirag127)                     | #6501, #6506           |
+| [@developerjillur](https://github.com/developerjillur)         | direct commit / report |
+| [@dilneiss](https://github.com/dilneiss)                       | #6499                  |
+| [@dtybnrj](https://github.com/dtybnrj)                         | direct commit / report |
+| [@Forcerecon](https://github.com/Forcerecon)                   | direct commit / report |
+| [@hao3039032](https://github.com/hao3039032)                   | direct commit / report |
+| [@Iammilansoni](https://github.com/Iammilansoni)               | #6150, #6245           |
+| [@jmengit](https://github.com/jmengit)                         | direct commit / report |
+| [@jordansilly77-stack](https://github.com/jordansilly77-stack) | direct commit / report |
+| [@JxnLexn](https://github.com/JxnLexn)                         | direct commit / report |
+| [@KooshaPari](https://github.com/KooshaPari)                   | #6166                  |
+| [@loopyd](https://github.com/loopyd)                           | direct commit / report |
+| [@luweiCN](https://github.com/luweiCN)                         | #6221                  |
+| [@makcimbx](https://github.com/makcimbx)                       | direct commit / report |
+| [@muflifadla38](https://github.com/muflifadla38)               | direct commit / report |
+| [@newnol](https://github.com/newnol)                           | direct commit / report |
+| [@ofekbetzalel](https://github.com/ofekbetzalel)               | direct commit / report |
+| [@ohahe52-dot](https://github.com/ohahe52-dot)                 | direct commit / report |
+| [@phidinhmanh](https://github.com/phidinhmanh)                 | direct commit / report |
+| [@powellnorma](https://github.com/powellnorma)                 | direct commit / report |
+| [@qpeyba](https://github.com/qpeyba)                           | direct commit / report |
+| [@RaviTharuma](https://github.com/RaviTharuma)                 | direct commit / report |
+| [@RCrushMe](https://github.com/RCrushMe)                       | direct commit / report |
+| [@rianonehub](https://github.com/rianonehub)                   | #6134, #6204           |
+| [@serverless83](https://github.com/serverless83)               | #6212                  |
+| [@swingtempo](https://github.com/swingtempo)                   | direct commit / report |
+| [@tenshiak](https://github.com/tenshiak)                       | direct commit / report |
+| [@ThongAccount](https://github.com/ThongAccount)               | direct commit / report |
+| [@UnrealAryan](https://github.com/UnrealAryan)                 | direct commit / report |
+| [@vinayakkulkarni](https://github.com/vinayakkulkarni)         | direct commit / report |
+| [@vittoroliveira-dev](https://github.com/vittoroliveira-dev)   | #6233                  |
+| [@warelik](https://github.com/warelik)                         | direct commit / report |
+| [@xxy9468615](https://github.com/xxy9468615)                   | direct commit / report |
+| [@xz-dev](https://github.com/xz-dev)                           | direct commit / report |
+| [@yanpaing007](https://github.com/yanpaing007)                 | direct commit / report |
+| [@diegosouzapw](https://github.com/diegosouzapw)               | maintainer             |
+
+---
+
 ## [3.8.45] — 2026-07-06
 
 ### ✨ New Features

docs/i18n/pl/CHANGELOG.md

@@ -6,6 +6,213 @@
 
 ## [3.8.31] — 2026-06-20
 
+## [3.8.46] — TBD
+
+### ✨ New Features
+
+- **feat(sse):** **hide paid-only models from `auto/*` routing** when `hidePaidModels` is on ([#6512](https://github.com/diegosouzapw/OmniRoute/issues/6512)) — follow-up to #6328/#6495. PR #6495 hid paid-only models from the `GET /v1/models` listing, but `auto/*` combos (`auto/best-coding`, `auto/glm`, …) could still pick a paid-only backend into their candidate pool → a 402/403 at request time. `createVirtualAutoCombo` now filters the candidate pool through the new pure `open-sse/services/autoCombo/paidModelFilter.ts` (`filterPaidOnlyCandidates`), applying the same free-model predicate #6495 uses in `catalog.ts` (`providerHasFreeModels(provider) && isFreeModel(provider, {id})`) whenever `settings.hidePaidModels === true`. Applied before the category/tier/family narrowing, so it covers every `auto/*` combo; an all-paid pool degrades to the existing graceful empty-pool path. **Opt-in — default OFF leaves the pool unchanged** (identity). Regression guard: `tests/unit/autoCombo/paid-model-filter-6512.test.ts` (4, incl. the default-off identity guard).
+- **feat(sse):** **provider-family auto combos** — `auto/glm`, `auto/minimax`, `auto/mimo`, `auto/zai`, `auto/gemma`, `auto/llama`, `auto/gemini` ([#6453](https://github.com/diegosouzapw/OmniRoute/issues/6453)) — new routable ids that materialize an on-demand virtual combo spanning whatever installed backends currently expose that model family, degrading gracefully as backends rotate. A new pure `open-sse/services/autoCombo/modelFamily.ts` (`detectModelFamily`) classifies by model-id prefix for six families; `zai` is instead resolved by provider id (z.ai's hosted API serves the same `glm-*` model ids as every other GLM backend, so `auto/zai` means "route to my z.ai backend specifically" vs `auto/glm`'s "any connected GLM backend"). Reuses the existing `createVirtualAutoCombo` on-demand materialization path (no DB writes) and the `/v1/models` catalog advertising loop. Regression guard: `tests/unit/autoCombo/provider-family-combos.test.ts` (11).
+- **feat(proxy):** native **proxy-pool round-robin / egress IP rotation** ([#6365](https://github.com/diegosouzapw/OmniRoute/issues/6365)) — a scope (global / provider / account) can now hold **multiple** proxies as a pool with a rotation strategy, so outbound requests cycle their egress IP instead of pinning one proxy per scope. Migration `117_proxy_pool_rotation.sql` lifts the `UNIQUE(scope, scope_id)` constraint (rebuild via the canonical rename/copy/drop; existing single assignments become 1-element pools) and adds a `proxy_scope_rotation` companion table holding the per-scope strategy + a persisted monotonic round-robin cursor. Strategies: `round-robin` (default, monotonic cursor — never `Math.random`), `random`, and `sticky-per-N-min`. Resolution (`resolveProxyForScopeFromRegistry` / `resolveProxyForConnectionFromRegistry`) now fetches the alive, position-ordered candidate set (unchanged `PROXY_ALIVE_PREDICATE`) and applies the strategy; an empty / all-dead pool still returns `null` — the #6246 fail-closed guard is **untouched** (never falls through to direct egress). Backend + DB only; dashboard pool-builder UI is a follow-up. Regression guard: `tests/unit/proxy-pool-rotation-6365.test.ts` (8, incl. fail-closed + backward-compat).
+- **feat(providers):** end-to-end **tool/function calling on the native Gemini `/v1beta` endpoint** ([#6222](https://github.com/diegosouzapw/OmniRoute/issues/6222)) — both directions of the Gemini↔OpenAI conversion now preserve tool data (previously silently dropped). Request side: `convertGeminiToInternal` (extracted to its own testable module) maps `tools[].functionDeclarations` → OpenAI `tools`, prior `functionCall` parts → assistant `tool_calls`, and `functionResponse` parts → `tool`-role messages. Response side: `convertOpenAIResponseToGemini` emits `parts[].functionCall {name,args}` from `message.tool_calls`, and the streaming `openAIChunkToGeminiChunk` accumulates fragmented `tool_calls` deltas by index into complete `functionCall` parts. The non-Gemini client paths (Claude, OpenAI-Responses) already preserved tool calls — this closes the gap specific to the native Gemini surface. Regression guard: `tests/unit/v1beta-gemini-tool-calling-6222.test.ts` (6, incl. a streaming SSE round-trip).
+- **feat(providers):** copilot-m365-web **enterprise / work tier** support ([#6334](https://github.com/diegosouzapw/OmniRoute/issues/6334)) — mirrors the EDU-tier pattern (#6210): `M365ConnectionParams` gains an `agent` field, a new opt-in `M365_ENTERPRISE_OVERRIDES` preset (`agent=work`, `scenario=officeweb`, `licenseType=Premium`) applies via `providerSpecificData.tier="enterprise"` (alias `"work"`), and `agent` is also overridable directly via `providerSpecificData.agent`. `buildWsUrl` was hardcoding `agent="web"` (the one enterprise-distinguishing param with no override path), so a Premium work account handshook then returned an empty stream. The individual and EDU paths are untouched. Kilo's dup flag vs #6210 (EDU tier) was a false positive — different tier. Regression guard: `tests/unit/copilot-m365-enterprise-6334.test.ts` (7). End-to-end confirmation on a real Premium work account is a live-VPS validation follow-up (Hard Rule #18). (thanks @Forcerecon)
+- **feat(api):** standardized, provider-agnostic **`effort` + `thinking` request params** ([#6241](https://github.com/diegosouzapw/OmniRoute/issues/6241)) — a thin standardization layer over the existing mature per-provider reasoning plumbing (no provider mapper touched). `providerChatCompletionSchema` gains a canonical `effort` (reusing the shared `none/low/medium/high/xhigh` vocabulary — the UI tiers `extra`/`max` collapse onto `xhigh`) and a boolean `thinking`. A pure `normalizeReasoningRequest` (wired once in `src/sse/handlers/chat.ts`, before any reasoning field is read) folds them onto the fields the translators already consume (`reasoning_effort` / `reasoning.effort` / `thinking`), so they fan out to Anthropic / Gemini / xAI / Responses — an explicit client `reasoning_effort` / object-shaped `thinking` always wins (backward-compatible). `/models` additively exposes `supportsThinking` + `effort_tiers` so the frontend can render the toggles (UI component is a follow-up). Regression guard: `tests/unit/effort-thinking-standardization-6241.test.ts` (12). (thanks @Iammilansoni, @shabeer)
+- **feat(combo):** new **`pipeline` (sequential) combo strategy** ([#6297](https://github.com/diegosouzapw/OmniRoute/issues/6297)) — the 18th routing strategy runs targets **in order**, threading each step's output into the next step's input, with an optional per-step `prompt` (system instruction); only the final step's response is returned. Distinct from `fusion` (parallel fan-out + judge). Implemented as a self-contained `open-sse/services/pipeline.ts` (sibling to `fusion.ts`), dispatched from `combo.ts`; the step list reuses `combo.models` order and reads an optional `prompt` off each target (backward-compatible — ignored by every other strategy). Intermediate steps run non-streaming with tools stripped (complete prose to thread forward); the final step keeps the client's `stream` flag + tools. A failing/empty/unparseable intermediate step fails the whole pipeline explicitly via a sanitized error (never silently swallowed). Kilo's dup flag vs #563 was a false positive (that's model→chain selection; this is a sequential chain). Regression guard: `tests/unit/combo-pipeline-strategy.test.ts` (5). (thanks @ofekbetzalel)
+- **feat(ci):** `check:test-masking` now flags **inline-reimplemented prod conditions** ([#6348](https://github.com/diegosouzapw/OmniRoute/issues/6348)) — a new **report-only** subcheck (v2, 6A.10 family) catches the wrong-shape contract test: a test that recomputes the condition under test inline instead of importing/exercising the real function (the #6216 class, where `=== 500` → `>= 500` stayed green because the test re-implemented the branch). For each added/modified test file it warns when the file textually duplicates a ≥3-token conditional from a production file touched in the same PR **and** does not import the symbol/module owning it, via a pure, fixture-tested `findReimplementedConditions()` with an allowlist mirroring `assertReductionAllowlist`. Report-only for now (does not fail the gate) — to be promoted to blocking after a triage cycle. Regression guard: `tests/unit/check-test-masking.test.ts` (45).
+- **feat(sse):** per-connection routing override (native vs CLIProxyAPI) ([#6339](https://github.com/diegosouzapw/OmniRoute/issues/6339)) — the previously-dead `isCliproxyapiDeepModeEnabled` helper is now wired into `resolveExecutorWithProxy`: a single connection can opt itself into the CLIProxyAPI passthrough executor via `providerSpecificData.cliproxyapiMode="claude-native"`, with precedence **connection override > provider `upstream_proxy_config` mode > default**. `resolveExecutorWithProxy` now receives the resolved connection's `providerSpecificData` (threaded from `chatCore.ts`), so one connection can deep-route while the provider's other connections stay native — no DB schema change (the toggle rides in `providerSpecificData`). Also resolves the same-provider mixing ask in #6340. Regression guard: `tests/unit/chatcore-executor-proxy.test.ts` (9). (thanks @RaviTharuma)
+- **feat(dashboard):** "Add session cookie" modal now shows a prominent **"Open ‹host› →"** link to the provider's own site ([#6268](https://github.com/diegosouzapw/OmniRoute/issues/6268)) — every `-web` cookie-session provider (chatgpt-web, claude-web, gemini-web, kimi-web, lmarena, qwen-web, m365-copilot-web, …) renders a one-click external link (opening the provider's login/home page in a new tab) so operators no longer tab away to retype the URL mid-setup. The host resolves from a pure, unit-tested `resolveWebProviderHost()` (prefers `WEB_COOKIE_PROVIDERS[id].website`, falls back to the registry `baseUrl` origin); non-web providers render exactly as before. Kilo's dup flag vs #6265 (modal-too-small-on-1080p) was a false positive — distinct concern. Regression guard: `tests/unit/resolve-web-provider-host.test.ts` (5). (thanks @chirag127)
+- **feat(providers):** add **DigitalOcean** AI (serverless inference) as an OpenAI-compatible API-key provider ([#6373](https://github.com/diegosouzapw/OmniRoute/pull/6373)) — base `https://inference.do-ai.run/v1`, wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/digitalocean/`, `src/shared/constants/providers/apikey/inference-hosts.ts`). Regression guard: `tests/unit/digitalocean-provider.test.ts`. (thanks @newnol)
+- **feat(providers):** add **Huancheng Public API** (`hcnsec`) as an OpenAI-compatible regional provider ([#6410](https://github.com/diegosouzapw/OmniRoute/pull/6410)) — Xinjiang Huancheng Cybersecurity's public LLM platform (base `https://api.hcnsec.cn/v1`, free credits via daily check-ins), wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/hcnsec/`, `src/shared/constants/providers/apikey/regional.ts`). Regression guard: `tests/unit/hcnsec-provider.test.ts`. (thanks @UnrealAryan)
+- **feat(dashboard):** the web-session credential guide now shows an **"Open {host}" link** ([#6316](https://github.com/diegosouzapw/OmniRoute/pull/6316)) to the provider's sign-in site (derived from the provider `website` via `getProviderWebsiteHost`), so you can jump straight to the page where the cookie/session must be captured. Regression guard: `tests/unit/web-session-provider-link-6316.test.ts`. (thanks @jordansilly77-stack)
+- **feat(cerebras):** add the **Gemma 4 31B** model (`gemma-4-31b`) to the Cerebras registry + pricing table ([#6331](https://github.com/diegosouzapw/OmniRoute/pull/6331)). Regression guard extends `tests/unit/t28-model-catalog-updates.test.ts`. (thanks @backryun)
+- **feat(providers):** add **Yuanbao (web)** as a cookie-session provider ([#6196](https://github.com/diegosouzapw/OmniRoute/issues/6196)) — `yuanbao-web` (Tencent Yuanbao, `yuanbao.tencent.com`) with cookie-only auth (`hy_user`/`hy_token` + public agent id), SSE→OpenAI translation incl. `reasoning_content`, exposing DeepSeek V3/R1 + Hunyuan / Hunyuan-T1. Regression guard: `tests/unit/providers-yuanbao-web.test.ts`. `together-web` was **deferred** (no verifiable web-session endpoint — needs a captured request) and `huggingchat-web` **dropped** (the existing `huggingchat` already is a web-cookie provider). (thanks @chirag127)
+- **feat(providers):** route the built-in **agentrouter** through the dynamic Claude-Code wire image ([#6056](https://github.com/diegosouzapw/OmniRoute/issues/6056)) — a small static allow-set (`CC_WIRE_IMAGE_BUILTINS` in `open-sse/services/ccWireImageBuiltins.ts`), consulted by `isClaudeCodeCompatible` / `isClaudeCodeCompatibleProvider` / `applyFingerprint`, makes agentrouter adopt the CC wire-image headers + fingerprint **while guarding the CC baseUrl/auth branches** so it keeps its own registry `baseUrl` and `x-api-key` auth. Regression guard: `tests/unit/agentrouter-cc-wire-image.test.ts` (asserts the wire image is applied AND agentrouter's baseUrl/auth are preserved). Live WAF-acceptance against agentrouter.org is a VPS validation follow-up (Hard Rule #18).
+- **feat(providers):** **bulk-add API keys for Cloudflare Workers AI** ([#6174](https://github.com/diegosouzapw/OmniRoute/issues/6174)) — `cloudflare-ai` is removed from the bulk-add exclusion list and the bulk parser gains a 3-field `name|accountId|apiKey` mode; the bulk route now builds a **per-entry** `providerSpecificData` so each key carries its own `accountId` (fixing the previous shared-object reuse), and both the create + key-validation paths receive it. Regression guard: `tests/unit/bulk-api-key-parser-cloudflare.test.ts`. (thanks @muflifadla38)
+- **feat(dashboard):** routing/settings UX clarity ([#6147](https://github.com/diegosouzapw/OmniRoute/issues/6147)) — (1) weighted combos show the **effective routing share %** next to each weight when weights don't sum to 100 (`WeightTotalBar.tsx`); (2) the status widget's user-facing **"Cloud Sync" label is renamed** to "Remote Settings Sync" (`CloudSyncStatus.tsx`; internal ids/state untouched); (3) built-in providers gain an **opt-in advanced base-URL override** (`isBaseUrlOverrideEligibleProvider`, hidden behind an "Advanced" toggle, reusing the existing `providerSpecificData.baseUrl` persistence — not globally widened). Regression guard: `tests/unit/routing-settings-ux-6147.test.ts`.
+- **feat(combo):** add an option to **disable session stickiness**, per-combo or globally — round-robin / random combos can rotate to a different connection on every request instead of pinning a whole conversation to one connection by its first-message hash. Resolution precedence per-combo `config.disableSessionStickiness` → global `settings.disableSessionStickiness` → default `false` (preserves the #3825 prompt-cache/504 fix); gates **both** stickiness call sites in `open-sse/services/combo.ts`. Exposed as a global toggle (Combo Defaults) and a per-combo Inherit/on/off control. ([#6168](https://github.com/diegosouzapw/OmniRoute/issues/6168)) Regression guard: `tests/unit/combo-disable-session-stickiness.test.ts`. (thanks @RCrushMe)
+- **feat(docker):** add the `OMNIROUTE_NO_SUDO` env flag for root-less / user-namespaced deployments — the MITM cert-trust command path (`resolveSudoSpawn` in `src/mitm/systemCommands.ts`) now strips the leading `sudo` when the flag is truthy, in addition to the existing root / sudo-missing cases, so the Proxy Agent runs without `sudo` (the operator trusts the CA manually, e.g. via `NODE_EXTRA_CA_CERTS`). Argv-array `spawn` preserved — no shell interpolation (Hard Rule #13). ([#6122](https://github.com/diegosouzapw/OmniRoute/issues/6122)) Regression guard: `tests/unit/mitm-systemCommands-no-sudo.test.ts`. (thanks @powellnorma)
+- **feat(providers):** add **Requesty** as an OpenAI-compatible gateway provider (BYOK, base `https://router.requesty.ai/v1`, ~200 free requests/day) — wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/requesty/`, `src/shared/constants/providers/apikey/gateways.ts`). ([#6120](https://github.com/diegosouzapw/OmniRoute/issues/6120)) Regression guard: `tests/unit/requesty-provider.test.ts`. (thanks @chirag127)
+- **feat(dashboard):** add **configured-only / available-only filters** to the Free Provider Rankings page ([#6150](https://github.com/diegosouzapw/OmniRoute/issues/6150)) — hide providers you haven't configured, or whose connections are all rate-limited / out of quota, via server-side query params (`?configuredOnly` / `?availableOnly` on `GET /api/free-provider-rankings`) backed by a testable lib helper reusing the in-process connection state (no Redis). Both filters default off, so the default view is unchanged; this supersedes the earlier client-side "Configured Only" toggle (#6245) with an available-only dimension and unit-tested logic. Regression guard: `tests/unit/freeProviderRankings-filters.test.ts`.
+
+### 🔧 Bug Fixes
+
+- **fix(dashboard):** adding a second API key connection for the same provider no longer silently overwrites the first — the Add-API-key modal now derives a unique default connection name (`main`, then `main-2`, `main-3`, …) so the backend name-based upsert can't collide (#6499 — thanks @dilneiss).
+- **fix(compression):** the session-dedup engine now also deduplicates a large multi-line block repeated **within a single message** (intra-message dedup), not just across turns; the compression-preview API surfaces a `fallbackReason`, and the fusion panel reports how many models were rate-limited vs failed on total-panel failure (#6501 — thanks @chirag127).
+- **fix(compression):** a stacked-pipeline step naming an unregistered engine now surfaces a `validationErrors` entry instead of silently no-op'ing, so misconfigured pipelines are visible in the preview API (#6506 — thanks @chirag127).
+- **feat(usage):** add a **Codex reset-credit redemption** flow to the Provider Limits UI ([#6361](https://github.com/diegosouzapw/OmniRoute/pull/6361)) — a `useCodexResetCreditRedemption` hook + `/api/usage/codex-reset-credit` route + `codexResetCredits` lib let you redeem banked Codex reset credits from the quota card. Regression guard: `tests/unit/codex-reset-credits.test.ts`. (thanks @JxnLexn)
+- **feat(glm):** add **team-plan quota settings** for `glm-cn` connections ([#6351](https://github.com/diegosouzapw/OmniRoute/pull/6351)) — a dedicated `GlmTeamQuotaFields` form section (team quota id / limits) threaded through the Add/Edit connection modals, persisted via `providerSpecificData`, with the GLM usage service reading the team quota. Regression guards: `tests/unit/glm-team-quota.test.ts`, `provider-specific-data-schema.test.ts`. (thanks @hao3039032)
+- **feat(providers):** add **TinyFish** web-fetch/search support ([#6349](https://github.com/diegosouzapw/OmniRoute/pull/6349)) — a `tinyfish-fetch` executor + `/v1/web/fetch` route + MCP web-fetch tool, registered as a specialty-media provider with request-validation and a search-provider catalog entry. Regression guards: `tests/unit/executor-tinyfish-fetch.test.ts`, `web-fetch-handler.test.ts`, `mcp-web-fetch-tool.test.ts`, `provider-validation-tinyfish.test.ts`. (thanks @dtybnrj)
+- **fix(cli):** `omniroute launch-codex` now spawns `codex.cmd` through a shell on Windows (the npm `.cmd` shim is unresolvable by bare `spawn` → ENOENT), mirroring the qodercli Windows fix (#6263) ([#6312](https://github.com/diegosouzapw/OmniRoute/pull/6312)). Regression guard: `tests/unit/launch-codex-windows-spawn-6312.test.ts`. (thanks @swingtempo)
+- **fix(codex):** isolate the **Spark** quota from the shared Codex quota and stabilize the quota UI ordering / hydration so per-scope limits render consistently ([#6336](https://github.com/diegosouzapw/OmniRoute/pull/6336)). Regression guards: `tests/unit/codex-quota-selection-hydration.test.ts`, `provider-limits-ui.test.ts` + 3 more. (thanks @xz-dev)
+- **feat(api):** add a `hidePaidModels` setting that filters paid-only models out of the `/v1/models` catalog. Regression guard: `tests/unit/models-catalog-hide-paid.test.ts`. (thanks @chirag127)
+- **fix(api-manager):** the fallback model picker now preserves combos instead of dropping them when a primary model is unavailable ([#6443](https://github.com/diegosouzapw/OmniRoute/pull/6443)). Regression guard: `tests/unit/api-manager-page-static.test.ts`. (thanks @jmengit)
+- **fix(providers):** recoverable Antigravity / Cloud-Code (Gemini Code Assist) `403` responses ([#6452](https://github.com/diegosouzapw/OmniRoute/pull/6452)) are now classified as a retryable project-config error instead of a terminal account ban, so a fixable project/API-disabled 403 no longer forces a ~1-year cooldown / full OAuth reconnect. Regression guard: `tests/unit/errorclassifier-antigravity-403.test.ts`. (thanks @developerjillur)
+- **fix(mitm):** `sanitizeHeaders` now redacts `Set-Cookie` response headers so upstream session cookies never leak into logs / diagnostics ([#6451](https://github.com/diegosouzapw/OmniRoute/pull/6451)). Regression guard: `tests/unit/mitm-sanitize-headers.test.ts`. (thanks @developerjillur)
+- **fix(api):** `/api/compression/preview` now accepts `mode: "caveman"` and correctly handles stacked / zero-compression previews ([#6425](https://github.com/diegosouzapw/OmniRoute/issues/6425)). Regression guard: `tests/unit/api/compression-preview-caveman-and-stacked-6425.test.ts`. (thanks @chirag127)
+- **feat(providers):** add **Zed** hosted LLM aggregator as a native-app provider ([#6118](https://github.com/diegosouzapw/OmniRoute/pull/6118)) — OAuth sign-in via the Zed hosted flow, registered through the shared provider registry + executor. Regression guards: `tests/unit/zed-oauth-provider.test.ts`, `zed-import-utils.test.ts`, `zed-docker-detect.test.ts`, `mitm-handler-zed.test.ts`. VPS-validated via live operator login (Hard Rule #18).
+- **fix(oauth):** the Kiro SSO-cache auto-import now **preserves the IDC region** — cross-region Amazon Q / Kiro profiles imported from the SSO cache are no longer collapsed to the default region ([#6113](https://github.com/diegosouzapw/OmniRoute/pull/6113)). Regression guard: `tests/unit/kiro-auto-import-idc-2059.test.ts`. VPS-validated via live operator login (Hard Rule #18).
+- **fix(dashboard):** passthrough model aliases no longer collide when two namespaced model ids share a last segment (port from 9router#1850, [#6431](https://github.com/diegosouzapw/OmniRoute/pull/6431)). `enx/gpt-5.5` and `enx/codebuddy/gpt-5.5` both auto-generated the alias `gpt-5.5`, so the second model could never be added (the UI just alerted "alias already exists"). Aliases are now disambiguated deterministically — bare last segment when free, then parent-qualified (`codebuddy-gpt-5.5`), then a numeric suffix — while re-adding the exact same model id is still blocked. Regression guard: `tests/unit/passthrough-alias-1850.test.ts`. (thanks @arpicato)
+- **fix(translator):** preserve a Gemini `functionResponse` co-located with other parts (another `functionCall`, or trailing `text`) in the same content when translating **Gemini → OpenAI** ([#6376](https://github.com/diegosouzapw/OmniRoute/pull/6376)). `convertGeminiContent()` early-returned the tool message on the first `functionResponse` part, dropping any co-located parts; such contents are now pre-split (one tool message per `functionResponse`, emitted first, plus one message for the remaining parts). Regression guard: `tests/unit/gemini-to-openai-function-response.test.ts`. (thanks @warelik)
+- **fix(headroom):** detect a python interpreter managed by **mise / pyenv / asdf / conda** (port from 9router#2353, [#6382](https://github.com/diegosouzapw/OmniRoute/pull/6382)). Headroom's python probe (`src/lib/headroom/detect.ts`) searched a hardcoded `PATH`, but version managers expose their interpreters via shim dirs that only join `PATH` through interactive-shell activation — which the non-interactive server never runs, so a managed python (≥3.10) was invisible and Headroom reported it missing. The search path now prepends the well-known shim/bin dirs (`~/.local/share/mise/shims`, `~/.pyenv/shims`, `~/.asdf/shims`, `$CONDA_PREFIX/bin`, `~/.local/bin`, respecting `MISE_DATA_DIR`/`PYENV_ROOT`/`ASDF_DATA_DIR` when set), and a new `HEADROOM_PYTHON` env override lets operators point straight at their interpreter (mirroring `HEADROOM_URL`). Still shell-free (`execFileSync`). Regression guard: `tests/unit/headroom-detect.test.ts` (5). (thanks @loopyd)
+- **fix(executors):** strip the OpenAI-Codex/Claude-CLI `client_metadata` passthrough field for **NVIDIA** requests (port from 9router#1887, [#6411](https://github.com/diegosouzapw/OmniRoute/pull/6411)). NVIDIA's OpenAI-compatible wrapper rejects it with `400 Unsupported parameter`, the same class already handled for `cerebras`/`mistral`; `nvidia` (executor `default`) was missing from the strip allowlist so Codex/Claude-Code passthrough requests 400'd. Regression guard: `tests/unit/executor-default-strip-client-metadata.test.ts` (+nvidia case). (thanks @phidinhmanh)
+- **fix(translator):** strip the Claude-style `thinking` field for **NVIDIA `z-ai/glm-5.2`** (port from 9router#2023, [#6413](https://github.com/diegosouzapw/OmniRoute/pull/6413)). NVIDIA's OpenAI-compatible wrapper 400s on `thinking` (a Claude-format client routed here leaves a `thinking:{type:"adaptive"}`); the existing strip rule only dropped `reasoning`. Same class already handled for `minimax-m2.7`. Regression guard: `tests/unit/nvidia-minimax-thinking-strip.test.ts` (+glm-5.2 case). (thanks @phidinhmanh)
+- **fix(translator):** suppress the streamed `</think>` close marker for the **Antigravity IDE** client (port from 9router#1061, [#6415](https://github.com/diegosouzapw/OmniRoute/pull/6415)). On thinking-only turns Antigravity rendered a bare `</think>` as the sole visible content, tripping its loop-detection and wasting requests. Antigravity's UA (`vscode/<v> (Antigravity/<v>)`) is added to the marker-suppress allowlist (alongside OpenCode); Claude Code / Cursor still get the marker, and `x-omniroute-thinking-marker: on` force-restores it. Regression guard: `tests/unit/think-close-marker-suppress-5245.test.ts`. (thanks @abdofallah)
+- **fix(executors):** strip nested `reasoning_content` from messages for **Mistral** (port from 9router#1649, [#6417](https://github.com/diegosouzapw/OmniRoute/pull/6417)). Mistral's API returns `422 extra_forbidden` when an assistant message carries `reasoning_content` (replayed thinking from a prior turn, e.g. via the Codex `/responses` path); the generic top-level 400 field-downgrade retry never covered the nested per-message field. `DefaultExecutor` now strips it for provider `mistral` only, so DeepSeek (which requires replayed `reasoning_content`) is unaffected. Regression guard: `tests/unit/mistral-strip-reasoning-content-1649.test.ts`. (thanks @xxy9468615)
+- **fix(executors):** strip the `client_metadata` passthrough field on the **OpenCode** path (port from 9router#1442, [#6418](https://github.com/diegosouzapw/OmniRoute/pull/6418)). OpenCode upstreams (e.g. `kimi-k2.6` via opencode-go) reject it with `400 "Extra inputs are not permitted, field: 'client_metadata'"`; the DefaultExecutor strip only covered cerebras/mistral and `OpencodeExecutor` extends `BaseExecutor` directly, so nothing removed it there. Regression guard: `tests/unit/opencode-strip-client-metadata-1442.test.ts`. (thanks @yanpaing007)
+- **fix(executors):** inject the `reasoning_content` echo for the native **Moonshot Kimi** provider (port from 9router#1480, [#6419](https://github.com/diegosouzapw/OmniRoute/pull/6419)). Kimi (executor `default`) is a thinking-mode upstream that 400s with "reasoning_content must be passed back" when a prior assistant turn lacks it; the placeholder injection was only wired into the OpenCode meta-provider, so direct multi-turn Kimi conversations failed. Scoped to `kimi` (gateway-served models matching the thinking-model name pattern are unaffected). Regression guard: `tests/unit/kimi-native-reasoning-injected-1480.test.ts`. (thanks @2220258345)
+- **fix(executors):** recover from a strict gateway's `context_management: Extra inputs are not permitted` 400 (port from 9router#1468, [#6420](https://github.com/diegosouzapw/OmniRoute/pull/6420)). **Claude Code** always sends a top-level `context_management` field; strict anthropic-compatible gateways reject it. The dedicated context-editing 400-fallback only fired when OmniRoute's own `contextEditing` feature was enabled (default off), so a client-sent field passed through untouched and 400'd. `context_management` is now in the generic reactive field-strip list, so it's stripped-and-retried once regardless of the feature flag (with correct request re-signing for claude-compatible relays). Regression guard: `tests/unit/provider-field-strips.test.ts`. (thanks @ohahe52-dot)
+- **fix(network):** enable **RFC 8305 Happy Eyeballs** (`autoSelectFamily`) on the direct-egress undici dispatcher (port from 9router#1237, [#6423](https://github.com/diegosouzapw/OmniRoute/pull/6423)). When DNS returns both IPv6 (AAAA) and IPv4 (A) and the IPv6 route is broken (e.g. a NAT64 `64:ff9b::` prefix without routing), undici tried IPv6 first and hung until `ETIMEDOUT` (then a 502 + account lockout), even though `curl` reached the same host. The direct dispatcher now races both families and uses whichever connects first. Proxy paths pin family via `proxyTls` and are unaffected. Regression guard: `tests/unit/direct-dispatcher-pipelining-4580.test.ts`. (thanks @adentdk)
+- **fix(combo):** round-robin now advances the rotation pointer past the model that **actually served**, not the eagerly-scheduled one (port from 9router#948, [#6428](https://github.com/diegosouzapw/OmniRoute/pull/6428)). With `stickyLimit: 1` (true round-robin), when the scheduled model failed and a _different_ model served via fallback, the counter had already advanced +1 from the scheduled index — so the next request reused the fallback-served model, degrading round-robin into hot-spotting on whichever model was healthy. The pointer now advances to the served index + 1 (mirroring the sticky-limit>1 path). Session-stickiness (#3825) and distribution are preserved. Regression guard: `tests/unit/combo-rr-fallback-advance-948.test.ts`. (thanks @binsarjr)
+- **fix(sse):** a non-string `model` field is now rejected with a `400` before the resolver, instead of crashing downstream `.toLowerCase()`/`.split()` calls into an empty-body `500` that escapes the error sanitizer ([#6407](https://github.com/diegosouzapw/OmniRoute/issues/6407)). Regression guard: `tests/unit/chat-non-string-model-6407.test.ts`. (thanks @chirag127)
+- **fix(api):** unknown `/api/*` routes now return a JSON `404` (instead of the dashboard HTML shell) and scalar chat params (`model`/`temperature`/etc.) are validated **before** the provider lookup so malformed requests fail fast with a clear `400` ([#6424](https://github.com/diegosouzapw/OmniRoute/issues/6424), [#6412](https://github.com/diegosouzapw/OmniRoute/issues/6412)). Regression guards: `tests/unit/api/api-catchall-json-404.test.ts`, `tests/unit/chat-early-schema-validation-6412.test.ts`. (thanks @chirag127)
+- **fix(api):** `/v1/chat/completions` now rejects a non-JSON `Content-Type` with a `400` before parsing the body ([#6414](https://github.com/diegosouzapw/OmniRoute/issues/6414)). Regression guard: `tests/unit/v1-chat-completions-content-type-6414.test.ts`. (thanks @chirag127)
+- **fix(api):** the `X-OmniRoute-Compression` response header is now echoed on `/v1/chat/completions` and `/v1/completions` ([#6422](https://github.com/diegosouzapw/OmniRoute/issues/6422)). Regression guard: `tests/unit/compression-header-echo-6422.test.ts`. (thanks @chirag127)
+- **fix(api):** concurrent `GET /v1/models` requests are coalesced into a single catalog build ([#6408](https://github.com/diegosouzapw/OmniRoute/issues/6408)). Regression guard: `tests/unit/v1-models-concurrent-6408.test.ts`. (thanks @chirag127)
+- **fix(api):** `/v1/completions` now echoes the requested `body.model` in its JSON + streamed responses ([#6429](https://github.com/diegosouzapw/OmniRoute/pull/6429)). Regression guard: `tests/unit/completions-body-model-echo.test.ts`. (thanks @chirag127)
+- **fix(api):** env-var master keys now see the full `/v1/models` catalog ([#6406](https://github.com/diegosouzapw/OmniRoute/issues/6406)). Regression guard: `tests/unit/models-catalog-envkey-6406.test.ts`. (thanks @chirag127)
+- **fix(api):** non-streaming `/v1/completions` responses now echo `body.model` aligned with the `X-OmniRoute-Model` header ([#6426](https://github.com/diegosouzapw/OmniRoute/issues/6426)). Regression guard: `tests/unit/v1-completions-model-header-match-6426.test.ts`. (thanks @chirag127)
+- **fix(api):** unknown `/v1/*` routes now return a JSON `404 not_found` instead of the Next.js dashboard HTML shell ([#6405](https://github.com/diegosouzapw/OmniRoute/issues/6405)). Regression guard: `tests/unit/api/v1-catchall-json-404.test.ts`. (thanks @chirag127)
+- **fix(api):** the per-connection provider models route now degrades to the shipped catalog when a provider's `/models` endpoint answers with a redirect ([#6267](https://github.com/diegosouzapw/OmniRoute/issues/6267)) — a `qwen-web` import failed with a raw `Redirect blocked … (307)` 503. `safeOutboundFetch` throws `REDIRECT_BLOCKED` on the 307, `getSafeOutboundFetchErrorStatus` maps it to 503, and `buildDiscoveryErrorFallbackResponse` treated every 503 as a hard error — so the non-empty `getModelsByProviderId("qwen-web")` catalog was never surfaced. A models-endpoint redirect is not a fixable-config error (unlike `URL_GUARD_BLOCKED`/`INVALID_URL`, which stay hard errors), so it now falls back to the local/cached catalog before the 503 short-circuit. General fix — covers any config-driven provider that 307s. Regression guard: `tests/unit/provider-models-qwen-web-redirect-6267.test.ts`. (thanks @chirag127)
+- **fix(api):** the per-connection provider models route (MCP `list_models_catalog` + the dashboard import view) now merges USER-ADDED custom models into its response ([#6247](https://github.com/diegosouzapw/OmniRoute/issues/6247)) — custom models live in the `key_value` namespace `customModels`, which the live REST `/api/v1/models` already merges, but `src/app/api/providers/[id]/models/route.ts` never read `getCustomModels`, so custom models were dropped on both the discovery-success and local_catalog paths. They are now folded into the returned model list (deduped by id, stamped `owned_by: provider`), fixing MCP + the dashboard import view in one place. Regression guard: `tests/unit/provider-models-custom-merge-6247.test.ts`. (thanks @RCrushMe)
+
+- **fix(providers):** GitLab Duo tool-calling follow-up turns no longer fail upstream with `422 {"detail":"Validation error"}` (tokens 0/0, rejected pre-inference). The #6234 tool-result-feedback fix serialized the **entire** multi-turn conversation into GitLab's single-file `code_suggestions` (`small_file`) generation endpoint — folded history that turn-N sent as an oversized `current_file.content_above_cursor` **and** duplicated verbatim into `user_instruction`, tripping the AI-Gateway's `small_file` validation guard. The executor now **bounds** that prompt: it keeps system + latest user message + the most-recent tool round (dropping older turns), caps oversized tool results, and stops duplicating the full prompt into `user_instruction` (which now carries only the short latest user message) — while still feeding the most-recent tool result back so the agent continues (`open-sse/executors/gitlab.ts`, [#6220](https://github.com/diegosouzapw/OmniRoute/issues/6220)). The unit test covers the bounding logic; the upstream 422→200 clearing is VPS-only (Hard Rule #18). Regression guard: `tests/unit/gitlab-tool-exchange-bounded-6220.test.ts`.
+
+- **fix(i18n):** the provider-detail (`/dashboard/providers/[id]`) connection-status filter labels no longer render as `__MISSING__:All` / `__MISSING__:Active` / `__MISSING__:Error` / `__MISSING__:Banned` / `__MISSING__:CreditsExhausted` in non-English locales (notably pt-BR) ([#6290](https://github.com/diegosouzapw/OmniRoute/issues/6290)). Root cause was **not** the namespace mismatch the issue guessed — the `providers.filter*` keys resolve correctly in `en.json`; the debt lived in the locale mirrors (`src/i18n/messages/*.json`), where these five keys carried the `__MISSING__:` sync sentinel in ~15 locales and were absent entirely in ~26 others, so next-intl found the key and echoed the sentinel verbatim. All 40 non-English/-Chinese mirrors now ship real translations for the five `providers.filter*` labels. Regression guard: `tests/unit/i18n-provider-filter-keys-6290.test.ts`. (thanks @diegosouzapw)
+
+- **fix(providers):** the `copilot-m365-web` streaming executor now emits `debug`-level WebSocket diagnostics ([#6210](https://github.com/diegosouzapw/OmniRoute/issues/6210)) — the outbound WS URL (with the `access_token` **redacted** via `redactWsUrl()`), handshake success/failure, and each received SignalR frame's `type`/`target`. Previously the streaming path logged nothing, so an empty `content:null` response (the M365 Education / Starter tier symptom fixed in #6234) was undiagnosable even at `APP_LOG_LEVEL=debug`. The change is debug-level and side-effect-free — it does not alter streaming behavior or the frame parser, and the token never reaches the logs. Regression guard: `tests/unit/copilot-m365-web-logging-6210.test.ts` (thanks @qpeyba)
+
+- **fix(resilience):** a round-robin combo no longer returns `503 all upstream accounts are unavailable` when a compatibility-rejected target is actually healthy ([#6238](https://github.com/diegosouzapw/OmniRoute/issues/6238)). `filterTargetsByRequestCompatibility` drops request-incompatible targets (tool/vision/structured-output unsupported, or below the required context window) **before** any availability check runs, and its `compatible.length === 0` safety net only fired when _all_ targets were filtered — not when the kept targets later all turned out runtime-unavailable (circuit-open / cooldown / no credentials). So a combo could 503 while a compat-rejected-but-healthy provider sat unused. `handleRoundRobinCombo` now keeps the compat-rejected set and, when every compat-kept target was skipped without a single real attempt, probes those rejected targets as a **last-resort fallback tier** (via the new pure `open-sse/services/combo/comboCompatFallback.ts`) before crystallizing the 503. Regression guard: `tests/unit/combo-roundrobin-compat-fallback-6238.test.ts`. (thanks @ThongAccount)
+
+- **fix(startup):** best-effort self-heal for a corrupted Turbopack dev cache on Windows ([#6289](https://github.com/diegosouzapw/OmniRoute/issues/6289)). On Windows, `pnpm dev` can fail at startup when Turbopack `mmap`s a persistent-cache SST file and the OS refuses the mapping (`os error 1455` — "paging file too small"), which Turbopack surfaces as a misleading `Module not found: Can't resolve '@/shared/utils/machine'`. This is a **known upstream Turbopack cache-corruption bug — not our code**. The dev launcher (`scripts/dev/run-next.mjs`) now wraps `nextApp.prepare()` and, when it rejects with that signature (`isTurbopackCacheCorruption` in the new `scripts/dev/turbopackCacheHeal.mjs`), purges `.build/next/**/cache/turbopack` and retries **once** with a clear log. **Caveat — best-effort only:** the corruption often surfaces as a runtime overlay rather than a `prepare()` rejection, so this cannot always intercept it; the reliable remedy remains manually deleting the Turbopack cache dir. Regression guard: `tests/unit/turbopack-cache-heal-6289.test.ts`. (thanks @chirag127)
+
+- **fix(providers):** qodercli PAT auth no longer fails with `spawn qodercli ENOENT` on Windows ([#6263](https://github.com/diegosouzapw/OmniRoute/issues/6263)) — `spawnQoderCli` spawned the bare `qodercli` name with `shell:false` and an unenriched env, so the npm `.cmd` wrapper under `%APPDATA%\npm` (a user-PATH directory) was never resolved. It now resolves the absolute `.cmd`/`.exe` path through the existing `getCliRuntimeStatus("qoder")` resolver in `src/shared/services/cliRuntime.ts` (memoized), spawns with `shell` when the target is a `.cmd`/`.bat`, and uses the cliRuntime-enriched env (PATH + PATHEXT + APPDATA); the ENOENT error now lists the searched paths plus the `CLI_QODER_BIN` override. End-to-end spawn on a real Windows host is host-only (Hard Rule #18); the path-resolution logic is unit-tested. Regression guard: `tests/unit/qodercli-windows-resolve-6263.test.ts`. (thanks @chirag127)
+
+- **fix(sse):** the reasoning-token buffer no longer inflates **probe-sized `max_tokens`** ([#6274](https://github.com/diegosouzapw/OmniRoute/issues/6274)) — Claude Code's `/model` capability check sends `max_tokens: 1`, but for a thinking-capable model with a large output cap (e.g. `glm-5.2`) the #3587 headroom heuristic (`max(current + 1000, ceil(current * 1.5))`) rewrote it to `1001` and forwarded that upstream, wasting tokens on a request that was never a genuine reasoning budget. `resolveReasoningBufferedMaxTokens()` (`open-sse/services/reasoningTokenBuffer.ts`) now short-circuits and returns the caller's value verbatim when it is below the new `REASONING_BUFFER_MIN_TRIGGER` (256) threshold — a tiny explicit limit is a probe, not a reasoning request. Real budgets still receive the #3587 headroom unchanged, and the guard runs after the existing capability checks so unknown / non-reasoning models keep returning `null`. Regression guard: `tests/unit/reasoning-token-buffer-6274.test.ts`. (thanks @brightfiscalband)
+
+- **fix(cli):** `omniroute reset-password` now works as a real subcommand, and password resets over piped (non-TTY) stdin actually apply ([#6261](https://github.com/diegosouzapw/OmniRoute/issues/6261), [#6258](https://github.com/diegosouzapw/OmniRoute/issues/6258)). Two coupled defects: (1) **#6261** — `bin/omniroute.mjs` routed everything through Commander with only two pre-Commander bypasses (`--mcp`, `reset-encrypted-columns`), so `omniroute reset-password` was rejected as an unknown command; only the separate `omniroute-reset-password` bin worked, while the docs falsely advertised the subcommand (incl. a bogus "legacy alias still works"). A pre-Commander bypass mirroring `reset-encrypted-columns` now dynamically imports `bin/reset-password.mjs` (which self-executes) before Commander parses; the three doc lines were corrected. (2) **#6258** — `bin/reset-password.mjs` issued two sequential `rl.question` prompts; under piped stdin the second read never settled at EOF, so `main()` never reached `resetManagementPassword` and the reset was a silent no-op (both prompts printed, no success, password unchanged). The CLI now detects non-TTY stdin and reads it once (first line = password, second line = confirm if present, else reused), adds a `--password-stdin` flag (entire stdin is the password, no confirmation), and exits `0` explicitly so the success line always flushes; interactive TTY behavior is unchanged. Regression guard: `tests/unit/reset-password-cli-6261-6258.test.ts` (3). (thanks @chirag127)
+
+- **fix(db):** the mass-migration **safety abort** now tells the operator how to bypass it and stops flooding the log ([#6260](https://github.com/diegosouzapw/OmniRoute/issues/6260)) — after restoring a backup that wiped the migration tracking table, `runMigrations()` threw the abort on every downstream `ensureDbInitialized()`, re-logging the full banner 11+ times, and the message never mentioned the existing `OMNIROUTE_MAX_PENDING_MIGRATIONS` escape hatch. The abort text now appends a bypass hint (set `OMNIROUTE_MAX_PENDING_MIGRATIONS=0` in `server.env` / `DATA_DIR/.env`), and a new `MigrationSafetyAbortError` is memoized so repeated calls in the same process throw the same instance and emit a single concise line instead of the full cascade. Regression guard: `tests/unit/migration-safety-abort-6260.test.ts`. (thanks @chirag127)
+
+- **fix(auth):** importing a **distinct** Codex/ChatGPT OAuth `auth.json` is no longer falsely rejected as "already exists" when it belongs to a different user in the same workspace ([#6301](https://github.com/diegosouzapw/OmniRoute/issues/6301)). `findExistingCodexConnection` (in `src/lib/oauth/utils/codexAuthImport.ts`) deduped **only** on `providerSpecificData.workspaceId === accountId`, where `accountId` is the shared `chatgpt_account_id`/`tokens.account_id` — so two members of the same ChatGPT Team collapsed onto a single connection (409 `duplicate_account`). The id_token's `https://api.openai.com/auth` claim carries a per-user `chatgpt_user_id` alongside the workspace id (the device-flow path already persisted it as `chatgptUserId`, but the import path did not). Now `parseAndValidateCodexAuth` extracts `userId` (`chatgpt_user_id` → `user_id` → JWT `sub`) into `ParsedCodexAuth`, the create/update paths persist `chatgptUserId` in `providerSpecificData` (mirroring `codex.ts`), and dedup keys on `workspaceId` **AND** `chatgptUserId` — with a backward-compat fallback to legacy accountId-only matching when no stored connection for that workspace records a `chatgptUserId`, so genuinely-same accounts still dedup. Regression guard: `tests/unit/codex-auth-import-userid-dedup-6301.test.ts` (4). (thanks @anungma)
+
+- **fix(providers):** importing models for the **venice-web** provider no longer fails with a red "Provider venice-web does not support models listing" ([#6269](https://github.com/diegosouzapw/OmniRoute/issues/6269)). `venice-web` is a web-cookie provider with an executor but no upstream `/v1/models` endpoint and no registry `models`, so the models route fell through to the tail `400`. Mirroring the `jules`/`linkup-search`/`ollama-search` fix (#5569), it now ships a static local catalog entry in `src/lib/providers/staticModels.ts` — seeding the current Venice lineup (`venice-uncensored`, `llama-3.3-70b`, `qwen3-235b`, `qwen3-4b`, `deepseek-r1-671b`; Venice rotates its catalog, see docs.venice.ai/models/overview) — so the route returns `200` with `source:"local_catalog"`, `intentional:true`. Regression guard: `tests/unit/static-models-venice-web-6269.test.ts`. (thanks @chirag127)
+
+- **fix(api):** the specialty model catalogs (`/v1/embeddings`, `/v1/images`, `/v1/music`, `/v1/videos` model lists) are now derived from the **unified catalog filtered by a predicate** (`getSpecialtyModelsResponse`) instead of ad-hoc per-route logic, so they consistently respect active-credential visibility and stay in sync with the main catalog ([#6303](https://github.com/diegosouzapw/OmniRoute/pull/6303)). Regression guard: `tests/unit/specialty-model-catalog-routes.test.ts`. (thanks @makcimbx)
+- **fix(api):** the agent-bridge server route now resolves the MITM manager via a **dynamic `import("@/mitm/manager.runtime")`** so Turbopack does not statically pull the stub (or over-bundle the manager), and the agent-skills generator anchors its output base path with `path.join(process.cwd(), …)` so Turbopack's static analyzer stops tracing the whole project root ([#6329](https://github.com/diegosouzapw/OmniRoute/issues/6329), [#6366](https://github.com/diegosouzapw/OmniRoute/pull/6366)). Regression guard: `tests/unit/agent-bridge-server-route-dynamic-import.test.ts`. (thanks @Iammilansoni)
+- **fix(api):** internal probes (combo-test, cloud-sync verify) now pick a **management-scoped / allow-all API key** instead of naively grabbing `getApiKeys()[0]` — a restricted `self:usage` first row made the probe fail with "Model X is not allowed for this API key" even when the combo path was healthy (`pickApiKeyForInternalUse` in `src/lib/db/apiKeys.ts`). The API-manager model editor also falls back to `/api/models?all=true` when `/v1/models` is catalog-protected ([#6372](https://github.com/diegosouzapw/OmniRoute/pull/6372)). Regression guard: `tests/unit/pick-internal-api-key-6372.test.ts`. (thanks @jmengit)
+- **fix(live-ws):** the Live Dashboard WebSocket server now **rejects on bind failure** (e.g. `EADDRINUSE` when the API bridge already holds the port) instead of letting the error surface as an unhandled `error` event that crash-loops the process — the `error` listener is attached to `wss` (not `server`) and releases the EventBus subscription on a failed start ([#6324](https://github.com/diegosouzapw/OmniRoute/issues/6324)). Regression guard: `tests/unit/live-ws-eaddrinuse-6324.test.ts`. (thanks @vinayakkulkarni)
+- **fix(dashboard):** the Home provider-topology widget now trusts the live provider-metrics snapshot — it uses `topology.errorProvider` and live `activeRequests` directly instead of re-deriving state from a stale `lastErrorAt` or applying a frontend timeout filter, so the topology reflects real-time provider health ([#6322](https://github.com/diegosouzapw/OmniRoute/pull/6322)). Regression guard: `tests/unit/home-provider-topology-live-state.test.ts`. (thanks @xz-dev)
+- **fix(sse):** strip zero-width markers from streamed **tool-call arguments** — a follow-up to [#5857](https://github.com/diegosouzapw/OmniRoute/pull/5857). That PR removed injected zero-width joiners (U+200D) from streamed assistant text/reasoning but deliberately left tool-call argument JSON byte-exact. The request-side obfuscation (`open-sse/services/claudeCodeObfuscation.ts`) injects ZWJ into agent words — including the temp path inside the Bash tool description — and Claude models copy that verbatim into generated commands, which are delivered as tool-call arguments rather than assistant text. As a result the ZWJ survived and corrupted code blocks (e.g. a temp path rendered with an invisible joiner). Now `open-sse/handlers/responseSanitizer.ts` strips zero-width code points from tool-call argument strings at every emit site (OpenAI non-stream/stream chat `tool_calls` + legacy `function_call`, native Responses `function_call` items, the OpenAI→Responses conversion, and the native Responses streaming `response.function_call_arguments.delta/.done` events). Only zero-width code points are removed; JSON structure and all other bytes stay identical (no parse/restringify), so normal arguments remain byte-exact. Regression guard: 6 new cases in `tests/unit/response-sanitizer.test.ts` (suite 50/50).
+
+- **fix(nodejs):** the default app log path now resolves under `DATA_DIR` (`~/.omniroute/logs/application/app.log`) instead of `process.cwd()` ([#6197](https://github.com/diegosouzapw/OmniRoute/issues/6197)) — the globally-installed CLI runs from an arbitrary working directory, so anchoring the default to cwd made file logging silently write to (or no-op under) an unrelated directory, contradicting the documented `.env.example` default. `getAppLogFilePath()` now computes the default lazily via the pure `resolveDataDir()` resolver (honours a per-process `DATA_DIR`, no directory-creation side effect); an explicit `APP_LOG_FILE_PATH` still wins. Regression guard: `tests/unit/logenv-datadir-path-6197.test.ts` (3).
+
+- **fix(docker):** AgentBridge/`startMitm` no longer aborts in containers/headless when the Antigravity-default DNS step can't write `/etc/hosts` ([#6127](https://github.com/diegosouzapw/OmniRoute/issues/6127)), and the privileged command's stderr now reaches `app.log` instead of only a bare exit code hitting the toast ([#6198](https://github.com/diegosouzapw/OmniRoute/issues/6198)). The default DNS step (`addDNSEntry`) was called unguarded while cert install and the two sibling DNS steps were each best-effort — in the runtime Docker image (`USER node`, no `sudo`, read-only `/etc/hosts`) it threw `Command failed with code 1` out of `startMitmInternal` and killed the whole start, discarding the stderr. The three DNS steps are extracted into a best-effort `provisionDnsEntries()` where each failure is logged with the full `err` (stderr included, folded in by `systemCommands.ts`) and never aborts the start. Regression guard: `tests/unit/mitm-dns-graceful-degrade-6127.test.ts` (4).
+
+- **fix(providers):** copilot-m365-web now supports the M365 Education "Starter / OfficeWebIncludedCopilot" tier and no longer returns an empty `content:null` stream ([#6210](https://github.com/diegosouzapw/OmniRoute/issues/6210)). Two gaps: (1) `buildWsUrl()` hardcoded the individual-consumer scenario (`OfficeWebPaidConsumerCopilot`, `isEdu=false`) — the EDU tier is now opt-in via `providerSpecificData.tier="edu"`, emitting `scenario=OfficeWebIncludedCopilot`/`isEdu=true` (the individual path is unchanged); (2) the EDU/GPT-5.5 path streams deltas via `arguments[0].writeAtCursor` (incremental) instead of only `messages[].text` (accumulated snapshots), which the parser dropped — a new `accumulateBotContent()` folds both formats, with `type:2 item.result.message` as a last-resort fallback. Regression guard: `tests/unit/copilot-m365-edu-writeatcursor-6210.test.ts` (10). (thanks @qpeyba)
+
+- **fix(providers):** GitLab Duo executor now feeds tool results back into the prompt instead of looping ([#6220](https://github.com/diegosouzapw/OmniRoute/issues/6220)) — `buildPrompt()` branched only on `system`/`user` and took `userParts.at(-1)`, silently dropping the `assistant{tool_calls}` + `tool{result}` turns the client appended, so the reconstructed prompt was byte-identical to turn 1 and the model re-emitted the same `<tool>` call forever. When a tool exchange is present the full conversation is now serialized, folding each tool result back keyed by its `tool_call_id`; simple conversations keep the legacy shape. Complements the tool_call emission from [#6051](https://github.com/diegosouzapw/OmniRoute/issues/6051) (the `kilo-duplicate` label was a false positive — different, sequential defect). Regression guard: `tests/unit/gitlab-tool-result-feedback-6220.test.ts` (4).
+
+- **fix(providers):** opencode-go/opencode-zen can now synthesize the OpenCode CLI identity headers Cloudflare requires on VPS egress ([#5997](https://github.com/diegosouzapw/OmniRoute/issues/5997)) — on a datacenter VPS, `opencode.ai/zen/go/v1/chat/completions` 403s (HTML challenge) requests lacking CLI identity, while the reporter's control curl proved that `User-Agent: opencode-cli/1.0.0` + `x-opencode-client: cli` + `x-opencode-project: default` + fresh request/session UUIDs succeed. Opt-in via `OPENCODE_SYNTHESIZE_CLI_HEADERS=true` (values overridable via `OPENCODE_GO_USER_AGENT`/`OPENCODE_USER_AGENT`/`OPENCODE_CLIENT`/`OPENCODE_PROJECT`); it fills only headers the client did not already send. Kept **off by default** — the forward-only path is deliberate (fabricating a wrong value risks upstream rejection; a prior dedup regressed with `opencode/local`), so this replaces the fragile local header-injection shim without changing default behavior. Regression guard: `tests/unit/opencode-cli-headers-synthesis-5997.test.ts` (6). (thanks @aleksesipenko)
+- fix(resilience): sticky session affinity now evicts and fails over to another account when the pinned account is exhausted/unavailable (#6219)
+
+- fix(sse): Responses API passthrough now drops internal commentary-phase output before forwarding to clients (gated by RESPONSES_PASSTHROUGH_DROP_COMMENTARY, default on) (#6199)
+
+- **fix(sse):** tool-call function schemas with a root `type: null` are now coerced to `type: "object"` before dispatch ([#6359](https://github.com/diegosouzapw/OmniRoute/issues/6359)) — clients like the Codex app emit `parameters: { type: null, ... }` for some tools, which OpenAI-compatible upstreams reject with `400 Invalid schema for function '...': schema must be a JSON Schema of 'type: "object"', got 'type: null'`, failing the whole request. `toolSchemaSanitizer` already stripped the null; it now re-adds the mandatory root `"object"` type (and empty `properties`/open `additionalProperties` when absent). Combinator roots (`anyOf`/`oneOf`/`allOf`) and explicit root types are left untouched. Regression guard: 5 new cases in `tests/unit/tool-schema-sanitizer.test.mjs`.
+
+- **fix(docker):** AgentBridge no longer fails to start on npm/Electron/VPS installs with "MITM manager stub reached at runtime" ([#6344](https://github.com/diegosouzapw/OmniRoute/issues/6344)) — v3.8.45 flipped the production bundler default to Turbopack, but `next.config.mjs` aliased `@/mitm/manager` to its Docker-only degraded stub **unconditionally**. That was harmless while Docker (which sets the alias intentionally for #3390 graceful degradation) was the sole Turbopack consumer, but once every artifact built with Turbopack the stub shipped to all non-Docker users and `startMitm` threw on the first Agent-Bridge start. The alias is now opt-in via `OMNIROUTE_MITM_STUB=1` (set only by the Dockerfile) through the shared `scripts/build/mitm-stub-flag.mjs` helper; default builds bundle the real manager. Regression guard: `tests/unit/mitm-stub-alias-6344.test.mjs` (4).
+
+- **fix(proxy):** stop the v3.8.44 proxy regression that leaked the real IP and disabled healthy proxies ([#6246](https://github.com/diegosouzapw/OmniRoute/issues/6246)). Two coupled defects from the new health scheduler: (1) **IP leak** — when a proxy assigned to a connection was marked `inactive`, resolution fell through to a **direct** egress instead of blocking, exposing the operator's real IP; (2) **over-deactivation** — the sweep flipped a proxy to `inactive` on the **first** failed probe and counted our own 5s timeout / a probe-target `5xx` as the proxy's fault, so healthy paid proxies vanished from egress selection ("my proxies are not being used anymore"). Fix: the sweep decision is extracted into a pure, network-free `decideProxyHealthAction` (`src/lib/proxyHealth/decision.ts`) — by default the health check now **only counts/logs and never downgrades status** (a proxy is downgraded/removed only with `PROXY_AUTO_REMOVE=true`, after `PROXY_AUTO_REMOVE_AFTER` **consecutive** conclusive failures); probes are classified tri-state so an inconclusive result (our timeout, or a `5xx` from the probe target) never penalizes the proxy, and the probe timeout is raised 5s→15s. Separately, `safeResolveProxy` now **fails closed** via the existing policy: a connection whose assigned proxy is dead is blocked instead of leaking direct (`hasBlockingProxyAssignment`), honoring the explicit `proxy off` toggles and the `PROXY_FAIL_OPEN=true` opt-out. Existing proxies stuck `inactive` by the old behavior need a one-time manual re-activate (the operator owns proxy status). Regression guards: `tests/unit/proxy-health-decide-action-6246.test.ts`, `tests/unit/proxy-assigned-unavailable-6246.test.ts`.
+
+- **fix(proxy):** make "Test All" read-only and add bulk enable/disable ([#6246](https://github.com/diegosouzapw/OmniRoute/issues/6246)). Complements the core fail-closed / scheduler fix (#6296) with the two remaining reporter asks. (1) The **"Test All" button** (`POST /api/settings/proxies/auto-test`) used to flip a proxy to `inactive` on a failed reachability probe; since the egress selector excludes `inactive` proxies, a flaky probe (an unreachable `httpbin.org`, a proxy that blocks `HEAD`, or a slow paid proxy) silently disabled every proxy that failed — "Test All" is now **read-only by default** (only the operator sets a proxy active/inactive; opt back into the legacy test-and-set with `PROXY_HEALTH_AUTO_DEACTIVATE=true`). (2) Adds a **bulk enable/disable** proxies endpoint + toolbar action (`POST /api/settings/proxies/batch-activate`) so an operator can re-activate proxies in one click. Regression guard: `tests/unit/proxy-health-6246.test.ts`. (thanks @tenshiak)
+
+- **chatcore (tools): stop the default 128-tool cap from silently dropping opencode's `task`/MCP tools.** opencode (used as an MCP/agent host) sends a large tool list; when it exceeds the speculative `MAX_TOOLS_LIMIT` (128) default, `truncateToolList` did a blind `tools.slice(0, 128)`, dropping every tool past index 128 — including opencode's built-in `task` tool (subagent launch) and many MCP tools, so models routed through OmniRoute could no longer spawn subagents or reach part of their tools. The cap exists to avoid upstream `400`s for providers with real hard limits (e.g. grok-cli 200), so it is kept for those: detection of the opencode client (`isOpencodeClient` — any `x-opencode-*` header, or `opencode` in the user-agent) now only bypasses the **speculative 128 default**, never a known provider ceiling. Precedence is explicit — a proactive/detected provider limit always truncates (even for opencode); otherwise opencode forwards its full tool list; otherwise the unchanged 128 default applies to every other client. Refactors `getEffectiveToolLimit` into `getKnownToolLimit(provider) ?? DEFAULT_LIMIT` (byte-identical for existing callers) and fixes a cosmetic debug-log that reported the truncated count instead of the original. Regression guard: `tests/unit/tool-limit-detector.test.ts`.
+
+- **fix(mitm):** the macOS MITM-cert install check now matches the system keychain again. `security find-certificate -a -Z` prints the SHA-1 as a colon-less hex string, but the installed-check compared it against `getCertFingerprint()`'s colon-separated form, so the substring match never hit — the cert was reported as not-installed and re-prompted for the sudo install on every run. Fingerprints are now normalized (colons stripped, upper-cased) on both sides via the extracted `macCertOutputHasFingerprint` helper. Regression guard: `tests/unit/mitm-cert-mac-fingerprint.test.ts`. ([#6204](https://github.com/diegosouzapw/OmniRoute/pull/6204), closes [#6134](https://github.com/diegosouzapw/OmniRoute/issues/6134) — thanks @rianonehub)
+
+- **fix(api):** `/v1/messages/count_tokens` now counts `tool_use`, `tool_result` and `thinking` content blocks (and array-form `system` prompts) in the local-estimation path, instead of only `text`. Real agentic conversations keep ~95% of their tokens inside tool results; the previous estimate returned near-zero for them, which silently broke Claude Code's auto-compaction (context grew past the window with no compaction until the upstream API rejected the request). The real provider-side count path is unchanged. Regression guard: `tests/unit/messages-count-tokens-route.test.ts`. ([#6221](https://github.com/diegosouzapw/OmniRoute/pull/6221) — thanks @luweiCN)
+
+- **fix(antigravity):** strip a trailing assistant prefill turn for Vertex Claude models to avoid upstream 400s ([#6114](https://github.com/diegosouzapw/OmniRoute/pull/6114)). Regression guard: `tests/unit/antigravity-claude-prefill-strip.test.ts`. (thanks @anki1kr)
+
+- **fix(security):** the mutable cloud-agent routes (`/api/cloud/credentials/update`, `/api/cloud/models/alias`) now require management auth instead of being treated as public. They were classified as public API routes, so a request without management credentials could update stored cloud-agent credentials and model aliases. They are removed from the public-route set, classified as management routes in the authz pipeline, and gated by `requireManagementAuth`; cloud **read**/auth routes stay public. Regression guards: `tests/unit/cloud-write-auth.test.ts`, `tests/unit/authz/classify.test.ts`, `tests/unit/public-api-routes.test.ts`. ([#6233](https://github.com/diegosouzapw/OmniRoute/pull/6233) — thanks @vittoroliveira-dev)
+
+- **refactor(dashboard):** extract the onboarding-wizard "Open provider details" link target into a pure, unit-tested `buildProviderDetailsHref(connection)` helper. The wizard already routes by `connection.id` (the node UUID) rather than the provider category slug (#6144/#6145); this hardens that behavior behind a tested helper that guards a missing id/connection. Regression guard: `tests/unit/provider-onboarding-href.test.ts`. ([#6166](https://github.com/diegosouzapw/OmniRoute/pull/6166) — thanks @KooshaPari)
+
+- **fix(security):** the doubao synthetic device-id generator now derives its digits via an unbiased crypto-random draw (rejection sampling over `crypto.randomBytes()`) instead of a `% 10` reduction, closing a CodeQL `js/biased-cryptographic-random` finding.
+- **fix(agentSkills):** the GitHub-skills generator now resolves `outputDir` to an absolute path before writing, fixing a regression introduced by #6366 (relative-to-cwd base path) that could write generated skill files to the wrong directory.
+- **fix(security):** `/api/keys/{id}/devices` now checks the HTTP method before auth/validation, returning a `405` for non-GET/DELETE verbs instead of a misleading `401`/`500` (closes a `dast-smoke` QUERY-method finding).
+- **fix(quality):** clear the last 2 heavy quality-gate reds on the release tip (cycle pre-flight).
+- **fix(mitm):** the test suite and CI can never mutate the OS trust store — `OMNIROUTE_SKIP_SYSTEM_TRUST=1` is set globally for tests/CI so `installCert`/`uninstallCert`/`installTproxyCa` skip the privileged OS dispatch ([#6310](https://github.com/diegosouzapw/OmniRoute/pull/6310); full detail is under the [3.8.45] section below — this branch received it via the parallel-cycle sync-back).
+- **fix(api):** `POST /api/github-skills` now Zod-validates its request body; documented the new quality-gate env vars and pinned the merge-integrity GitHub Actions to a commit SHA.
+- **fix(skills):** generate the missing `omni-github-skills` registry entry and align the agent-skills catalog-count tests (follow-up to #6186).
+- **fix(quality):** clear the cycle's 11 net-new ESLint errors and make `validate-release-green` suppressions-aware.
+
+### 📝 Maintenance
+
+- **i18n(it):** add 118 missing Italian (`it`) translations (net-additive — no existing keys dropped, valid JSON), improving Italian UI coverage. ([#6212](https://github.com/diegosouzapw/OmniRoute/pull/6212) — thanks @serverless83)
+- **chore(providers):** remove deprecated **MiMo V2** model entries from the catalogs (xiaomi-mimo, opencode-go, zenmux-free, audio TTS) — the upstream V2 line is superseded by MiMo V2.5; drops `mimo-v2-tts`, `mimo-v2-pro`, `mimo-v2-omni`, `mimo-v2-flash`, `mimo-v2-flash-free` and realigns the provider-catalog tests. ([#6248](https://github.com/diegosouzapw/OmniRoute/pull/6248) — thanks @backryun)
+- **chore(release):** ~50 commits on this branch are v3.8.45 pre-flight/hardening fixes and CI-perf work that landed here via the parallel-cycle sync-back (`sync-next-cycle.mjs`, Hard Rule #21) **after** the `v3.8.45` git tag was cut, and are already fully documented under the **[3.8.45]** section below — listed here only so the per-cycle commit-coverage check (`npm run release:uncovered`) doesn't flag them as gaps. Provider/catalog/UX/backend fixes: #6041, #6078, #6108, #6135, #6148, #6149, #6154, #6158, #6161, #6162, #6163, #6164, #6165, #6170, #6177, #6178, #6181, #6186, #6187, #6191, #6193, #6194, #6195, #6200, #6205, #6208, #6209, #6211, #6213, #6223, #6224, #6225, #6226, #6227, #6228, #6229, #6230, #6235, #6291, #6292. CI/release-pipeline work: #6167, #6203, #6214, #6215, #6218, #6273, #6275, #6283, #6284, #6285, #6300, #6305.
+- **chore(release):** additional zero-ref release-cycle plumbing on this branch, kept out of `release:uncovered` on purpose (no `#N` in the commit subject to cite): opening the v3.8.46 cycle, opening/closing the v3.8.45 cycle, the finalized [3.8.45] CHANGELOG i18n sync-back to 42 mirrors, the v3.8.45 cognitive/cyclomatic and file-size drift rebaselines, ESLint stale-suppression pruning (4,273 → 4,233), and clearing test-masking/docs-all pre-flight reds for v3.8.45.
+
+### ⚡ Performance & Infrastructure
+
+- **perf(release-green):** the pre-flight validator (`scripts/quality/validate-release-green.mjs`) now runs its 4 slow suites (unit / vitest / integration / pack-artifact) **concurrently** via `Promise.all` — pre-flight wall time drops from ~the sum of the suites to ~the slowest one (~30min saved per round; Phase 0 was the nº1 bottleneck of the v3.8.45 release benchmark, 2h54 of 6h34 e2e). Guard: `tests/unit/validate-release-green.test.ts` ("runs the slow suites CONCURRENTLY"). ([#6319](https://github.com/diegosouzapw/OmniRoute/pull/6319))
+- **fix(ci):** `scripts/release/sync-next-cycle.mjs` — two defects found live in its first production run (v3.8.45 Phase 5): (1) the `git()` helper's default 1 MiB `maxBuffer` crashed with `ENOBUFS` on `git show origin/main:CHANGELOG.md` (the CHANGELOG alone is >1 MiB) — widened to 64 MiB; (2) the i18n resync only propagated the `[NEXT]` (TBD) section, leaving the just-shipped finalized section as "— TBD" in all 42 mirrors — it now also syncs `[prevVersion]` bounded by the heading below it (new exported pure helper `versionAfter`). Guards: +5 tests in `tests/unit/sync-next-cycle.test.ts` (8/8). ([#6327](https://github.com/diegosouzapw/OmniRoute/pull/6327))
+- **test(ci):** concurrency-sensitive flaky tests are **quarantined into a serial pass** (`tests/unit/serial/`, `--test-concurrency=1`, appended to every unit runner incl. sharded variants — the serial pass is sharded too so concurrent shard jobs never self-collide). Initial set: `glm-coding-plan-monthly-3580`, `quota-division-blocks`, `provider-health-autopilot`, `combo-health-autopilot` — the class behind the ~28min CI wedges/re-runs (two live 1h+ wedges cancelled during this PR's own validation). Discovery + TIA gates track the new glob; systemic root cause (async logger writing after teardown) tracked in [#6360](https://github.com/diegosouzapw/OmniRoute/issues/6360). Guard: `tests/unit/test-serial-quarantine.test.ts` (4). ([#6347](https://github.com/diegosouzapw/OmniRoute/pull/6347))
+
+### 🙌 Contributors
+
+Thanks to everyone whose work landed in v3.8.46:
+
+| Contributor                                                    | PRs / Issues           |
+| -------------------------------------------------------------- | ---------------------- |
+| [@2220258345](https://github.com/2220258345)                   | direct commit / report |
+| [@abdofallah](https://github.com/abdofallah)                   | direct commit / report |
+| [@adentdk](https://github.com/adentdk)                         | direct commit / report |
+| [@aleksesipenko](https://github.com/aleksesipenko)             | direct commit / report |
+| [@anki1kr](https://github.com/anki1kr)                         | direct commit / report |
+| [@anungma](https://github.com/anungma)                         | direct commit / report |
+| [@arpicato](https://github.com/arpicato)                       | direct commit / report |
+| [@backryun](https://github.com/backryun)                       | #6248                  |
+| [@binsarjr](https://github.com/binsarjr)                       | direct commit / report |
+| [@brightfiscalband](https://github.com/brightfiscalband)       | direct commit / report |
+| [@chirag127](https://github.com/chirag127)                     | #6501, #6506           |
+| [@developerjillur](https://github.com/developerjillur)         | direct commit / report |
+| [@dilneiss](https://github.com/dilneiss)                       | #6499                  |
+| [@dtybnrj](https://github.com/dtybnrj)                         | direct commit / report |
+| [@Forcerecon](https://github.com/Forcerecon)                   | direct commit / report |
+| [@hao3039032](https://github.com/hao3039032)                   | direct commit / report |
+| [@Iammilansoni](https://github.com/Iammilansoni)               | #6150, #6245           |
+| [@jmengit](https://github.com/jmengit)                         | direct commit / report |
+| [@jordansilly77-stack](https://github.com/jordansilly77-stack) | direct commit / report |
+| [@JxnLexn](https://github.com/JxnLexn)                         | direct commit / report |
+| [@KooshaPari](https://github.com/KooshaPari)                   | #6166                  |
+| [@loopyd](https://github.com/loopyd)                           | direct commit / report |
+| [@luweiCN](https://github.com/luweiCN)                         | #6221                  |
+| [@makcimbx](https://github.com/makcimbx)                       | direct commit / report |
+| [@muflifadla38](https://github.com/muflifadla38)               | direct commit / report |
+| [@newnol](https://github.com/newnol)                           | direct commit / report |
+| [@ofekbetzalel](https://github.com/ofekbetzalel)               | direct commit / report |
+| [@ohahe52-dot](https://github.com/ohahe52-dot)                 | direct commit / report |
+| [@phidinhmanh](https://github.com/phidinhmanh)                 | direct commit / report |
+| [@powellnorma](https://github.com/powellnorma)                 | direct commit / report |
+| [@qpeyba](https://github.com/qpeyba)                           | direct commit / report |
+| [@RaviTharuma](https://github.com/RaviTharuma)                 | direct commit / report |
+| [@RCrushMe](https://github.com/RCrushMe)                       | direct commit / report |
+| [@rianonehub](https://github.com/rianonehub)                   | #6134, #6204           |
+| [@serverless83](https://github.com/serverless83)               | #6212                  |
+| [@swingtempo](https://github.com/swingtempo)                   | direct commit / report |
+| [@tenshiak](https://github.com/tenshiak)                       | direct commit / report |
+| [@ThongAccount](https://github.com/ThongAccount)               | direct commit / report |
+| [@UnrealAryan](https://github.com/UnrealAryan)                 | direct commit / report |
+| [@vinayakkulkarni](https://github.com/vinayakkulkarni)         | direct commit / report |
+| [@vittoroliveira-dev](https://github.com/vittoroliveira-dev)   | #6233                  |
+| [@warelik](https://github.com/warelik)                         | direct commit / report |
+| [@xxy9468615](https://github.com/xxy9468615)                   | direct commit / report |
+| [@xz-dev](https://github.com/xz-dev)                           | direct commit / report |
+| [@yanpaing007](https://github.com/yanpaing007)                 | direct commit / report |
+| [@diegosouzapw](https://github.com/diegosouzapw)               | maintainer             |
+
+---
+
 ## [3.8.45] — 2026-07-06
 
 ### ✨ New Features

docs/i18n/pt-BR/CHANGELOG.md

@@ -6,6 +6,213 @@
 
 ## [3.8.31] — 2026-06-20
 
+## [3.8.46] — TBD
+
+### ✨ New Features
+
+- **feat(sse):** **hide paid-only models from `auto/*` routing** when `hidePaidModels` is on ([#6512](https://github.com/diegosouzapw/OmniRoute/issues/6512)) — follow-up to #6328/#6495. PR #6495 hid paid-only models from the `GET /v1/models` listing, but `auto/*` combos (`auto/best-coding`, `auto/glm`, …) could still pick a paid-only backend into their candidate pool → a 402/403 at request time. `createVirtualAutoCombo` now filters the candidate pool through the new pure `open-sse/services/autoCombo/paidModelFilter.ts` (`filterPaidOnlyCandidates`), applying the same free-model predicate #6495 uses in `catalog.ts` (`providerHasFreeModels(provider) && isFreeModel(provider, {id})`) whenever `settings.hidePaidModels === true`. Applied before the category/tier/family narrowing, so it covers every `auto/*` combo; an all-paid pool degrades to the existing graceful empty-pool path. **Opt-in — default OFF leaves the pool unchanged** (identity). Regression guard: `tests/unit/autoCombo/paid-model-filter-6512.test.ts` (4, incl. the default-off identity guard).
+- **feat(sse):** **provider-family auto combos** — `auto/glm`, `auto/minimax`, `auto/mimo`, `auto/zai`, `auto/gemma`, `auto/llama`, `auto/gemini` ([#6453](https://github.com/diegosouzapw/OmniRoute/issues/6453)) — new routable ids that materialize an on-demand virtual combo spanning whatever installed backends currently expose that model family, degrading gracefully as backends rotate. A new pure `open-sse/services/autoCombo/modelFamily.ts` (`detectModelFamily`) classifies by model-id prefix for six families; `zai` is instead resolved by provider id (z.ai's hosted API serves the same `glm-*` model ids as every other GLM backend, so `auto/zai` means "route to my z.ai backend specifically" vs `auto/glm`'s "any connected GLM backend"). Reuses the existing `createVirtualAutoCombo` on-demand materialization path (no DB writes) and the `/v1/models` catalog advertising loop. Regression guard: `tests/unit/autoCombo/provider-family-combos.test.ts` (11).
+- **feat(proxy):** native **proxy-pool round-robin / egress IP rotation** ([#6365](https://github.com/diegosouzapw/OmniRoute/issues/6365)) — a scope (global / provider / account) can now hold **multiple** proxies as a pool with a rotation strategy, so outbound requests cycle their egress IP instead of pinning one proxy per scope. Migration `117_proxy_pool_rotation.sql` lifts the `UNIQUE(scope, scope_id)` constraint (rebuild via the canonical rename/copy/drop; existing single assignments become 1-element pools) and adds a `proxy_scope_rotation` companion table holding the per-scope strategy + a persisted monotonic round-robin cursor. Strategies: `round-robin` (default, monotonic cursor — never `Math.random`), `random`, and `sticky-per-N-min`. Resolution (`resolveProxyForScopeFromRegistry` / `resolveProxyForConnectionFromRegistry`) now fetches the alive, position-ordered candidate set (unchanged `PROXY_ALIVE_PREDICATE`) and applies the strategy; an empty / all-dead pool still returns `null` — the #6246 fail-closed guard is **untouched** (never falls through to direct egress). Backend + DB only; dashboard pool-builder UI is a follow-up. Regression guard: `tests/unit/proxy-pool-rotation-6365.test.ts` (8, incl. fail-closed + backward-compat).
+- **feat(providers):** end-to-end **tool/function calling on the native Gemini `/v1beta` endpoint** ([#6222](https://github.com/diegosouzapw/OmniRoute/issues/6222)) — both directions of the Gemini↔OpenAI conversion now preserve tool data (previously silently dropped). Request side: `convertGeminiToInternal` (extracted to its own testable module) maps `tools[].functionDeclarations` → OpenAI `tools`, prior `functionCall` parts → assistant `tool_calls`, and `functionResponse` parts → `tool`-role messages. Response side: `convertOpenAIResponseToGemini` emits `parts[].functionCall {name,args}` from `message.tool_calls`, and the streaming `openAIChunkToGeminiChunk` accumulates fragmented `tool_calls` deltas by index into complete `functionCall` parts. The non-Gemini client paths (Claude, OpenAI-Responses) already preserved tool calls — this closes the gap specific to the native Gemini surface. Regression guard: `tests/unit/v1beta-gemini-tool-calling-6222.test.ts` (6, incl. a streaming SSE round-trip).
+- **feat(providers):** copilot-m365-web **enterprise / work tier** support ([#6334](https://github.com/diegosouzapw/OmniRoute/issues/6334)) — mirrors the EDU-tier pattern (#6210): `M365ConnectionParams` gains an `agent` field, a new opt-in `M365_ENTERPRISE_OVERRIDES` preset (`agent=work`, `scenario=officeweb`, `licenseType=Premium`) applies via `providerSpecificData.tier="enterprise"` (alias `"work"`), and `agent` is also overridable directly via `providerSpecificData.agent`. `buildWsUrl` was hardcoding `agent="web"` (the one enterprise-distinguishing param with no override path), so a Premium work account handshook then returned an empty stream. The individual and EDU paths are untouched. Kilo's dup flag vs #6210 (EDU tier) was a false positive — different tier. Regression guard: `tests/unit/copilot-m365-enterprise-6334.test.ts` (7). End-to-end confirmation on a real Premium work account is a live-VPS validation follow-up (Hard Rule #18). (thanks @Forcerecon)
+- **feat(api):** standardized, provider-agnostic **`effort` + `thinking` request params** ([#6241](https://github.com/diegosouzapw/OmniRoute/issues/6241)) — a thin standardization layer over the existing mature per-provider reasoning plumbing (no provider mapper touched). `providerChatCompletionSchema` gains a canonical `effort` (reusing the shared `none/low/medium/high/xhigh` vocabulary — the UI tiers `extra`/`max` collapse onto `xhigh`) and a boolean `thinking`. A pure `normalizeReasoningRequest` (wired once in `src/sse/handlers/chat.ts`, before any reasoning field is read) folds them onto the fields the translators already consume (`reasoning_effort` / `reasoning.effort` / `thinking`), so they fan out to Anthropic / Gemini / xAI / Responses — an explicit client `reasoning_effort` / object-shaped `thinking` always wins (backward-compatible). `/models` additively exposes `supportsThinking` + `effort_tiers` so the frontend can render the toggles (UI component is a follow-up). Regression guard: `tests/unit/effort-thinking-standardization-6241.test.ts` (12). (thanks @Iammilansoni, @shabeer)
+- **feat(combo):** new **`pipeline` (sequential) combo strategy** ([#6297](https://github.com/diegosouzapw/OmniRoute/issues/6297)) — the 18th routing strategy runs targets **in order**, threading each step's output into the next step's input, with an optional per-step `prompt` (system instruction); only the final step's response is returned. Distinct from `fusion` (parallel fan-out + judge). Implemented as a self-contained `open-sse/services/pipeline.ts` (sibling to `fusion.ts`), dispatched from `combo.ts`; the step list reuses `combo.models` order and reads an optional `prompt` off each target (backward-compatible — ignored by every other strategy). Intermediate steps run non-streaming with tools stripped (complete prose to thread forward); the final step keeps the client's `stream` flag + tools. A failing/empty/unparseable intermediate step fails the whole pipeline explicitly via a sanitized error (never silently swallowed). Kilo's dup flag vs #563 was a false positive (that's model→chain selection; this is a sequential chain). Regression guard: `tests/unit/combo-pipeline-strategy.test.ts` (5). (thanks @ofekbetzalel)
+- **feat(ci):** `check:test-masking` now flags **inline-reimplemented prod conditions** ([#6348](https://github.com/diegosouzapw/OmniRoute/issues/6348)) — a new **report-only** subcheck (v2, 6A.10 family) catches the wrong-shape contract test: a test that recomputes the condition under test inline instead of importing/exercising the real function (the #6216 class, where `=== 500` → `>= 500` stayed green because the test re-implemented the branch). For each added/modified test file it warns when the file textually duplicates a ≥3-token conditional from a production file touched in the same PR **and** does not import the symbol/module owning it, via a pure, fixture-tested `findReimplementedConditions()` with an allowlist mirroring `assertReductionAllowlist`. Report-only for now (does not fail the gate) — to be promoted to blocking after a triage cycle. Regression guard: `tests/unit/check-test-masking.test.ts` (45).
+- **feat(sse):** per-connection routing override (native vs CLIProxyAPI) ([#6339](https://github.com/diegosouzapw/OmniRoute/issues/6339)) — the previously-dead `isCliproxyapiDeepModeEnabled` helper is now wired into `resolveExecutorWithProxy`: a single connection can opt itself into the CLIProxyAPI passthrough executor via `providerSpecificData.cliproxyapiMode="claude-native"`, with precedence **connection override > provider `upstream_proxy_config` mode > default**. `resolveExecutorWithProxy` now receives the resolved connection's `providerSpecificData` (threaded from `chatCore.ts`), so one connection can deep-route while the provider's other connections stay native — no DB schema change (the toggle rides in `providerSpecificData`). Also resolves the same-provider mixing ask in #6340. Regression guard: `tests/unit/chatcore-executor-proxy.test.ts` (9). (thanks @RaviTharuma)
+- **feat(dashboard):** "Add session cookie" modal now shows a prominent **"Open ‹host› →"** link to the provider's own site ([#6268](https://github.com/diegosouzapw/OmniRoute/issues/6268)) — every `-web` cookie-session provider (chatgpt-web, claude-web, gemini-web, kimi-web, lmarena, qwen-web, m365-copilot-web, …) renders a one-click external link (opening the provider's login/home page in a new tab) so operators no longer tab away to retype the URL mid-setup. The host resolves from a pure, unit-tested `resolveWebProviderHost()` (prefers `WEB_COOKIE_PROVIDERS[id].website`, falls back to the registry `baseUrl` origin); non-web providers render exactly as before. Kilo's dup flag vs #6265 (modal-too-small-on-1080p) was a false positive — distinct concern. Regression guard: `tests/unit/resolve-web-provider-host.test.ts` (5). (thanks @chirag127)
+- **feat(providers):** add **DigitalOcean** AI (serverless inference) as an OpenAI-compatible API-key provider ([#6373](https://github.com/diegosouzapw/OmniRoute/pull/6373)) — base `https://inference.do-ai.run/v1`, wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/digitalocean/`, `src/shared/constants/providers/apikey/inference-hosts.ts`). Regression guard: `tests/unit/digitalocean-provider.test.ts`. (thanks @newnol)
+- **feat(providers):** add **Huancheng Public API** (`hcnsec`) as an OpenAI-compatible regional provider ([#6410](https://github.com/diegosouzapw/OmniRoute/pull/6410)) — Xinjiang Huancheng Cybersecurity's public LLM platform (base `https://api.hcnsec.cn/v1`, free credits via daily check-ins), wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/hcnsec/`, `src/shared/constants/providers/apikey/regional.ts`). Regression guard: `tests/unit/hcnsec-provider.test.ts`. (thanks @UnrealAryan)
+- **feat(dashboard):** the web-session credential guide now shows an **"Open {host}" link** ([#6316](https://github.com/diegosouzapw/OmniRoute/pull/6316)) to the provider's sign-in site (derived from the provider `website` via `getProviderWebsiteHost`), so you can jump straight to the page where the cookie/session must be captured. Regression guard: `tests/unit/web-session-provider-link-6316.test.ts`. (thanks @jordansilly77-stack)
+- **feat(cerebras):** add the **Gemma 4 31B** model (`gemma-4-31b`) to the Cerebras registry + pricing table ([#6331](https://github.com/diegosouzapw/OmniRoute/pull/6331)). Regression guard extends `tests/unit/t28-model-catalog-updates.test.ts`. (thanks @backryun)
+- **feat(providers):** add **Yuanbao (web)** as a cookie-session provider ([#6196](https://github.com/diegosouzapw/OmniRoute/issues/6196)) — `yuanbao-web` (Tencent Yuanbao, `yuanbao.tencent.com`) with cookie-only auth (`hy_user`/`hy_token` + public agent id), SSE→OpenAI translation incl. `reasoning_content`, exposing DeepSeek V3/R1 + Hunyuan / Hunyuan-T1. Regression guard: `tests/unit/providers-yuanbao-web.test.ts`. `together-web` was **deferred** (no verifiable web-session endpoint — needs a captured request) and `huggingchat-web` **dropped** (the existing `huggingchat` already is a web-cookie provider). (thanks @chirag127)
+- **feat(providers):** route the built-in **agentrouter** through the dynamic Claude-Code wire image ([#6056](https://github.com/diegosouzapw/OmniRoute/issues/6056)) — a small static allow-set (`CC_WIRE_IMAGE_BUILTINS` in `open-sse/services/ccWireImageBuiltins.ts`), consulted by `isClaudeCodeCompatible` / `isClaudeCodeCompatibleProvider` / `applyFingerprint`, makes agentrouter adopt the CC wire-image headers + fingerprint **while guarding the CC baseUrl/auth branches** so it keeps its own registry `baseUrl` and `x-api-key` auth. Regression guard: `tests/unit/agentrouter-cc-wire-image.test.ts` (asserts the wire image is applied AND agentrouter's baseUrl/auth are preserved). Live WAF-acceptance against agentrouter.org is a VPS validation follow-up (Hard Rule #18).
+- **feat(providers):** **bulk-add API keys for Cloudflare Workers AI** ([#6174](https://github.com/diegosouzapw/OmniRoute/issues/6174)) — `cloudflare-ai` is removed from the bulk-add exclusion list and the bulk parser gains a 3-field `name|accountId|apiKey` mode; the bulk route now builds a **per-entry** `providerSpecificData` so each key carries its own `accountId` (fixing the previous shared-object reuse), and both the create + key-validation paths receive it. Regression guard: `tests/unit/bulk-api-key-parser-cloudflare.test.ts`. (thanks @muflifadla38)
+- **feat(dashboard):** routing/settings UX clarity ([#6147](https://github.com/diegosouzapw/OmniRoute/issues/6147)) — (1) weighted combos show the **effective routing share %** next to each weight when weights don't sum to 100 (`WeightTotalBar.tsx`); (2) the status widget's user-facing **"Cloud Sync" label is renamed** to "Remote Settings Sync" (`CloudSyncStatus.tsx`; internal ids/state untouched); (3) built-in providers gain an **opt-in advanced base-URL override** (`isBaseUrlOverrideEligibleProvider`, hidden behind an "Advanced" toggle, reusing the existing `providerSpecificData.baseUrl` persistence — not globally widened). Regression guard: `tests/unit/routing-settings-ux-6147.test.ts`.
+- **feat(combo):** add an option to **disable session stickiness**, per-combo or globally — round-robin / random combos can rotate to a different connection on every request instead of pinning a whole conversation to one connection by its first-message hash. Resolution precedence per-combo `config.disableSessionStickiness` → global `settings.disableSessionStickiness` → default `false` (preserves the #3825 prompt-cache/504 fix); gates **both** stickiness call sites in `open-sse/services/combo.ts`. Exposed as a global toggle (Combo Defaults) and a per-combo Inherit/on/off control. ([#6168](https://github.com/diegosouzapw/OmniRoute/issues/6168)) Regression guard: `tests/unit/combo-disable-session-stickiness.test.ts`. (thanks @RCrushMe)
+- **feat(docker):** add the `OMNIROUTE_NO_SUDO` env flag for root-less / user-namespaced deployments — the MITM cert-trust command path (`resolveSudoSpawn` in `src/mitm/systemCommands.ts`) now strips the leading `sudo` when the flag is truthy, in addition to the existing root / sudo-missing cases, so the Proxy Agent runs without `sudo` (the operator trusts the CA manually, e.g. via `NODE_EXTRA_CA_CERTS`). Argv-array `spawn` preserved — no shell interpolation (Hard Rule #13). ([#6122](https://github.com/diegosouzapw/OmniRoute/issues/6122)) Regression guard: `tests/unit/mitm-systemCommands-no-sudo.test.ts`. (thanks @powellnorma)
+- **feat(providers):** add **Requesty** as an OpenAI-compatible gateway provider (BYOK, base `https://router.requesty.ai/v1`, ~200 free requests/day) — wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/requesty/`, `src/shared/constants/providers/apikey/gateways.ts`). ([#6120](https://github.com/diegosouzapw/OmniRoute/issues/6120)) Regression guard: `tests/unit/requesty-provider.test.ts`. (thanks @chirag127)
+- **feat(dashboard):** add **configured-only / available-only filters** to the Free Provider Rankings page ([#6150](https://github.com/diegosouzapw/OmniRoute/issues/6150)) — hide providers you haven't configured, or whose connections are all rate-limited / out of quota, via server-side query params (`?configuredOnly` / `?availableOnly` on `GET /api/free-provider-rankings`) backed by a testable lib helper reusing the in-process connection state (no Redis). Both filters default off, so the default view is unchanged; this supersedes the earlier client-side "Configured Only" toggle (#6245) with an available-only dimension and unit-tested logic. Regression guard: `tests/unit/freeProviderRankings-filters.test.ts`.
+
+### 🔧 Bug Fixes
+
+- **fix(dashboard):** adding a second API key connection for the same provider no longer silently overwrites the first — the Add-API-key modal now derives a unique default connection name (`main`, then `main-2`, `main-3`, …) so the backend name-based upsert can't collide (#6499 — thanks @dilneiss).
+- **fix(compression):** the session-dedup engine now also deduplicates a large multi-line block repeated **within a single message** (intra-message dedup), not just across turns; the compression-preview API surfaces a `fallbackReason`, and the fusion panel reports how many models were rate-limited vs failed on total-panel failure (#6501 — thanks @chirag127).
+- **fix(compression):** a stacked-pipeline step naming an unregistered engine now surfaces a `validationErrors` entry instead of silently no-op'ing, so misconfigured pipelines are visible in the preview API (#6506 — thanks @chirag127).
+- **feat(usage):** add a **Codex reset-credit redemption** flow to the Provider Limits UI ([#6361](https://github.com/diegosouzapw/OmniRoute/pull/6361)) — a `useCodexResetCreditRedemption` hook + `/api/usage/codex-reset-credit` route + `codexResetCredits` lib let you redeem banked Codex reset credits from the quota card. Regression guard: `tests/unit/codex-reset-credits.test.ts`. (thanks @JxnLexn)
+- **feat(glm):** add **team-plan quota settings** for `glm-cn` connections ([#6351](https://github.com/diegosouzapw/OmniRoute/pull/6351)) — a dedicated `GlmTeamQuotaFields` form section (team quota id / limits) threaded through the Add/Edit connection modals, persisted via `providerSpecificData`, with the GLM usage service reading the team quota. Regression guards: `tests/unit/glm-team-quota.test.ts`, `provider-specific-data-schema.test.ts`. (thanks @hao3039032)
+- **feat(providers):** add **TinyFish** web-fetch/search support ([#6349](https://github.com/diegosouzapw/OmniRoute/pull/6349)) — a `tinyfish-fetch` executor + `/v1/web/fetch` route + MCP web-fetch tool, registered as a specialty-media provider with request-validation and a search-provider catalog entry. Regression guards: `tests/unit/executor-tinyfish-fetch.test.ts`, `web-fetch-handler.test.ts`, `mcp-web-fetch-tool.test.ts`, `provider-validation-tinyfish.test.ts`. (thanks @dtybnrj)
+- **fix(cli):** `omniroute launch-codex` now spawns `codex.cmd` through a shell on Windows (the npm `.cmd` shim is unresolvable by bare `spawn` → ENOENT), mirroring the qodercli Windows fix (#6263) ([#6312](https://github.com/diegosouzapw/OmniRoute/pull/6312)). Regression guard: `tests/unit/launch-codex-windows-spawn-6312.test.ts`. (thanks @swingtempo)
+- **fix(codex):** isolate the **Spark** quota from the shared Codex quota and stabilize the quota UI ordering / hydration so per-scope limits render consistently ([#6336](https://github.com/diegosouzapw/OmniRoute/pull/6336)). Regression guards: `tests/unit/codex-quota-selection-hydration.test.ts`, `provider-limits-ui.test.ts` + 3 more. (thanks @xz-dev)
+- **feat(api):** add a `hidePaidModels` setting that filters paid-only models out of the `/v1/models` catalog. Regression guard: `tests/unit/models-catalog-hide-paid.test.ts`. (thanks @chirag127)
+- **fix(api-manager):** the fallback model picker now preserves combos instead of dropping them when a primary model is unavailable ([#6443](https://github.com/diegosouzapw/OmniRoute/pull/6443)). Regression guard: `tests/unit/api-manager-page-static.test.ts`. (thanks @jmengit)
+- **fix(providers):** recoverable Antigravity / Cloud-Code (Gemini Code Assist) `403` responses ([#6452](https://github.com/diegosouzapw/OmniRoute/pull/6452)) are now classified as a retryable project-config error instead of a terminal account ban, so a fixable project/API-disabled 403 no longer forces a ~1-year cooldown / full OAuth reconnect. Regression guard: `tests/unit/errorclassifier-antigravity-403.test.ts`. (thanks @developerjillur)
+- **fix(mitm):** `sanitizeHeaders` now redacts `Set-Cookie` response headers so upstream session cookies never leak into logs / diagnostics ([#6451](https://github.com/diegosouzapw/OmniRoute/pull/6451)). Regression guard: `tests/unit/mitm-sanitize-headers.test.ts`. (thanks @developerjillur)
+- **fix(api):** `/api/compression/preview` now accepts `mode: "caveman"` and correctly handles stacked / zero-compression previews ([#6425](https://github.com/diegosouzapw/OmniRoute/issues/6425)). Regression guard: `tests/unit/api/compression-preview-caveman-and-stacked-6425.test.ts`. (thanks @chirag127)
+- **feat(providers):** add **Zed** hosted LLM aggregator as a native-app provider ([#6118](https://github.com/diegosouzapw/OmniRoute/pull/6118)) — OAuth sign-in via the Zed hosted flow, registered through the shared provider registry + executor. Regression guards: `tests/unit/zed-oauth-provider.test.ts`, `zed-import-utils.test.ts`, `zed-docker-detect.test.ts`, `mitm-handler-zed.test.ts`. VPS-validated via live operator login (Hard Rule #18).
+- **fix(oauth):** the Kiro SSO-cache auto-import now **preserves the IDC region** — cross-region Amazon Q / Kiro profiles imported from the SSO cache are no longer collapsed to the default region ([#6113](https://github.com/diegosouzapw/OmniRoute/pull/6113)). Regression guard: `tests/unit/kiro-auto-import-idc-2059.test.ts`. VPS-validated via live operator login (Hard Rule #18).
+- **fix(dashboard):** passthrough model aliases no longer collide when two namespaced model ids share a last segment (port from 9router#1850, [#6431](https://github.com/diegosouzapw/OmniRoute/pull/6431)). `enx/gpt-5.5` and `enx/codebuddy/gpt-5.5` both auto-generated the alias `gpt-5.5`, so the second model could never be added (the UI just alerted "alias already exists"). Aliases are now disambiguated deterministically — bare last segment when free, then parent-qualified (`codebuddy-gpt-5.5`), then a numeric suffix — while re-adding the exact same model id is still blocked. Regression guard: `tests/unit/passthrough-alias-1850.test.ts`. (thanks @arpicato)
+- **fix(translator):** preserve a Gemini `functionResponse` co-located with other parts (another `functionCall`, or trailing `text`) in the same content when translating **Gemini → OpenAI** ([#6376](https://github.com/diegosouzapw/OmniRoute/pull/6376)). `convertGeminiContent()` early-returned the tool message on the first `functionResponse` part, dropping any co-located parts; such contents are now pre-split (one tool message per `functionResponse`, emitted first, plus one message for the remaining parts). Regression guard: `tests/unit/gemini-to-openai-function-response.test.ts`. (thanks @warelik)
+- **fix(headroom):** detect a python interpreter managed by **mise / pyenv / asdf / conda** (port from 9router#2353, [#6382](https://github.com/diegosouzapw/OmniRoute/pull/6382)). Headroom's python probe (`src/lib/headroom/detect.ts`) searched a hardcoded `PATH`, but version managers expose their interpreters via shim dirs that only join `PATH` through interactive-shell activation — which the non-interactive server never runs, so a managed python (≥3.10) was invisible and Headroom reported it missing. The search path now prepends the well-known shim/bin dirs (`~/.local/share/mise/shims`, `~/.pyenv/shims`, `~/.asdf/shims`, `$CONDA_PREFIX/bin`, `~/.local/bin`, respecting `MISE_DATA_DIR`/`PYENV_ROOT`/`ASDF_DATA_DIR` when set), and a new `HEADROOM_PYTHON` env override lets operators point straight at their interpreter (mirroring `HEADROOM_URL`). Still shell-free (`execFileSync`). Regression guard: `tests/unit/headroom-detect.test.ts` (5). (thanks @loopyd)
+- **fix(executors):** strip the OpenAI-Codex/Claude-CLI `client_metadata` passthrough field for **NVIDIA** requests (port from 9router#1887, [#6411](https://github.com/diegosouzapw/OmniRoute/pull/6411)). NVIDIA's OpenAI-compatible wrapper rejects it with `400 Unsupported parameter`, the same class already handled for `cerebras`/`mistral`; `nvidia` (executor `default`) was missing from the strip allowlist so Codex/Claude-Code passthrough requests 400'd. Regression guard: `tests/unit/executor-default-strip-client-metadata.test.ts` (+nvidia case). (thanks @phidinhmanh)
+- **fix(translator):** strip the Claude-style `thinking` field for **NVIDIA `z-ai/glm-5.2`** (port from 9router#2023, [#6413](https://github.com/diegosouzapw/OmniRoute/pull/6413)). NVIDIA's OpenAI-compatible wrapper 400s on `thinking` (a Claude-format client routed here leaves a `thinking:{type:"adaptive"}`); the existing strip rule only dropped `reasoning`. Same class already handled for `minimax-m2.7`. Regression guard: `tests/unit/nvidia-minimax-thinking-strip.test.ts` (+glm-5.2 case). (thanks @phidinhmanh)
+- **fix(translator):** suppress the streamed `</think>` close marker for the **Antigravity IDE** client (port from 9router#1061, [#6415](https://github.com/diegosouzapw/OmniRoute/pull/6415)). On thinking-only turns Antigravity rendered a bare `</think>` as the sole visible content, tripping its loop-detection and wasting requests. Antigravity's UA (`vscode/<v> (Antigravity/<v>)`) is added to the marker-suppress allowlist (alongside OpenCode); Claude Code / Cursor still get the marker, and `x-omniroute-thinking-marker: on` force-restores it. Regression guard: `tests/unit/think-close-marker-suppress-5245.test.ts`. (thanks @abdofallah)
+- **fix(executors):** strip nested `reasoning_content` from messages for **Mistral** (port from 9router#1649, [#6417](https://github.com/diegosouzapw/OmniRoute/pull/6417)). Mistral's API returns `422 extra_forbidden` when an assistant message carries `reasoning_content` (replayed thinking from a prior turn, e.g. via the Codex `/responses` path); the generic top-level 400 field-downgrade retry never covered the nested per-message field. `DefaultExecutor` now strips it for provider `mistral` only, so DeepSeek (which requires replayed `reasoning_content`) is unaffected. Regression guard: `tests/unit/mistral-strip-reasoning-content-1649.test.ts`. (thanks @xxy9468615)
+- **fix(executors):** strip the `client_metadata` passthrough field on the **OpenCode** path (port from 9router#1442, [#6418](https://github.com/diegosouzapw/OmniRoute/pull/6418)). OpenCode upstreams (e.g. `kimi-k2.6` via opencode-go) reject it with `400 "Extra inputs are not permitted, field: 'client_metadata'"`; the DefaultExecutor strip only covered cerebras/mistral and `OpencodeExecutor` extends `BaseExecutor` directly, so nothing removed it there. Regression guard: `tests/unit/opencode-strip-client-metadata-1442.test.ts`. (thanks @yanpaing007)
+- **fix(executors):** inject the `reasoning_content` echo for the native **Moonshot Kimi** provider (port from 9router#1480, [#6419](https://github.com/diegosouzapw/OmniRoute/pull/6419)). Kimi (executor `default`) is a thinking-mode upstream that 400s with "reasoning_content must be passed back" when a prior assistant turn lacks it; the placeholder injection was only wired into the OpenCode meta-provider, so direct multi-turn Kimi conversations failed. Scoped to `kimi` (gateway-served models matching the thinking-model name pattern are unaffected). Regression guard: `tests/unit/kimi-native-reasoning-injected-1480.test.ts`. (thanks @2220258345)
+- **fix(executors):** recover from a strict gateway's `context_management: Extra inputs are not permitted` 400 (port from 9router#1468, [#6420](https://github.com/diegosouzapw/OmniRoute/pull/6420)). **Claude Code** always sends a top-level `context_management` field; strict anthropic-compatible gateways reject it. The dedicated context-editing 400-fallback only fired when OmniRoute's own `contextEditing` feature was enabled (default off), so a client-sent field passed through untouched and 400'd. `context_management` is now in the generic reactive field-strip list, so it's stripped-and-retried once regardless of the feature flag (with correct request re-signing for claude-compatible relays). Regression guard: `tests/unit/provider-field-strips.test.ts`. (thanks @ohahe52-dot)
+- **fix(network):** enable **RFC 8305 Happy Eyeballs** (`autoSelectFamily`) on the direct-egress undici dispatcher (port from 9router#1237, [#6423](https://github.com/diegosouzapw/OmniRoute/pull/6423)). When DNS returns both IPv6 (AAAA) and IPv4 (A) and the IPv6 route is broken (e.g. a NAT64 `64:ff9b::` prefix without routing), undici tried IPv6 first and hung until `ETIMEDOUT` (then a 502 + account lockout), even though `curl` reached the same host. The direct dispatcher now races both families and uses whichever connects first. Proxy paths pin family via `proxyTls` and are unaffected. Regression guard: `tests/unit/direct-dispatcher-pipelining-4580.test.ts`. (thanks @adentdk)
+- **fix(combo):** round-robin now advances the rotation pointer past the model that **actually served**, not the eagerly-scheduled one (port from 9router#948, [#6428](https://github.com/diegosouzapw/OmniRoute/pull/6428)). With `stickyLimit: 1` (true round-robin), when the scheduled model failed and a _different_ model served via fallback, the counter had already advanced +1 from the scheduled index — so the next request reused the fallback-served model, degrading round-robin into hot-spotting on whichever model was healthy. The pointer now advances to the served index + 1 (mirroring the sticky-limit>1 path). Session-stickiness (#3825) and distribution are preserved. Regression guard: `tests/unit/combo-rr-fallback-advance-948.test.ts`. (thanks @binsarjr)
+- **fix(sse):** a non-string `model` field is now rejected with a `400` before the resolver, instead of crashing downstream `.toLowerCase()`/`.split()` calls into an empty-body `500` that escapes the error sanitizer ([#6407](https://github.com/diegosouzapw/OmniRoute/issues/6407)). Regression guard: `tests/unit/chat-non-string-model-6407.test.ts`. (thanks @chirag127)
+- **fix(api):** unknown `/api/*` routes now return a JSON `404` (instead of the dashboard HTML shell) and scalar chat params (`model`/`temperature`/etc.) are validated **before** the provider lookup so malformed requests fail fast with a clear `400` ([#6424](https://github.com/diegosouzapw/OmniRoute/issues/6424), [#6412](https://github.com/diegosouzapw/OmniRoute/issues/6412)). Regression guards: `tests/unit/api/api-catchall-json-404.test.ts`, `tests/unit/chat-early-schema-validation-6412.test.ts`. (thanks @chirag127)
+- **fix(api):** `/v1/chat/completions` now rejects a non-JSON `Content-Type` with a `400` before parsing the body ([#6414](https://github.com/diegosouzapw/OmniRoute/issues/6414)). Regression guard: `tests/unit/v1-chat-completions-content-type-6414.test.ts`. (thanks @chirag127)
+- **fix(api):** the `X-OmniRoute-Compression` response header is now echoed on `/v1/chat/completions` and `/v1/completions` ([#6422](https://github.com/diegosouzapw/OmniRoute/issues/6422)). Regression guard: `tests/unit/compression-header-echo-6422.test.ts`. (thanks @chirag127)
+- **fix(api):** concurrent `GET /v1/models` requests are coalesced into a single catalog build ([#6408](https://github.com/diegosouzapw/OmniRoute/issues/6408)). Regression guard: `tests/unit/v1-models-concurrent-6408.test.ts`. (thanks @chirag127)
+- **fix(api):** `/v1/completions` now echoes the requested `body.model` in its JSON + streamed responses ([#6429](https://github.com/diegosouzapw/OmniRoute/pull/6429)). Regression guard: `tests/unit/completions-body-model-echo.test.ts`. (thanks @chirag127)
+- **fix(api):** env-var master keys now see the full `/v1/models` catalog ([#6406](https://github.com/diegosouzapw/OmniRoute/issues/6406)). Regression guard: `tests/unit/models-catalog-envkey-6406.test.ts`. (thanks @chirag127)
+- **fix(api):** non-streaming `/v1/completions` responses now echo `body.model` aligned with the `X-OmniRoute-Model` header ([#6426](https://github.com/diegosouzapw/OmniRoute/issues/6426)). Regression guard: `tests/unit/v1-completions-model-header-match-6426.test.ts`. (thanks @chirag127)
+- **fix(api):** unknown `/v1/*` routes now return a JSON `404 not_found` instead of the Next.js dashboard HTML shell ([#6405](https://github.com/diegosouzapw/OmniRoute/issues/6405)). Regression guard: `tests/unit/api/v1-catchall-json-404.test.ts`. (thanks @chirag127)
+- **fix(api):** the per-connection provider models route now degrades to the shipped catalog when a provider's `/models` endpoint answers with a redirect ([#6267](https://github.com/diegosouzapw/OmniRoute/issues/6267)) — a `qwen-web` import failed with a raw `Redirect blocked … (307)` 503. `safeOutboundFetch` throws `REDIRECT_BLOCKED` on the 307, `getSafeOutboundFetchErrorStatus` maps it to 503, and `buildDiscoveryErrorFallbackResponse` treated every 503 as a hard error — so the non-empty `getModelsByProviderId("qwen-web")` catalog was never surfaced. A models-endpoint redirect is not a fixable-config error (unlike `URL_GUARD_BLOCKED`/`INVALID_URL`, which stay hard errors), so it now falls back to the local/cached catalog before the 503 short-circuit. General fix — covers any config-driven provider that 307s. Regression guard: `tests/unit/provider-models-qwen-web-redirect-6267.test.ts`. (thanks @chirag127)
+- **fix(api):** the per-connection provider models route (MCP `list_models_catalog` + the dashboard import view) now merges USER-ADDED custom models into its response ([#6247](https://github.com/diegosouzapw/OmniRoute/issues/6247)) — custom models live in the `key_value` namespace `customModels`, which the live REST `/api/v1/models` already merges, but `src/app/api/providers/[id]/models/route.ts` never read `getCustomModels`, so custom models were dropped on both the discovery-success and local_catalog paths. They are now folded into the returned model list (deduped by id, stamped `owned_by: provider`), fixing MCP + the dashboard import view in one place. Regression guard: `tests/unit/provider-models-custom-merge-6247.test.ts`. (thanks @RCrushMe)
+
+- **fix(providers):** GitLab Duo tool-calling follow-up turns no longer fail upstream with `422 {"detail":"Validation error"}` (tokens 0/0, rejected pre-inference). The #6234 tool-result-feedback fix serialized the **entire** multi-turn conversation into GitLab's single-file `code_suggestions` (`small_file`) generation endpoint — folded history that turn-N sent as an oversized `current_file.content_above_cursor` **and** duplicated verbatim into `user_instruction`, tripping the AI-Gateway's `small_file` validation guard. The executor now **bounds** that prompt: it keeps system + latest user message + the most-recent tool round (dropping older turns), caps oversized tool results, and stops duplicating the full prompt into `user_instruction` (which now carries only the short latest user message) — while still feeding the most-recent tool result back so the agent continues (`open-sse/executors/gitlab.ts`, [#6220](https://github.com/diegosouzapw/OmniRoute/issues/6220)). The unit test covers the bounding logic; the upstream 422→200 clearing is VPS-only (Hard Rule #18). Regression guard: `tests/unit/gitlab-tool-exchange-bounded-6220.test.ts`.
+
+- **fix(i18n):** the provider-detail (`/dashboard/providers/[id]`) connection-status filter labels no longer render as `__MISSING__:All` / `__MISSING__:Active` / `__MISSING__:Error` / `__MISSING__:Banned` / `__MISSING__:CreditsExhausted` in non-English locales (notably pt-BR) ([#6290](https://github.com/diegosouzapw/OmniRoute/issues/6290)). Root cause was **not** the namespace mismatch the issue guessed — the `providers.filter*` keys resolve correctly in `en.json`; the debt lived in the locale mirrors (`src/i18n/messages/*.json`), where these five keys carried the `__MISSING__:` sync sentinel in ~15 locales and were absent entirely in ~26 others, so next-intl found the key and echoed the sentinel verbatim. All 40 non-English/-Chinese mirrors now ship real translations for the five `providers.filter*` labels. Regression guard: `tests/unit/i18n-provider-filter-keys-6290.test.ts`. (thanks @diegosouzapw)
+
+- **fix(providers):** the `copilot-m365-web` streaming executor now emits `debug`-level WebSocket diagnostics ([#6210](https://github.com/diegosouzapw/OmniRoute/issues/6210)) — the outbound WS URL (with the `access_token` **redacted** via `redactWsUrl()`), handshake success/failure, and each received SignalR frame's `type`/`target`. Previously the streaming path logged nothing, so an empty `content:null` response (the M365 Education / Starter tier symptom fixed in #6234) was undiagnosable even at `APP_LOG_LEVEL=debug`. The change is debug-level and side-effect-free — it does not alter streaming behavior or the frame parser, and the token never reaches the logs. Regression guard: `tests/unit/copilot-m365-web-logging-6210.test.ts` (thanks @qpeyba)
+
+- **fix(resilience):** a round-robin combo no longer returns `503 all upstream accounts are unavailable` when a compatibility-rejected target is actually healthy ([#6238](https://github.com/diegosouzapw/OmniRoute/issues/6238)). `filterTargetsByRequestCompatibility` drops request-incompatible targets (tool/vision/structured-output unsupported, or below the required context window) **before** any availability check runs, and its `compatible.length === 0` safety net only fired when _all_ targets were filtered — not when the kept targets later all turned out runtime-unavailable (circuit-open / cooldown / no credentials). So a combo could 503 while a compat-rejected-but-healthy provider sat unused. `handleRoundRobinCombo` now keeps the compat-rejected set and, when every compat-kept target was skipped without a single real attempt, probes those rejected targets as a **last-resort fallback tier** (via the new pure `open-sse/services/combo/comboCompatFallback.ts`) before crystallizing the 503. Regression guard: `tests/unit/combo-roundrobin-compat-fallback-6238.test.ts`. (thanks @ThongAccount)
+
+- **fix(startup):** best-effort self-heal for a corrupted Turbopack dev cache on Windows ([#6289](https://github.com/diegosouzapw/OmniRoute/issues/6289)). On Windows, `pnpm dev` can fail at startup when Turbopack `mmap`s a persistent-cache SST file and the OS refuses the mapping (`os error 1455` — "paging file too small"), which Turbopack surfaces as a misleading `Module not found: Can't resolve '@/shared/utils/machine'`. This is a **known upstream Turbopack cache-corruption bug — not our code**. The dev launcher (`scripts/dev/run-next.mjs`) now wraps `nextApp.prepare()` and, when it rejects with that signature (`isTurbopackCacheCorruption` in the new `scripts/dev/turbopackCacheHeal.mjs`), purges `.build/next/**/cache/turbopack` and retries **once** with a clear log. **Caveat — best-effort only:** the corruption often surfaces as a runtime overlay rather than a `prepare()` rejection, so this cannot always intercept it; the reliable remedy remains manually deleting the Turbopack cache dir. Regression guard: `tests/unit/turbopack-cache-heal-6289.test.ts`. (thanks @chirag127)
+
+- **fix(providers):** qodercli PAT auth no longer fails with `spawn qodercli ENOENT` on Windows ([#6263](https://github.com/diegosouzapw/OmniRoute/issues/6263)) — `spawnQoderCli` spawned the bare `qodercli` name with `shell:false` and an unenriched env, so the npm `.cmd` wrapper under `%APPDATA%\npm` (a user-PATH directory) was never resolved. It now resolves the absolute `.cmd`/`.exe` path through the existing `getCliRuntimeStatus("qoder")` resolver in `src/shared/services/cliRuntime.ts` (memoized), spawns with `shell` when the target is a `.cmd`/`.bat`, and uses the cliRuntime-enriched env (PATH + PATHEXT + APPDATA); the ENOENT error now lists the searched paths plus the `CLI_QODER_BIN` override. End-to-end spawn on a real Windows host is host-only (Hard Rule #18); the path-resolution logic is unit-tested. Regression guard: `tests/unit/qodercli-windows-resolve-6263.test.ts`. (thanks @chirag127)
+
+- **fix(sse):** the reasoning-token buffer no longer inflates **probe-sized `max_tokens`** ([#6274](https://github.com/diegosouzapw/OmniRoute/issues/6274)) — Claude Code's `/model` capability check sends `max_tokens: 1`, but for a thinking-capable model with a large output cap (e.g. `glm-5.2`) the #3587 headroom heuristic (`max(current + 1000, ceil(current * 1.5))`) rewrote it to `1001` and forwarded that upstream, wasting tokens on a request that was never a genuine reasoning budget. `resolveReasoningBufferedMaxTokens()` (`open-sse/services/reasoningTokenBuffer.ts`) now short-circuits and returns the caller's value verbatim when it is below the new `REASONING_BUFFER_MIN_TRIGGER` (256) threshold — a tiny explicit limit is a probe, not a reasoning request. Real budgets still receive the #3587 headroom unchanged, and the guard runs after the existing capability checks so unknown / non-reasoning models keep returning `null`. Regression guard: `tests/unit/reasoning-token-buffer-6274.test.ts`. (thanks @brightfiscalband)
+
+- **fix(cli):** `omniroute reset-password` now works as a real subcommand, and password resets over piped (non-TTY) stdin actually apply ([#6261](https://github.com/diegosouzapw/OmniRoute/issues/6261), [#6258](https://github.com/diegosouzapw/OmniRoute/issues/6258)). Two coupled defects: (1) **#6261** — `bin/omniroute.mjs` routed everything through Commander with only two pre-Commander bypasses (`--mcp`, `reset-encrypted-columns`), so `omniroute reset-password` was rejected as an unknown command; only the separate `omniroute-reset-password` bin worked, while the docs falsely advertised the subcommand (incl. a bogus "legacy alias still works"). A pre-Commander bypass mirroring `reset-encrypted-columns` now dynamically imports `bin/reset-password.mjs` (which self-executes) before Commander parses; the three doc lines were corrected. (2) **#6258** — `bin/reset-password.mjs` issued two sequential `rl.question` prompts; under piped stdin the second read never settled at EOF, so `main()` never reached `resetManagementPassword` and the reset was a silent no-op (both prompts printed, no success, password unchanged). The CLI now detects non-TTY stdin and reads it once (first line = password, second line = confirm if present, else reused), adds a `--password-stdin` flag (entire stdin is the password, no confirmation), and exits `0` explicitly so the success line always flushes; interactive TTY behavior is unchanged. Regression guard: `tests/unit/reset-password-cli-6261-6258.test.ts` (3). (thanks @chirag127)
+
+- **fix(db):** the mass-migration **safety abort** now tells the operator how to bypass it and stops flooding the log ([#6260](https://github.com/diegosouzapw/OmniRoute/issues/6260)) — after restoring a backup that wiped the migration tracking table, `runMigrations()` threw the abort on every downstream `ensureDbInitialized()`, re-logging the full banner 11+ times, and the message never mentioned the existing `OMNIROUTE_MAX_PENDING_MIGRATIONS` escape hatch. The abort text now appends a bypass hint (set `OMNIROUTE_MAX_PENDING_MIGRATIONS=0` in `server.env` / `DATA_DIR/.env`), and a new `MigrationSafetyAbortError` is memoized so repeated calls in the same process throw the same instance and emit a single concise line instead of the full cascade. Regression guard: `tests/unit/migration-safety-abort-6260.test.ts`. (thanks @chirag127)
+
+- **fix(auth):** importing a **distinct** Codex/ChatGPT OAuth `auth.json` is no longer falsely rejected as "already exists" when it belongs to a different user in the same workspace ([#6301](https://github.com/diegosouzapw/OmniRoute/issues/6301)). `findExistingCodexConnection` (in `src/lib/oauth/utils/codexAuthImport.ts`) deduped **only** on `providerSpecificData.workspaceId === accountId`, where `accountId` is the shared `chatgpt_account_id`/`tokens.account_id` — so two members of the same ChatGPT Team collapsed onto a single connection (409 `duplicate_account`). The id_token's `https://api.openai.com/auth` claim carries a per-user `chatgpt_user_id` alongside the workspace id (the device-flow path already persisted it as `chatgptUserId`, but the import path did not). Now `parseAndValidateCodexAuth` extracts `userId` (`chatgpt_user_id` → `user_id` → JWT `sub`) into `ParsedCodexAuth`, the create/update paths persist `chatgptUserId` in `providerSpecificData` (mirroring `codex.ts`), and dedup keys on `workspaceId` **AND** `chatgptUserId` — with a backward-compat fallback to legacy accountId-only matching when no stored connection for that workspace records a `chatgptUserId`, so genuinely-same accounts still dedup. Regression guard: `tests/unit/codex-auth-import-userid-dedup-6301.test.ts` (4). (thanks @anungma)
+
+- **fix(providers):** importing models for the **venice-web** provider no longer fails with a red "Provider venice-web does not support models listing" ([#6269](https://github.com/diegosouzapw/OmniRoute/issues/6269)). `venice-web` is a web-cookie provider with an executor but no upstream `/v1/models` endpoint and no registry `models`, so the models route fell through to the tail `400`. Mirroring the `jules`/`linkup-search`/`ollama-search` fix (#5569), it now ships a static local catalog entry in `src/lib/providers/staticModels.ts` — seeding the current Venice lineup (`venice-uncensored`, `llama-3.3-70b`, `qwen3-235b`, `qwen3-4b`, `deepseek-r1-671b`; Venice rotates its catalog, see docs.venice.ai/models/overview) — so the route returns `200` with `source:"local_catalog"`, `intentional:true`. Regression guard: `tests/unit/static-models-venice-web-6269.test.ts`. (thanks @chirag127)
+
+- **fix(api):** the specialty model catalogs (`/v1/embeddings`, `/v1/images`, `/v1/music`, `/v1/videos` model lists) are now derived from the **unified catalog filtered by a predicate** (`getSpecialtyModelsResponse`) instead of ad-hoc per-route logic, so they consistently respect active-credential visibility and stay in sync with the main catalog ([#6303](https://github.com/diegosouzapw/OmniRoute/pull/6303)). Regression guard: `tests/unit/specialty-model-catalog-routes.test.ts`. (thanks @makcimbx)
+- **fix(api):** the agent-bridge server route now resolves the MITM manager via a **dynamic `import("@/mitm/manager.runtime")`** so Turbopack does not statically pull the stub (or over-bundle the manager), and the agent-skills generator anchors its output base path with `path.join(process.cwd(), …)` so Turbopack's static analyzer stops tracing the whole project root ([#6329](https://github.com/diegosouzapw/OmniRoute/issues/6329), [#6366](https://github.com/diegosouzapw/OmniRoute/pull/6366)). Regression guard: `tests/unit/agent-bridge-server-route-dynamic-import.test.ts`. (thanks @Iammilansoni)
+- **fix(api):** internal probes (combo-test, cloud-sync verify) now pick a **management-scoped / allow-all API key** instead of naively grabbing `getApiKeys()[0]` — a restricted `self:usage` first row made the probe fail with "Model X is not allowed for this API key" even when the combo path was healthy (`pickApiKeyForInternalUse` in `src/lib/db/apiKeys.ts`). The API-manager model editor also falls back to `/api/models?all=true` when `/v1/models` is catalog-protected ([#6372](https://github.com/diegosouzapw/OmniRoute/pull/6372)). Regression guard: `tests/unit/pick-internal-api-key-6372.test.ts`. (thanks @jmengit)
+- **fix(live-ws):** the Live Dashboard WebSocket server now **rejects on bind failure** (e.g. `EADDRINUSE` when the API bridge already holds the port) instead of letting the error surface as an unhandled `error` event that crash-loops the process — the `error` listener is attached to `wss` (not `server`) and releases the EventBus subscription on a failed start ([#6324](https://github.com/diegosouzapw/OmniRoute/issues/6324)). Regression guard: `tests/unit/live-ws-eaddrinuse-6324.test.ts`. (thanks @vinayakkulkarni)
+- **fix(dashboard):** the Home provider-topology widget now trusts the live provider-metrics snapshot — it uses `topology.errorProvider` and live `activeRequests` directly instead of re-deriving state from a stale `lastErrorAt` or applying a frontend timeout filter, so the topology reflects real-time provider health ([#6322](https://github.com/diegosouzapw/OmniRoute/pull/6322)). Regression guard: `tests/unit/home-provider-topology-live-state.test.ts`. (thanks @xz-dev)
+- **fix(sse):** strip zero-width markers from streamed **tool-call arguments** — a follow-up to [#5857](https://github.com/diegosouzapw/OmniRoute/pull/5857). That PR removed injected zero-width joiners (U+200D) from streamed assistant text/reasoning but deliberately left tool-call argument JSON byte-exact. The request-side obfuscation (`open-sse/services/claudeCodeObfuscation.ts`) injects ZWJ into agent words — including the temp path inside the Bash tool description — and Claude models copy that verbatim into generated commands, which are delivered as tool-call arguments rather than assistant text. As a result the ZWJ survived and corrupted code blocks (e.g. a temp path rendered with an invisible joiner). Now `open-sse/handlers/responseSanitizer.ts` strips zero-width code points from tool-call argument strings at every emit site (OpenAI non-stream/stream chat `tool_calls` + legacy `function_call`, native Responses `function_call` items, the OpenAI→Responses conversion, and the native Responses streaming `response.function_call_arguments.delta/.done` events). Only zero-width code points are removed; JSON structure and all other bytes stay identical (no parse/restringify), so normal arguments remain byte-exact. Regression guard: 6 new cases in `tests/unit/response-sanitizer.test.ts` (suite 50/50).
+
+- **fix(nodejs):** the default app log path now resolves under `DATA_DIR` (`~/.omniroute/logs/application/app.log`) instead of `process.cwd()` ([#6197](https://github.com/diegosouzapw/OmniRoute/issues/6197)) — the globally-installed CLI runs from an arbitrary working directory, so anchoring the default to cwd made file logging silently write to (or no-op under) an unrelated directory, contradicting the documented `.env.example` default. `getAppLogFilePath()` now computes the default lazily via the pure `resolveDataDir()` resolver (honours a per-process `DATA_DIR`, no directory-creation side effect); an explicit `APP_LOG_FILE_PATH` still wins. Regression guard: `tests/unit/logenv-datadir-path-6197.test.ts` (3).
+
+- **fix(docker):** AgentBridge/`startMitm` no longer aborts in containers/headless when the Antigravity-default DNS step can't write `/etc/hosts` ([#6127](https://github.com/diegosouzapw/OmniRoute/issues/6127)), and the privileged command's stderr now reaches `app.log` instead of only a bare exit code hitting the toast ([#6198](https://github.com/diegosouzapw/OmniRoute/issues/6198)). The default DNS step (`addDNSEntry`) was called unguarded while cert install and the two sibling DNS steps were each best-effort — in the runtime Docker image (`USER node`, no `sudo`, read-only `/etc/hosts`) it threw `Command failed with code 1` out of `startMitmInternal` and killed the whole start, discarding the stderr. The three DNS steps are extracted into a best-effort `provisionDnsEntries()` where each failure is logged with the full `err` (stderr included, folded in by `systemCommands.ts`) and never aborts the start. Regression guard: `tests/unit/mitm-dns-graceful-degrade-6127.test.ts` (4).
+
+- **fix(providers):** copilot-m365-web now supports the M365 Education "Starter / OfficeWebIncludedCopilot" tier and no longer returns an empty `content:null` stream ([#6210](https://github.com/diegosouzapw/OmniRoute/issues/6210)). Two gaps: (1) `buildWsUrl()` hardcoded the individual-consumer scenario (`OfficeWebPaidConsumerCopilot`, `isEdu=false`) — the EDU tier is now opt-in via `providerSpecificData.tier="edu"`, emitting `scenario=OfficeWebIncludedCopilot`/`isEdu=true` (the individual path is unchanged); (2) the EDU/GPT-5.5 path streams deltas via `arguments[0].writeAtCursor` (incremental) instead of only `messages[].text` (accumulated snapshots), which the parser dropped — a new `accumulateBotContent()` folds both formats, with `type:2 item.result.message` as a last-resort fallback. Regression guard: `tests/unit/copilot-m365-edu-writeatcursor-6210.test.ts` (10). (thanks @qpeyba)
+
+- **fix(providers):** GitLab Duo executor now feeds tool results back into the prompt instead of looping ([#6220](https://github.com/diegosouzapw/OmniRoute/issues/6220)) — `buildPrompt()` branched only on `system`/`user` and took `userParts.at(-1)`, silently dropping the `assistant{tool_calls}` + `tool{result}` turns the client appended, so the reconstructed prompt was byte-identical to turn 1 and the model re-emitted the same `<tool>` call forever. When a tool exchange is present the full conversation is now serialized, folding each tool result back keyed by its `tool_call_id`; simple conversations keep the legacy shape. Complements the tool_call emission from [#6051](https://github.com/diegosouzapw/OmniRoute/issues/6051) (the `kilo-duplicate` label was a false positive — different, sequential defect). Regression guard: `tests/unit/gitlab-tool-result-feedback-6220.test.ts` (4).
+
+- **fix(providers):** opencode-go/opencode-zen can now synthesize the OpenCode CLI identity headers Cloudflare requires on VPS egress ([#5997](https://github.com/diegosouzapw/OmniRoute/issues/5997)) — on a datacenter VPS, `opencode.ai/zen/go/v1/chat/completions` 403s (HTML challenge) requests lacking CLI identity, while the reporter's control curl proved that `User-Agent: opencode-cli/1.0.0` + `x-opencode-client: cli` + `x-opencode-project: default` + fresh request/session UUIDs succeed. Opt-in via `OPENCODE_SYNTHESIZE_CLI_HEADERS=true` (values overridable via `OPENCODE_GO_USER_AGENT`/`OPENCODE_USER_AGENT`/`OPENCODE_CLIENT`/`OPENCODE_PROJECT`); it fills only headers the client did not already send. Kept **off by default** — the forward-only path is deliberate (fabricating a wrong value risks upstream rejection; a prior dedup regressed with `opencode/local`), so this replaces the fragile local header-injection shim without changing default behavior. Regression guard: `tests/unit/opencode-cli-headers-synthesis-5997.test.ts` (6). (thanks @aleksesipenko)
+- fix(resilience): sticky session affinity now evicts and fails over to another account when the pinned account is exhausted/unavailable (#6219)
+
+- fix(sse): Responses API passthrough now drops internal commentary-phase output before forwarding to clients (gated by RESPONSES_PASSTHROUGH_DROP_COMMENTARY, default on) (#6199)
+
+- **fix(sse):** tool-call function schemas with a root `type: null` are now coerced to `type: "object"` before dispatch ([#6359](https://github.com/diegosouzapw/OmniRoute/issues/6359)) — clients like the Codex app emit `parameters: { type: null, ... }` for some tools, which OpenAI-compatible upstreams reject with `400 Invalid schema for function '...': schema must be a JSON Schema of 'type: "object"', got 'type: null'`, failing the whole request. `toolSchemaSanitizer` already stripped the null; it now re-adds the mandatory root `"object"` type (and empty `properties`/open `additionalProperties` when absent). Combinator roots (`anyOf`/`oneOf`/`allOf`) and explicit root types are left untouched. Regression guard: 5 new cases in `tests/unit/tool-schema-sanitizer.test.mjs`.
+
+- **fix(docker):** AgentBridge no longer fails to start on npm/Electron/VPS installs with "MITM manager stub reached at runtime" ([#6344](https://github.com/diegosouzapw/OmniRoute/issues/6344)) — v3.8.45 flipped the production bundler default to Turbopack, but `next.config.mjs` aliased `@/mitm/manager` to its Docker-only degraded stub **unconditionally**. That was harmless while Docker (which sets the alias intentionally for #3390 graceful degradation) was the sole Turbopack consumer, but once every artifact built with Turbopack the stub shipped to all non-Docker users and `startMitm` threw on the first Agent-Bridge start. The alias is now opt-in via `OMNIROUTE_MITM_STUB=1` (set only by the Dockerfile) through the shared `scripts/build/mitm-stub-flag.mjs` helper; default builds bundle the real manager. Regression guard: `tests/unit/mitm-stub-alias-6344.test.mjs` (4).
+
+- **fix(proxy):** stop the v3.8.44 proxy regression that leaked the real IP and disabled healthy proxies ([#6246](https://github.com/diegosouzapw/OmniRoute/issues/6246)). Two coupled defects from the new health scheduler: (1) **IP leak** — when a proxy assigned to a connection was marked `inactive`, resolution fell through to a **direct** egress instead of blocking, exposing the operator's real IP; (2) **over-deactivation** — the sweep flipped a proxy to `inactive` on the **first** failed probe and counted our own 5s timeout / a probe-target `5xx` as the proxy's fault, so healthy paid proxies vanished from egress selection ("my proxies are not being used anymore"). Fix: the sweep decision is extracted into a pure, network-free `decideProxyHealthAction` (`src/lib/proxyHealth/decision.ts`) — by default the health check now **only counts/logs and never downgrades status** (a proxy is downgraded/removed only with `PROXY_AUTO_REMOVE=true`, after `PROXY_AUTO_REMOVE_AFTER` **consecutive** conclusive failures); probes are classified tri-state so an inconclusive result (our timeout, or a `5xx` from the probe target) never penalizes the proxy, and the probe timeout is raised 5s→15s. Separately, `safeResolveProxy` now **fails closed** via the existing policy: a connection whose assigned proxy is dead is blocked instead of leaking direct (`hasBlockingProxyAssignment`), honoring the explicit `proxy off` toggles and the `PROXY_FAIL_OPEN=true` opt-out. Existing proxies stuck `inactive` by the old behavior need a one-time manual re-activate (the operator owns proxy status). Regression guards: `tests/unit/proxy-health-decide-action-6246.test.ts`, `tests/unit/proxy-assigned-unavailable-6246.test.ts`.
+
+- **fix(proxy):** make "Test All" read-only and add bulk enable/disable ([#6246](https://github.com/diegosouzapw/OmniRoute/issues/6246)). Complements the core fail-closed / scheduler fix (#6296) with the two remaining reporter asks. (1) The **"Test All" button** (`POST /api/settings/proxies/auto-test`) used to flip a proxy to `inactive` on a failed reachability probe; since the egress selector excludes `inactive` proxies, a flaky probe (an unreachable `httpbin.org`, a proxy that blocks `HEAD`, or a slow paid proxy) silently disabled every proxy that failed — "Test All" is now **read-only by default** (only the operator sets a proxy active/inactive; opt back into the legacy test-and-set with `PROXY_HEALTH_AUTO_DEACTIVATE=true`). (2) Adds a **bulk enable/disable** proxies endpoint + toolbar action (`POST /api/settings/proxies/batch-activate`) so an operator can re-activate proxies in one click. Regression guard: `tests/unit/proxy-health-6246.test.ts`. (thanks @tenshiak)
+
+- **chatcore (tools): stop the default 128-tool cap from silently dropping opencode's `task`/MCP tools.** opencode (used as an MCP/agent host) sends a large tool list; when it exceeds the speculative `MAX_TOOLS_LIMIT` (128) default, `truncateToolList` did a blind `tools.slice(0, 128)`, dropping every tool past index 128 — including opencode's built-in `task` tool (subagent launch) and many MCP tools, so models routed through OmniRoute could no longer spawn subagents or reach part of their tools. The cap exists to avoid upstream `400`s for providers with real hard limits (e.g. grok-cli 200), so it is kept for those: detection of the opencode client (`isOpencodeClient` — any `x-opencode-*` header, or `opencode` in the user-agent) now only bypasses the **speculative 128 default**, never a known provider ceiling. Precedence is explicit — a proactive/detected provider limit always truncates (even for opencode); otherwise opencode forwards its full tool list; otherwise the unchanged 128 default applies to every other client. Refactors `getEffectiveToolLimit` into `getKnownToolLimit(provider) ?? DEFAULT_LIMIT` (byte-identical for existing callers) and fixes a cosmetic debug-log that reported the truncated count instead of the original. Regression guard: `tests/unit/tool-limit-detector.test.ts`.
+
+- **fix(mitm):** the macOS MITM-cert install check now matches the system keychain again. `security find-certificate -a -Z` prints the SHA-1 as a colon-less hex string, but the installed-check compared it against `getCertFingerprint()`'s colon-separated form, so the substring match never hit — the cert was reported as not-installed and re-prompted for the sudo install on every run. Fingerprints are now normalized (colons stripped, upper-cased) on both sides via the extracted `macCertOutputHasFingerprint` helper. Regression guard: `tests/unit/mitm-cert-mac-fingerprint.test.ts`. ([#6204](https://github.com/diegosouzapw/OmniRoute/pull/6204), closes [#6134](https://github.com/diegosouzapw/OmniRoute/issues/6134) — thanks @rianonehub)
+
+- **fix(api):** `/v1/messages/count_tokens` now counts `tool_use`, `tool_result` and `thinking` content blocks (and array-form `system` prompts) in the local-estimation path, instead of only `text`. Real agentic conversations keep ~95% of their tokens inside tool results; the previous estimate returned near-zero for them, which silently broke Claude Code's auto-compaction (context grew past the window with no compaction until the upstream API rejected the request). The real provider-side count path is unchanged. Regression guard: `tests/unit/messages-count-tokens-route.test.ts`. ([#6221](https://github.com/diegosouzapw/OmniRoute/pull/6221) — thanks @luweiCN)
+
+- **fix(antigravity):** strip a trailing assistant prefill turn for Vertex Claude models to avoid upstream 400s ([#6114](https://github.com/diegosouzapw/OmniRoute/pull/6114)). Regression guard: `tests/unit/antigravity-claude-prefill-strip.test.ts`. (thanks @anki1kr)
+
+- **fix(security):** the mutable cloud-agent routes (`/api/cloud/credentials/update`, `/api/cloud/models/alias`) now require management auth instead of being treated as public. They were classified as public API routes, so a request without management credentials could update stored cloud-agent credentials and model aliases. They are removed from the public-route set, classified as management routes in the authz pipeline, and gated by `requireManagementAuth`; cloud **read**/auth routes stay public. Regression guards: `tests/unit/cloud-write-auth.test.ts`, `tests/unit/authz/classify.test.ts`, `tests/unit/public-api-routes.test.ts`. ([#6233](https://github.com/diegosouzapw/OmniRoute/pull/6233) — thanks @vittoroliveira-dev)
+
+- **refactor(dashboard):** extract the onboarding-wizard "Open provider details" link target into a pure, unit-tested `buildProviderDetailsHref(connection)` helper. The wizard already routes by `connection.id` (the node UUID) rather than the provider category slug (#6144/#6145); this hardens that behavior behind a tested helper that guards a missing id/connection. Regression guard: `tests/unit/provider-onboarding-href.test.ts`. ([#6166](https://github.com/diegosouzapw/OmniRoute/pull/6166) — thanks @KooshaPari)
+
+- **fix(security):** the doubao synthetic device-id generator now derives its digits via an unbiased crypto-random draw (rejection sampling over `crypto.randomBytes()`) instead of a `% 10` reduction, closing a CodeQL `js/biased-cryptographic-random` finding.
+- **fix(agentSkills):** the GitHub-skills generator now resolves `outputDir` to an absolute path before writing, fixing a regression introduced by #6366 (relative-to-cwd base path) that could write generated skill files to the wrong directory.
+- **fix(security):** `/api/keys/{id}/devices` now checks the HTTP method before auth/validation, returning a `405` for non-GET/DELETE verbs instead of a misleading `401`/`500` (closes a `dast-smoke` QUERY-method finding).
+- **fix(quality):** clear the last 2 heavy quality-gate reds on the release tip (cycle pre-flight).
+- **fix(mitm):** the test suite and CI can never mutate the OS trust store — `OMNIROUTE_SKIP_SYSTEM_TRUST=1` is set globally for tests/CI so `installCert`/`uninstallCert`/`installTproxyCa` skip the privileged OS dispatch ([#6310](https://github.com/diegosouzapw/OmniRoute/pull/6310); full detail is under the [3.8.45] section below — this branch received it via the parallel-cycle sync-back).
+- **fix(api):** `POST /api/github-skills` now Zod-validates its request body; documented the new quality-gate env vars and pinned the merge-integrity GitHub Actions to a commit SHA.
+- **fix(skills):** generate the missing `omni-github-skills` registry entry and align the agent-skills catalog-count tests (follow-up to #6186).
+- **fix(quality):** clear the cycle's 11 net-new ESLint errors and make `validate-release-green` suppressions-aware.
+
+### 📝 Maintenance
+
+- **i18n(it):** add 118 missing Italian (`it`) translations (net-additive — no existing keys dropped, valid JSON), improving Italian UI coverage. ([#6212](https://github.com/diegosouzapw/OmniRoute/pull/6212) — thanks @serverless83)
+- **chore(providers):** remove deprecated **MiMo V2** model entries from the catalogs (xiaomi-mimo, opencode-go, zenmux-free, audio TTS) — the upstream V2 line is superseded by MiMo V2.5; drops `mimo-v2-tts`, `mimo-v2-pro`, `mimo-v2-omni`, `mimo-v2-flash`, `mimo-v2-flash-free` and realigns the provider-catalog tests. ([#6248](https://github.com/diegosouzapw/OmniRoute/pull/6248) — thanks @backryun)
+- **chore(release):** ~50 commits on this branch are v3.8.45 pre-flight/hardening fixes and CI-perf work that landed here via the parallel-cycle sync-back (`sync-next-cycle.mjs`, Hard Rule #21) **after** the `v3.8.45` git tag was cut, and are already fully documented under the **[3.8.45]** section below — listed here only so the per-cycle commit-coverage check (`npm run release:uncovered`) doesn't flag them as gaps. Provider/catalog/UX/backend fixes: #6041, #6078, #6108, #6135, #6148, #6149, #6154, #6158, #6161, #6162, #6163, #6164, #6165, #6170, #6177, #6178, #6181, #6186, #6187, #6191, #6193, #6194, #6195, #6200, #6205, #6208, #6209, #6211, #6213, #6223, #6224, #6225, #6226, #6227, #6228, #6229, #6230, #6235, #6291, #6292. CI/release-pipeline work: #6167, #6203, #6214, #6215, #6218, #6273, #6275, #6283, #6284, #6285, #6300, #6305.
+- **chore(release):** additional zero-ref release-cycle plumbing on this branch, kept out of `release:uncovered` on purpose (no `#N` in the commit subject to cite): opening the v3.8.46 cycle, opening/closing the v3.8.45 cycle, the finalized [3.8.45] CHANGELOG i18n sync-back to 42 mirrors, the v3.8.45 cognitive/cyclomatic and file-size drift rebaselines, ESLint stale-suppression pruning (4,273 → 4,233), and clearing test-masking/docs-all pre-flight reds for v3.8.45.
+
+### ⚡ Performance & Infrastructure
+
+- **perf(release-green):** the pre-flight validator (`scripts/quality/validate-release-green.mjs`) now runs its 4 slow suites (unit / vitest / integration / pack-artifact) **concurrently** via `Promise.all` — pre-flight wall time drops from ~the sum of the suites to ~the slowest one (~30min saved per round; Phase 0 was the nº1 bottleneck of the v3.8.45 release benchmark, 2h54 of 6h34 e2e). Guard: `tests/unit/validate-release-green.test.ts` ("runs the slow suites CONCURRENTLY"). ([#6319](https://github.com/diegosouzapw/OmniRoute/pull/6319))
+- **fix(ci):** `scripts/release/sync-next-cycle.mjs` — two defects found live in its first production run (v3.8.45 Phase 5): (1) the `git()` helper's default 1 MiB `maxBuffer` crashed with `ENOBUFS` on `git show origin/main:CHANGELOG.md` (the CHANGELOG alone is >1 MiB) — widened to 64 MiB; (2) the i18n resync only propagated the `[NEXT]` (TBD) section, leaving the just-shipped finalized section as "— TBD" in all 42 mirrors — it now also syncs `[prevVersion]` bounded by the heading below it (new exported pure helper `versionAfter`). Guards: +5 tests in `tests/unit/sync-next-cycle.test.ts` (8/8). ([#6327](https://github.com/diegosouzapw/OmniRoute/pull/6327))
+- **test(ci):** concurrency-sensitive flaky tests are **quarantined into a serial pass** (`tests/unit/serial/`, `--test-concurrency=1`, appended to every unit runner incl. sharded variants — the serial pass is sharded too so concurrent shard jobs never self-collide). Initial set: `glm-coding-plan-monthly-3580`, `quota-division-blocks`, `provider-health-autopilot`, `combo-health-autopilot` — the class behind the ~28min CI wedges/re-runs (two live 1h+ wedges cancelled during this PR's own validation). Discovery + TIA gates track the new glob; systemic root cause (async logger writing after teardown) tracked in [#6360](https://github.com/diegosouzapw/OmniRoute/issues/6360). Guard: `tests/unit/test-serial-quarantine.test.ts` (4). ([#6347](https://github.com/diegosouzapw/OmniRoute/pull/6347))
+
+### 🙌 Contributors
+
+Thanks to everyone whose work landed in v3.8.46:
+
+| Contributor                                                    | PRs / Issues           |
+| -------------------------------------------------------------- | ---------------------- |
+| [@2220258345](https://github.com/2220258345)                   | direct commit / report |
+| [@abdofallah](https://github.com/abdofallah)                   | direct commit / report |
+| [@adentdk](https://github.com/adentdk)                         | direct commit / report |
+| [@aleksesipenko](https://github.com/aleksesipenko)             | direct commit / report |
+| [@anki1kr](https://github.com/anki1kr)                         | direct commit / report |
+| [@anungma](https://github.com/anungma)                         | direct commit / report |
+| [@arpicato](https://github.com/arpicato)                       | direct commit / report |
+| [@backryun](https://github.com/backryun)                       | #6248                  |
+| [@binsarjr](https://github.com/binsarjr)                       | direct commit / report |
+| [@brightfiscalband](https://github.com/brightfiscalband)       | direct commit / report |
+| [@chirag127](https://github.com/chirag127)                     | #6501, #6506           |
+| [@developerjillur](https://github.com/developerjillur)         | direct commit / report |
+| [@dilneiss](https://github.com/dilneiss)                       | #6499                  |
+| [@dtybnrj](https://github.com/dtybnrj)                         | direct commit / report |
+| [@Forcerecon](https://github.com/Forcerecon)                   | direct commit / report |
+| [@hao3039032](https://github.com/hao3039032)                   | direct commit / report |
+| [@Iammilansoni](https://github.com/Iammilansoni)               | #6150, #6245           |
+| [@jmengit](https://github.com/jmengit)                         | direct commit / report |
+| [@jordansilly77-stack](https://github.com/jordansilly77-stack) | direct commit / report |
+| [@JxnLexn](https://github.com/JxnLexn)                         | direct commit / report |
+| [@KooshaPari](https://github.com/KooshaPari)                   | #6166                  |
+| [@loopyd](https://github.com/loopyd)                           | direct commit / report |
+| [@luweiCN](https://github.com/luweiCN)                         | #6221                  |
+| [@makcimbx](https://github.com/makcimbx)                       | direct commit / report |
+| [@muflifadla38](https://github.com/muflifadla38)               | direct commit / report |
+| [@newnol](https://github.com/newnol)                           | direct commit / report |
+| [@ofekbetzalel](https://github.com/ofekbetzalel)               | direct commit / report |
+| [@ohahe52-dot](https://github.com/ohahe52-dot)                 | direct commit / report |
+| [@phidinhmanh](https://github.com/phidinhmanh)                 | direct commit / report |
+| [@powellnorma](https://github.com/powellnorma)                 | direct commit / report |
+| [@qpeyba](https://github.com/qpeyba)                           | direct commit / report |
+| [@RaviTharuma](https://github.com/RaviTharuma)                 | direct commit / report |
+| [@RCrushMe](https://github.com/RCrushMe)                       | direct commit / report |
+| [@rianonehub](https://github.com/rianonehub)                   | #6134, #6204           |
+| [@serverless83](https://github.com/serverless83)               | #6212                  |
+| [@swingtempo](https://github.com/swingtempo)                   | direct commit / report |
+| [@tenshiak](https://github.com/tenshiak)                       | direct commit / report |
+| [@ThongAccount](https://github.com/ThongAccount)               | direct commit / report |
+| [@UnrealAryan](https://github.com/UnrealAryan)                 | direct commit / report |
+| [@vinayakkulkarni](https://github.com/vinayakkulkarni)         | direct commit / report |
+| [@vittoroliveira-dev](https://github.com/vittoroliveira-dev)   | #6233                  |
+| [@warelik](https://github.com/warelik)                         | direct commit / report |
+| [@xxy9468615](https://github.com/xxy9468615)                   | direct commit / report |
+| [@xz-dev](https://github.com/xz-dev)                           | direct commit / report |
+| [@yanpaing007](https://github.com/yanpaing007)                 | direct commit / report |
+| [@diegosouzapw](https://github.com/diegosouzapw)               | maintainer             |
+
+---
+
 ## [3.8.45] — 2026-07-06
 
 ### ✨ New Features

docs/i18n/pt/CHANGELOG.md

@@ -6,6 +6,213 @@
 
 ## [3.8.31] — 2026-06-20
 
+## [3.8.46] — TBD
+
+### ✨ New Features
+
+- **feat(sse):** **hide paid-only models from `auto/*` routing** when `hidePaidModels` is on ([#6512](https://github.com/diegosouzapw/OmniRoute/issues/6512)) — follow-up to #6328/#6495. PR #6495 hid paid-only models from the `GET /v1/models` listing, but `auto/*` combos (`auto/best-coding`, `auto/glm`, …) could still pick a paid-only backend into their candidate pool → a 402/403 at request time. `createVirtualAutoCombo` now filters the candidate pool through the new pure `open-sse/services/autoCombo/paidModelFilter.ts` (`filterPaidOnlyCandidates`), applying the same free-model predicate #6495 uses in `catalog.ts` (`providerHasFreeModels(provider) && isFreeModel(provider, {id})`) whenever `settings.hidePaidModels === true`. Applied before the category/tier/family narrowing, so it covers every `auto/*` combo; an all-paid pool degrades to the existing graceful empty-pool path. **Opt-in — default OFF leaves the pool unchanged** (identity). Regression guard: `tests/unit/autoCombo/paid-model-filter-6512.test.ts` (4, incl. the default-off identity guard).
+- **feat(sse):** **provider-family auto combos** — `auto/glm`, `auto/minimax`, `auto/mimo`, `auto/zai`, `auto/gemma`, `auto/llama`, `auto/gemini` ([#6453](https://github.com/diegosouzapw/OmniRoute/issues/6453)) — new routable ids that materialize an on-demand virtual combo spanning whatever installed backends currently expose that model family, degrading gracefully as backends rotate. A new pure `open-sse/services/autoCombo/modelFamily.ts` (`detectModelFamily`) classifies by model-id prefix for six families; `zai` is instead resolved by provider id (z.ai's hosted API serves the same `glm-*` model ids as every other GLM backend, so `auto/zai` means "route to my z.ai backend specifically" vs `auto/glm`'s "any connected GLM backend"). Reuses the existing `createVirtualAutoCombo` on-demand materialization path (no DB writes) and the `/v1/models` catalog advertising loop. Regression guard: `tests/unit/autoCombo/provider-family-combos.test.ts` (11).
+- **feat(proxy):** native **proxy-pool round-robin / egress IP rotation** ([#6365](https://github.com/diegosouzapw/OmniRoute/issues/6365)) — a scope (global / provider / account) can now hold **multiple** proxies as a pool with a rotation strategy, so outbound requests cycle their egress IP instead of pinning one proxy per scope. Migration `117_proxy_pool_rotation.sql` lifts the `UNIQUE(scope, scope_id)` constraint (rebuild via the canonical rename/copy/drop; existing single assignments become 1-element pools) and adds a `proxy_scope_rotation` companion table holding the per-scope strategy + a persisted monotonic round-robin cursor. Strategies: `round-robin` (default, monotonic cursor — never `Math.random`), `random`, and `sticky-per-N-min`. Resolution (`resolveProxyForScopeFromRegistry` / `resolveProxyForConnectionFromRegistry`) now fetches the alive, position-ordered candidate set (unchanged `PROXY_ALIVE_PREDICATE`) and applies the strategy; an empty / all-dead pool still returns `null` — the #6246 fail-closed guard is **untouched** (never falls through to direct egress). Backend + DB only; dashboard pool-builder UI is a follow-up. Regression guard: `tests/unit/proxy-pool-rotation-6365.test.ts` (8, incl. fail-closed + backward-compat).
+- **feat(providers):** end-to-end **tool/function calling on the native Gemini `/v1beta` endpoint** ([#6222](https://github.com/diegosouzapw/OmniRoute/issues/6222)) — both directions of the Gemini↔OpenAI conversion now preserve tool data (previously silently dropped). Request side: `convertGeminiToInternal` (extracted to its own testable module) maps `tools[].functionDeclarations` → OpenAI `tools`, prior `functionCall` parts → assistant `tool_calls`, and `functionResponse` parts → `tool`-role messages. Response side: `convertOpenAIResponseToGemini` emits `parts[].functionCall {name,args}` from `message.tool_calls`, and the streaming `openAIChunkToGeminiChunk` accumulates fragmented `tool_calls` deltas by index into complete `functionCall` parts. The non-Gemini client paths (Claude, OpenAI-Responses) already preserved tool calls — this closes the gap specific to the native Gemini surface. Regression guard: `tests/unit/v1beta-gemini-tool-calling-6222.test.ts` (6, incl. a streaming SSE round-trip).
+- **feat(providers):** copilot-m365-web **enterprise / work tier** support ([#6334](https://github.com/diegosouzapw/OmniRoute/issues/6334)) — mirrors the EDU-tier pattern (#6210): `M365ConnectionParams` gains an `agent` field, a new opt-in `M365_ENTERPRISE_OVERRIDES` preset (`agent=work`, `scenario=officeweb`, `licenseType=Premium`) applies via `providerSpecificData.tier="enterprise"` (alias `"work"`), and `agent` is also overridable directly via `providerSpecificData.agent`. `buildWsUrl` was hardcoding `agent="web"` (the one enterprise-distinguishing param with no override path), so a Premium work account handshook then returned an empty stream. The individual and EDU paths are untouched. Kilo's dup flag vs #6210 (EDU tier) was a false positive — different tier. Regression guard: `tests/unit/copilot-m365-enterprise-6334.test.ts` (7). End-to-end confirmation on a real Premium work account is a live-VPS validation follow-up (Hard Rule #18). (thanks @Forcerecon)
+- **feat(api):** standardized, provider-agnostic **`effort` + `thinking` request params** ([#6241](https://github.com/diegosouzapw/OmniRoute/issues/6241)) — a thin standardization layer over the existing mature per-provider reasoning plumbing (no provider mapper touched). `providerChatCompletionSchema` gains a canonical `effort` (reusing the shared `none/low/medium/high/xhigh` vocabulary — the UI tiers `extra`/`max` collapse onto `xhigh`) and a boolean `thinking`. A pure `normalizeReasoningRequest` (wired once in `src/sse/handlers/chat.ts`, before any reasoning field is read) folds them onto the fields the translators already consume (`reasoning_effort` / `reasoning.effort` / `thinking`), so they fan out to Anthropic / Gemini / xAI / Responses — an explicit client `reasoning_effort` / object-shaped `thinking` always wins (backward-compatible). `/models` additively exposes `supportsThinking` + `effort_tiers` so the frontend can render the toggles (UI component is a follow-up). Regression guard: `tests/unit/effort-thinking-standardization-6241.test.ts` (12). (thanks @Iammilansoni, @shabeer)
+- **feat(combo):** new **`pipeline` (sequential) combo strategy** ([#6297](https://github.com/diegosouzapw/OmniRoute/issues/6297)) — the 18th routing strategy runs targets **in order**, threading each step's output into the next step's input, with an optional per-step `prompt` (system instruction); only the final step's response is returned. Distinct from `fusion` (parallel fan-out + judge). Implemented as a self-contained `open-sse/services/pipeline.ts` (sibling to `fusion.ts`), dispatched from `combo.ts`; the step list reuses `combo.models` order and reads an optional `prompt` off each target (backward-compatible — ignored by every other strategy). Intermediate steps run non-streaming with tools stripped (complete prose to thread forward); the final step keeps the client's `stream` flag + tools. A failing/empty/unparseable intermediate step fails the whole pipeline explicitly via a sanitized error (never silently swallowed). Kilo's dup flag vs #563 was a false positive (that's model→chain selection; this is a sequential chain). Regression guard: `tests/unit/combo-pipeline-strategy.test.ts` (5). (thanks @ofekbetzalel)
+- **feat(ci):** `check:test-masking` now flags **inline-reimplemented prod conditions** ([#6348](https://github.com/diegosouzapw/OmniRoute/issues/6348)) — a new **report-only** subcheck (v2, 6A.10 family) catches the wrong-shape contract test: a test that recomputes the condition under test inline instead of importing/exercising the real function (the #6216 class, where `=== 500` → `>= 500` stayed green because the test re-implemented the branch). For each added/modified test file it warns when the file textually duplicates a ≥3-token conditional from a production file touched in the same PR **and** does not import the symbol/module owning it, via a pure, fixture-tested `findReimplementedConditions()` with an allowlist mirroring `assertReductionAllowlist`. Report-only for now (does not fail the gate) — to be promoted to blocking after a triage cycle. Regression guard: `tests/unit/check-test-masking.test.ts` (45).
+- **feat(sse):** per-connection routing override (native vs CLIProxyAPI) ([#6339](https://github.com/diegosouzapw/OmniRoute/issues/6339)) — the previously-dead `isCliproxyapiDeepModeEnabled` helper is now wired into `resolveExecutorWithProxy`: a single connection can opt itself into the CLIProxyAPI passthrough executor via `providerSpecificData.cliproxyapiMode="claude-native"`, with precedence **connection override > provider `upstream_proxy_config` mode > default**. `resolveExecutorWithProxy` now receives the resolved connection's `providerSpecificData` (threaded from `chatCore.ts`), so one connection can deep-route while the provider's other connections stay native — no DB schema change (the toggle rides in `providerSpecificData`). Also resolves the same-provider mixing ask in #6340. Regression guard: `tests/unit/chatcore-executor-proxy.test.ts` (9). (thanks @RaviTharuma)
+- **feat(dashboard):** "Add session cookie" modal now shows a prominent **"Open ‹host› →"** link to the provider's own site ([#6268](https://github.com/diegosouzapw/OmniRoute/issues/6268)) — every `-web` cookie-session provider (chatgpt-web, claude-web, gemini-web, kimi-web, lmarena, qwen-web, m365-copilot-web, …) renders a one-click external link (opening the provider's login/home page in a new tab) so operators no longer tab away to retype the URL mid-setup. The host resolves from a pure, unit-tested `resolveWebProviderHost()` (prefers `WEB_COOKIE_PROVIDERS[id].website`, falls back to the registry `baseUrl` origin); non-web providers render exactly as before. Kilo's dup flag vs #6265 (modal-too-small-on-1080p) was a false positive — distinct concern. Regression guard: `tests/unit/resolve-web-provider-host.test.ts` (5). (thanks @chirag127)
+- **feat(providers):** add **DigitalOcean** AI (serverless inference) as an OpenAI-compatible API-key provider ([#6373](https://github.com/diegosouzapw/OmniRoute/pull/6373)) — base `https://inference.do-ai.run/v1`, wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/digitalocean/`, `src/shared/constants/providers/apikey/inference-hosts.ts`). Regression guard: `tests/unit/digitalocean-provider.test.ts`. (thanks @newnol)
+- **feat(providers):** add **Huancheng Public API** (`hcnsec`) as an OpenAI-compatible regional provider ([#6410](https://github.com/diegosouzapw/OmniRoute/pull/6410)) — Xinjiang Huancheng Cybersecurity's public LLM platform (base `https://api.hcnsec.cn/v1`, free credits via daily check-ins), wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/hcnsec/`, `src/shared/constants/providers/apikey/regional.ts`). Regression guard: `tests/unit/hcnsec-provider.test.ts`. (thanks @UnrealAryan)
+- **feat(dashboard):** the web-session credential guide now shows an **"Open {host}" link** ([#6316](https://github.com/diegosouzapw/OmniRoute/pull/6316)) to the provider's sign-in site (derived from the provider `website` via `getProviderWebsiteHost`), so you can jump straight to the page where the cookie/session must be captured. Regression guard: `tests/unit/web-session-provider-link-6316.test.ts`. (thanks @jordansilly77-stack)
+- **feat(cerebras):** add the **Gemma 4 31B** model (`gemma-4-31b`) to the Cerebras registry + pricing table ([#6331](https://github.com/diegosouzapw/OmniRoute/pull/6331)). Regression guard extends `tests/unit/t28-model-catalog-updates.test.ts`. (thanks @backryun)
+- **feat(providers):** add **Yuanbao (web)** as a cookie-session provider ([#6196](https://github.com/diegosouzapw/OmniRoute/issues/6196)) — `yuanbao-web` (Tencent Yuanbao, `yuanbao.tencent.com`) with cookie-only auth (`hy_user`/`hy_token` + public agent id), SSE→OpenAI translation incl. `reasoning_content`, exposing DeepSeek V3/R1 + Hunyuan / Hunyuan-T1. Regression guard: `tests/unit/providers-yuanbao-web.test.ts`. `together-web` was **deferred** (no verifiable web-session endpoint — needs a captured request) and `huggingchat-web` **dropped** (the existing `huggingchat` already is a web-cookie provider). (thanks @chirag127)
+- **feat(providers):** route the built-in **agentrouter** through the dynamic Claude-Code wire image ([#6056](https://github.com/diegosouzapw/OmniRoute/issues/6056)) — a small static allow-set (`CC_WIRE_IMAGE_BUILTINS` in `open-sse/services/ccWireImageBuiltins.ts`), consulted by `isClaudeCodeCompatible` / `isClaudeCodeCompatibleProvider` / `applyFingerprint`, makes agentrouter adopt the CC wire-image headers + fingerprint **while guarding the CC baseUrl/auth branches** so it keeps its own registry `baseUrl` and `x-api-key` auth. Regression guard: `tests/unit/agentrouter-cc-wire-image.test.ts` (asserts the wire image is applied AND agentrouter's baseUrl/auth are preserved). Live WAF-acceptance against agentrouter.org is a VPS validation follow-up (Hard Rule #18).
+- **feat(providers):** **bulk-add API keys for Cloudflare Workers AI** ([#6174](https://github.com/diegosouzapw/OmniRoute/issues/6174)) — `cloudflare-ai` is removed from the bulk-add exclusion list and the bulk parser gains a 3-field `name|accountId|apiKey` mode; the bulk route now builds a **per-entry** `providerSpecificData` so each key carries its own `accountId` (fixing the previous shared-object reuse), and both the create + key-validation paths receive it. Regression guard: `tests/unit/bulk-api-key-parser-cloudflare.test.ts`. (thanks @muflifadla38)
+- **feat(dashboard):** routing/settings UX clarity ([#6147](https://github.com/diegosouzapw/OmniRoute/issues/6147)) — (1) weighted combos show the **effective routing share %** next to each weight when weights don't sum to 100 (`WeightTotalBar.tsx`); (2) the status widget's user-facing **"Cloud Sync" label is renamed** to "Remote Settings Sync" (`CloudSyncStatus.tsx`; internal ids/state untouched); (3) built-in providers gain an **opt-in advanced base-URL override** (`isBaseUrlOverrideEligibleProvider`, hidden behind an "Advanced" toggle, reusing the existing `providerSpecificData.baseUrl` persistence — not globally widened). Regression guard: `tests/unit/routing-settings-ux-6147.test.ts`.
+- **feat(combo):** add an option to **disable session stickiness**, per-combo or globally — round-robin / random combos can rotate to a different connection on every request instead of pinning a whole conversation to one connection by its first-message hash. Resolution precedence per-combo `config.disableSessionStickiness` → global `settings.disableSessionStickiness` → default `false` (preserves the #3825 prompt-cache/504 fix); gates **both** stickiness call sites in `open-sse/services/combo.ts`. Exposed as a global toggle (Combo Defaults) and a per-combo Inherit/on/off control. ([#6168](https://github.com/diegosouzapw/OmniRoute/issues/6168)) Regression guard: `tests/unit/combo-disable-session-stickiness.test.ts`. (thanks @RCrushMe)
+- **feat(docker):** add the `OMNIROUTE_NO_SUDO` env flag for root-less / user-namespaced deployments — the MITM cert-trust command path (`resolveSudoSpawn` in `src/mitm/systemCommands.ts`) now strips the leading `sudo` when the flag is truthy, in addition to the existing root / sudo-missing cases, so the Proxy Agent runs without `sudo` (the operator trusts the CA manually, e.g. via `NODE_EXTRA_CA_CERTS`). Argv-array `spawn` preserved — no shell interpolation (Hard Rule #13). ([#6122](https://github.com/diegosouzapw/OmniRoute/issues/6122)) Regression guard: `tests/unit/mitm-systemCommands-no-sudo.test.ts`. (thanks @powellnorma)
+- **feat(providers):** add **Requesty** as an OpenAI-compatible gateway provider (BYOK, base `https://router.requesty.ai/v1`, ~200 free requests/day) — wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/requesty/`, `src/shared/constants/providers/apikey/gateways.ts`). ([#6120](https://github.com/diegosouzapw/OmniRoute/issues/6120)) Regression guard: `tests/unit/requesty-provider.test.ts`. (thanks @chirag127)
+- **feat(dashboard):** add **configured-only / available-only filters** to the Free Provider Rankings page ([#6150](https://github.com/diegosouzapw/OmniRoute/issues/6150)) — hide providers you haven't configured, or whose connections are all rate-limited / out of quota, via server-side query params (`?configuredOnly` / `?availableOnly` on `GET /api/free-provider-rankings`) backed by a testable lib helper reusing the in-process connection state (no Redis). Both filters default off, so the default view is unchanged; this supersedes the earlier client-side "Configured Only" toggle (#6245) with an available-only dimension and unit-tested logic. Regression guard: `tests/unit/freeProviderRankings-filters.test.ts`.
+
+### 🔧 Bug Fixes
+
+- **fix(dashboard):** adding a second API key connection for the same provider no longer silently overwrites the first — the Add-API-key modal now derives a unique default connection name (`main`, then `main-2`, `main-3`, …) so the backend name-based upsert can't collide (#6499 — thanks @dilneiss).
+- **fix(compression):** the session-dedup engine now also deduplicates a large multi-line block repeated **within a single message** (intra-message dedup), not just across turns; the compression-preview API surfaces a `fallbackReason`, and the fusion panel reports how many models were rate-limited vs failed on total-panel failure (#6501 — thanks @chirag127).
+- **fix(compression):** a stacked-pipeline step naming an unregistered engine now surfaces a `validationErrors` entry instead of silently no-op'ing, so misconfigured pipelines are visible in the preview API (#6506 — thanks @chirag127).
+- **feat(usage):** add a **Codex reset-credit redemption** flow to the Provider Limits UI ([#6361](https://github.com/diegosouzapw/OmniRoute/pull/6361)) — a `useCodexResetCreditRedemption` hook + `/api/usage/codex-reset-credit` route + `codexResetCredits` lib let you redeem banked Codex reset credits from the quota card. Regression guard: `tests/unit/codex-reset-credits.test.ts`. (thanks @JxnLexn)
+- **feat(glm):** add **team-plan quota settings** for `glm-cn` connections ([#6351](https://github.com/diegosouzapw/OmniRoute/pull/6351)) — a dedicated `GlmTeamQuotaFields` form section (team quota id / limits) threaded through the Add/Edit connection modals, persisted via `providerSpecificData`, with the GLM usage service reading the team quota. Regression guards: `tests/unit/glm-team-quota.test.ts`, `provider-specific-data-schema.test.ts`. (thanks @hao3039032)
+- **feat(providers):** add **TinyFish** web-fetch/search support ([#6349](https://github.com/diegosouzapw/OmniRoute/pull/6349)) — a `tinyfish-fetch` executor + `/v1/web/fetch` route + MCP web-fetch tool, registered as a specialty-media provider with request-validation and a search-provider catalog entry. Regression guards: `tests/unit/executor-tinyfish-fetch.test.ts`, `web-fetch-handler.test.ts`, `mcp-web-fetch-tool.test.ts`, `provider-validation-tinyfish.test.ts`. (thanks @dtybnrj)
+- **fix(cli):** `omniroute launch-codex` now spawns `codex.cmd` through a shell on Windows (the npm `.cmd` shim is unresolvable by bare `spawn` → ENOENT), mirroring the qodercli Windows fix (#6263) ([#6312](https://github.com/diegosouzapw/OmniRoute/pull/6312)). Regression guard: `tests/unit/launch-codex-windows-spawn-6312.test.ts`. (thanks @swingtempo)
+- **fix(codex):** isolate the **Spark** quota from the shared Codex quota and stabilize the quota UI ordering / hydration so per-scope limits render consistently ([#6336](https://github.com/diegosouzapw/OmniRoute/pull/6336)). Regression guards: `tests/unit/codex-quota-selection-hydration.test.ts`, `provider-limits-ui.test.ts` + 3 more. (thanks @xz-dev)
+- **feat(api):** add a `hidePaidModels` setting that filters paid-only models out of the `/v1/models` catalog. Regression guard: `tests/unit/models-catalog-hide-paid.test.ts`. (thanks @chirag127)
+- **fix(api-manager):** the fallback model picker now preserves combos instead of dropping them when a primary model is unavailable ([#6443](https://github.com/diegosouzapw/OmniRoute/pull/6443)). Regression guard: `tests/unit/api-manager-page-static.test.ts`. (thanks @jmengit)
+- **fix(providers):** recoverable Antigravity / Cloud-Code (Gemini Code Assist) `403` responses ([#6452](https://github.com/diegosouzapw/OmniRoute/pull/6452)) are now classified as a retryable project-config error instead of a terminal account ban, so a fixable project/API-disabled 403 no longer forces a ~1-year cooldown / full OAuth reconnect. Regression guard: `tests/unit/errorclassifier-antigravity-403.test.ts`. (thanks @developerjillur)
+- **fix(mitm):** `sanitizeHeaders` now redacts `Set-Cookie` response headers so upstream session cookies never leak into logs / diagnostics ([#6451](https://github.com/diegosouzapw/OmniRoute/pull/6451)). Regression guard: `tests/unit/mitm-sanitize-headers.test.ts`. (thanks @developerjillur)
+- **fix(api):** `/api/compression/preview` now accepts `mode: "caveman"` and correctly handles stacked / zero-compression previews ([#6425](https://github.com/diegosouzapw/OmniRoute/issues/6425)). Regression guard: `tests/unit/api/compression-preview-caveman-and-stacked-6425.test.ts`. (thanks @chirag127)
+- **feat(providers):** add **Zed** hosted LLM aggregator as a native-app provider ([#6118](https://github.com/diegosouzapw/OmniRoute/pull/6118)) — OAuth sign-in via the Zed hosted flow, registered through the shared provider registry + executor. Regression guards: `tests/unit/zed-oauth-provider.test.ts`, `zed-import-utils.test.ts`, `zed-docker-detect.test.ts`, `mitm-handler-zed.test.ts`. VPS-validated via live operator login (Hard Rule #18).
+- **fix(oauth):** the Kiro SSO-cache auto-import now **preserves the IDC region** — cross-region Amazon Q / Kiro profiles imported from the SSO cache are no longer collapsed to the default region ([#6113](https://github.com/diegosouzapw/OmniRoute/pull/6113)). Regression guard: `tests/unit/kiro-auto-import-idc-2059.test.ts`. VPS-validated via live operator login (Hard Rule #18).
+- **fix(dashboard):** passthrough model aliases no longer collide when two namespaced model ids share a last segment (port from 9router#1850, [#6431](https://github.com/diegosouzapw/OmniRoute/pull/6431)). `enx/gpt-5.5` and `enx/codebuddy/gpt-5.5` both auto-generated the alias `gpt-5.5`, so the second model could never be added (the UI just alerted "alias already exists"). Aliases are now disambiguated deterministically — bare last segment when free, then parent-qualified (`codebuddy-gpt-5.5`), then a numeric suffix — while re-adding the exact same model id is still blocked. Regression guard: `tests/unit/passthrough-alias-1850.test.ts`. (thanks @arpicato)
+- **fix(translator):** preserve a Gemini `functionResponse` co-located with other parts (another `functionCall`, or trailing `text`) in the same content when translating **Gemini → OpenAI** ([#6376](https://github.com/diegosouzapw/OmniRoute/pull/6376)). `convertGeminiContent()` early-returned the tool message on the first `functionResponse` part, dropping any co-located parts; such contents are now pre-split (one tool message per `functionResponse`, emitted first, plus one message for the remaining parts). Regression guard: `tests/unit/gemini-to-openai-function-response.test.ts`. (thanks @warelik)
+- **fix(headroom):** detect a python interpreter managed by **mise / pyenv / asdf / conda** (port from 9router#2353, [#6382](https://github.com/diegosouzapw/OmniRoute/pull/6382)). Headroom's python probe (`src/lib/headroom/detect.ts`) searched a hardcoded `PATH`, but version managers expose their interpreters via shim dirs that only join `PATH` through interactive-shell activation — which the non-interactive server never runs, so a managed python (≥3.10) was invisible and Headroom reported it missing. The search path now prepends the well-known shim/bin dirs (`~/.local/share/mise/shims`, `~/.pyenv/shims`, `~/.asdf/shims`, `$CONDA_PREFIX/bin`, `~/.local/bin`, respecting `MISE_DATA_DIR`/`PYENV_ROOT`/`ASDF_DATA_DIR` when set), and a new `HEADROOM_PYTHON` env override lets operators point straight at their interpreter (mirroring `HEADROOM_URL`). Still shell-free (`execFileSync`). Regression guard: `tests/unit/headroom-detect.test.ts` (5). (thanks @loopyd)
+- **fix(executors):** strip the OpenAI-Codex/Claude-CLI `client_metadata` passthrough field for **NVIDIA** requests (port from 9router#1887, [#6411](https://github.com/diegosouzapw/OmniRoute/pull/6411)). NVIDIA's OpenAI-compatible wrapper rejects it with `400 Unsupported parameter`, the same class already handled for `cerebras`/`mistral`; `nvidia` (executor `default`) was missing from the strip allowlist so Codex/Claude-Code passthrough requests 400'd. Regression guard: `tests/unit/executor-default-strip-client-metadata.test.ts` (+nvidia case). (thanks @phidinhmanh)
+- **fix(translator):** strip the Claude-style `thinking` field for **NVIDIA `z-ai/glm-5.2`** (port from 9router#2023, [#6413](https://github.com/diegosouzapw/OmniRoute/pull/6413)). NVIDIA's OpenAI-compatible wrapper 400s on `thinking` (a Claude-format client routed here leaves a `thinking:{type:"adaptive"}`); the existing strip rule only dropped `reasoning`. Same class already handled for `minimax-m2.7`. Regression guard: `tests/unit/nvidia-minimax-thinking-strip.test.ts` (+glm-5.2 case). (thanks @phidinhmanh)
+- **fix(translator):** suppress the streamed `</think>` close marker for the **Antigravity IDE** client (port from 9router#1061, [#6415](https://github.com/diegosouzapw/OmniRoute/pull/6415)). On thinking-only turns Antigravity rendered a bare `</think>` as the sole visible content, tripping its loop-detection and wasting requests. Antigravity's UA (`vscode/<v> (Antigravity/<v>)`) is added to the marker-suppress allowlist (alongside OpenCode); Claude Code / Cursor still get the marker, and `x-omniroute-thinking-marker: on` force-restores it. Regression guard: `tests/unit/think-close-marker-suppress-5245.test.ts`. (thanks @abdofallah)
+- **fix(executors):** strip nested `reasoning_content` from messages for **Mistral** (port from 9router#1649, [#6417](https://github.com/diegosouzapw/OmniRoute/pull/6417)). Mistral's API returns `422 extra_forbidden` when an assistant message carries `reasoning_content` (replayed thinking from a prior turn, e.g. via the Codex `/responses` path); the generic top-level 400 field-downgrade retry never covered the nested per-message field. `DefaultExecutor` now strips it for provider `mistral` only, so DeepSeek (which requires replayed `reasoning_content`) is unaffected. Regression guard: `tests/unit/mistral-strip-reasoning-content-1649.test.ts`. (thanks @xxy9468615)
+- **fix(executors):** strip the `client_metadata` passthrough field on the **OpenCode** path (port from 9router#1442, [#6418](https://github.com/diegosouzapw/OmniRoute/pull/6418)). OpenCode upstreams (e.g. `kimi-k2.6` via opencode-go) reject it with `400 "Extra inputs are not permitted, field: 'client_metadata'"`; the DefaultExecutor strip only covered cerebras/mistral and `OpencodeExecutor` extends `BaseExecutor` directly, so nothing removed it there. Regression guard: `tests/unit/opencode-strip-client-metadata-1442.test.ts`. (thanks @yanpaing007)
+- **fix(executors):** inject the `reasoning_content` echo for the native **Moonshot Kimi** provider (port from 9router#1480, [#6419](https://github.com/diegosouzapw/OmniRoute/pull/6419)). Kimi (executor `default`) is a thinking-mode upstream that 400s with "reasoning_content must be passed back" when a prior assistant turn lacks it; the placeholder injection was only wired into the OpenCode meta-provider, so direct multi-turn Kimi conversations failed. Scoped to `kimi` (gateway-served models matching the thinking-model name pattern are unaffected). Regression guard: `tests/unit/kimi-native-reasoning-injected-1480.test.ts`. (thanks @2220258345)
+- **fix(executors):** recover from a strict gateway's `context_management: Extra inputs are not permitted` 400 (port from 9router#1468, [#6420](https://github.com/diegosouzapw/OmniRoute/pull/6420)). **Claude Code** always sends a top-level `context_management` field; strict anthropic-compatible gateways reject it. The dedicated context-editing 400-fallback only fired when OmniRoute's own `contextEditing` feature was enabled (default off), so a client-sent field passed through untouched and 400'd. `context_management` is now in the generic reactive field-strip list, so it's stripped-and-retried once regardless of the feature flag (with correct request re-signing for claude-compatible relays). Regression guard: `tests/unit/provider-field-strips.test.ts`. (thanks @ohahe52-dot)
+- **fix(network):** enable **RFC 8305 Happy Eyeballs** (`autoSelectFamily`) on the direct-egress undici dispatcher (port from 9router#1237, [#6423](https://github.com/diegosouzapw/OmniRoute/pull/6423)). When DNS returns both IPv6 (AAAA) and IPv4 (A) and the IPv6 route is broken (e.g. a NAT64 `64:ff9b::` prefix without routing), undici tried IPv6 first and hung until `ETIMEDOUT` (then a 502 + account lockout), even though `curl` reached the same host. The direct dispatcher now races both families and uses whichever connects first. Proxy paths pin family via `proxyTls` and are unaffected. Regression guard: `tests/unit/direct-dispatcher-pipelining-4580.test.ts`. (thanks @adentdk)
+- **fix(combo):** round-robin now advances the rotation pointer past the model that **actually served**, not the eagerly-scheduled one (port from 9router#948, [#6428](https://github.com/diegosouzapw/OmniRoute/pull/6428)). With `stickyLimit: 1` (true round-robin), when the scheduled model failed and a _different_ model served via fallback, the counter had already advanced +1 from the scheduled index — so the next request reused the fallback-served model, degrading round-robin into hot-spotting on whichever model was healthy. The pointer now advances to the served index + 1 (mirroring the sticky-limit>1 path). Session-stickiness (#3825) and distribution are preserved. Regression guard: `tests/unit/combo-rr-fallback-advance-948.test.ts`. (thanks @binsarjr)
+- **fix(sse):** a non-string `model` field is now rejected with a `400` before the resolver, instead of crashing downstream `.toLowerCase()`/`.split()` calls into an empty-body `500` that escapes the error sanitizer ([#6407](https://github.com/diegosouzapw/OmniRoute/issues/6407)). Regression guard: `tests/unit/chat-non-string-model-6407.test.ts`. (thanks @chirag127)
+- **fix(api):** unknown `/api/*` routes now return a JSON `404` (instead of the dashboard HTML shell) and scalar chat params (`model`/`temperature`/etc.) are validated **before** the provider lookup so malformed requests fail fast with a clear `400` ([#6424](https://github.com/diegosouzapw/OmniRoute/issues/6424), [#6412](https://github.com/diegosouzapw/OmniRoute/issues/6412)). Regression guards: `tests/unit/api/api-catchall-json-404.test.ts`, `tests/unit/chat-early-schema-validation-6412.test.ts`. (thanks @chirag127)
+- **fix(api):** `/v1/chat/completions` now rejects a non-JSON `Content-Type` with a `400` before parsing the body ([#6414](https://github.com/diegosouzapw/OmniRoute/issues/6414)). Regression guard: `tests/unit/v1-chat-completions-content-type-6414.test.ts`. (thanks @chirag127)
+- **fix(api):** the `X-OmniRoute-Compression` response header is now echoed on `/v1/chat/completions` and `/v1/completions` ([#6422](https://github.com/diegosouzapw/OmniRoute/issues/6422)). Regression guard: `tests/unit/compression-header-echo-6422.test.ts`. (thanks @chirag127)
+- **fix(api):** concurrent `GET /v1/models` requests are coalesced into a single catalog build ([#6408](https://github.com/diegosouzapw/OmniRoute/issues/6408)). Regression guard: `tests/unit/v1-models-concurrent-6408.test.ts`. (thanks @chirag127)
+- **fix(api):** `/v1/completions` now echoes the requested `body.model` in its JSON + streamed responses ([#6429](https://github.com/diegosouzapw/OmniRoute/pull/6429)). Regression guard: `tests/unit/completions-body-model-echo.test.ts`. (thanks @chirag127)
+- **fix(api):** env-var master keys now see the full `/v1/models` catalog ([#6406](https://github.com/diegosouzapw/OmniRoute/issues/6406)). Regression guard: `tests/unit/models-catalog-envkey-6406.test.ts`. (thanks @chirag127)
+- **fix(api):** non-streaming `/v1/completions` responses now echo `body.model` aligned with the `X-OmniRoute-Model` header ([#6426](https://github.com/diegosouzapw/OmniRoute/issues/6426)). Regression guard: `tests/unit/v1-completions-model-header-match-6426.test.ts`. (thanks @chirag127)
+- **fix(api):** unknown `/v1/*` routes now return a JSON `404 not_found` instead of the Next.js dashboard HTML shell ([#6405](https://github.com/diegosouzapw/OmniRoute/issues/6405)). Regression guard: `tests/unit/api/v1-catchall-json-404.test.ts`. (thanks @chirag127)
+- **fix(api):** the per-connection provider models route now degrades to the shipped catalog when a provider's `/models` endpoint answers with a redirect ([#6267](https://github.com/diegosouzapw/OmniRoute/issues/6267)) — a `qwen-web` import failed with a raw `Redirect blocked … (307)` 503. `safeOutboundFetch` throws `REDIRECT_BLOCKED` on the 307, `getSafeOutboundFetchErrorStatus` maps it to 503, and `buildDiscoveryErrorFallbackResponse` treated every 503 as a hard error — so the non-empty `getModelsByProviderId("qwen-web")` catalog was never surfaced. A models-endpoint redirect is not a fixable-config error (unlike `URL_GUARD_BLOCKED`/`INVALID_URL`, which stay hard errors), so it now falls back to the local/cached catalog before the 503 short-circuit. General fix — covers any config-driven provider that 307s. Regression guard: `tests/unit/provider-models-qwen-web-redirect-6267.test.ts`. (thanks @chirag127)
+- **fix(api):** the per-connection provider models route (MCP `list_models_catalog` + the dashboard import view) now merges USER-ADDED custom models into its response ([#6247](https://github.com/diegosouzapw/OmniRoute/issues/6247)) — custom models live in the `key_value` namespace `customModels`, which the live REST `/api/v1/models` already merges, but `src/app/api/providers/[id]/models/route.ts` never read `getCustomModels`, so custom models were dropped on both the discovery-success and local_catalog paths. They are now folded into the returned model list (deduped by id, stamped `owned_by: provider`), fixing MCP + the dashboard import view in one place. Regression guard: `tests/unit/provider-models-custom-merge-6247.test.ts`. (thanks @RCrushMe)
+
+- **fix(providers):** GitLab Duo tool-calling follow-up turns no longer fail upstream with `422 {"detail":"Validation error"}` (tokens 0/0, rejected pre-inference). The #6234 tool-result-feedback fix serialized the **entire** multi-turn conversation into GitLab's single-file `code_suggestions` (`small_file`) generation endpoint — folded history that turn-N sent as an oversized `current_file.content_above_cursor` **and** duplicated verbatim into `user_instruction`, tripping the AI-Gateway's `small_file` validation guard. The executor now **bounds** that prompt: it keeps system + latest user message + the most-recent tool round (dropping older turns), caps oversized tool results, and stops duplicating the full prompt into `user_instruction` (which now carries only the short latest user message) — while still feeding the most-recent tool result back so the agent continues (`open-sse/executors/gitlab.ts`, [#6220](https://github.com/diegosouzapw/OmniRoute/issues/6220)). The unit test covers the bounding logic; the upstream 422→200 clearing is VPS-only (Hard Rule #18). Regression guard: `tests/unit/gitlab-tool-exchange-bounded-6220.test.ts`.
+
+- **fix(i18n):** the provider-detail (`/dashboard/providers/[id]`) connection-status filter labels no longer render as `__MISSING__:All` / `__MISSING__:Active` / `__MISSING__:Error` / `__MISSING__:Banned` / `__MISSING__:CreditsExhausted` in non-English locales (notably pt-BR) ([#6290](https://github.com/diegosouzapw/OmniRoute/issues/6290)). Root cause was **not** the namespace mismatch the issue guessed — the `providers.filter*` keys resolve correctly in `en.json`; the debt lived in the locale mirrors (`src/i18n/messages/*.json`), where these five keys carried the `__MISSING__:` sync sentinel in ~15 locales and were absent entirely in ~26 others, so next-intl found the key and echoed the sentinel verbatim. All 40 non-English/-Chinese mirrors now ship real translations for the five `providers.filter*` labels. Regression guard: `tests/unit/i18n-provider-filter-keys-6290.test.ts`. (thanks @diegosouzapw)
+
+- **fix(providers):** the `copilot-m365-web` streaming executor now emits `debug`-level WebSocket diagnostics ([#6210](https://github.com/diegosouzapw/OmniRoute/issues/6210)) — the outbound WS URL (with the `access_token` **redacted** via `redactWsUrl()`), handshake success/failure, and each received SignalR frame's `type`/`target`. Previously the streaming path logged nothing, so an empty `content:null` response (the M365 Education / Starter tier symptom fixed in #6234) was undiagnosable even at `APP_LOG_LEVEL=debug`. The change is debug-level and side-effect-free — it does not alter streaming behavior or the frame parser, and the token never reaches the logs. Regression guard: `tests/unit/copilot-m365-web-logging-6210.test.ts` (thanks @qpeyba)
+
+- **fix(resilience):** a round-robin combo no longer returns `503 all upstream accounts are unavailable` when a compatibility-rejected target is actually healthy ([#6238](https://github.com/diegosouzapw/OmniRoute/issues/6238)). `filterTargetsByRequestCompatibility` drops request-incompatible targets (tool/vision/structured-output unsupported, or below the required context window) **before** any availability check runs, and its `compatible.length === 0` safety net only fired when _all_ targets were filtered — not when the kept targets later all turned out runtime-unavailable (circuit-open / cooldown / no credentials). So a combo could 503 while a compat-rejected-but-healthy provider sat unused. `handleRoundRobinCombo` now keeps the compat-rejected set and, when every compat-kept target was skipped without a single real attempt, probes those rejected targets as a **last-resort fallback tier** (via the new pure `open-sse/services/combo/comboCompatFallback.ts`) before crystallizing the 503. Regression guard: `tests/unit/combo-roundrobin-compat-fallback-6238.test.ts`. (thanks @ThongAccount)
+
+- **fix(startup):** best-effort self-heal for a corrupted Turbopack dev cache on Windows ([#6289](https://github.com/diegosouzapw/OmniRoute/issues/6289)). On Windows, `pnpm dev` can fail at startup when Turbopack `mmap`s a persistent-cache SST file and the OS refuses the mapping (`os error 1455` — "paging file too small"), which Turbopack surfaces as a misleading `Module not found: Can't resolve '@/shared/utils/machine'`. This is a **known upstream Turbopack cache-corruption bug — not our code**. The dev launcher (`scripts/dev/run-next.mjs`) now wraps `nextApp.prepare()` and, when it rejects with that signature (`isTurbopackCacheCorruption` in the new `scripts/dev/turbopackCacheHeal.mjs`), purges `.build/next/**/cache/turbopack` and retries **once** with a clear log. **Caveat — best-effort only:** the corruption often surfaces as a runtime overlay rather than a `prepare()` rejection, so this cannot always intercept it; the reliable remedy remains manually deleting the Turbopack cache dir. Regression guard: `tests/unit/turbopack-cache-heal-6289.test.ts`. (thanks @chirag127)
+
+- **fix(providers):** qodercli PAT auth no longer fails with `spawn qodercli ENOENT` on Windows ([#6263](https://github.com/diegosouzapw/OmniRoute/issues/6263)) — `spawnQoderCli` spawned the bare `qodercli` name with `shell:false` and an unenriched env, so the npm `.cmd` wrapper under `%APPDATA%\npm` (a user-PATH directory) was never resolved. It now resolves the absolute `.cmd`/`.exe` path through the existing `getCliRuntimeStatus("qoder")` resolver in `src/shared/services/cliRuntime.ts` (memoized), spawns with `shell` when the target is a `.cmd`/`.bat`, and uses the cliRuntime-enriched env (PATH + PATHEXT + APPDATA); the ENOENT error now lists the searched paths plus the `CLI_QODER_BIN` override. End-to-end spawn on a real Windows host is host-only (Hard Rule #18); the path-resolution logic is unit-tested. Regression guard: `tests/unit/qodercli-windows-resolve-6263.test.ts`. (thanks @chirag127)
+
+- **fix(sse):** the reasoning-token buffer no longer inflates **probe-sized `max_tokens`** ([#6274](https://github.com/diegosouzapw/OmniRoute/issues/6274)) — Claude Code's `/model` capability check sends `max_tokens: 1`, but for a thinking-capable model with a large output cap (e.g. `glm-5.2`) the #3587 headroom heuristic (`max(current + 1000, ceil(current * 1.5))`) rewrote it to `1001` and forwarded that upstream, wasting tokens on a request that was never a genuine reasoning budget. `resolveReasoningBufferedMaxTokens()` (`open-sse/services/reasoningTokenBuffer.ts`) now short-circuits and returns the caller's value verbatim when it is below the new `REASONING_BUFFER_MIN_TRIGGER` (256) threshold — a tiny explicit limit is a probe, not a reasoning request. Real budgets still receive the #3587 headroom unchanged, and the guard runs after the existing capability checks so unknown / non-reasoning models keep returning `null`. Regression guard: `tests/unit/reasoning-token-buffer-6274.test.ts`. (thanks @brightfiscalband)
+
+- **fix(cli):** `omniroute reset-password` now works as a real subcommand, and password resets over piped (non-TTY) stdin actually apply ([#6261](https://github.com/diegosouzapw/OmniRoute/issues/6261), [#6258](https://github.com/diegosouzapw/OmniRoute/issues/6258)). Two coupled defects: (1) **#6261** — `bin/omniroute.mjs` routed everything through Commander with only two pre-Commander bypasses (`--mcp`, `reset-encrypted-columns`), so `omniroute reset-password` was rejected as an unknown command; only the separate `omniroute-reset-password` bin worked, while the docs falsely advertised the subcommand (incl. a bogus "legacy alias still works"). A pre-Commander bypass mirroring `reset-encrypted-columns` now dynamically imports `bin/reset-password.mjs` (which self-executes) before Commander parses; the three doc lines were corrected. (2) **#6258** — `bin/reset-password.mjs` issued two sequential `rl.question` prompts; under piped stdin the second read never settled at EOF, so `main()` never reached `resetManagementPassword` and the reset was a silent no-op (both prompts printed, no success, password unchanged). The CLI now detects non-TTY stdin and reads it once (first line = password, second line = confirm if present, else reused), adds a `--password-stdin` flag (entire stdin is the password, no confirmation), and exits `0` explicitly so the success line always flushes; interactive TTY behavior is unchanged. Regression guard: `tests/unit/reset-password-cli-6261-6258.test.ts` (3). (thanks @chirag127)
+
+- **fix(db):** the mass-migration **safety abort** now tells the operator how to bypass it and stops flooding the log ([#6260](https://github.com/diegosouzapw/OmniRoute/issues/6260)) — after restoring a backup that wiped the migration tracking table, `runMigrations()` threw the abort on every downstream `ensureDbInitialized()`, re-logging the full banner 11+ times, and the message never mentioned the existing `OMNIROUTE_MAX_PENDING_MIGRATIONS` escape hatch. The abort text now appends a bypass hint (set `OMNIROUTE_MAX_PENDING_MIGRATIONS=0` in `server.env` / `DATA_DIR/.env`), and a new `MigrationSafetyAbortError` is memoized so repeated calls in the same process throw the same instance and emit a single concise line instead of the full cascade. Regression guard: `tests/unit/migration-safety-abort-6260.test.ts`. (thanks @chirag127)
+
+- **fix(auth):** importing a **distinct** Codex/ChatGPT OAuth `auth.json` is no longer falsely rejected as "already exists" when it belongs to a different user in the same workspace ([#6301](https://github.com/diegosouzapw/OmniRoute/issues/6301)). `findExistingCodexConnection` (in `src/lib/oauth/utils/codexAuthImport.ts`) deduped **only** on `providerSpecificData.workspaceId === accountId`, where `accountId` is the shared `chatgpt_account_id`/`tokens.account_id` — so two members of the same ChatGPT Team collapsed onto a single connection (409 `duplicate_account`). The id_token's `https://api.openai.com/auth` claim carries a per-user `chatgpt_user_id` alongside the workspace id (the device-flow path already persisted it as `chatgptUserId`, but the import path did not). Now `parseAndValidateCodexAuth` extracts `userId` (`chatgpt_user_id` → `user_id` → JWT `sub`) into `ParsedCodexAuth`, the create/update paths persist `chatgptUserId` in `providerSpecificData` (mirroring `codex.ts`), and dedup keys on `workspaceId` **AND** `chatgptUserId` — with a backward-compat fallback to legacy accountId-only matching when no stored connection for that workspace records a `chatgptUserId`, so genuinely-same accounts still dedup. Regression guard: `tests/unit/codex-auth-import-userid-dedup-6301.test.ts` (4). (thanks @anungma)
+
+- **fix(providers):** importing models for the **venice-web** provider no longer fails with a red "Provider venice-web does not support models listing" ([#6269](https://github.com/diegosouzapw/OmniRoute/issues/6269)). `venice-web` is a web-cookie provider with an executor but no upstream `/v1/models` endpoint and no registry `models`, so the models route fell through to the tail `400`. Mirroring the `jules`/`linkup-search`/`ollama-search` fix (#5569), it now ships a static local catalog entry in `src/lib/providers/staticModels.ts` — seeding the current Venice lineup (`venice-uncensored`, `llama-3.3-70b`, `qwen3-235b`, `qwen3-4b`, `deepseek-r1-671b`; Venice rotates its catalog, see docs.venice.ai/models/overview) — so the route returns `200` with `source:"local_catalog"`, `intentional:true`. Regression guard: `tests/unit/static-models-venice-web-6269.test.ts`. (thanks @chirag127)
+
+- **fix(api):** the specialty model catalogs (`/v1/embeddings`, `/v1/images`, `/v1/music`, `/v1/videos` model lists) are now derived from the **unified catalog filtered by a predicate** (`getSpecialtyModelsResponse`) instead of ad-hoc per-route logic, so they consistently respect active-credential visibility and stay in sync with the main catalog ([#6303](https://github.com/diegosouzapw/OmniRoute/pull/6303)). Regression guard: `tests/unit/specialty-model-catalog-routes.test.ts`. (thanks @makcimbx)
+- **fix(api):** the agent-bridge server route now resolves the MITM manager via a **dynamic `import("@/mitm/manager.runtime")`** so Turbopack does not statically pull the stub (or over-bundle the manager), and the agent-skills generator anchors its output base path with `path.join(process.cwd(), …)` so Turbopack's static analyzer stops tracing the whole project root ([#6329](https://github.com/diegosouzapw/OmniRoute/issues/6329), [#6366](https://github.com/diegosouzapw/OmniRoute/pull/6366)). Regression guard: `tests/unit/agent-bridge-server-route-dynamic-import.test.ts`. (thanks @Iammilansoni)
+- **fix(api):** internal probes (combo-test, cloud-sync verify) now pick a **management-scoped / allow-all API key** instead of naively grabbing `getApiKeys()[0]` — a restricted `self:usage` first row made the probe fail with "Model X is not allowed for this API key" even when the combo path was healthy (`pickApiKeyForInternalUse` in `src/lib/db/apiKeys.ts`). The API-manager model editor also falls back to `/api/models?all=true` when `/v1/models` is catalog-protected ([#6372](https://github.com/diegosouzapw/OmniRoute/pull/6372)). Regression guard: `tests/unit/pick-internal-api-key-6372.test.ts`. (thanks @jmengit)
+- **fix(live-ws):** the Live Dashboard WebSocket server now **rejects on bind failure** (e.g. `EADDRINUSE` when the API bridge already holds the port) instead of letting the error surface as an unhandled `error` event that crash-loops the process — the `error` listener is attached to `wss` (not `server`) and releases the EventBus subscription on a failed start ([#6324](https://github.com/diegosouzapw/OmniRoute/issues/6324)). Regression guard: `tests/unit/live-ws-eaddrinuse-6324.test.ts`. (thanks @vinayakkulkarni)
+- **fix(dashboard):** the Home provider-topology widget now trusts the live provider-metrics snapshot — it uses `topology.errorProvider` and live `activeRequests` directly instead of re-deriving state from a stale `lastErrorAt` or applying a frontend timeout filter, so the topology reflects real-time provider health ([#6322](https://github.com/diegosouzapw/OmniRoute/pull/6322)). Regression guard: `tests/unit/home-provider-topology-live-state.test.ts`. (thanks @xz-dev)
+- **fix(sse):** strip zero-width markers from streamed **tool-call arguments** — a follow-up to [#5857](https://github.com/diegosouzapw/OmniRoute/pull/5857). That PR removed injected zero-width joiners (U+200D) from streamed assistant text/reasoning but deliberately left tool-call argument JSON byte-exact. The request-side obfuscation (`open-sse/services/claudeCodeObfuscation.ts`) injects ZWJ into agent words — including the temp path inside the Bash tool description — and Claude models copy that verbatim into generated commands, which are delivered as tool-call arguments rather than assistant text. As a result the ZWJ survived and corrupted code blocks (e.g. a temp path rendered with an invisible joiner). Now `open-sse/handlers/responseSanitizer.ts` strips zero-width code points from tool-call argument strings at every emit site (OpenAI non-stream/stream chat `tool_calls` + legacy `function_call`, native Responses `function_call` items, the OpenAI→Responses conversion, and the native Responses streaming `response.function_call_arguments.delta/.done` events). Only zero-width code points are removed; JSON structure and all other bytes stay identical (no parse/restringify), so normal arguments remain byte-exact. Regression guard: 6 new cases in `tests/unit/response-sanitizer.test.ts` (suite 50/50).
+
+- **fix(nodejs):** the default app log path now resolves under `DATA_DIR` (`~/.omniroute/logs/application/app.log`) instead of `process.cwd()` ([#6197](https://github.com/diegosouzapw/OmniRoute/issues/6197)) — the globally-installed CLI runs from an arbitrary working directory, so anchoring the default to cwd made file logging silently write to (or no-op under) an unrelated directory, contradicting the documented `.env.example` default. `getAppLogFilePath()` now computes the default lazily via the pure `resolveDataDir()` resolver (honours a per-process `DATA_DIR`, no directory-creation side effect); an explicit `APP_LOG_FILE_PATH` still wins. Regression guard: `tests/unit/logenv-datadir-path-6197.test.ts` (3).
+
+- **fix(docker):** AgentBridge/`startMitm` no longer aborts in containers/headless when the Antigravity-default DNS step can't write `/etc/hosts` ([#6127](https://github.com/diegosouzapw/OmniRoute/issues/6127)), and the privileged command's stderr now reaches `app.log` instead of only a bare exit code hitting the toast ([#6198](https://github.com/diegosouzapw/OmniRoute/issues/6198)). The default DNS step (`addDNSEntry`) was called unguarded while cert install and the two sibling DNS steps were each best-effort — in the runtime Docker image (`USER node`, no `sudo`, read-only `/etc/hosts`) it threw `Command failed with code 1` out of `startMitmInternal` and killed the whole start, discarding the stderr. The three DNS steps are extracted into a best-effort `provisionDnsEntries()` where each failure is logged with the full `err` (stderr included, folded in by `systemCommands.ts`) and never aborts the start. Regression guard: `tests/unit/mitm-dns-graceful-degrade-6127.test.ts` (4).
+
+- **fix(providers):** copilot-m365-web now supports the M365 Education "Starter / OfficeWebIncludedCopilot" tier and no longer returns an empty `content:null` stream ([#6210](https://github.com/diegosouzapw/OmniRoute/issues/6210)). Two gaps: (1) `buildWsUrl()` hardcoded the individual-consumer scenario (`OfficeWebPaidConsumerCopilot`, `isEdu=false`) — the EDU tier is now opt-in via `providerSpecificData.tier="edu"`, emitting `scenario=OfficeWebIncludedCopilot`/`isEdu=true` (the individual path is unchanged); (2) the EDU/GPT-5.5 path streams deltas via `arguments[0].writeAtCursor` (incremental) instead of only `messages[].text` (accumulated snapshots), which the parser dropped — a new `accumulateBotContent()` folds both formats, with `type:2 item.result.message` as a last-resort fallback. Regression guard: `tests/unit/copilot-m365-edu-writeatcursor-6210.test.ts` (10). (thanks @qpeyba)
+
+- **fix(providers):** GitLab Duo executor now feeds tool results back into the prompt instead of looping ([#6220](https://github.com/diegosouzapw/OmniRoute/issues/6220)) — `buildPrompt()` branched only on `system`/`user` and took `userParts.at(-1)`, silently dropping the `assistant{tool_calls}` + `tool{result}` turns the client appended, so the reconstructed prompt was byte-identical to turn 1 and the model re-emitted the same `<tool>` call forever. When a tool exchange is present the full conversation is now serialized, folding each tool result back keyed by its `tool_call_id`; simple conversations keep the legacy shape. Complements the tool_call emission from [#6051](https://github.com/diegosouzapw/OmniRoute/issues/6051) (the `kilo-duplicate` label was a false positive — different, sequential defect). Regression guard: `tests/unit/gitlab-tool-result-feedback-6220.test.ts` (4).
+
+- **fix(providers):** opencode-go/opencode-zen can now synthesize the OpenCode CLI identity headers Cloudflare requires on VPS egress ([#5997](https://github.com/diegosouzapw/OmniRoute/issues/5997)) — on a datacenter VPS, `opencode.ai/zen/go/v1/chat/completions` 403s (HTML challenge) requests lacking CLI identity, while the reporter's control curl proved that `User-Agent: opencode-cli/1.0.0` + `x-opencode-client: cli` + `x-opencode-project: default` + fresh request/session UUIDs succeed. Opt-in via `OPENCODE_SYNTHESIZE_CLI_HEADERS=true` (values overridable via `OPENCODE_GO_USER_AGENT`/`OPENCODE_USER_AGENT`/`OPENCODE_CLIENT`/`OPENCODE_PROJECT`); it fills only headers the client did not already send. Kept **off by default** — the forward-only path is deliberate (fabricating a wrong value risks upstream rejection; a prior dedup regressed with `opencode/local`), so this replaces the fragile local header-injection shim without changing default behavior. Regression guard: `tests/unit/opencode-cli-headers-synthesis-5997.test.ts` (6). (thanks @aleksesipenko)
+- fix(resilience): sticky session affinity now evicts and fails over to another account when the pinned account is exhausted/unavailable (#6219)
+
+- fix(sse): Responses API passthrough now drops internal commentary-phase output before forwarding to clients (gated by RESPONSES_PASSTHROUGH_DROP_COMMENTARY, default on) (#6199)
+
+- **fix(sse):** tool-call function schemas with a root `type: null` are now coerced to `type: "object"` before dispatch ([#6359](https://github.com/diegosouzapw/OmniRoute/issues/6359)) — clients like the Codex app emit `parameters: { type: null, ... }` for some tools, which OpenAI-compatible upstreams reject with `400 Invalid schema for function '...': schema must be a JSON Schema of 'type: "object"', got 'type: null'`, failing the whole request. `toolSchemaSanitizer` already stripped the null; it now re-adds the mandatory root `"object"` type (and empty `properties`/open `additionalProperties` when absent). Combinator roots (`anyOf`/`oneOf`/`allOf`) and explicit root types are left untouched. Regression guard: 5 new cases in `tests/unit/tool-schema-sanitizer.test.mjs`.
+
+- **fix(docker):** AgentBridge no longer fails to start on npm/Electron/VPS installs with "MITM manager stub reached at runtime" ([#6344](https://github.com/diegosouzapw/OmniRoute/issues/6344)) — v3.8.45 flipped the production bundler default to Turbopack, but `next.config.mjs` aliased `@/mitm/manager` to its Docker-only degraded stub **unconditionally**. That was harmless while Docker (which sets the alias intentionally for #3390 graceful degradation) was the sole Turbopack consumer, but once every artifact built with Turbopack the stub shipped to all non-Docker users and `startMitm` threw on the first Agent-Bridge start. The alias is now opt-in via `OMNIROUTE_MITM_STUB=1` (set only by the Dockerfile) through the shared `scripts/build/mitm-stub-flag.mjs` helper; default builds bundle the real manager. Regression guard: `tests/unit/mitm-stub-alias-6344.test.mjs` (4).
+
+- **fix(proxy):** stop the v3.8.44 proxy regression that leaked the real IP and disabled healthy proxies ([#6246](https://github.com/diegosouzapw/OmniRoute/issues/6246)). Two coupled defects from the new health scheduler: (1) **IP leak** — when a proxy assigned to a connection was marked `inactive`, resolution fell through to a **direct** egress instead of blocking, exposing the operator's real IP; (2) **over-deactivation** — the sweep flipped a proxy to `inactive` on the **first** failed probe and counted our own 5s timeout / a probe-target `5xx` as the proxy's fault, so healthy paid proxies vanished from egress selection ("my proxies are not being used anymore"). Fix: the sweep decision is extracted into a pure, network-free `decideProxyHealthAction` (`src/lib/proxyHealth/decision.ts`) — by default the health check now **only counts/logs and never downgrades status** (a proxy is downgraded/removed only with `PROXY_AUTO_REMOVE=true`, after `PROXY_AUTO_REMOVE_AFTER` **consecutive** conclusive failures); probes are classified tri-state so an inconclusive result (our timeout, or a `5xx` from the probe target) never penalizes the proxy, and the probe timeout is raised 5s→15s. Separately, `safeResolveProxy` now **fails closed** via the existing policy: a connection whose assigned proxy is dead is blocked instead of leaking direct (`hasBlockingProxyAssignment`), honoring the explicit `proxy off` toggles and the `PROXY_FAIL_OPEN=true` opt-out. Existing proxies stuck `inactive` by the old behavior need a one-time manual re-activate (the operator owns proxy status). Regression guards: `tests/unit/proxy-health-decide-action-6246.test.ts`, `tests/unit/proxy-assigned-unavailable-6246.test.ts`.
+
+- **fix(proxy):** make "Test All" read-only and add bulk enable/disable ([#6246](https://github.com/diegosouzapw/OmniRoute/issues/6246)). Complements the core fail-closed / scheduler fix (#6296) with the two remaining reporter asks. (1) The **"Test All" button** (`POST /api/settings/proxies/auto-test`) used to flip a proxy to `inactive` on a failed reachability probe; since the egress selector excludes `inactive` proxies, a flaky probe (an unreachable `httpbin.org`, a proxy that blocks `HEAD`, or a slow paid proxy) silently disabled every proxy that failed — "Test All" is now **read-only by default** (only the operator sets a proxy active/inactive; opt back into the legacy test-and-set with `PROXY_HEALTH_AUTO_DEACTIVATE=true`). (2) Adds a **bulk enable/disable** proxies endpoint + toolbar action (`POST /api/settings/proxies/batch-activate`) so an operator can re-activate proxies in one click. Regression guard: `tests/unit/proxy-health-6246.test.ts`. (thanks @tenshiak)
+
+- **chatcore (tools): stop the default 128-tool cap from silently dropping opencode's `task`/MCP tools.** opencode (used as an MCP/agent host) sends a large tool list; when it exceeds the speculative `MAX_TOOLS_LIMIT` (128) default, `truncateToolList` did a blind `tools.slice(0, 128)`, dropping every tool past index 128 — including opencode's built-in `task` tool (subagent launch) and many MCP tools, so models routed through OmniRoute could no longer spawn subagents or reach part of their tools. The cap exists to avoid upstream `400`s for providers with real hard limits (e.g. grok-cli 200), so it is kept for those: detection of the opencode client (`isOpencodeClient` — any `x-opencode-*` header, or `opencode` in the user-agent) now only bypasses the **speculative 128 default**, never a known provider ceiling. Precedence is explicit — a proactive/detected provider limit always truncates (even for opencode); otherwise opencode forwards its full tool list; otherwise the unchanged 128 default applies to every other client. Refactors `getEffectiveToolLimit` into `getKnownToolLimit(provider) ?? DEFAULT_LIMIT` (byte-identical for existing callers) and fixes a cosmetic debug-log that reported the truncated count instead of the original. Regression guard: `tests/unit/tool-limit-detector.test.ts`.
+
+- **fix(mitm):** the macOS MITM-cert install check now matches the system keychain again. `security find-certificate -a -Z` prints the SHA-1 as a colon-less hex string, but the installed-check compared it against `getCertFingerprint()`'s colon-separated form, so the substring match never hit — the cert was reported as not-installed and re-prompted for the sudo install on every run. Fingerprints are now normalized (colons stripped, upper-cased) on both sides via the extracted `macCertOutputHasFingerprint` helper. Regression guard: `tests/unit/mitm-cert-mac-fingerprint.test.ts`. ([#6204](https://github.com/diegosouzapw/OmniRoute/pull/6204), closes [#6134](https://github.com/diegosouzapw/OmniRoute/issues/6134) — thanks @rianonehub)
+
+- **fix(api):** `/v1/messages/count_tokens` now counts `tool_use`, `tool_result` and `thinking` content blocks (and array-form `system` prompts) in the local-estimation path, instead of only `text`. Real agentic conversations keep ~95% of their tokens inside tool results; the previous estimate returned near-zero for them, which silently broke Claude Code's auto-compaction (context grew past the window with no compaction until the upstream API rejected the request). The real provider-side count path is unchanged. Regression guard: `tests/unit/messages-count-tokens-route.test.ts`. ([#6221](https://github.com/diegosouzapw/OmniRoute/pull/6221) — thanks @luweiCN)
+
+- **fix(antigravity):** strip a trailing assistant prefill turn for Vertex Claude models to avoid upstream 400s ([#6114](https://github.com/diegosouzapw/OmniRoute/pull/6114)). Regression guard: `tests/unit/antigravity-claude-prefill-strip.test.ts`. (thanks @anki1kr)
+
+- **fix(security):** the mutable cloud-agent routes (`/api/cloud/credentials/update`, `/api/cloud/models/alias`) now require management auth instead of being treated as public. They were classified as public API routes, so a request without management credentials could update stored cloud-agent credentials and model aliases. They are removed from the public-route set, classified as management routes in the authz pipeline, and gated by `requireManagementAuth`; cloud **read**/auth routes stay public. Regression guards: `tests/unit/cloud-write-auth.test.ts`, `tests/unit/authz/classify.test.ts`, `tests/unit/public-api-routes.test.ts`. ([#6233](https://github.com/diegosouzapw/OmniRoute/pull/6233) — thanks @vittoroliveira-dev)
+
+- **refactor(dashboard):** extract the onboarding-wizard "Open provider details" link target into a pure, unit-tested `buildProviderDetailsHref(connection)` helper. The wizard already routes by `connection.id` (the node UUID) rather than the provider category slug (#6144/#6145); this hardens that behavior behind a tested helper that guards a missing id/connection. Regression guard: `tests/unit/provider-onboarding-href.test.ts`. ([#6166](https://github.com/diegosouzapw/OmniRoute/pull/6166) — thanks @KooshaPari)
+
+- **fix(security):** the doubao synthetic device-id generator now derives its digits via an unbiased crypto-random draw (rejection sampling over `crypto.randomBytes()`) instead of a `% 10` reduction, closing a CodeQL `js/biased-cryptographic-random` finding.
+- **fix(agentSkills):** the GitHub-skills generator now resolves `outputDir` to an absolute path before writing, fixing a regression introduced by #6366 (relative-to-cwd base path) that could write generated skill files to the wrong directory.
+- **fix(security):** `/api/keys/{id}/devices` now checks the HTTP method before auth/validation, returning a `405` for non-GET/DELETE verbs instead of a misleading `401`/`500` (closes a `dast-smoke` QUERY-method finding).
+- **fix(quality):** clear the last 2 heavy quality-gate reds on the release tip (cycle pre-flight).
+- **fix(mitm):** the test suite and CI can never mutate the OS trust store — `OMNIROUTE_SKIP_SYSTEM_TRUST=1` is set globally for tests/CI so `installCert`/`uninstallCert`/`installTproxyCa` skip the privileged OS dispatch ([#6310](https://github.com/diegosouzapw/OmniRoute/pull/6310); full detail is under the [3.8.45] section below — this branch received it via the parallel-cycle sync-back).
+- **fix(api):** `POST /api/github-skills` now Zod-validates its request body; documented the new quality-gate env vars and pinned the merge-integrity GitHub Actions to a commit SHA.
+- **fix(skills):** generate the missing `omni-github-skills` registry entry and align the agent-skills catalog-count tests (follow-up to #6186).
+- **fix(quality):** clear the cycle's 11 net-new ESLint errors and make `validate-release-green` suppressions-aware.
+
+### 📝 Maintenance
+
+- **i18n(it):** add 118 missing Italian (`it`) translations (net-additive — no existing keys dropped, valid JSON), improving Italian UI coverage. ([#6212](https://github.com/diegosouzapw/OmniRoute/pull/6212) — thanks @serverless83)
+- **chore(providers):** remove deprecated **MiMo V2** model entries from the catalogs (xiaomi-mimo, opencode-go, zenmux-free, audio TTS) — the upstream V2 line is superseded by MiMo V2.5; drops `mimo-v2-tts`, `mimo-v2-pro`, `mimo-v2-omni`, `mimo-v2-flash`, `mimo-v2-flash-free` and realigns the provider-catalog tests. ([#6248](https://github.com/diegosouzapw/OmniRoute/pull/6248) — thanks @backryun)
+- **chore(release):** ~50 commits on this branch are v3.8.45 pre-flight/hardening fixes and CI-perf work that landed here via the parallel-cycle sync-back (`sync-next-cycle.mjs`, Hard Rule #21) **after** the `v3.8.45` git tag was cut, and are already fully documented under the **[3.8.45]** section below — listed here only so the per-cycle commit-coverage check (`npm run release:uncovered`) doesn't flag them as gaps. Provider/catalog/UX/backend fixes: #6041, #6078, #6108, #6135, #6148, #6149, #6154, #6158, #6161, #6162, #6163, #6164, #6165, #6170, #6177, #6178, #6181, #6186, #6187, #6191, #6193, #6194, #6195, #6200, #6205, #6208, #6209, #6211, #6213, #6223, #6224, #6225, #6226, #6227, #6228, #6229, #6230, #6235, #6291, #6292. CI/release-pipeline work: #6167, #6203, #6214, #6215, #6218, #6273, #6275, #6283, #6284, #6285, #6300, #6305.
+- **chore(release):** additional zero-ref release-cycle plumbing on this branch, kept out of `release:uncovered` on purpose (no `#N` in the commit subject to cite): opening the v3.8.46 cycle, opening/closing the v3.8.45 cycle, the finalized [3.8.45] CHANGELOG i18n sync-back to 42 mirrors, the v3.8.45 cognitive/cyclomatic and file-size drift rebaselines, ESLint stale-suppression pruning (4,273 → 4,233), and clearing test-masking/docs-all pre-flight reds for v3.8.45.
+
+### ⚡ Performance & Infrastructure
+
+- **perf(release-green):** the pre-flight validator (`scripts/quality/validate-release-green.mjs`) now runs its 4 slow suites (unit / vitest / integration / pack-artifact) **concurrently** via `Promise.all` — pre-flight wall time drops from ~the sum of the suites to ~the slowest one (~30min saved per round; Phase 0 was the nº1 bottleneck of the v3.8.45 release benchmark, 2h54 of 6h34 e2e). Guard: `tests/unit/validate-release-green.test.ts` ("runs the slow suites CONCURRENTLY"). ([#6319](https://github.com/diegosouzapw/OmniRoute/pull/6319))
+- **fix(ci):** `scripts/release/sync-next-cycle.mjs` — two defects found live in its first production run (v3.8.45 Phase 5): (1) the `git()` helper's default 1 MiB `maxBuffer` crashed with `ENOBUFS` on `git show origin/main:CHANGELOG.md` (the CHANGELOG alone is >1 MiB) — widened to 64 MiB; (2) the i18n resync only propagated the `[NEXT]` (TBD) section, leaving the just-shipped finalized section as "— TBD" in all 42 mirrors — it now also syncs `[prevVersion]` bounded by the heading below it (new exported pure helper `versionAfter`). Guards: +5 tests in `tests/unit/sync-next-cycle.test.ts` (8/8). ([#6327](https://github.com/diegosouzapw/OmniRoute/pull/6327))
+- **test(ci):** concurrency-sensitive flaky tests are **quarantined into a serial pass** (`tests/unit/serial/`, `--test-concurrency=1`, appended to every unit runner incl. sharded variants — the serial pass is sharded too so concurrent shard jobs never self-collide). Initial set: `glm-coding-plan-monthly-3580`, `quota-division-blocks`, `provider-health-autopilot`, `combo-health-autopilot` — the class behind the ~28min CI wedges/re-runs (two live 1h+ wedges cancelled during this PR's own validation). Discovery + TIA gates track the new glob; systemic root cause (async logger writing after teardown) tracked in [#6360](https://github.com/diegosouzapw/OmniRoute/issues/6360). Guard: `tests/unit/test-serial-quarantine.test.ts` (4). ([#6347](https://github.com/diegosouzapw/OmniRoute/pull/6347))
+
+### 🙌 Contributors
+
+Thanks to everyone whose work landed in v3.8.46:
+
+| Contributor                                                    | PRs / Issues           |
+| -------------------------------------------------------------- | ---------------------- |
+| [@2220258345](https://github.com/2220258345)                   | direct commit / report |
+| [@abdofallah](https://github.com/abdofallah)                   | direct commit / report |
+| [@adentdk](https://github.com/adentdk)                         | direct commit / report |
+| [@aleksesipenko](https://github.com/aleksesipenko)             | direct commit / report |
+| [@anki1kr](https://github.com/anki1kr)                         | direct commit / report |
+| [@anungma](https://github.com/anungma)                         | direct commit / report |
+| [@arpicato](https://github.com/arpicato)                       | direct commit / report |
+| [@backryun](https://github.com/backryun)                       | #6248                  |
+| [@binsarjr](https://github.com/binsarjr)                       | direct commit / report |
+| [@brightfiscalband](https://github.com/brightfiscalband)       | direct commit / report |
+| [@chirag127](https://github.com/chirag127)                     | #6501, #6506           |
+| [@developerjillur](https://github.com/developerjillur)         | direct commit / report |
+| [@dilneiss](https://github.com/dilneiss)                       | #6499                  |
+| [@dtybnrj](https://github.com/dtybnrj)                         | direct commit / report |
+| [@Forcerecon](https://github.com/Forcerecon)                   | direct commit / report |
+| [@hao3039032](https://github.com/hao3039032)                   | direct commit / report |
+| [@Iammilansoni](https://github.com/Iammilansoni)               | #6150, #6245           |
+| [@jmengit](https://github.com/jmengit)                         | direct commit / report |
+| [@jordansilly77-stack](https://github.com/jordansilly77-stack) | direct commit / report |
+| [@JxnLexn](https://github.com/JxnLexn)                         | direct commit / report |
+| [@KooshaPari](https://github.com/KooshaPari)                   | #6166                  |
+| [@loopyd](https://github.com/loopyd)                           | direct commit / report |
+| [@luweiCN](https://github.com/luweiCN)                         | #6221                  |
+| [@makcimbx](https://github.com/makcimbx)                       | direct commit / report |
+| [@muflifadla38](https://github.com/muflifadla38)               | direct commit / report |
+| [@newnol](https://github.com/newnol)                           | direct commit / report |
+| [@ofekbetzalel](https://github.com/ofekbetzalel)               | direct commit / report |
+| [@ohahe52-dot](https://github.com/ohahe52-dot)                 | direct commit / report |
+| [@phidinhmanh](https://github.com/phidinhmanh)                 | direct commit / report |
+| [@powellnorma](https://github.com/powellnorma)                 | direct commit / report |
+| [@qpeyba](https://github.com/qpeyba)                           | direct commit / report |
+| [@RaviTharuma](https://github.com/RaviTharuma)                 | direct commit / report |
+| [@RCrushMe](https://github.com/RCrushMe)                       | direct commit / report |
+| [@rianonehub](https://github.com/rianonehub)                   | #6134, #6204           |
+| [@serverless83](https://github.com/serverless83)               | #6212                  |
+| [@swingtempo](https://github.com/swingtempo)                   | direct commit / report |
+| [@tenshiak](https://github.com/tenshiak)                       | direct commit / report |
+| [@ThongAccount](https://github.com/ThongAccount)               | direct commit / report |
+| [@UnrealAryan](https://github.com/UnrealAryan)                 | direct commit / report |
+| [@vinayakkulkarni](https://github.com/vinayakkulkarni)         | direct commit / report |
+| [@vittoroliveira-dev](https://github.com/vittoroliveira-dev)   | #6233                  |
+| [@warelik](https://github.com/warelik)                         | direct commit / report |
+| [@xxy9468615](https://github.com/xxy9468615)                   | direct commit / report |
+| [@xz-dev](https://github.com/xz-dev)                           | direct commit / report |
+| [@yanpaing007](https://github.com/yanpaing007)                 | direct commit / report |
+| [@diegosouzapw](https://github.com/diegosouzapw)               | maintainer             |
+
+---
+
 ## [3.8.45] — 2026-07-06
 
 ### ✨ New Features

docs/i18n/ro/CHANGELOG.md

@@ -6,6 +6,213 @@
 
 ## [3.8.31] — 2026-06-20
 
+## [3.8.46] — TBD
+
+### ✨ New Features
+
+- **feat(sse):** **hide paid-only models from `auto/*` routing** when `hidePaidModels` is on ([#6512](https://github.com/diegosouzapw/OmniRoute/issues/6512)) — follow-up to #6328/#6495. PR #6495 hid paid-only models from the `GET /v1/models` listing, but `auto/*` combos (`auto/best-coding`, `auto/glm`, …) could still pick a paid-only backend into their candidate pool → a 402/403 at request time. `createVirtualAutoCombo` now filters the candidate pool through the new pure `open-sse/services/autoCombo/paidModelFilter.ts` (`filterPaidOnlyCandidates`), applying the same free-model predicate #6495 uses in `catalog.ts` (`providerHasFreeModels(provider) && isFreeModel(provider, {id})`) whenever `settings.hidePaidModels === true`. Applied before the category/tier/family narrowing, so it covers every `auto/*` combo; an all-paid pool degrades to the existing graceful empty-pool path. **Opt-in — default OFF leaves the pool unchanged** (identity). Regression guard: `tests/unit/autoCombo/paid-model-filter-6512.test.ts` (4, incl. the default-off identity guard).
+- **feat(sse):** **provider-family auto combos** — `auto/glm`, `auto/minimax`, `auto/mimo`, `auto/zai`, `auto/gemma`, `auto/llama`, `auto/gemini` ([#6453](https://github.com/diegosouzapw/OmniRoute/issues/6453)) — new routable ids that materialize an on-demand virtual combo spanning whatever installed backends currently expose that model family, degrading gracefully as backends rotate. A new pure `open-sse/services/autoCombo/modelFamily.ts` (`detectModelFamily`) classifies by model-id prefix for six families; `zai` is instead resolved by provider id (z.ai's hosted API serves the same `glm-*` model ids as every other GLM backend, so `auto/zai` means "route to my z.ai backend specifically" vs `auto/glm`'s "any connected GLM backend"). Reuses the existing `createVirtualAutoCombo` on-demand materialization path (no DB writes) and the `/v1/models` catalog advertising loop. Regression guard: `tests/unit/autoCombo/provider-family-combos.test.ts` (11).
+- **feat(proxy):** native **proxy-pool round-robin / egress IP rotation** ([#6365](https://github.com/diegosouzapw/OmniRoute/issues/6365)) — a scope (global / provider / account) can now hold **multiple** proxies as a pool with a rotation strategy, so outbound requests cycle their egress IP instead of pinning one proxy per scope. Migration `117_proxy_pool_rotation.sql` lifts the `UNIQUE(scope, scope_id)` constraint (rebuild via the canonical rename/copy/drop; existing single assignments become 1-element pools) and adds a `proxy_scope_rotation` companion table holding the per-scope strategy + a persisted monotonic round-robin cursor. Strategies: `round-robin` (default, monotonic cursor — never `Math.random`), `random`, and `sticky-per-N-min`. Resolution (`resolveProxyForScopeFromRegistry` / `resolveProxyForConnectionFromRegistry`) now fetches the alive, position-ordered candidate set (unchanged `PROXY_ALIVE_PREDICATE`) and applies the strategy; an empty / all-dead pool still returns `null` — the #6246 fail-closed guard is **untouched** (never falls through to direct egress). Backend + DB only; dashboard pool-builder UI is a follow-up. Regression guard: `tests/unit/proxy-pool-rotation-6365.test.ts` (8, incl. fail-closed + backward-compat).
+- **feat(providers):** end-to-end **tool/function calling on the native Gemini `/v1beta` endpoint** ([#6222](https://github.com/diegosouzapw/OmniRoute/issues/6222)) — both directions of the Gemini↔OpenAI conversion now preserve tool data (previously silently dropped). Request side: `convertGeminiToInternal` (extracted to its own testable module) maps `tools[].functionDeclarations` → OpenAI `tools`, prior `functionCall` parts → assistant `tool_calls`, and `functionResponse` parts → `tool`-role messages. Response side: `convertOpenAIResponseToGemini` emits `parts[].functionCall {name,args}` from `message.tool_calls`, and the streaming `openAIChunkToGeminiChunk` accumulates fragmented `tool_calls` deltas by index into complete `functionCall` parts. The non-Gemini client paths (Claude, OpenAI-Responses) already preserved tool calls — this closes the gap specific to the native Gemini surface. Regression guard: `tests/unit/v1beta-gemini-tool-calling-6222.test.ts` (6, incl. a streaming SSE round-trip).
+- **feat(providers):** copilot-m365-web **enterprise / work tier** support ([#6334](https://github.com/diegosouzapw/OmniRoute/issues/6334)) — mirrors the EDU-tier pattern (#6210): `M365ConnectionParams` gains an `agent` field, a new opt-in `M365_ENTERPRISE_OVERRIDES` preset (`agent=work`, `scenario=officeweb`, `licenseType=Premium`) applies via `providerSpecificData.tier="enterprise"` (alias `"work"`), and `agent` is also overridable directly via `providerSpecificData.agent`. `buildWsUrl` was hardcoding `agent="web"` (the one enterprise-distinguishing param with no override path), so a Premium work account handshook then returned an empty stream. The individual and EDU paths are untouched. Kilo's dup flag vs #6210 (EDU tier) was a false positive — different tier. Regression guard: `tests/unit/copilot-m365-enterprise-6334.test.ts` (7). End-to-end confirmation on a real Premium work account is a live-VPS validation follow-up (Hard Rule #18). (thanks @Forcerecon)
+- **feat(api):** standardized, provider-agnostic **`effort` + `thinking` request params** ([#6241](https://github.com/diegosouzapw/OmniRoute/issues/6241)) — a thin standardization layer over the existing mature per-provider reasoning plumbing (no provider mapper touched). `providerChatCompletionSchema` gains a canonical `effort` (reusing the shared `none/low/medium/high/xhigh` vocabulary — the UI tiers `extra`/`max` collapse onto `xhigh`) and a boolean `thinking`. A pure `normalizeReasoningRequest` (wired once in `src/sse/handlers/chat.ts`, before any reasoning field is read) folds them onto the fields the translators already consume (`reasoning_effort` / `reasoning.effort` / `thinking`), so they fan out to Anthropic / Gemini / xAI / Responses — an explicit client `reasoning_effort` / object-shaped `thinking` always wins (backward-compatible). `/models` additively exposes `supportsThinking` + `effort_tiers` so the frontend can render the toggles (UI component is a follow-up). Regression guard: `tests/unit/effort-thinking-standardization-6241.test.ts` (12). (thanks @Iammilansoni, @shabeer)
+- **feat(combo):** new **`pipeline` (sequential) combo strategy** ([#6297](https://github.com/diegosouzapw/OmniRoute/issues/6297)) — the 18th routing strategy runs targets **in order**, threading each step's output into the next step's input, with an optional per-step `prompt` (system instruction); only the final step's response is returned. Distinct from `fusion` (parallel fan-out + judge). Implemented as a self-contained `open-sse/services/pipeline.ts` (sibling to `fusion.ts`), dispatched from `combo.ts`; the step list reuses `combo.models` order and reads an optional `prompt` off each target (backward-compatible — ignored by every other strategy). Intermediate steps run non-streaming with tools stripped (complete prose to thread forward); the final step keeps the client's `stream` flag + tools. A failing/empty/unparseable intermediate step fails the whole pipeline explicitly via a sanitized error (never silently swallowed). Kilo's dup flag vs #563 was a false positive (that's model→chain selection; this is a sequential chain). Regression guard: `tests/unit/combo-pipeline-strategy.test.ts` (5). (thanks @ofekbetzalel)
+- **feat(ci):** `check:test-masking` now flags **inline-reimplemented prod conditions** ([#6348](https://github.com/diegosouzapw/OmniRoute/issues/6348)) — a new **report-only** subcheck (v2, 6A.10 family) catches the wrong-shape contract test: a test that recomputes the condition under test inline instead of importing/exercising the real function (the #6216 class, where `=== 500` → `>= 500` stayed green because the test re-implemented the branch). For each added/modified test file it warns when the file textually duplicates a ≥3-token conditional from a production file touched in the same PR **and** does not import the symbol/module owning it, via a pure, fixture-tested `findReimplementedConditions()` with an allowlist mirroring `assertReductionAllowlist`. Report-only for now (does not fail the gate) — to be promoted to blocking after a triage cycle. Regression guard: `tests/unit/check-test-masking.test.ts` (45).
+- **feat(sse):** per-connection routing override (native vs CLIProxyAPI) ([#6339](https://github.com/diegosouzapw/OmniRoute/issues/6339)) — the previously-dead `isCliproxyapiDeepModeEnabled` helper is now wired into `resolveExecutorWithProxy`: a single connection can opt itself into the CLIProxyAPI passthrough executor via `providerSpecificData.cliproxyapiMode="claude-native"`, with precedence **connection override > provider `upstream_proxy_config` mode > default**. `resolveExecutorWithProxy` now receives the resolved connection's `providerSpecificData` (threaded from `chatCore.ts`), so one connection can deep-route while the provider's other connections stay native — no DB schema change (the toggle rides in `providerSpecificData`). Also resolves the same-provider mixing ask in #6340. Regression guard: `tests/unit/chatcore-executor-proxy.test.ts` (9). (thanks @RaviTharuma)
+- **feat(dashboard):** "Add session cookie" modal now shows a prominent **"Open ‹host› →"** link to the provider's own site ([#6268](https://github.com/diegosouzapw/OmniRoute/issues/6268)) — every `-web` cookie-session provider (chatgpt-web, claude-web, gemini-web, kimi-web, lmarena, qwen-web, m365-copilot-web, …) renders a one-click external link (opening the provider's login/home page in a new tab) so operators no longer tab away to retype the URL mid-setup. The host resolves from a pure, unit-tested `resolveWebProviderHost()` (prefers `WEB_COOKIE_PROVIDERS[id].website`, falls back to the registry `baseUrl` origin); non-web providers render exactly as before. Kilo's dup flag vs #6265 (modal-too-small-on-1080p) was a false positive — distinct concern. Regression guard: `tests/unit/resolve-web-provider-host.test.ts` (5). (thanks @chirag127)
+- **feat(providers):** add **DigitalOcean** AI (serverless inference) as an OpenAI-compatible API-key provider ([#6373](https://github.com/diegosouzapw/OmniRoute/pull/6373)) — base `https://inference.do-ai.run/v1`, wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/digitalocean/`, `src/shared/constants/providers/apikey/inference-hosts.ts`). Regression guard: `tests/unit/digitalocean-provider.test.ts`. (thanks @newnol)
+- **feat(providers):** add **Huancheng Public API** (`hcnsec`) as an OpenAI-compatible regional provider ([#6410](https://github.com/diegosouzapw/OmniRoute/pull/6410)) — Xinjiang Huancheng Cybersecurity's public LLM platform (base `https://api.hcnsec.cn/v1`, free credits via daily check-ins), wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/hcnsec/`, `src/shared/constants/providers/apikey/regional.ts`). Regression guard: `tests/unit/hcnsec-provider.test.ts`. (thanks @UnrealAryan)
+- **feat(dashboard):** the web-session credential guide now shows an **"Open {host}" link** ([#6316](https://github.com/diegosouzapw/OmniRoute/pull/6316)) to the provider's sign-in site (derived from the provider `website` via `getProviderWebsiteHost`), so you can jump straight to the page where the cookie/session must be captured. Regression guard: `tests/unit/web-session-provider-link-6316.test.ts`. (thanks @jordansilly77-stack)
+- **feat(cerebras):** add the **Gemma 4 31B** model (`gemma-4-31b`) to the Cerebras registry + pricing table ([#6331](https://github.com/diegosouzapw/OmniRoute/pull/6331)). Regression guard extends `tests/unit/t28-model-catalog-updates.test.ts`. (thanks @backryun)
+- **feat(providers):** add **Yuanbao (web)** as a cookie-session provider ([#6196](https://github.com/diegosouzapw/OmniRoute/issues/6196)) — `yuanbao-web` (Tencent Yuanbao, `yuanbao.tencent.com`) with cookie-only auth (`hy_user`/`hy_token` + public agent id), SSE→OpenAI translation incl. `reasoning_content`, exposing DeepSeek V3/R1 + Hunyuan / Hunyuan-T1. Regression guard: `tests/unit/providers-yuanbao-web.test.ts`. `together-web` was **deferred** (no verifiable web-session endpoint — needs a captured request) and `huggingchat-web` **dropped** (the existing `huggingchat` already is a web-cookie provider). (thanks @chirag127)
+- **feat(providers):** route the built-in **agentrouter** through the dynamic Claude-Code wire image ([#6056](https://github.com/diegosouzapw/OmniRoute/issues/6056)) — a small static allow-set (`CC_WIRE_IMAGE_BUILTINS` in `open-sse/services/ccWireImageBuiltins.ts`), consulted by `isClaudeCodeCompatible` / `isClaudeCodeCompatibleProvider` / `applyFingerprint`, makes agentrouter adopt the CC wire-image headers + fingerprint **while guarding the CC baseUrl/auth branches** so it keeps its own registry `baseUrl` and `x-api-key` auth. Regression guard: `tests/unit/agentrouter-cc-wire-image.test.ts` (asserts the wire image is applied AND agentrouter's baseUrl/auth are preserved). Live WAF-acceptance against agentrouter.org is a VPS validation follow-up (Hard Rule #18).
+- **feat(providers):** **bulk-add API keys for Cloudflare Workers AI** ([#6174](https://github.com/diegosouzapw/OmniRoute/issues/6174)) — `cloudflare-ai` is removed from the bulk-add exclusion list and the bulk parser gains a 3-field `name|accountId|apiKey` mode; the bulk route now builds a **per-entry** `providerSpecificData` so each key carries its own `accountId` (fixing the previous shared-object reuse), and both the create + key-validation paths receive it. Regression guard: `tests/unit/bulk-api-key-parser-cloudflare.test.ts`. (thanks @muflifadla38)
+- **feat(dashboard):** routing/settings UX clarity ([#6147](https://github.com/diegosouzapw/OmniRoute/issues/6147)) — (1) weighted combos show the **effective routing share %** next to each weight when weights don't sum to 100 (`WeightTotalBar.tsx`); (2) the status widget's user-facing **"Cloud Sync" label is renamed** to "Remote Settings Sync" (`CloudSyncStatus.tsx`; internal ids/state untouched); (3) built-in providers gain an **opt-in advanced base-URL override** (`isBaseUrlOverrideEligibleProvider`, hidden behind an "Advanced" toggle, reusing the existing `providerSpecificData.baseUrl` persistence — not globally widened). Regression guard: `tests/unit/routing-settings-ux-6147.test.ts`.
+- **feat(combo):** add an option to **disable session stickiness**, per-combo or globally — round-robin / random combos can rotate to a different connection on every request instead of pinning a whole conversation to one connection by its first-message hash. Resolution precedence per-combo `config.disableSessionStickiness` → global `settings.disableSessionStickiness` → default `false` (preserves the #3825 prompt-cache/504 fix); gates **both** stickiness call sites in `open-sse/services/combo.ts`. Exposed as a global toggle (Combo Defaults) and a per-combo Inherit/on/off control. ([#6168](https://github.com/diegosouzapw/OmniRoute/issues/6168)) Regression guard: `tests/unit/combo-disable-session-stickiness.test.ts`. (thanks @RCrushMe)
+- **feat(docker):** add the `OMNIROUTE_NO_SUDO` env flag for root-less / user-namespaced deployments — the MITM cert-trust command path (`resolveSudoSpawn` in `src/mitm/systemCommands.ts`) now strips the leading `sudo` when the flag is truthy, in addition to the existing root / sudo-missing cases, so the Proxy Agent runs without `sudo` (the operator trusts the CA manually, e.g. via `NODE_EXTRA_CA_CERTS`). Argv-array `spawn` preserved — no shell interpolation (Hard Rule #13). ([#6122](https://github.com/diegosouzapw/OmniRoute/issues/6122)) Regression guard: `tests/unit/mitm-systemCommands-no-sudo.test.ts`. (thanks @powellnorma)
+- **feat(providers):** add **Requesty** as an OpenAI-compatible gateway provider (BYOK, base `https://router.requesty.ai/v1`, ~200 free requests/day) — wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/requesty/`, `src/shared/constants/providers/apikey/gateways.ts`). ([#6120](https://github.com/diegosouzapw/OmniRoute/issues/6120)) Regression guard: `tests/unit/requesty-provider.test.ts`. (thanks @chirag127)
+- **feat(dashboard):** add **configured-only / available-only filters** to the Free Provider Rankings page ([#6150](https://github.com/diegosouzapw/OmniRoute/issues/6150)) — hide providers you haven't configured, or whose connections are all rate-limited / out of quota, via server-side query params (`?configuredOnly` / `?availableOnly` on `GET /api/free-provider-rankings`) backed by a testable lib helper reusing the in-process connection state (no Redis). Both filters default off, so the default view is unchanged; this supersedes the earlier client-side "Configured Only" toggle (#6245) with an available-only dimension and unit-tested logic. Regression guard: `tests/unit/freeProviderRankings-filters.test.ts`.
+
+### 🔧 Bug Fixes
+
+- **fix(dashboard):** adding a second API key connection for the same provider no longer silently overwrites the first — the Add-API-key modal now derives a unique default connection name (`main`, then `main-2`, `main-3`, …) so the backend name-based upsert can't collide (#6499 — thanks @dilneiss).
+- **fix(compression):** the session-dedup engine now also deduplicates a large multi-line block repeated **within a single message** (intra-message dedup), not just across turns; the compression-preview API surfaces a `fallbackReason`, and the fusion panel reports how many models were rate-limited vs failed on total-panel failure (#6501 — thanks @chirag127).
+- **fix(compression):** a stacked-pipeline step naming an unregistered engine now surfaces a `validationErrors` entry instead of silently no-op'ing, so misconfigured pipelines are visible in the preview API (#6506 — thanks @chirag127).
+- **feat(usage):** add a **Codex reset-credit redemption** flow to the Provider Limits UI ([#6361](https://github.com/diegosouzapw/OmniRoute/pull/6361)) — a `useCodexResetCreditRedemption` hook + `/api/usage/codex-reset-credit` route + `codexResetCredits` lib let you redeem banked Codex reset credits from the quota card. Regression guard: `tests/unit/codex-reset-credits.test.ts`. (thanks @JxnLexn)
+- **feat(glm):** add **team-plan quota settings** for `glm-cn` connections ([#6351](https://github.com/diegosouzapw/OmniRoute/pull/6351)) — a dedicated `GlmTeamQuotaFields` form section (team quota id / limits) threaded through the Add/Edit connection modals, persisted via `providerSpecificData`, with the GLM usage service reading the team quota. Regression guards: `tests/unit/glm-team-quota.test.ts`, `provider-specific-data-schema.test.ts`. (thanks @hao3039032)
+- **feat(providers):** add **TinyFish** web-fetch/search support ([#6349](https://github.com/diegosouzapw/OmniRoute/pull/6349)) — a `tinyfish-fetch` executor + `/v1/web/fetch` route + MCP web-fetch tool, registered as a specialty-media provider with request-validation and a search-provider catalog entry. Regression guards: `tests/unit/executor-tinyfish-fetch.test.ts`, `web-fetch-handler.test.ts`, `mcp-web-fetch-tool.test.ts`, `provider-validation-tinyfish.test.ts`. (thanks @dtybnrj)
+- **fix(cli):** `omniroute launch-codex` now spawns `codex.cmd` through a shell on Windows (the npm `.cmd` shim is unresolvable by bare `spawn` → ENOENT), mirroring the qodercli Windows fix (#6263) ([#6312](https://github.com/diegosouzapw/OmniRoute/pull/6312)). Regression guard: `tests/unit/launch-codex-windows-spawn-6312.test.ts`. (thanks @swingtempo)
+- **fix(codex):** isolate the **Spark** quota from the shared Codex quota and stabilize the quota UI ordering / hydration so per-scope limits render consistently ([#6336](https://github.com/diegosouzapw/OmniRoute/pull/6336)). Regression guards: `tests/unit/codex-quota-selection-hydration.test.ts`, `provider-limits-ui.test.ts` + 3 more. (thanks @xz-dev)
+- **feat(api):** add a `hidePaidModels` setting that filters paid-only models out of the `/v1/models` catalog. Regression guard: `tests/unit/models-catalog-hide-paid.test.ts`. (thanks @chirag127)
+- **fix(api-manager):** the fallback model picker now preserves combos instead of dropping them when a primary model is unavailable ([#6443](https://github.com/diegosouzapw/OmniRoute/pull/6443)). Regression guard: `tests/unit/api-manager-page-static.test.ts`. (thanks @jmengit)
+- **fix(providers):** recoverable Antigravity / Cloud-Code (Gemini Code Assist) `403` responses ([#6452](https://github.com/diegosouzapw/OmniRoute/pull/6452)) are now classified as a retryable project-config error instead of a terminal account ban, so a fixable project/API-disabled 403 no longer forces a ~1-year cooldown / full OAuth reconnect. Regression guard: `tests/unit/errorclassifier-antigravity-403.test.ts`. (thanks @developerjillur)
+- **fix(mitm):** `sanitizeHeaders` now redacts `Set-Cookie` response headers so upstream session cookies never leak into logs / diagnostics ([#6451](https://github.com/diegosouzapw/OmniRoute/pull/6451)). Regression guard: `tests/unit/mitm-sanitize-headers.test.ts`. (thanks @developerjillur)
+- **fix(api):** `/api/compression/preview` now accepts `mode: "caveman"` and correctly handles stacked / zero-compression previews ([#6425](https://github.com/diegosouzapw/OmniRoute/issues/6425)). Regression guard: `tests/unit/api/compression-preview-caveman-and-stacked-6425.test.ts`. (thanks @chirag127)
+- **feat(providers):** add **Zed** hosted LLM aggregator as a native-app provider ([#6118](https://github.com/diegosouzapw/OmniRoute/pull/6118)) — OAuth sign-in via the Zed hosted flow, registered through the shared provider registry + executor. Regression guards: `tests/unit/zed-oauth-provider.test.ts`, `zed-import-utils.test.ts`, `zed-docker-detect.test.ts`, `mitm-handler-zed.test.ts`. VPS-validated via live operator login (Hard Rule #18).
+- **fix(oauth):** the Kiro SSO-cache auto-import now **preserves the IDC region** — cross-region Amazon Q / Kiro profiles imported from the SSO cache are no longer collapsed to the default region ([#6113](https://github.com/diegosouzapw/OmniRoute/pull/6113)). Regression guard: `tests/unit/kiro-auto-import-idc-2059.test.ts`. VPS-validated via live operator login (Hard Rule #18).
+- **fix(dashboard):** passthrough model aliases no longer collide when two namespaced model ids share a last segment (port from 9router#1850, [#6431](https://github.com/diegosouzapw/OmniRoute/pull/6431)). `enx/gpt-5.5` and `enx/codebuddy/gpt-5.5` both auto-generated the alias `gpt-5.5`, so the second model could never be added (the UI just alerted "alias already exists"). Aliases are now disambiguated deterministically — bare last segment when free, then parent-qualified (`codebuddy-gpt-5.5`), then a numeric suffix — while re-adding the exact same model id is still blocked. Regression guard: `tests/unit/passthrough-alias-1850.test.ts`. (thanks @arpicato)
+- **fix(translator):** preserve a Gemini `functionResponse` co-located with other parts (another `functionCall`, or trailing `text`) in the same content when translating **Gemini → OpenAI** ([#6376](https://github.com/diegosouzapw/OmniRoute/pull/6376)). `convertGeminiContent()` early-returned the tool message on the first `functionResponse` part, dropping any co-located parts; such contents are now pre-split (one tool message per `functionResponse`, emitted first, plus one message for the remaining parts). Regression guard: `tests/unit/gemini-to-openai-function-response.test.ts`. (thanks @warelik)
+- **fix(headroom):** detect a python interpreter managed by **mise / pyenv / asdf / conda** (port from 9router#2353, [#6382](https://github.com/diegosouzapw/OmniRoute/pull/6382)). Headroom's python probe (`src/lib/headroom/detect.ts`) searched a hardcoded `PATH`, but version managers expose their interpreters via shim dirs that only join `PATH` through interactive-shell activation — which the non-interactive server never runs, so a managed python (≥3.10) was invisible and Headroom reported it missing. The search path now prepends the well-known shim/bin dirs (`~/.local/share/mise/shims`, `~/.pyenv/shims`, `~/.asdf/shims`, `$CONDA_PREFIX/bin`, `~/.local/bin`, respecting `MISE_DATA_DIR`/`PYENV_ROOT`/`ASDF_DATA_DIR` when set), and a new `HEADROOM_PYTHON` env override lets operators point straight at their interpreter (mirroring `HEADROOM_URL`). Still shell-free (`execFileSync`). Regression guard: `tests/unit/headroom-detect.test.ts` (5). (thanks @loopyd)
+- **fix(executors):** strip the OpenAI-Codex/Claude-CLI `client_metadata` passthrough field for **NVIDIA** requests (port from 9router#1887, [#6411](https://github.com/diegosouzapw/OmniRoute/pull/6411)). NVIDIA's OpenAI-compatible wrapper rejects it with `400 Unsupported parameter`, the same class already handled for `cerebras`/`mistral`; `nvidia` (executor `default`) was missing from the strip allowlist so Codex/Claude-Code passthrough requests 400'd. Regression guard: `tests/unit/executor-default-strip-client-metadata.test.ts` (+nvidia case). (thanks @phidinhmanh)
+- **fix(translator):** strip the Claude-style `thinking` field for **NVIDIA `z-ai/glm-5.2`** (port from 9router#2023, [#6413](https://github.com/diegosouzapw/OmniRoute/pull/6413)). NVIDIA's OpenAI-compatible wrapper 400s on `thinking` (a Claude-format client routed here leaves a `thinking:{type:"adaptive"}`); the existing strip rule only dropped `reasoning`. Same class already handled for `minimax-m2.7`. Regression guard: `tests/unit/nvidia-minimax-thinking-strip.test.ts` (+glm-5.2 case). (thanks @phidinhmanh)
+- **fix(translator):** suppress the streamed `</think>` close marker for the **Antigravity IDE** client (port from 9router#1061, [#6415](https://github.com/diegosouzapw/OmniRoute/pull/6415)). On thinking-only turns Antigravity rendered a bare `</think>` as the sole visible content, tripping its loop-detection and wasting requests. Antigravity's UA (`vscode/<v> (Antigravity/<v>)`) is added to the marker-suppress allowlist (alongside OpenCode); Claude Code / Cursor still get the marker, and `x-omniroute-thinking-marker: on` force-restores it. Regression guard: `tests/unit/think-close-marker-suppress-5245.test.ts`. (thanks @abdofallah)
+- **fix(executors):** strip nested `reasoning_content` from messages for **Mistral** (port from 9router#1649, [#6417](https://github.com/diegosouzapw/OmniRoute/pull/6417)). Mistral's API returns `422 extra_forbidden` when an assistant message carries `reasoning_content` (replayed thinking from a prior turn, e.g. via the Codex `/responses` path); the generic top-level 400 field-downgrade retry never covered the nested per-message field. `DefaultExecutor` now strips it for provider `mistral` only, so DeepSeek (which requires replayed `reasoning_content`) is unaffected. Regression guard: `tests/unit/mistral-strip-reasoning-content-1649.test.ts`. (thanks @xxy9468615)
+- **fix(executors):** strip the `client_metadata` passthrough field on the **OpenCode** path (port from 9router#1442, [#6418](https://github.com/diegosouzapw/OmniRoute/pull/6418)). OpenCode upstreams (e.g. `kimi-k2.6` via opencode-go) reject it with `400 "Extra inputs are not permitted, field: 'client_metadata'"`; the DefaultExecutor strip only covered cerebras/mistral and `OpencodeExecutor` extends `BaseExecutor` directly, so nothing removed it there. Regression guard: `tests/unit/opencode-strip-client-metadata-1442.test.ts`. (thanks @yanpaing007)
+- **fix(executors):** inject the `reasoning_content` echo for the native **Moonshot Kimi** provider (port from 9router#1480, [#6419](https://github.com/diegosouzapw/OmniRoute/pull/6419)). Kimi (executor `default`) is a thinking-mode upstream that 400s with "reasoning_content must be passed back" when a prior assistant turn lacks it; the placeholder injection was only wired into the OpenCode meta-provider, so direct multi-turn Kimi conversations failed. Scoped to `kimi` (gateway-served models matching the thinking-model name pattern are unaffected). Regression guard: `tests/unit/kimi-native-reasoning-injected-1480.test.ts`. (thanks @2220258345)
+- **fix(executors):** recover from a strict gateway's `context_management: Extra inputs are not permitted` 400 (port from 9router#1468, [#6420](https://github.com/diegosouzapw/OmniRoute/pull/6420)). **Claude Code** always sends a top-level `context_management` field; strict anthropic-compatible gateways reject it. The dedicated context-editing 400-fallback only fired when OmniRoute's own `contextEditing` feature was enabled (default off), so a client-sent field passed through untouched and 400'd. `context_management` is now in the generic reactive field-strip list, so it's stripped-and-retried once regardless of the feature flag (with correct request re-signing for claude-compatible relays). Regression guard: `tests/unit/provider-field-strips.test.ts`. (thanks @ohahe52-dot)
+- **fix(network):** enable **RFC 8305 Happy Eyeballs** (`autoSelectFamily`) on the direct-egress undici dispatcher (port from 9router#1237, [#6423](https://github.com/diegosouzapw/OmniRoute/pull/6423)). When DNS returns both IPv6 (AAAA) and IPv4 (A) and the IPv6 route is broken (e.g. a NAT64 `64:ff9b::` prefix without routing), undici tried IPv6 first and hung until `ETIMEDOUT` (then a 502 + account lockout), even though `curl` reached the same host. The direct dispatcher now races both families and uses whichever connects first. Proxy paths pin family via `proxyTls` and are unaffected. Regression guard: `tests/unit/direct-dispatcher-pipelining-4580.test.ts`. (thanks @adentdk)
+- **fix(combo):** round-robin now advances the rotation pointer past the model that **actually served**, not the eagerly-scheduled one (port from 9router#948, [#6428](https://github.com/diegosouzapw/OmniRoute/pull/6428)). With `stickyLimit: 1` (true round-robin), when the scheduled model failed and a _different_ model served via fallback, the counter had already advanced +1 from the scheduled index — so the next request reused the fallback-served model, degrading round-robin into hot-spotting on whichever model was healthy. The pointer now advances to the served index + 1 (mirroring the sticky-limit>1 path). Session-stickiness (#3825) and distribution are preserved. Regression guard: `tests/unit/combo-rr-fallback-advance-948.test.ts`. (thanks @binsarjr)
+- **fix(sse):** a non-string `model` field is now rejected with a `400` before the resolver, instead of crashing downstream `.toLowerCase()`/`.split()` calls into an empty-body `500` that escapes the error sanitizer ([#6407](https://github.com/diegosouzapw/OmniRoute/issues/6407)). Regression guard: `tests/unit/chat-non-string-model-6407.test.ts`. (thanks @chirag127)
+- **fix(api):** unknown `/api/*` routes now return a JSON `404` (instead of the dashboard HTML shell) and scalar chat params (`model`/`temperature`/etc.) are validated **before** the provider lookup so malformed requests fail fast with a clear `400` ([#6424](https://github.com/diegosouzapw/OmniRoute/issues/6424), [#6412](https://github.com/diegosouzapw/OmniRoute/issues/6412)). Regression guards: `tests/unit/api/api-catchall-json-404.test.ts`, `tests/unit/chat-early-schema-validation-6412.test.ts`. (thanks @chirag127)
+- **fix(api):** `/v1/chat/completions` now rejects a non-JSON `Content-Type` with a `400` before parsing the body ([#6414](https://github.com/diegosouzapw/OmniRoute/issues/6414)). Regression guard: `tests/unit/v1-chat-completions-content-type-6414.test.ts`. (thanks @chirag127)
+- **fix(api):** the `X-OmniRoute-Compression` response header is now echoed on `/v1/chat/completions` and `/v1/completions` ([#6422](https://github.com/diegosouzapw/OmniRoute/issues/6422)). Regression guard: `tests/unit/compression-header-echo-6422.test.ts`. (thanks @chirag127)
+- **fix(api):** concurrent `GET /v1/models` requests are coalesced into a single catalog build ([#6408](https://github.com/diegosouzapw/OmniRoute/issues/6408)). Regression guard: `tests/unit/v1-models-concurrent-6408.test.ts`. (thanks @chirag127)
+- **fix(api):** `/v1/completions` now echoes the requested `body.model` in its JSON + streamed responses ([#6429](https://github.com/diegosouzapw/OmniRoute/pull/6429)). Regression guard: `tests/unit/completions-body-model-echo.test.ts`. (thanks @chirag127)
+- **fix(api):** env-var master keys now see the full `/v1/models` catalog ([#6406](https://github.com/diegosouzapw/OmniRoute/issues/6406)). Regression guard: `tests/unit/models-catalog-envkey-6406.test.ts`. (thanks @chirag127)
+- **fix(api):** non-streaming `/v1/completions` responses now echo `body.model` aligned with the `X-OmniRoute-Model` header ([#6426](https://github.com/diegosouzapw/OmniRoute/issues/6426)). Regression guard: `tests/unit/v1-completions-model-header-match-6426.test.ts`. (thanks @chirag127)
+- **fix(api):** unknown `/v1/*` routes now return a JSON `404 not_found` instead of the Next.js dashboard HTML shell ([#6405](https://github.com/diegosouzapw/OmniRoute/issues/6405)). Regression guard: `tests/unit/api/v1-catchall-json-404.test.ts`. (thanks @chirag127)
+- **fix(api):** the per-connection provider models route now degrades to the shipped catalog when a provider's `/models` endpoint answers with a redirect ([#6267](https://github.com/diegosouzapw/OmniRoute/issues/6267)) — a `qwen-web` import failed with a raw `Redirect blocked … (307)` 503. `safeOutboundFetch` throws `REDIRECT_BLOCKED` on the 307, `getSafeOutboundFetchErrorStatus` maps it to 503, and `buildDiscoveryErrorFallbackResponse` treated every 503 as a hard error — so the non-empty `getModelsByProviderId("qwen-web")` catalog was never surfaced. A models-endpoint redirect is not a fixable-config error (unlike `URL_GUARD_BLOCKED`/`INVALID_URL`, which stay hard errors), so it now falls back to the local/cached catalog before the 503 short-circuit. General fix — covers any config-driven provider that 307s. Regression guard: `tests/unit/provider-models-qwen-web-redirect-6267.test.ts`. (thanks @chirag127)
+- **fix(api):** the per-connection provider models route (MCP `list_models_catalog` + the dashboard import view) now merges USER-ADDED custom models into its response ([#6247](https://github.com/diegosouzapw/OmniRoute/issues/6247)) — custom models live in the `key_value` namespace `customModels`, which the live REST `/api/v1/models` already merges, but `src/app/api/providers/[id]/models/route.ts` never read `getCustomModels`, so custom models were dropped on both the discovery-success and local_catalog paths. They are now folded into the returned model list (deduped by id, stamped `owned_by: provider`), fixing MCP + the dashboard import view in one place. Regression guard: `tests/unit/provider-models-custom-merge-6247.test.ts`. (thanks @RCrushMe)
+
+- **fix(providers):** GitLab Duo tool-calling follow-up turns no longer fail upstream with `422 {"detail":"Validation error"}` (tokens 0/0, rejected pre-inference). The #6234 tool-result-feedback fix serialized the **entire** multi-turn conversation into GitLab's single-file `code_suggestions` (`small_file`) generation endpoint — folded history that turn-N sent as an oversized `current_file.content_above_cursor` **and** duplicated verbatim into `user_instruction`, tripping the AI-Gateway's `small_file` validation guard. The executor now **bounds** that prompt: it keeps system + latest user message + the most-recent tool round (dropping older turns), caps oversized tool results, and stops duplicating the full prompt into `user_instruction` (which now carries only the short latest user message) — while still feeding the most-recent tool result back so the agent continues (`open-sse/executors/gitlab.ts`, [#6220](https://github.com/diegosouzapw/OmniRoute/issues/6220)). The unit test covers the bounding logic; the upstream 422→200 clearing is VPS-only (Hard Rule #18). Regression guard: `tests/unit/gitlab-tool-exchange-bounded-6220.test.ts`.
+
+- **fix(i18n):** the provider-detail (`/dashboard/providers/[id]`) connection-status filter labels no longer render as `__MISSING__:All` / `__MISSING__:Active` / `__MISSING__:Error` / `__MISSING__:Banned` / `__MISSING__:CreditsExhausted` in non-English locales (notably pt-BR) ([#6290](https://github.com/diegosouzapw/OmniRoute/issues/6290)). Root cause was **not** the namespace mismatch the issue guessed — the `providers.filter*` keys resolve correctly in `en.json`; the debt lived in the locale mirrors (`src/i18n/messages/*.json`), where these five keys carried the `__MISSING__:` sync sentinel in ~15 locales and were absent entirely in ~26 others, so next-intl found the key and echoed the sentinel verbatim. All 40 non-English/-Chinese mirrors now ship real translations for the five `providers.filter*` labels. Regression guard: `tests/unit/i18n-provider-filter-keys-6290.test.ts`. (thanks @diegosouzapw)
+
+- **fix(providers):** the `copilot-m365-web` streaming executor now emits `debug`-level WebSocket diagnostics ([#6210](https://github.com/diegosouzapw/OmniRoute/issues/6210)) — the outbound WS URL (with the `access_token` **redacted** via `redactWsUrl()`), handshake success/failure, and each received SignalR frame's `type`/`target`. Previously the streaming path logged nothing, so an empty `content:null` response (the M365 Education / Starter tier symptom fixed in #6234) was undiagnosable even at `APP_LOG_LEVEL=debug`. The change is debug-level and side-effect-free — it does not alter streaming behavior or the frame parser, and the token never reaches the logs. Regression guard: `tests/unit/copilot-m365-web-logging-6210.test.ts` (thanks @qpeyba)
+
+- **fix(resilience):** a round-robin combo no longer returns `503 all upstream accounts are unavailable` when a compatibility-rejected target is actually healthy ([#6238](https://github.com/diegosouzapw/OmniRoute/issues/6238)). `filterTargetsByRequestCompatibility` drops request-incompatible targets (tool/vision/structured-output unsupported, or below the required context window) **before** any availability check runs, and its `compatible.length === 0` safety net only fired when _all_ targets were filtered — not when the kept targets later all turned out runtime-unavailable (circuit-open / cooldown / no credentials). So a combo could 503 while a compat-rejected-but-healthy provider sat unused. `handleRoundRobinCombo` now keeps the compat-rejected set and, when every compat-kept target was skipped without a single real attempt, probes those rejected targets as a **last-resort fallback tier** (via the new pure `open-sse/services/combo/comboCompatFallback.ts`) before crystallizing the 503. Regression guard: `tests/unit/combo-roundrobin-compat-fallback-6238.test.ts`. (thanks @ThongAccount)
+
+- **fix(startup):** best-effort self-heal for a corrupted Turbopack dev cache on Windows ([#6289](https://github.com/diegosouzapw/OmniRoute/issues/6289)). On Windows, `pnpm dev` can fail at startup when Turbopack `mmap`s a persistent-cache SST file and the OS refuses the mapping (`os error 1455` — "paging file too small"), which Turbopack surfaces as a misleading `Module not found: Can't resolve '@/shared/utils/machine'`. This is a **known upstream Turbopack cache-corruption bug — not our code**. The dev launcher (`scripts/dev/run-next.mjs`) now wraps `nextApp.prepare()` and, when it rejects with that signature (`isTurbopackCacheCorruption` in the new `scripts/dev/turbopackCacheHeal.mjs`), purges `.build/next/**/cache/turbopack` and retries **once** with a clear log. **Caveat — best-effort only:** the corruption often surfaces as a runtime overlay rather than a `prepare()` rejection, so this cannot always intercept it; the reliable remedy remains manually deleting the Turbopack cache dir. Regression guard: `tests/unit/turbopack-cache-heal-6289.test.ts`. (thanks @chirag127)
+
+- **fix(providers):** qodercli PAT auth no longer fails with `spawn qodercli ENOENT` on Windows ([#6263](https://github.com/diegosouzapw/OmniRoute/issues/6263)) — `spawnQoderCli` spawned the bare `qodercli` name with `shell:false` and an unenriched env, so the npm `.cmd` wrapper under `%APPDATA%\npm` (a user-PATH directory) was never resolved. It now resolves the absolute `.cmd`/`.exe` path through the existing `getCliRuntimeStatus("qoder")` resolver in `src/shared/services/cliRuntime.ts` (memoized), spawns with `shell` when the target is a `.cmd`/`.bat`, and uses the cliRuntime-enriched env (PATH + PATHEXT + APPDATA); the ENOENT error now lists the searched paths plus the `CLI_QODER_BIN` override. End-to-end spawn on a real Windows host is host-only (Hard Rule #18); the path-resolution logic is unit-tested. Regression guard: `tests/unit/qodercli-windows-resolve-6263.test.ts`. (thanks @chirag127)
+
+- **fix(sse):** the reasoning-token buffer no longer inflates **probe-sized `max_tokens`** ([#6274](https://github.com/diegosouzapw/OmniRoute/issues/6274)) — Claude Code's `/model` capability check sends `max_tokens: 1`, but for a thinking-capable model with a large output cap (e.g. `glm-5.2`) the #3587 headroom heuristic (`max(current + 1000, ceil(current * 1.5))`) rewrote it to `1001` and forwarded that upstream, wasting tokens on a request that was never a genuine reasoning budget. `resolveReasoningBufferedMaxTokens()` (`open-sse/services/reasoningTokenBuffer.ts`) now short-circuits and returns the caller's value verbatim when it is below the new `REASONING_BUFFER_MIN_TRIGGER` (256) threshold — a tiny explicit limit is a probe, not a reasoning request. Real budgets still receive the #3587 headroom unchanged, and the guard runs after the existing capability checks so unknown / non-reasoning models keep returning `null`. Regression guard: `tests/unit/reasoning-token-buffer-6274.test.ts`. (thanks @brightfiscalband)
+
+- **fix(cli):** `omniroute reset-password` now works as a real subcommand, and password resets over piped (non-TTY) stdin actually apply ([#6261](https://github.com/diegosouzapw/OmniRoute/issues/6261), [#6258](https://github.com/diegosouzapw/OmniRoute/issues/6258)). Two coupled defects: (1) **#6261** — `bin/omniroute.mjs` routed everything through Commander with only two pre-Commander bypasses (`--mcp`, `reset-encrypted-columns`), so `omniroute reset-password` was rejected as an unknown command; only the separate `omniroute-reset-password` bin worked, while the docs falsely advertised the subcommand (incl. a bogus "legacy alias still works"). A pre-Commander bypass mirroring `reset-encrypted-columns` now dynamically imports `bin/reset-password.mjs` (which self-executes) before Commander parses; the three doc lines were corrected. (2) **#6258** — `bin/reset-password.mjs` issued two sequential `rl.question` prompts; under piped stdin the second read never settled at EOF, so `main()` never reached `resetManagementPassword` and the reset was a silent no-op (both prompts printed, no success, password unchanged). The CLI now detects non-TTY stdin and reads it once (first line = password, second line = confirm if present, else reused), adds a `--password-stdin` flag (entire stdin is the password, no confirmation), and exits `0` explicitly so the success line always flushes; interactive TTY behavior is unchanged. Regression guard: `tests/unit/reset-password-cli-6261-6258.test.ts` (3). (thanks @chirag127)
+
+- **fix(db):** the mass-migration **safety abort** now tells the operator how to bypass it and stops flooding the log ([#6260](https://github.com/diegosouzapw/OmniRoute/issues/6260)) — after restoring a backup that wiped the migration tracking table, `runMigrations()` threw the abort on every downstream `ensureDbInitialized()`, re-logging the full banner 11+ times, and the message never mentioned the existing `OMNIROUTE_MAX_PENDING_MIGRATIONS` escape hatch. The abort text now appends a bypass hint (set `OMNIROUTE_MAX_PENDING_MIGRATIONS=0` in `server.env` / `DATA_DIR/.env`), and a new `MigrationSafetyAbortError` is memoized so repeated calls in the same process throw the same instance and emit a single concise line instead of the full cascade. Regression guard: `tests/unit/migration-safety-abort-6260.test.ts`. (thanks @chirag127)
+
+- **fix(auth):** importing a **distinct** Codex/ChatGPT OAuth `auth.json` is no longer falsely rejected as "already exists" when it belongs to a different user in the same workspace ([#6301](https://github.com/diegosouzapw/OmniRoute/issues/6301)). `findExistingCodexConnection` (in `src/lib/oauth/utils/codexAuthImport.ts`) deduped **only** on `providerSpecificData.workspaceId === accountId`, where `accountId` is the shared `chatgpt_account_id`/`tokens.account_id` — so two members of the same ChatGPT Team collapsed onto a single connection (409 `duplicate_account`). The id_token's `https://api.openai.com/auth` claim carries a per-user `chatgpt_user_id` alongside the workspace id (the device-flow path already persisted it as `chatgptUserId`, but the import path did not). Now `parseAndValidateCodexAuth` extracts `userId` (`chatgpt_user_id` → `user_id` → JWT `sub`) into `ParsedCodexAuth`, the create/update paths persist `chatgptUserId` in `providerSpecificData` (mirroring `codex.ts`), and dedup keys on `workspaceId` **AND** `chatgptUserId` — with a backward-compat fallback to legacy accountId-only matching when no stored connection for that workspace records a `chatgptUserId`, so genuinely-same accounts still dedup. Regression guard: `tests/unit/codex-auth-import-userid-dedup-6301.test.ts` (4). (thanks @anungma)
+
+- **fix(providers):** importing models for the **venice-web** provider no longer fails with a red "Provider venice-web does not support models listing" ([#6269](https://github.com/diegosouzapw/OmniRoute/issues/6269)). `venice-web` is a web-cookie provider with an executor but no upstream `/v1/models` endpoint and no registry `models`, so the models route fell through to the tail `400`. Mirroring the `jules`/`linkup-search`/`ollama-search` fix (#5569), it now ships a static local catalog entry in `src/lib/providers/staticModels.ts` — seeding the current Venice lineup (`venice-uncensored`, `llama-3.3-70b`, `qwen3-235b`, `qwen3-4b`, `deepseek-r1-671b`; Venice rotates its catalog, see docs.venice.ai/models/overview) — so the route returns `200` with `source:"local_catalog"`, `intentional:true`. Regression guard: `tests/unit/static-models-venice-web-6269.test.ts`. (thanks @chirag127)
+
+- **fix(api):** the specialty model catalogs (`/v1/embeddings`, `/v1/images`, `/v1/music`, `/v1/videos` model lists) are now derived from the **unified catalog filtered by a predicate** (`getSpecialtyModelsResponse`) instead of ad-hoc per-route logic, so they consistently respect active-credential visibility and stay in sync with the main catalog ([#6303](https://github.com/diegosouzapw/OmniRoute/pull/6303)). Regression guard: `tests/unit/specialty-model-catalog-routes.test.ts`. (thanks @makcimbx)
+- **fix(api):** the agent-bridge server route now resolves the MITM manager via a **dynamic `import("@/mitm/manager.runtime")`** so Turbopack does not statically pull the stub (or over-bundle the manager), and the agent-skills generator anchors its output base path with `path.join(process.cwd(), …)` so Turbopack's static analyzer stops tracing the whole project root ([#6329](https://github.com/diegosouzapw/OmniRoute/issues/6329), [#6366](https://github.com/diegosouzapw/OmniRoute/pull/6366)). Regression guard: `tests/unit/agent-bridge-server-route-dynamic-import.test.ts`. (thanks @Iammilansoni)
+- **fix(api):** internal probes (combo-test, cloud-sync verify) now pick a **management-scoped / allow-all API key** instead of naively grabbing `getApiKeys()[0]` — a restricted `self:usage` first row made the probe fail with "Model X is not allowed for this API key" even when the combo path was healthy (`pickApiKeyForInternalUse` in `src/lib/db/apiKeys.ts`). The API-manager model editor also falls back to `/api/models?all=true` when `/v1/models` is catalog-protected ([#6372](https://github.com/diegosouzapw/OmniRoute/pull/6372)). Regression guard: `tests/unit/pick-internal-api-key-6372.test.ts`. (thanks @jmengit)
+- **fix(live-ws):** the Live Dashboard WebSocket server now **rejects on bind failure** (e.g. `EADDRINUSE` when the API bridge already holds the port) instead of letting the error surface as an unhandled `error` event that crash-loops the process — the `error` listener is attached to `wss` (not `server`) and releases the EventBus subscription on a failed start ([#6324](https://github.com/diegosouzapw/OmniRoute/issues/6324)). Regression guard: `tests/unit/live-ws-eaddrinuse-6324.test.ts`. (thanks @vinayakkulkarni)
+- **fix(dashboard):** the Home provider-topology widget now trusts the live provider-metrics snapshot — it uses `topology.errorProvider` and live `activeRequests` directly instead of re-deriving state from a stale `lastErrorAt` or applying a frontend timeout filter, so the topology reflects real-time provider health ([#6322](https://github.com/diegosouzapw/OmniRoute/pull/6322)). Regression guard: `tests/unit/home-provider-topology-live-state.test.ts`. (thanks @xz-dev)
+- **fix(sse):** strip zero-width markers from streamed **tool-call arguments** — a follow-up to [#5857](https://github.com/diegosouzapw/OmniRoute/pull/5857). That PR removed injected zero-width joiners (U+200D) from streamed assistant text/reasoning but deliberately left tool-call argument JSON byte-exact. The request-side obfuscation (`open-sse/services/claudeCodeObfuscation.ts`) injects ZWJ into agent words — including the temp path inside the Bash tool description — and Claude models copy that verbatim into generated commands, which are delivered as tool-call arguments rather than assistant text. As a result the ZWJ survived and corrupted code blocks (e.g. a temp path rendered with an invisible joiner). Now `open-sse/handlers/responseSanitizer.ts` strips zero-width code points from tool-call argument strings at every emit site (OpenAI non-stream/stream chat `tool_calls` + legacy `function_call`, native Responses `function_call` items, the OpenAI→Responses conversion, and the native Responses streaming `response.function_call_arguments.delta/.done` events). Only zero-width code points are removed; JSON structure and all other bytes stay identical (no parse/restringify), so normal arguments remain byte-exact. Regression guard: 6 new cases in `tests/unit/response-sanitizer.test.ts` (suite 50/50).
+
+- **fix(nodejs):** the default app log path now resolves under `DATA_DIR` (`~/.omniroute/logs/application/app.log`) instead of `process.cwd()` ([#6197](https://github.com/diegosouzapw/OmniRoute/issues/6197)) — the globally-installed CLI runs from an arbitrary working directory, so anchoring the default to cwd made file logging silently write to (or no-op under) an unrelated directory, contradicting the documented `.env.example` default. `getAppLogFilePath()` now computes the default lazily via the pure `resolveDataDir()` resolver (honours a per-process `DATA_DIR`, no directory-creation side effect); an explicit `APP_LOG_FILE_PATH` still wins. Regression guard: `tests/unit/logenv-datadir-path-6197.test.ts` (3).
+
+- **fix(docker):** AgentBridge/`startMitm` no longer aborts in containers/headless when the Antigravity-default DNS step can't write `/etc/hosts` ([#6127](https://github.com/diegosouzapw/OmniRoute/issues/6127)), and the privileged command's stderr now reaches `app.log` instead of only a bare exit code hitting the toast ([#6198](https://github.com/diegosouzapw/OmniRoute/issues/6198)). The default DNS step (`addDNSEntry`) was called unguarded while cert install and the two sibling DNS steps were each best-effort — in the runtime Docker image (`USER node`, no `sudo`, read-only `/etc/hosts`) it threw `Command failed with code 1` out of `startMitmInternal` and killed the whole start, discarding the stderr. The three DNS steps are extracted into a best-effort `provisionDnsEntries()` where each failure is logged with the full `err` (stderr included, folded in by `systemCommands.ts`) and never aborts the start. Regression guard: `tests/unit/mitm-dns-graceful-degrade-6127.test.ts` (4).
+
+- **fix(providers):** copilot-m365-web now supports the M365 Education "Starter / OfficeWebIncludedCopilot" tier and no longer returns an empty `content:null` stream ([#6210](https://github.com/diegosouzapw/OmniRoute/issues/6210)). Two gaps: (1) `buildWsUrl()` hardcoded the individual-consumer scenario (`OfficeWebPaidConsumerCopilot`, `isEdu=false`) — the EDU tier is now opt-in via `providerSpecificData.tier="edu"`, emitting `scenario=OfficeWebIncludedCopilot`/`isEdu=true` (the individual path is unchanged); (2) the EDU/GPT-5.5 path streams deltas via `arguments[0].writeAtCursor` (incremental) instead of only `messages[].text` (accumulated snapshots), which the parser dropped — a new `accumulateBotContent()` folds both formats, with `type:2 item.result.message` as a last-resort fallback. Regression guard: `tests/unit/copilot-m365-edu-writeatcursor-6210.test.ts` (10). (thanks @qpeyba)
+
+- **fix(providers):** GitLab Duo executor now feeds tool results back into the prompt instead of looping ([#6220](https://github.com/diegosouzapw/OmniRoute/issues/6220)) — `buildPrompt()` branched only on `system`/`user` and took `userParts.at(-1)`, silently dropping the `assistant{tool_calls}` + `tool{result}` turns the client appended, so the reconstructed prompt was byte-identical to turn 1 and the model re-emitted the same `<tool>` call forever. When a tool exchange is present the full conversation is now serialized, folding each tool result back keyed by its `tool_call_id`; simple conversations keep the legacy shape. Complements the tool_call emission from [#6051](https://github.com/diegosouzapw/OmniRoute/issues/6051) (the `kilo-duplicate` label was a false positive — different, sequential defect). Regression guard: `tests/unit/gitlab-tool-result-feedback-6220.test.ts` (4).
+
+- **fix(providers):** opencode-go/opencode-zen can now synthesize the OpenCode CLI identity headers Cloudflare requires on VPS egress ([#5997](https://github.com/diegosouzapw/OmniRoute/issues/5997)) — on a datacenter VPS, `opencode.ai/zen/go/v1/chat/completions` 403s (HTML challenge) requests lacking CLI identity, while the reporter's control curl proved that `User-Agent: opencode-cli/1.0.0` + `x-opencode-client: cli` + `x-opencode-project: default` + fresh request/session UUIDs succeed. Opt-in via `OPENCODE_SYNTHESIZE_CLI_HEADERS=true` (values overridable via `OPENCODE_GO_USER_AGENT`/`OPENCODE_USER_AGENT`/`OPENCODE_CLIENT`/`OPENCODE_PROJECT`); it fills only headers the client did not already send. Kept **off by default** — the forward-only path is deliberate (fabricating a wrong value risks upstream rejection; a prior dedup regressed with `opencode/local`), so this replaces the fragile local header-injection shim without changing default behavior. Regression guard: `tests/unit/opencode-cli-headers-synthesis-5997.test.ts` (6). (thanks @aleksesipenko)
+- fix(resilience): sticky session affinity now evicts and fails over to another account when the pinned account is exhausted/unavailable (#6219)
+
+- fix(sse): Responses API passthrough now drops internal commentary-phase output before forwarding to clients (gated by RESPONSES_PASSTHROUGH_DROP_COMMENTARY, default on) (#6199)
+
+- **fix(sse):** tool-call function schemas with a root `type: null` are now coerced to `type: "object"` before dispatch ([#6359](https://github.com/diegosouzapw/OmniRoute/issues/6359)) — clients like the Codex app emit `parameters: { type: null, ... }` for some tools, which OpenAI-compatible upstreams reject with `400 Invalid schema for function '...': schema must be a JSON Schema of 'type: "object"', got 'type: null'`, failing the whole request. `toolSchemaSanitizer` already stripped the null; it now re-adds the mandatory root `"object"` type (and empty `properties`/open `additionalProperties` when absent). Combinator roots (`anyOf`/`oneOf`/`allOf`) and explicit root types are left untouched. Regression guard: 5 new cases in `tests/unit/tool-schema-sanitizer.test.mjs`.
+
+- **fix(docker):** AgentBridge no longer fails to start on npm/Electron/VPS installs with "MITM manager stub reached at runtime" ([#6344](https://github.com/diegosouzapw/OmniRoute/issues/6344)) — v3.8.45 flipped the production bundler default to Turbopack, but `next.config.mjs` aliased `@/mitm/manager` to its Docker-only degraded stub **unconditionally**. That was harmless while Docker (which sets the alias intentionally for #3390 graceful degradation) was the sole Turbopack consumer, but once every artifact built with Turbopack the stub shipped to all non-Docker users and `startMitm` threw on the first Agent-Bridge start. The alias is now opt-in via `OMNIROUTE_MITM_STUB=1` (set only by the Dockerfile) through the shared `scripts/build/mitm-stub-flag.mjs` helper; default builds bundle the real manager. Regression guard: `tests/unit/mitm-stub-alias-6344.test.mjs` (4).
+
+- **fix(proxy):** stop the v3.8.44 proxy regression that leaked the real IP and disabled healthy proxies ([#6246](https://github.com/diegosouzapw/OmniRoute/issues/6246)). Two coupled defects from the new health scheduler: (1) **IP leak** — when a proxy assigned to a connection was marked `inactive`, resolution fell through to a **direct** egress instead of blocking, exposing the operator's real IP; (2) **over-deactivation** — the sweep flipped a proxy to `inactive` on the **first** failed probe and counted our own 5s timeout / a probe-target `5xx` as the proxy's fault, so healthy paid proxies vanished from egress selection ("my proxies are not being used anymore"). Fix: the sweep decision is extracted into a pure, network-free `decideProxyHealthAction` (`src/lib/proxyHealth/decision.ts`) — by default the health check now **only counts/logs and never downgrades status** (a proxy is downgraded/removed only with `PROXY_AUTO_REMOVE=true`, after `PROXY_AUTO_REMOVE_AFTER` **consecutive** conclusive failures); probes are classified tri-state so an inconclusive result (our timeout, or a `5xx` from the probe target) never penalizes the proxy, and the probe timeout is raised 5s→15s. Separately, `safeResolveProxy` now **fails closed** via the existing policy: a connection whose assigned proxy is dead is blocked instead of leaking direct (`hasBlockingProxyAssignment`), honoring the explicit `proxy off` toggles and the `PROXY_FAIL_OPEN=true` opt-out. Existing proxies stuck `inactive` by the old behavior need a one-time manual re-activate (the operator owns proxy status). Regression guards: `tests/unit/proxy-health-decide-action-6246.test.ts`, `tests/unit/proxy-assigned-unavailable-6246.test.ts`.
+
+- **fix(proxy):** make "Test All" read-only and add bulk enable/disable ([#6246](https://github.com/diegosouzapw/OmniRoute/issues/6246)). Complements the core fail-closed / scheduler fix (#6296) with the two remaining reporter asks. (1) The **"Test All" button** (`POST /api/settings/proxies/auto-test`) used to flip a proxy to `inactive` on a failed reachability probe; since the egress selector excludes `inactive` proxies, a flaky probe (an unreachable `httpbin.org`, a proxy that blocks `HEAD`, or a slow paid proxy) silently disabled every proxy that failed — "Test All" is now **read-only by default** (only the operator sets a proxy active/inactive; opt back into the legacy test-and-set with `PROXY_HEALTH_AUTO_DEACTIVATE=true`). (2) Adds a **bulk enable/disable** proxies endpoint + toolbar action (`POST /api/settings/proxies/batch-activate`) so an operator can re-activate proxies in one click. Regression guard: `tests/unit/proxy-health-6246.test.ts`. (thanks @tenshiak)
+
+- **chatcore (tools): stop the default 128-tool cap from silently dropping opencode's `task`/MCP tools.** opencode (used as an MCP/agent host) sends a large tool list; when it exceeds the speculative `MAX_TOOLS_LIMIT` (128) default, `truncateToolList` did a blind `tools.slice(0, 128)`, dropping every tool past index 128 — including opencode's built-in `task` tool (subagent launch) and many MCP tools, so models routed through OmniRoute could no longer spawn subagents or reach part of their tools. The cap exists to avoid upstream `400`s for providers with real hard limits (e.g. grok-cli 200), so it is kept for those: detection of the opencode client (`isOpencodeClient` — any `x-opencode-*` header, or `opencode` in the user-agent) now only bypasses the **speculative 128 default**, never a known provider ceiling. Precedence is explicit — a proactive/detected provider limit always truncates (even for opencode); otherwise opencode forwards its full tool list; otherwise the unchanged 128 default applies to every other client. Refactors `getEffectiveToolLimit` into `getKnownToolLimit(provider) ?? DEFAULT_LIMIT` (byte-identical for existing callers) and fixes a cosmetic debug-log that reported the truncated count instead of the original. Regression guard: `tests/unit/tool-limit-detector.test.ts`.
+
+- **fix(mitm):** the macOS MITM-cert install check now matches the system keychain again. `security find-certificate -a -Z` prints the SHA-1 as a colon-less hex string, but the installed-check compared it against `getCertFingerprint()`'s colon-separated form, so the substring match never hit — the cert was reported as not-installed and re-prompted for the sudo install on every run. Fingerprints are now normalized (colons stripped, upper-cased) on both sides via the extracted `macCertOutputHasFingerprint` helper. Regression guard: `tests/unit/mitm-cert-mac-fingerprint.test.ts`. ([#6204](https://github.com/diegosouzapw/OmniRoute/pull/6204), closes [#6134](https://github.com/diegosouzapw/OmniRoute/issues/6134) — thanks @rianonehub)
+
+- **fix(api):** `/v1/messages/count_tokens` now counts `tool_use`, `tool_result` and `thinking` content blocks (and array-form `system` prompts) in the local-estimation path, instead of only `text`. Real agentic conversations keep ~95% of their tokens inside tool results; the previous estimate returned near-zero for them, which silently broke Claude Code's auto-compaction (context grew past the window with no compaction until the upstream API rejected the request). The real provider-side count path is unchanged. Regression guard: `tests/unit/messages-count-tokens-route.test.ts`. ([#6221](https://github.com/diegosouzapw/OmniRoute/pull/6221) — thanks @luweiCN)
+
+- **fix(antigravity):** strip a trailing assistant prefill turn for Vertex Claude models to avoid upstream 400s ([#6114](https://github.com/diegosouzapw/OmniRoute/pull/6114)). Regression guard: `tests/unit/antigravity-claude-prefill-strip.test.ts`. (thanks @anki1kr)
+
+- **fix(security):** the mutable cloud-agent routes (`/api/cloud/credentials/update`, `/api/cloud/models/alias`) now require management auth instead of being treated as public. They were classified as public API routes, so a request without management credentials could update stored cloud-agent credentials and model aliases. They are removed from the public-route set, classified as management routes in the authz pipeline, and gated by `requireManagementAuth`; cloud **read**/auth routes stay public. Regression guards: `tests/unit/cloud-write-auth.test.ts`, `tests/unit/authz/classify.test.ts`, `tests/unit/public-api-routes.test.ts`. ([#6233](https://github.com/diegosouzapw/OmniRoute/pull/6233) — thanks @vittoroliveira-dev)
+
+- **refactor(dashboard):** extract the onboarding-wizard "Open provider details" link target into a pure, unit-tested `buildProviderDetailsHref(connection)` helper. The wizard already routes by `connection.id` (the node UUID) rather than the provider category slug (#6144/#6145); this hardens that behavior behind a tested helper that guards a missing id/connection. Regression guard: `tests/unit/provider-onboarding-href.test.ts`. ([#6166](https://github.com/diegosouzapw/OmniRoute/pull/6166) — thanks @KooshaPari)
+
+- **fix(security):** the doubao synthetic device-id generator now derives its digits via an unbiased crypto-random draw (rejection sampling over `crypto.randomBytes()`) instead of a `% 10` reduction, closing a CodeQL `js/biased-cryptographic-random` finding.
+- **fix(agentSkills):** the GitHub-skills generator now resolves `outputDir` to an absolute path before writing, fixing a regression introduced by #6366 (relative-to-cwd base path) that could write generated skill files to the wrong directory.
+- **fix(security):** `/api/keys/{id}/devices` now checks the HTTP method before auth/validation, returning a `405` for non-GET/DELETE verbs instead of a misleading `401`/`500` (closes a `dast-smoke` QUERY-method finding).
+- **fix(quality):** clear the last 2 heavy quality-gate reds on the release tip (cycle pre-flight).
+- **fix(mitm):** the test suite and CI can never mutate the OS trust store — `OMNIROUTE_SKIP_SYSTEM_TRUST=1` is set globally for tests/CI so `installCert`/`uninstallCert`/`installTproxyCa` skip the privileged OS dispatch ([#6310](https://github.com/diegosouzapw/OmniRoute/pull/6310); full detail is under the [3.8.45] section below — this branch received it via the parallel-cycle sync-back).
+- **fix(api):** `POST /api/github-skills` now Zod-validates its request body; documented the new quality-gate env vars and pinned the merge-integrity GitHub Actions to a commit SHA.
+- **fix(skills):** generate the missing `omni-github-skills` registry entry and align the agent-skills catalog-count tests (follow-up to #6186).
+- **fix(quality):** clear the cycle's 11 net-new ESLint errors and make `validate-release-green` suppressions-aware.
+
+### 📝 Maintenance
+
+- **i18n(it):** add 118 missing Italian (`it`) translations (net-additive — no existing keys dropped, valid JSON), improving Italian UI coverage. ([#6212](https://github.com/diegosouzapw/OmniRoute/pull/6212) — thanks @serverless83)
+- **chore(providers):** remove deprecated **MiMo V2** model entries from the catalogs (xiaomi-mimo, opencode-go, zenmux-free, audio TTS) — the upstream V2 line is superseded by MiMo V2.5; drops `mimo-v2-tts`, `mimo-v2-pro`, `mimo-v2-omni`, `mimo-v2-flash`, `mimo-v2-flash-free` and realigns the provider-catalog tests. ([#6248](https://github.com/diegosouzapw/OmniRoute/pull/6248) — thanks @backryun)
+- **chore(release):** ~50 commits on this branch are v3.8.45 pre-flight/hardening fixes and CI-perf work that landed here via the parallel-cycle sync-back (`sync-next-cycle.mjs`, Hard Rule #21) **after** the `v3.8.45` git tag was cut, and are already fully documented under the **[3.8.45]** section below — listed here only so the per-cycle commit-coverage check (`npm run release:uncovered`) doesn't flag them as gaps. Provider/catalog/UX/backend fixes: #6041, #6078, #6108, #6135, #6148, #6149, #6154, #6158, #6161, #6162, #6163, #6164, #6165, #6170, #6177, #6178, #6181, #6186, #6187, #6191, #6193, #6194, #6195, #6200, #6205, #6208, #6209, #6211, #6213, #6223, #6224, #6225, #6226, #6227, #6228, #6229, #6230, #6235, #6291, #6292. CI/release-pipeline work: #6167, #6203, #6214, #6215, #6218, #6273, #6275, #6283, #6284, #6285, #6300, #6305.
+- **chore(release):** additional zero-ref release-cycle plumbing on this branch, kept out of `release:uncovered` on purpose (no `#N` in the commit subject to cite): opening the v3.8.46 cycle, opening/closing the v3.8.45 cycle, the finalized [3.8.45] CHANGELOG i18n sync-back to 42 mirrors, the v3.8.45 cognitive/cyclomatic and file-size drift rebaselines, ESLint stale-suppression pruning (4,273 → 4,233), and clearing test-masking/docs-all pre-flight reds for v3.8.45.
+
+### ⚡ Performance & Infrastructure
+
+- **perf(release-green):** the pre-flight validator (`scripts/quality/validate-release-green.mjs`) now runs its 4 slow suites (unit / vitest / integration / pack-artifact) **concurrently** via `Promise.all` — pre-flight wall time drops from ~the sum of the suites to ~the slowest one (~30min saved per round; Phase 0 was the nº1 bottleneck of the v3.8.45 release benchmark, 2h54 of 6h34 e2e). Guard: `tests/unit/validate-release-green.test.ts` ("runs the slow suites CONCURRENTLY"). ([#6319](https://github.com/diegosouzapw/OmniRoute/pull/6319))
+- **fix(ci):** `scripts/release/sync-next-cycle.mjs` — two defects found live in its first production run (v3.8.45 Phase 5): (1) the `git()` helper's default 1 MiB `maxBuffer` crashed with `ENOBUFS` on `git show origin/main:CHANGELOG.md` (the CHANGELOG alone is >1 MiB) — widened to 64 MiB; (2) the i18n resync only propagated the `[NEXT]` (TBD) section, leaving the just-shipped finalized section as "— TBD" in all 42 mirrors — it now also syncs `[prevVersion]` bounded by the heading below it (new exported pure helper `versionAfter`). Guards: +5 tests in `tests/unit/sync-next-cycle.test.ts` (8/8). ([#6327](https://github.com/diegosouzapw/OmniRoute/pull/6327))
+- **test(ci):** concurrency-sensitive flaky tests are **quarantined into a serial pass** (`tests/unit/serial/`, `--test-concurrency=1`, appended to every unit runner incl. sharded variants — the serial pass is sharded too so concurrent shard jobs never self-collide). Initial set: `glm-coding-plan-monthly-3580`, `quota-division-blocks`, `provider-health-autopilot`, `combo-health-autopilot` — the class behind the ~28min CI wedges/re-runs (two live 1h+ wedges cancelled during this PR's own validation). Discovery + TIA gates track the new glob; systemic root cause (async logger writing after teardown) tracked in [#6360](https://github.com/diegosouzapw/OmniRoute/issues/6360). Guard: `tests/unit/test-serial-quarantine.test.ts` (4). ([#6347](https://github.com/diegosouzapw/OmniRoute/pull/6347))
+
+### 🙌 Contributors
+
+Thanks to everyone whose work landed in v3.8.46:
+
+| Contributor                                                    | PRs / Issues           |
+| -------------------------------------------------------------- | ---------------------- |
+| [@2220258345](https://github.com/2220258345)                   | direct commit / report |
+| [@abdofallah](https://github.com/abdofallah)                   | direct commit / report |
+| [@adentdk](https://github.com/adentdk)                         | direct commit / report |
+| [@aleksesipenko](https://github.com/aleksesipenko)             | direct commit / report |
+| [@anki1kr](https://github.com/anki1kr)                         | direct commit / report |
+| [@anungma](https://github.com/anungma)                         | direct commit / report |
+| [@arpicato](https://github.com/arpicato)                       | direct commit / report |
+| [@backryun](https://github.com/backryun)                       | #6248                  |
+| [@binsarjr](https://github.com/binsarjr)                       | direct commit / report |
+| [@brightfiscalband](https://github.com/brightfiscalband)       | direct commit / report |
+| [@chirag127](https://github.com/chirag127)                     | #6501, #6506           |
+| [@developerjillur](https://github.com/developerjillur)         | direct commit / report |
+| [@dilneiss](https://github.com/dilneiss)                       | #6499                  |
+| [@dtybnrj](https://github.com/dtybnrj)                         | direct commit / report |
+| [@Forcerecon](https://github.com/Forcerecon)                   | direct commit / report |
+| [@hao3039032](https://github.com/hao3039032)                   | direct commit / report |
+| [@Iammilansoni](https://github.com/Iammilansoni)               | #6150, #6245           |
+| [@jmengit](https://github.com/jmengit)                         | direct commit / report |
+| [@jordansilly77-stack](https://github.com/jordansilly77-stack) | direct commit / report |
+| [@JxnLexn](https://github.com/JxnLexn)                         | direct commit / report |
+| [@KooshaPari](https://github.com/KooshaPari)                   | #6166                  |
+| [@loopyd](https://github.com/loopyd)                           | direct commit / report |
+| [@luweiCN](https://github.com/luweiCN)                         | #6221                  |
+| [@makcimbx](https://github.com/makcimbx)                       | direct commit / report |
+| [@muflifadla38](https://github.com/muflifadla38)               | direct commit / report |
+| [@newnol](https://github.com/newnol)                           | direct commit / report |
+| [@ofekbetzalel](https://github.com/ofekbetzalel)               | direct commit / report |
+| [@ohahe52-dot](https://github.com/ohahe52-dot)                 | direct commit / report |
+| [@phidinhmanh](https://github.com/phidinhmanh)                 | direct commit / report |
+| [@powellnorma](https://github.com/powellnorma)                 | direct commit / report |
+| [@qpeyba](https://github.com/qpeyba)                           | direct commit / report |
+| [@RaviTharuma](https://github.com/RaviTharuma)                 | direct commit / report |
+| [@RCrushMe](https://github.com/RCrushMe)                       | direct commit / report |
+| [@rianonehub](https://github.com/rianonehub)                   | #6134, #6204           |
+| [@serverless83](https://github.com/serverless83)               | #6212                  |
+| [@swingtempo](https://github.com/swingtempo)                   | direct commit / report |
+| [@tenshiak](https://github.com/tenshiak)                       | direct commit / report |
+| [@ThongAccount](https://github.com/ThongAccount)               | direct commit / report |
+| [@UnrealAryan](https://github.com/UnrealAryan)                 | direct commit / report |
+| [@vinayakkulkarni](https://github.com/vinayakkulkarni)         | direct commit / report |
+| [@vittoroliveira-dev](https://github.com/vittoroliveira-dev)   | #6233                  |
+| [@warelik](https://github.com/warelik)                         | direct commit / report |
+| [@xxy9468615](https://github.com/xxy9468615)                   | direct commit / report |
+| [@xz-dev](https://github.com/xz-dev)                           | direct commit / report |
+| [@yanpaing007](https://github.com/yanpaing007)                 | direct commit / report |
+| [@diegosouzapw](https://github.com/diegosouzapw)               | maintainer             |
+
+---
+
 ## [3.8.45] — 2026-07-06
 
 ### ✨ New Features

docs/i18n/ru/CHANGELOG.md

@@ -6,6 +6,213 @@
 
 ## [3.8.31] — 2026-06-20
 
+## [3.8.46] — TBD
+
+### ✨ New Features
+
+- **feat(sse):** **hide paid-only models from `auto/*` routing** when `hidePaidModels` is on ([#6512](https://github.com/diegosouzapw/OmniRoute/issues/6512)) — follow-up to #6328/#6495. PR #6495 hid paid-only models from the `GET /v1/models` listing, but `auto/*` combos (`auto/best-coding`, `auto/glm`, …) could still pick a paid-only backend into their candidate pool → a 402/403 at request time. `createVirtualAutoCombo` now filters the candidate pool through the new pure `open-sse/services/autoCombo/paidModelFilter.ts` (`filterPaidOnlyCandidates`), applying the same free-model predicate #6495 uses in `catalog.ts` (`providerHasFreeModels(provider) && isFreeModel(provider, {id})`) whenever `settings.hidePaidModels === true`. Applied before the category/tier/family narrowing, so it covers every `auto/*` combo; an all-paid pool degrades to the existing graceful empty-pool path. **Opt-in — default OFF leaves the pool unchanged** (identity). Regression guard: `tests/unit/autoCombo/paid-model-filter-6512.test.ts` (4, incl. the default-off identity guard).
+- **feat(sse):** **provider-family auto combos** — `auto/glm`, `auto/minimax`, `auto/mimo`, `auto/zai`, `auto/gemma`, `auto/llama`, `auto/gemini` ([#6453](https://github.com/diegosouzapw/OmniRoute/issues/6453)) — new routable ids that materialize an on-demand virtual combo spanning whatever installed backends currently expose that model family, degrading gracefully as backends rotate. A new pure `open-sse/services/autoCombo/modelFamily.ts` (`detectModelFamily`) classifies by model-id prefix for six families; `zai` is instead resolved by provider id (z.ai's hosted API serves the same `glm-*` model ids as every other GLM backend, so `auto/zai` means "route to my z.ai backend specifically" vs `auto/glm`'s "any connected GLM backend"). Reuses the existing `createVirtualAutoCombo` on-demand materialization path (no DB writes) and the `/v1/models` catalog advertising loop. Regression guard: `tests/unit/autoCombo/provider-family-combos.test.ts` (11).
+- **feat(proxy):** native **proxy-pool round-robin / egress IP rotation** ([#6365](https://github.com/diegosouzapw/OmniRoute/issues/6365)) — a scope (global / provider / account) can now hold **multiple** proxies as a pool with a rotation strategy, so outbound requests cycle their egress IP instead of pinning one proxy per scope. Migration `117_proxy_pool_rotation.sql` lifts the `UNIQUE(scope, scope_id)` constraint (rebuild via the canonical rename/copy/drop; existing single assignments become 1-element pools) and adds a `proxy_scope_rotation` companion table holding the per-scope strategy + a persisted monotonic round-robin cursor. Strategies: `round-robin` (default, monotonic cursor — never `Math.random`), `random`, and `sticky-per-N-min`. Resolution (`resolveProxyForScopeFromRegistry` / `resolveProxyForConnectionFromRegistry`) now fetches the alive, position-ordered candidate set (unchanged `PROXY_ALIVE_PREDICATE`) and applies the strategy; an empty / all-dead pool still returns `null` — the #6246 fail-closed guard is **untouched** (never falls through to direct egress). Backend + DB only; dashboard pool-builder UI is a follow-up. Regression guard: `tests/unit/proxy-pool-rotation-6365.test.ts` (8, incl. fail-closed + backward-compat).
+- **feat(providers):** end-to-end **tool/function calling on the native Gemini `/v1beta` endpoint** ([#6222](https://github.com/diegosouzapw/OmniRoute/issues/6222)) — both directions of the Gemini↔OpenAI conversion now preserve tool data (previously silently dropped). Request side: `convertGeminiToInternal` (extracted to its own testable module) maps `tools[].functionDeclarations` → OpenAI `tools`, prior `functionCall` parts → assistant `tool_calls`, and `functionResponse` parts → `tool`-role messages. Response side: `convertOpenAIResponseToGemini` emits `parts[].functionCall {name,args}` from `message.tool_calls`, and the streaming `openAIChunkToGeminiChunk` accumulates fragmented `tool_calls` deltas by index into complete `functionCall` parts. The non-Gemini client paths (Claude, OpenAI-Responses) already preserved tool calls — this closes the gap specific to the native Gemini surface. Regression guard: `tests/unit/v1beta-gemini-tool-calling-6222.test.ts` (6, incl. a streaming SSE round-trip).
+- **feat(providers):** copilot-m365-web **enterprise / work tier** support ([#6334](https://github.com/diegosouzapw/OmniRoute/issues/6334)) — mirrors the EDU-tier pattern (#6210): `M365ConnectionParams` gains an `agent` field, a new opt-in `M365_ENTERPRISE_OVERRIDES` preset (`agent=work`, `scenario=officeweb`, `licenseType=Premium`) applies via `providerSpecificData.tier="enterprise"` (alias `"work"`), and `agent` is also overridable directly via `providerSpecificData.agent`. `buildWsUrl` was hardcoding `agent="web"` (the one enterprise-distinguishing param with no override path), so a Premium work account handshook then returned an empty stream. The individual and EDU paths are untouched. Kilo's dup flag vs #6210 (EDU tier) was a false positive — different tier. Regression guard: `tests/unit/copilot-m365-enterprise-6334.test.ts` (7). End-to-end confirmation on a real Premium work account is a live-VPS validation follow-up (Hard Rule #18). (thanks @Forcerecon)
+- **feat(api):** standardized, provider-agnostic **`effort` + `thinking` request params** ([#6241](https://github.com/diegosouzapw/OmniRoute/issues/6241)) — a thin standardization layer over the existing mature per-provider reasoning plumbing (no provider mapper touched). `providerChatCompletionSchema` gains a canonical `effort` (reusing the shared `none/low/medium/high/xhigh` vocabulary — the UI tiers `extra`/`max` collapse onto `xhigh`) and a boolean `thinking`. A pure `normalizeReasoningRequest` (wired once in `src/sse/handlers/chat.ts`, before any reasoning field is read) folds them onto the fields the translators already consume (`reasoning_effort` / `reasoning.effort` / `thinking`), so they fan out to Anthropic / Gemini / xAI / Responses — an explicit client `reasoning_effort` / object-shaped `thinking` always wins (backward-compatible). `/models` additively exposes `supportsThinking` + `effort_tiers` so the frontend can render the toggles (UI component is a follow-up). Regression guard: `tests/unit/effort-thinking-standardization-6241.test.ts` (12). (thanks @Iammilansoni, @shabeer)
+- **feat(combo):** new **`pipeline` (sequential) combo strategy** ([#6297](https://github.com/diegosouzapw/OmniRoute/issues/6297)) — the 18th routing strategy runs targets **in order**, threading each step's output into the next step's input, with an optional per-step `prompt` (system instruction); only the final step's response is returned. Distinct from `fusion` (parallel fan-out + judge). Implemented as a self-contained `open-sse/services/pipeline.ts` (sibling to `fusion.ts`), dispatched from `combo.ts`; the step list reuses `combo.models` order and reads an optional `prompt` off each target (backward-compatible — ignored by every other strategy). Intermediate steps run non-streaming with tools stripped (complete prose to thread forward); the final step keeps the client's `stream` flag + tools. A failing/empty/unparseable intermediate step fails the whole pipeline explicitly via a sanitized error (never silently swallowed). Kilo's dup flag vs #563 was a false positive (that's model→chain selection; this is a sequential chain). Regression guard: `tests/unit/combo-pipeline-strategy.test.ts` (5). (thanks @ofekbetzalel)
+- **feat(ci):** `check:test-masking` now flags **inline-reimplemented prod conditions** ([#6348](https://github.com/diegosouzapw/OmniRoute/issues/6348)) — a new **report-only** subcheck (v2, 6A.10 family) catches the wrong-shape contract test: a test that recomputes the condition under test inline instead of importing/exercising the real function (the #6216 class, where `=== 500` → `>= 500` stayed green because the test re-implemented the branch). For each added/modified test file it warns when the file textually duplicates a ≥3-token conditional from a production file touched in the same PR **and** does not import the symbol/module owning it, via a pure, fixture-tested `findReimplementedConditions()` with an allowlist mirroring `assertReductionAllowlist`. Report-only for now (does not fail the gate) — to be promoted to blocking after a triage cycle. Regression guard: `tests/unit/check-test-masking.test.ts` (45).
+- **feat(sse):** per-connection routing override (native vs CLIProxyAPI) ([#6339](https://github.com/diegosouzapw/OmniRoute/issues/6339)) — the previously-dead `isCliproxyapiDeepModeEnabled` helper is now wired into `resolveExecutorWithProxy`: a single connection can opt itself into the CLIProxyAPI passthrough executor via `providerSpecificData.cliproxyapiMode="claude-native"`, with precedence **connection override > provider `upstream_proxy_config` mode > default**. `resolveExecutorWithProxy` now receives the resolved connection's `providerSpecificData` (threaded from `chatCore.ts`), so one connection can deep-route while the provider's other connections stay native — no DB schema change (the toggle rides in `providerSpecificData`). Also resolves the same-provider mixing ask in #6340. Regression guard: `tests/unit/chatcore-executor-proxy.test.ts` (9). (thanks @RaviTharuma)
+- **feat(dashboard):** "Add session cookie" modal now shows a prominent **"Open ‹host› →"** link to the provider's own site ([#6268](https://github.com/diegosouzapw/OmniRoute/issues/6268)) — every `-web` cookie-session provider (chatgpt-web, claude-web, gemini-web, kimi-web, lmarena, qwen-web, m365-copilot-web, …) renders a one-click external link (opening the provider's login/home page in a new tab) so operators no longer tab away to retype the URL mid-setup. The host resolves from a pure, unit-tested `resolveWebProviderHost()` (prefers `WEB_COOKIE_PROVIDERS[id].website`, falls back to the registry `baseUrl` origin); non-web providers render exactly as before. Kilo's dup flag vs #6265 (modal-too-small-on-1080p) was a false positive — distinct concern. Regression guard: `tests/unit/resolve-web-provider-host.test.ts` (5). (thanks @chirag127)
+- **feat(providers):** add **DigitalOcean** AI (serverless inference) as an OpenAI-compatible API-key provider ([#6373](https://github.com/diegosouzapw/OmniRoute/pull/6373)) — base `https://inference.do-ai.run/v1`, wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/digitalocean/`, `src/shared/constants/providers/apikey/inference-hosts.ts`). Regression guard: `tests/unit/digitalocean-provider.test.ts`. (thanks @newnol)
+- **feat(providers):** add **Huancheng Public API** (`hcnsec`) as an OpenAI-compatible regional provider ([#6410](https://github.com/diegosouzapw/OmniRoute/pull/6410)) — Xinjiang Huancheng Cybersecurity's public LLM platform (base `https://api.hcnsec.cn/v1`, free credits via daily check-ins), wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/hcnsec/`, `src/shared/constants/providers/apikey/regional.ts`). Regression guard: `tests/unit/hcnsec-provider.test.ts`. (thanks @UnrealAryan)
+- **feat(dashboard):** the web-session credential guide now shows an **"Open {host}" link** ([#6316](https://github.com/diegosouzapw/OmniRoute/pull/6316)) to the provider's sign-in site (derived from the provider `website` via `getProviderWebsiteHost`), so you can jump straight to the page where the cookie/session must be captured. Regression guard: `tests/unit/web-session-provider-link-6316.test.ts`. (thanks @jordansilly77-stack)
+- **feat(cerebras):** add the **Gemma 4 31B** model (`gemma-4-31b`) to the Cerebras registry + pricing table ([#6331](https://github.com/diegosouzapw/OmniRoute/pull/6331)). Regression guard extends `tests/unit/t28-model-catalog-updates.test.ts`. (thanks @backryun)
+- **feat(providers):** add **Yuanbao (web)** as a cookie-session provider ([#6196](https://github.com/diegosouzapw/OmniRoute/issues/6196)) — `yuanbao-web` (Tencent Yuanbao, `yuanbao.tencent.com`) with cookie-only auth (`hy_user`/`hy_token` + public agent id), SSE→OpenAI translation incl. `reasoning_content`, exposing DeepSeek V3/R1 + Hunyuan / Hunyuan-T1. Regression guard: `tests/unit/providers-yuanbao-web.test.ts`. `together-web` was **deferred** (no verifiable web-session endpoint — needs a captured request) and `huggingchat-web` **dropped** (the existing `huggingchat` already is a web-cookie provider). (thanks @chirag127)
+- **feat(providers):** route the built-in **agentrouter** through the dynamic Claude-Code wire image ([#6056](https://github.com/diegosouzapw/OmniRoute/issues/6056)) — a small static allow-set (`CC_WIRE_IMAGE_BUILTINS` in `open-sse/services/ccWireImageBuiltins.ts`), consulted by `isClaudeCodeCompatible` / `isClaudeCodeCompatibleProvider` / `applyFingerprint`, makes agentrouter adopt the CC wire-image headers + fingerprint **while guarding the CC baseUrl/auth branches** so it keeps its own registry `baseUrl` and `x-api-key` auth. Regression guard: `tests/unit/agentrouter-cc-wire-image.test.ts` (asserts the wire image is applied AND agentrouter's baseUrl/auth are preserved). Live WAF-acceptance against agentrouter.org is a VPS validation follow-up (Hard Rule #18).
+- **feat(providers):** **bulk-add API keys for Cloudflare Workers AI** ([#6174](https://github.com/diegosouzapw/OmniRoute/issues/6174)) — `cloudflare-ai` is removed from the bulk-add exclusion list and the bulk parser gains a 3-field `name|accountId|apiKey` mode; the bulk route now builds a **per-entry** `providerSpecificData` so each key carries its own `accountId` (fixing the previous shared-object reuse), and both the create + key-validation paths receive it. Regression guard: `tests/unit/bulk-api-key-parser-cloudflare.test.ts`. (thanks @muflifadla38)
+- **feat(dashboard):** routing/settings UX clarity ([#6147](https://github.com/diegosouzapw/OmniRoute/issues/6147)) — (1) weighted combos show the **effective routing share %** next to each weight when weights don't sum to 100 (`WeightTotalBar.tsx`); (2) the status widget's user-facing **"Cloud Sync" label is renamed** to "Remote Settings Sync" (`CloudSyncStatus.tsx`; internal ids/state untouched); (3) built-in providers gain an **opt-in advanced base-URL override** (`isBaseUrlOverrideEligibleProvider`, hidden behind an "Advanced" toggle, reusing the existing `providerSpecificData.baseUrl` persistence — not globally widened). Regression guard: `tests/unit/routing-settings-ux-6147.test.ts`.
+- **feat(combo):** add an option to **disable session stickiness**, per-combo or globally — round-robin / random combos can rotate to a different connection on every request instead of pinning a whole conversation to one connection by its first-message hash. Resolution precedence per-combo `config.disableSessionStickiness` → global `settings.disableSessionStickiness` → default `false` (preserves the #3825 prompt-cache/504 fix); gates **both** stickiness call sites in `open-sse/services/combo.ts`. Exposed as a global toggle (Combo Defaults) and a per-combo Inherit/on/off control. ([#6168](https://github.com/diegosouzapw/OmniRoute/issues/6168)) Regression guard: `tests/unit/combo-disable-session-stickiness.test.ts`. (thanks @RCrushMe)
+- **feat(docker):** add the `OMNIROUTE_NO_SUDO` env flag for root-less / user-namespaced deployments — the MITM cert-trust command path (`resolveSudoSpawn` in `src/mitm/systemCommands.ts`) now strips the leading `sudo` when the flag is truthy, in addition to the existing root / sudo-missing cases, so the Proxy Agent runs without `sudo` (the operator trusts the CA manually, e.g. via `NODE_EXTRA_CA_CERTS`). Argv-array `spawn` preserved — no shell interpolation (Hard Rule #13). ([#6122](https://github.com/diegosouzapw/OmniRoute/issues/6122)) Regression guard: `tests/unit/mitm-systemCommands-no-sudo.test.ts`. (thanks @powellnorma)
+- **feat(providers):** add **Requesty** as an OpenAI-compatible gateway provider (BYOK, base `https://router.requesty.ai/v1`, ~200 free requests/day) — wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/requesty/`, `src/shared/constants/providers/apikey/gateways.ts`). ([#6120](https://github.com/diegosouzapw/OmniRoute/issues/6120)) Regression guard: `tests/unit/requesty-provider.test.ts`. (thanks @chirag127)
+- **feat(dashboard):** add **configured-only / available-only filters** to the Free Provider Rankings page ([#6150](https://github.com/diegosouzapw/OmniRoute/issues/6150)) — hide providers you haven't configured, or whose connections are all rate-limited / out of quota, via server-side query params (`?configuredOnly` / `?availableOnly` on `GET /api/free-provider-rankings`) backed by a testable lib helper reusing the in-process connection state (no Redis). Both filters default off, so the default view is unchanged; this supersedes the earlier client-side "Configured Only" toggle (#6245) with an available-only dimension and unit-tested logic. Regression guard: `tests/unit/freeProviderRankings-filters.test.ts`.
+
+### 🔧 Bug Fixes
+
+- **fix(dashboard):** adding a second API key connection for the same provider no longer silently overwrites the first — the Add-API-key modal now derives a unique default connection name (`main`, then `main-2`, `main-3`, …) so the backend name-based upsert can't collide (#6499 — thanks @dilneiss).
+- **fix(compression):** the session-dedup engine now also deduplicates a large multi-line block repeated **within a single message** (intra-message dedup), not just across turns; the compression-preview API surfaces a `fallbackReason`, and the fusion panel reports how many models were rate-limited vs failed on total-panel failure (#6501 — thanks @chirag127).
+- **fix(compression):** a stacked-pipeline step naming an unregistered engine now surfaces a `validationErrors` entry instead of silently no-op'ing, so misconfigured pipelines are visible in the preview API (#6506 — thanks @chirag127).
+- **feat(usage):** add a **Codex reset-credit redemption** flow to the Provider Limits UI ([#6361](https://github.com/diegosouzapw/OmniRoute/pull/6361)) — a `useCodexResetCreditRedemption` hook + `/api/usage/codex-reset-credit` route + `codexResetCredits` lib let you redeem banked Codex reset credits from the quota card. Regression guard: `tests/unit/codex-reset-credits.test.ts`. (thanks @JxnLexn)
+- **feat(glm):** add **team-plan quota settings** for `glm-cn` connections ([#6351](https://github.com/diegosouzapw/OmniRoute/pull/6351)) — a dedicated `GlmTeamQuotaFields` form section (team quota id / limits) threaded through the Add/Edit connection modals, persisted via `providerSpecificData`, with the GLM usage service reading the team quota. Regression guards: `tests/unit/glm-team-quota.test.ts`, `provider-specific-data-schema.test.ts`. (thanks @hao3039032)
+- **feat(providers):** add **TinyFish** web-fetch/search support ([#6349](https://github.com/diegosouzapw/OmniRoute/pull/6349)) — a `tinyfish-fetch` executor + `/v1/web/fetch` route + MCP web-fetch tool, registered as a specialty-media provider with request-validation and a search-provider catalog entry. Regression guards: `tests/unit/executor-tinyfish-fetch.test.ts`, `web-fetch-handler.test.ts`, `mcp-web-fetch-tool.test.ts`, `provider-validation-tinyfish.test.ts`. (thanks @dtybnrj)
+- **fix(cli):** `omniroute launch-codex` now spawns `codex.cmd` through a shell on Windows (the npm `.cmd` shim is unresolvable by bare `spawn` → ENOENT), mirroring the qodercli Windows fix (#6263) ([#6312](https://github.com/diegosouzapw/OmniRoute/pull/6312)). Regression guard: `tests/unit/launch-codex-windows-spawn-6312.test.ts`. (thanks @swingtempo)
+- **fix(codex):** isolate the **Spark** quota from the shared Codex quota and stabilize the quota UI ordering / hydration so per-scope limits render consistently ([#6336](https://github.com/diegosouzapw/OmniRoute/pull/6336)). Regression guards: `tests/unit/codex-quota-selection-hydration.test.ts`, `provider-limits-ui.test.ts` + 3 more. (thanks @xz-dev)
+- **feat(api):** add a `hidePaidModels` setting that filters paid-only models out of the `/v1/models` catalog. Regression guard: `tests/unit/models-catalog-hide-paid.test.ts`. (thanks @chirag127)
+- **fix(api-manager):** the fallback model picker now preserves combos instead of dropping them when a primary model is unavailable ([#6443](https://github.com/diegosouzapw/OmniRoute/pull/6443)). Regression guard: `tests/unit/api-manager-page-static.test.ts`. (thanks @jmengit)
+- **fix(providers):** recoverable Antigravity / Cloud-Code (Gemini Code Assist) `403` responses ([#6452](https://github.com/diegosouzapw/OmniRoute/pull/6452)) are now classified as a retryable project-config error instead of a terminal account ban, so a fixable project/API-disabled 403 no longer forces a ~1-year cooldown / full OAuth reconnect. Regression guard: `tests/unit/errorclassifier-antigravity-403.test.ts`. (thanks @developerjillur)
+- **fix(mitm):** `sanitizeHeaders` now redacts `Set-Cookie` response headers so upstream session cookies never leak into logs / diagnostics ([#6451](https://github.com/diegosouzapw/OmniRoute/pull/6451)). Regression guard: `tests/unit/mitm-sanitize-headers.test.ts`. (thanks @developerjillur)
+- **fix(api):** `/api/compression/preview` now accepts `mode: "caveman"` and correctly handles stacked / zero-compression previews ([#6425](https://github.com/diegosouzapw/OmniRoute/issues/6425)). Regression guard: `tests/unit/api/compression-preview-caveman-and-stacked-6425.test.ts`. (thanks @chirag127)
+- **feat(providers):** add **Zed** hosted LLM aggregator as a native-app provider ([#6118](https://github.com/diegosouzapw/OmniRoute/pull/6118)) — OAuth sign-in via the Zed hosted flow, registered through the shared provider registry + executor. Regression guards: `tests/unit/zed-oauth-provider.test.ts`, `zed-import-utils.test.ts`, `zed-docker-detect.test.ts`, `mitm-handler-zed.test.ts`. VPS-validated via live operator login (Hard Rule #18).
+- **fix(oauth):** the Kiro SSO-cache auto-import now **preserves the IDC region** — cross-region Amazon Q / Kiro profiles imported from the SSO cache are no longer collapsed to the default region ([#6113](https://github.com/diegosouzapw/OmniRoute/pull/6113)). Regression guard: `tests/unit/kiro-auto-import-idc-2059.test.ts`. VPS-validated via live operator login (Hard Rule #18).
+- **fix(dashboard):** passthrough model aliases no longer collide when two namespaced model ids share a last segment (port from 9router#1850, [#6431](https://github.com/diegosouzapw/OmniRoute/pull/6431)). `enx/gpt-5.5` and `enx/codebuddy/gpt-5.5` both auto-generated the alias `gpt-5.5`, so the second model could never be added (the UI just alerted "alias already exists"). Aliases are now disambiguated deterministically — bare last segment when free, then parent-qualified (`codebuddy-gpt-5.5`), then a numeric suffix — while re-adding the exact same model id is still blocked. Regression guard: `tests/unit/passthrough-alias-1850.test.ts`. (thanks @arpicato)
+- **fix(translator):** preserve a Gemini `functionResponse` co-located with other parts (another `functionCall`, or trailing `text`) in the same content when translating **Gemini → OpenAI** ([#6376](https://github.com/diegosouzapw/OmniRoute/pull/6376)). `convertGeminiContent()` early-returned the tool message on the first `functionResponse` part, dropping any co-located parts; such contents are now pre-split (one tool message per `functionResponse`, emitted first, plus one message for the remaining parts). Regression guard: `tests/unit/gemini-to-openai-function-response.test.ts`. (thanks @warelik)
+- **fix(headroom):** detect a python interpreter managed by **mise / pyenv / asdf / conda** (port from 9router#2353, [#6382](https://github.com/diegosouzapw/OmniRoute/pull/6382)). Headroom's python probe (`src/lib/headroom/detect.ts`) searched a hardcoded `PATH`, but version managers expose their interpreters via shim dirs that only join `PATH` through interactive-shell activation — which the non-interactive server never runs, so a managed python (≥3.10) was invisible and Headroom reported it missing. The search path now prepends the well-known shim/bin dirs (`~/.local/share/mise/shims`, `~/.pyenv/shims`, `~/.asdf/shims`, `$CONDA_PREFIX/bin`, `~/.local/bin`, respecting `MISE_DATA_DIR`/`PYENV_ROOT`/`ASDF_DATA_DIR` when set), and a new `HEADROOM_PYTHON` env override lets operators point straight at their interpreter (mirroring `HEADROOM_URL`). Still shell-free (`execFileSync`). Regression guard: `tests/unit/headroom-detect.test.ts` (5). (thanks @loopyd)
+- **fix(executors):** strip the OpenAI-Codex/Claude-CLI `client_metadata` passthrough field for **NVIDIA** requests (port from 9router#1887, [#6411](https://github.com/diegosouzapw/OmniRoute/pull/6411)). NVIDIA's OpenAI-compatible wrapper rejects it with `400 Unsupported parameter`, the same class already handled for `cerebras`/`mistral`; `nvidia` (executor `default`) was missing from the strip allowlist so Codex/Claude-Code passthrough requests 400'd. Regression guard: `tests/unit/executor-default-strip-client-metadata.test.ts` (+nvidia case). (thanks @phidinhmanh)
+- **fix(translator):** strip the Claude-style `thinking` field for **NVIDIA `z-ai/glm-5.2`** (port from 9router#2023, [#6413](https://github.com/diegosouzapw/OmniRoute/pull/6413)). NVIDIA's OpenAI-compatible wrapper 400s on `thinking` (a Claude-format client routed here leaves a `thinking:{type:"adaptive"}`); the existing strip rule only dropped `reasoning`. Same class already handled for `minimax-m2.7`. Regression guard: `tests/unit/nvidia-minimax-thinking-strip.test.ts` (+glm-5.2 case). (thanks @phidinhmanh)
+- **fix(translator):** suppress the streamed `</think>` close marker for the **Antigravity IDE** client (port from 9router#1061, [#6415](https://github.com/diegosouzapw/OmniRoute/pull/6415)). On thinking-only turns Antigravity rendered a bare `</think>` as the sole visible content, tripping its loop-detection and wasting requests. Antigravity's UA (`vscode/<v> (Antigravity/<v>)`) is added to the marker-suppress allowlist (alongside OpenCode); Claude Code / Cursor still get the marker, and `x-omniroute-thinking-marker: on` force-restores it. Regression guard: `tests/unit/think-close-marker-suppress-5245.test.ts`. (thanks @abdofallah)
+- **fix(executors):** strip nested `reasoning_content` from messages for **Mistral** (port from 9router#1649, [#6417](https://github.com/diegosouzapw/OmniRoute/pull/6417)). Mistral's API returns `422 extra_forbidden` when an assistant message carries `reasoning_content` (replayed thinking from a prior turn, e.g. via the Codex `/responses` path); the generic top-level 400 field-downgrade retry never covered the nested per-message field. `DefaultExecutor` now strips it for provider `mistral` only, so DeepSeek (which requires replayed `reasoning_content`) is unaffected. Regression guard: `tests/unit/mistral-strip-reasoning-content-1649.test.ts`. (thanks @xxy9468615)
+- **fix(executors):** strip the `client_metadata` passthrough field on the **OpenCode** path (port from 9router#1442, [#6418](https://github.com/diegosouzapw/OmniRoute/pull/6418)). OpenCode upstreams (e.g. `kimi-k2.6` via opencode-go) reject it with `400 "Extra inputs are not permitted, field: 'client_metadata'"`; the DefaultExecutor strip only covered cerebras/mistral and `OpencodeExecutor` extends `BaseExecutor` directly, so nothing removed it there. Regression guard: `tests/unit/opencode-strip-client-metadata-1442.test.ts`. (thanks @yanpaing007)
+- **fix(executors):** inject the `reasoning_content` echo for the native **Moonshot Kimi** provider (port from 9router#1480, [#6419](https://github.com/diegosouzapw/OmniRoute/pull/6419)). Kimi (executor `default`) is a thinking-mode upstream that 400s with "reasoning_content must be passed back" when a prior assistant turn lacks it; the placeholder injection was only wired into the OpenCode meta-provider, so direct multi-turn Kimi conversations failed. Scoped to `kimi` (gateway-served models matching the thinking-model name pattern are unaffected). Regression guard: `tests/unit/kimi-native-reasoning-injected-1480.test.ts`. (thanks @2220258345)
+- **fix(executors):** recover from a strict gateway's `context_management: Extra inputs are not permitted` 400 (port from 9router#1468, [#6420](https://github.com/diegosouzapw/OmniRoute/pull/6420)). **Claude Code** always sends a top-level `context_management` field; strict anthropic-compatible gateways reject it. The dedicated context-editing 400-fallback only fired when OmniRoute's own `contextEditing` feature was enabled (default off), so a client-sent field passed through untouched and 400'd. `context_management` is now in the generic reactive field-strip list, so it's stripped-and-retried once regardless of the feature flag (with correct request re-signing for claude-compatible relays). Regression guard: `tests/unit/provider-field-strips.test.ts`. (thanks @ohahe52-dot)
+- **fix(network):** enable **RFC 8305 Happy Eyeballs** (`autoSelectFamily`) on the direct-egress undici dispatcher (port from 9router#1237, [#6423](https://github.com/diegosouzapw/OmniRoute/pull/6423)). When DNS returns both IPv6 (AAAA) and IPv4 (A) and the IPv6 route is broken (e.g. a NAT64 `64:ff9b::` prefix without routing), undici tried IPv6 first and hung until `ETIMEDOUT` (then a 502 + account lockout), even though `curl` reached the same host. The direct dispatcher now races both families and uses whichever connects first. Proxy paths pin family via `proxyTls` and are unaffected. Regression guard: `tests/unit/direct-dispatcher-pipelining-4580.test.ts`. (thanks @adentdk)
+- **fix(combo):** round-robin now advances the rotation pointer past the model that **actually served**, not the eagerly-scheduled one (port from 9router#948, [#6428](https://github.com/diegosouzapw/OmniRoute/pull/6428)). With `stickyLimit: 1` (true round-robin), when the scheduled model failed and a _different_ model served via fallback, the counter had already advanced +1 from the scheduled index — so the next request reused the fallback-served model, degrading round-robin into hot-spotting on whichever model was healthy. The pointer now advances to the served index + 1 (mirroring the sticky-limit>1 path). Session-stickiness (#3825) and distribution are preserved. Regression guard: `tests/unit/combo-rr-fallback-advance-948.test.ts`. (thanks @binsarjr)
+- **fix(sse):** a non-string `model` field is now rejected with a `400` before the resolver, instead of crashing downstream `.toLowerCase()`/`.split()` calls into an empty-body `500` that escapes the error sanitizer ([#6407](https://github.com/diegosouzapw/OmniRoute/issues/6407)). Regression guard: `tests/unit/chat-non-string-model-6407.test.ts`. (thanks @chirag127)
+- **fix(api):** unknown `/api/*` routes now return a JSON `404` (instead of the dashboard HTML shell) and scalar chat params (`model`/`temperature`/etc.) are validated **before** the provider lookup so malformed requests fail fast with a clear `400` ([#6424](https://github.com/diegosouzapw/OmniRoute/issues/6424), [#6412](https://github.com/diegosouzapw/OmniRoute/issues/6412)). Regression guards: `tests/unit/api/api-catchall-json-404.test.ts`, `tests/unit/chat-early-schema-validation-6412.test.ts`. (thanks @chirag127)
+- **fix(api):** `/v1/chat/completions` now rejects a non-JSON `Content-Type` with a `400` before parsing the body ([#6414](https://github.com/diegosouzapw/OmniRoute/issues/6414)). Regression guard: `tests/unit/v1-chat-completions-content-type-6414.test.ts`. (thanks @chirag127)
+- **fix(api):** the `X-OmniRoute-Compression` response header is now echoed on `/v1/chat/completions` and `/v1/completions` ([#6422](https://github.com/diegosouzapw/OmniRoute/issues/6422)). Regression guard: `tests/unit/compression-header-echo-6422.test.ts`. (thanks @chirag127)
+- **fix(api):** concurrent `GET /v1/models` requests are coalesced into a single catalog build ([#6408](https://github.com/diegosouzapw/OmniRoute/issues/6408)). Regression guard: `tests/unit/v1-models-concurrent-6408.test.ts`. (thanks @chirag127)
+- **fix(api):** `/v1/completions` now echoes the requested `body.model` in its JSON + streamed responses ([#6429](https://github.com/diegosouzapw/OmniRoute/pull/6429)). Regression guard: `tests/unit/completions-body-model-echo.test.ts`. (thanks @chirag127)
+- **fix(api):** env-var master keys now see the full `/v1/models` catalog ([#6406](https://github.com/diegosouzapw/OmniRoute/issues/6406)). Regression guard: `tests/unit/models-catalog-envkey-6406.test.ts`. (thanks @chirag127)
+- **fix(api):** non-streaming `/v1/completions` responses now echo `body.model` aligned with the `X-OmniRoute-Model` header ([#6426](https://github.com/diegosouzapw/OmniRoute/issues/6426)). Regression guard: `tests/unit/v1-completions-model-header-match-6426.test.ts`. (thanks @chirag127)
+- **fix(api):** unknown `/v1/*` routes now return a JSON `404 not_found` instead of the Next.js dashboard HTML shell ([#6405](https://github.com/diegosouzapw/OmniRoute/issues/6405)). Regression guard: `tests/unit/api/v1-catchall-json-404.test.ts`. (thanks @chirag127)
+- **fix(api):** the per-connection provider models route now degrades to the shipped catalog when a provider's `/models` endpoint answers with a redirect ([#6267](https://github.com/diegosouzapw/OmniRoute/issues/6267)) — a `qwen-web` import failed with a raw `Redirect blocked … (307)` 503. `safeOutboundFetch` throws `REDIRECT_BLOCKED` on the 307, `getSafeOutboundFetchErrorStatus` maps it to 503, and `buildDiscoveryErrorFallbackResponse` treated every 503 as a hard error — so the non-empty `getModelsByProviderId("qwen-web")` catalog was never surfaced. A models-endpoint redirect is not a fixable-config error (unlike `URL_GUARD_BLOCKED`/`INVALID_URL`, which stay hard errors), so it now falls back to the local/cached catalog before the 503 short-circuit. General fix — covers any config-driven provider that 307s. Regression guard: `tests/unit/provider-models-qwen-web-redirect-6267.test.ts`. (thanks @chirag127)
+- **fix(api):** the per-connection provider models route (MCP `list_models_catalog` + the dashboard import view) now merges USER-ADDED custom models into its response ([#6247](https://github.com/diegosouzapw/OmniRoute/issues/6247)) — custom models live in the `key_value` namespace `customModels`, which the live REST `/api/v1/models` already merges, but `src/app/api/providers/[id]/models/route.ts` never read `getCustomModels`, so custom models were dropped on both the discovery-success and local_catalog paths. They are now folded into the returned model list (deduped by id, stamped `owned_by: provider`), fixing MCP + the dashboard import view in one place. Regression guard: `tests/unit/provider-models-custom-merge-6247.test.ts`. (thanks @RCrushMe)
+
+- **fix(providers):** GitLab Duo tool-calling follow-up turns no longer fail upstream with `422 {"detail":"Validation error"}` (tokens 0/0, rejected pre-inference). The #6234 tool-result-feedback fix serialized the **entire** multi-turn conversation into GitLab's single-file `code_suggestions` (`small_file`) generation endpoint — folded history that turn-N sent as an oversized `current_file.content_above_cursor` **and** duplicated verbatim into `user_instruction`, tripping the AI-Gateway's `small_file` validation guard. The executor now **bounds** that prompt: it keeps system + latest user message + the most-recent tool round (dropping older turns), caps oversized tool results, and stops duplicating the full prompt into `user_instruction` (which now carries only the short latest user message) — while still feeding the most-recent tool result back so the agent continues (`open-sse/executors/gitlab.ts`, [#6220](https://github.com/diegosouzapw/OmniRoute/issues/6220)). The unit test covers the bounding logic; the upstream 422→200 clearing is VPS-only (Hard Rule #18). Regression guard: `tests/unit/gitlab-tool-exchange-bounded-6220.test.ts`.
+
+- **fix(i18n):** the provider-detail (`/dashboard/providers/[id]`) connection-status filter labels no longer render as `__MISSING__:All` / `__MISSING__:Active` / `__MISSING__:Error` / `__MISSING__:Banned` / `__MISSING__:CreditsExhausted` in non-English locales (notably pt-BR) ([#6290](https://github.com/diegosouzapw/OmniRoute/issues/6290)). Root cause was **not** the namespace mismatch the issue guessed — the `providers.filter*` keys resolve correctly in `en.json`; the debt lived in the locale mirrors (`src/i18n/messages/*.json`), where these five keys carried the `__MISSING__:` sync sentinel in ~15 locales and were absent entirely in ~26 others, so next-intl found the key and echoed the sentinel verbatim. All 40 non-English/-Chinese mirrors now ship real translations for the five `providers.filter*` labels. Regression guard: `tests/unit/i18n-provider-filter-keys-6290.test.ts`. (thanks @diegosouzapw)
+
+- **fix(providers):** the `copilot-m365-web` streaming executor now emits `debug`-level WebSocket diagnostics ([#6210](https://github.com/diegosouzapw/OmniRoute/issues/6210)) — the outbound WS URL (with the `access_token` **redacted** via `redactWsUrl()`), handshake success/failure, and each received SignalR frame's `type`/`target`. Previously the streaming path logged nothing, so an empty `content:null` response (the M365 Education / Starter tier symptom fixed in #6234) was undiagnosable even at `APP_LOG_LEVEL=debug`. The change is debug-level and side-effect-free — it does not alter streaming behavior or the frame parser, and the token never reaches the logs. Regression guard: `tests/unit/copilot-m365-web-logging-6210.test.ts` (thanks @qpeyba)
+
+- **fix(resilience):** a round-robin combo no longer returns `503 all upstream accounts are unavailable` when a compatibility-rejected target is actually healthy ([#6238](https://github.com/diegosouzapw/OmniRoute/issues/6238)). `filterTargetsByRequestCompatibility` drops request-incompatible targets (tool/vision/structured-output unsupported, or below the required context window) **before** any availability check runs, and its `compatible.length === 0` safety net only fired when _all_ targets were filtered — not when the kept targets later all turned out runtime-unavailable (circuit-open / cooldown / no credentials). So a combo could 503 while a compat-rejected-but-healthy provider sat unused. `handleRoundRobinCombo` now keeps the compat-rejected set and, when every compat-kept target was skipped without a single real attempt, probes those rejected targets as a **last-resort fallback tier** (via the new pure `open-sse/services/combo/comboCompatFallback.ts`) before crystallizing the 503. Regression guard: `tests/unit/combo-roundrobin-compat-fallback-6238.test.ts`. (thanks @ThongAccount)
+
+- **fix(startup):** best-effort self-heal for a corrupted Turbopack dev cache on Windows ([#6289](https://github.com/diegosouzapw/OmniRoute/issues/6289)). On Windows, `pnpm dev` can fail at startup when Turbopack `mmap`s a persistent-cache SST file and the OS refuses the mapping (`os error 1455` — "paging file too small"), which Turbopack surfaces as a misleading `Module not found: Can't resolve '@/shared/utils/machine'`. This is a **known upstream Turbopack cache-corruption bug — not our code**. The dev launcher (`scripts/dev/run-next.mjs`) now wraps `nextApp.prepare()` and, when it rejects with that signature (`isTurbopackCacheCorruption` in the new `scripts/dev/turbopackCacheHeal.mjs`), purges `.build/next/**/cache/turbopack` and retries **once** with a clear log. **Caveat — best-effort only:** the corruption often surfaces as a runtime overlay rather than a `prepare()` rejection, so this cannot always intercept it; the reliable remedy remains manually deleting the Turbopack cache dir. Regression guard: `tests/unit/turbopack-cache-heal-6289.test.ts`. (thanks @chirag127)
+
+- **fix(providers):** qodercli PAT auth no longer fails with `spawn qodercli ENOENT` on Windows ([#6263](https://github.com/diegosouzapw/OmniRoute/issues/6263)) — `spawnQoderCli` spawned the bare `qodercli` name with `shell:false` and an unenriched env, so the npm `.cmd` wrapper under `%APPDATA%\npm` (a user-PATH directory) was never resolved. It now resolves the absolute `.cmd`/`.exe` path through the existing `getCliRuntimeStatus("qoder")` resolver in `src/shared/services/cliRuntime.ts` (memoized), spawns with `shell` when the target is a `.cmd`/`.bat`, and uses the cliRuntime-enriched env (PATH + PATHEXT + APPDATA); the ENOENT error now lists the searched paths plus the `CLI_QODER_BIN` override. End-to-end spawn on a real Windows host is host-only (Hard Rule #18); the path-resolution logic is unit-tested. Regression guard: `tests/unit/qodercli-windows-resolve-6263.test.ts`. (thanks @chirag127)
+
+- **fix(sse):** the reasoning-token buffer no longer inflates **probe-sized `max_tokens`** ([#6274](https://github.com/diegosouzapw/OmniRoute/issues/6274)) — Claude Code's `/model` capability check sends `max_tokens: 1`, but for a thinking-capable model with a large output cap (e.g. `glm-5.2`) the #3587 headroom heuristic (`max(current + 1000, ceil(current * 1.5))`) rewrote it to `1001` and forwarded that upstream, wasting tokens on a request that was never a genuine reasoning budget. `resolveReasoningBufferedMaxTokens()` (`open-sse/services/reasoningTokenBuffer.ts`) now short-circuits and returns the caller's value verbatim when it is below the new `REASONING_BUFFER_MIN_TRIGGER` (256) threshold — a tiny explicit limit is a probe, not a reasoning request. Real budgets still receive the #3587 headroom unchanged, and the guard runs after the existing capability checks so unknown / non-reasoning models keep returning `null`. Regression guard: `tests/unit/reasoning-token-buffer-6274.test.ts`. (thanks @brightfiscalband)
+
+- **fix(cli):** `omniroute reset-password` now works as a real subcommand, and password resets over piped (non-TTY) stdin actually apply ([#6261](https://github.com/diegosouzapw/OmniRoute/issues/6261), [#6258](https://github.com/diegosouzapw/OmniRoute/issues/6258)). Two coupled defects: (1) **#6261** — `bin/omniroute.mjs` routed everything through Commander with only two pre-Commander bypasses (`--mcp`, `reset-encrypted-columns`), so `omniroute reset-password` was rejected as an unknown command; only the separate `omniroute-reset-password` bin worked, while the docs falsely advertised the subcommand (incl. a bogus "legacy alias still works"). A pre-Commander bypass mirroring `reset-encrypted-columns` now dynamically imports `bin/reset-password.mjs` (which self-executes) before Commander parses; the three doc lines were corrected. (2) **#6258** — `bin/reset-password.mjs` issued two sequential `rl.question` prompts; under piped stdin the second read never settled at EOF, so `main()` never reached `resetManagementPassword` and the reset was a silent no-op (both prompts printed, no success, password unchanged). The CLI now detects non-TTY stdin and reads it once (first line = password, second line = confirm if present, else reused), adds a `--password-stdin` flag (entire stdin is the password, no confirmation), and exits `0` explicitly so the success line always flushes; interactive TTY behavior is unchanged. Regression guard: `tests/unit/reset-password-cli-6261-6258.test.ts` (3). (thanks @chirag127)
+
+- **fix(db):** the mass-migration **safety abort** now tells the operator how to bypass it and stops flooding the log ([#6260](https://github.com/diegosouzapw/OmniRoute/issues/6260)) — after restoring a backup that wiped the migration tracking table, `runMigrations()` threw the abort on every downstream `ensureDbInitialized()`, re-logging the full banner 11+ times, and the message never mentioned the existing `OMNIROUTE_MAX_PENDING_MIGRATIONS` escape hatch. The abort text now appends a bypass hint (set `OMNIROUTE_MAX_PENDING_MIGRATIONS=0` in `server.env` / `DATA_DIR/.env`), and a new `MigrationSafetyAbortError` is memoized so repeated calls in the same process throw the same instance and emit a single concise line instead of the full cascade. Regression guard: `tests/unit/migration-safety-abort-6260.test.ts`. (thanks @chirag127)
+
+- **fix(auth):** importing a **distinct** Codex/ChatGPT OAuth `auth.json` is no longer falsely rejected as "already exists" when it belongs to a different user in the same workspace ([#6301](https://github.com/diegosouzapw/OmniRoute/issues/6301)). `findExistingCodexConnection` (in `src/lib/oauth/utils/codexAuthImport.ts`) deduped **only** on `providerSpecificData.workspaceId === accountId`, where `accountId` is the shared `chatgpt_account_id`/`tokens.account_id` — so two members of the same ChatGPT Team collapsed onto a single connection (409 `duplicate_account`). The id_token's `https://api.openai.com/auth` claim carries a per-user `chatgpt_user_id` alongside the workspace id (the device-flow path already persisted it as `chatgptUserId`, but the import path did not). Now `parseAndValidateCodexAuth` extracts `userId` (`chatgpt_user_id` → `user_id` → JWT `sub`) into `ParsedCodexAuth`, the create/update paths persist `chatgptUserId` in `providerSpecificData` (mirroring `codex.ts`), and dedup keys on `workspaceId` **AND** `chatgptUserId` — with a backward-compat fallback to legacy accountId-only matching when no stored connection for that workspace records a `chatgptUserId`, so genuinely-same accounts still dedup. Regression guard: `tests/unit/codex-auth-import-userid-dedup-6301.test.ts` (4). (thanks @anungma)
+
+- **fix(providers):** importing models for the **venice-web** provider no longer fails with a red "Provider venice-web does not support models listing" ([#6269](https://github.com/diegosouzapw/OmniRoute/issues/6269)). `venice-web` is a web-cookie provider with an executor but no upstream `/v1/models` endpoint and no registry `models`, so the models route fell through to the tail `400`. Mirroring the `jules`/`linkup-search`/`ollama-search` fix (#5569), it now ships a static local catalog entry in `src/lib/providers/staticModels.ts` — seeding the current Venice lineup (`venice-uncensored`, `llama-3.3-70b`, `qwen3-235b`, `qwen3-4b`, `deepseek-r1-671b`; Venice rotates its catalog, see docs.venice.ai/models/overview) — so the route returns `200` with `source:"local_catalog"`, `intentional:true`. Regression guard: `tests/unit/static-models-venice-web-6269.test.ts`. (thanks @chirag127)
+
+- **fix(api):** the specialty model catalogs (`/v1/embeddings`, `/v1/images`, `/v1/music`, `/v1/videos` model lists) are now derived from the **unified catalog filtered by a predicate** (`getSpecialtyModelsResponse`) instead of ad-hoc per-route logic, so they consistently respect active-credential visibility and stay in sync with the main catalog ([#6303](https://github.com/diegosouzapw/OmniRoute/pull/6303)). Regression guard: `tests/unit/specialty-model-catalog-routes.test.ts`. (thanks @makcimbx)
+- **fix(api):** the agent-bridge server route now resolves the MITM manager via a **dynamic `import("@/mitm/manager.runtime")`** so Turbopack does not statically pull the stub (or over-bundle the manager), and the agent-skills generator anchors its output base path with `path.join(process.cwd(), …)` so Turbopack's static analyzer stops tracing the whole project root ([#6329](https://github.com/diegosouzapw/OmniRoute/issues/6329), [#6366](https://github.com/diegosouzapw/OmniRoute/pull/6366)). Regression guard: `tests/unit/agent-bridge-server-route-dynamic-import.test.ts`. (thanks @Iammilansoni)
+- **fix(api):** internal probes (combo-test, cloud-sync verify) now pick a **management-scoped / allow-all API key** instead of naively grabbing `getApiKeys()[0]` — a restricted `self:usage` first row made the probe fail with "Model X is not allowed for this API key" even when the combo path was healthy (`pickApiKeyForInternalUse` in `src/lib/db/apiKeys.ts`). The API-manager model editor also falls back to `/api/models?all=true` when `/v1/models` is catalog-protected ([#6372](https://github.com/diegosouzapw/OmniRoute/pull/6372)). Regression guard: `tests/unit/pick-internal-api-key-6372.test.ts`. (thanks @jmengit)
+- **fix(live-ws):** the Live Dashboard WebSocket server now **rejects on bind failure** (e.g. `EADDRINUSE` when the API bridge already holds the port) instead of letting the error surface as an unhandled `error` event that crash-loops the process — the `error` listener is attached to `wss` (not `server`) and releases the EventBus subscription on a failed start ([#6324](https://github.com/diegosouzapw/OmniRoute/issues/6324)). Regression guard: `tests/unit/live-ws-eaddrinuse-6324.test.ts`. (thanks @vinayakkulkarni)
+- **fix(dashboard):** the Home provider-topology widget now trusts the live provider-metrics snapshot — it uses `topology.errorProvider` and live `activeRequests` directly instead of re-deriving state from a stale `lastErrorAt` or applying a frontend timeout filter, so the topology reflects real-time provider health ([#6322](https://github.com/diegosouzapw/OmniRoute/pull/6322)). Regression guard: `tests/unit/home-provider-topology-live-state.test.ts`. (thanks @xz-dev)
+- **fix(sse):** strip zero-width markers from streamed **tool-call arguments** — a follow-up to [#5857](https://github.com/diegosouzapw/OmniRoute/pull/5857). That PR removed injected zero-width joiners (U+200D) from streamed assistant text/reasoning but deliberately left tool-call argument JSON byte-exact. The request-side obfuscation (`open-sse/services/claudeCodeObfuscation.ts`) injects ZWJ into agent words — including the temp path inside the Bash tool description — and Claude models copy that verbatim into generated commands, which are delivered as tool-call arguments rather than assistant text. As a result the ZWJ survived and corrupted code blocks (e.g. a temp path rendered with an invisible joiner). Now `open-sse/handlers/responseSanitizer.ts` strips zero-width code points from tool-call argument strings at every emit site (OpenAI non-stream/stream chat `tool_calls` + legacy `function_call`, native Responses `function_call` items, the OpenAI→Responses conversion, and the native Responses streaming `response.function_call_arguments.delta/.done` events). Only zero-width code points are removed; JSON structure and all other bytes stay identical (no parse/restringify), so normal arguments remain byte-exact. Regression guard: 6 new cases in `tests/unit/response-sanitizer.test.ts` (suite 50/50).
+
+- **fix(nodejs):** the default app log path now resolves under `DATA_DIR` (`~/.omniroute/logs/application/app.log`) instead of `process.cwd()` ([#6197](https://github.com/diegosouzapw/OmniRoute/issues/6197)) — the globally-installed CLI runs from an arbitrary working directory, so anchoring the default to cwd made file logging silently write to (or no-op under) an unrelated directory, contradicting the documented `.env.example` default. `getAppLogFilePath()` now computes the default lazily via the pure `resolveDataDir()` resolver (honours a per-process `DATA_DIR`, no directory-creation side effect); an explicit `APP_LOG_FILE_PATH` still wins. Regression guard: `tests/unit/logenv-datadir-path-6197.test.ts` (3).
+
+- **fix(docker):** AgentBridge/`startMitm` no longer aborts in containers/headless when the Antigravity-default DNS step can't write `/etc/hosts` ([#6127](https://github.com/diegosouzapw/OmniRoute/issues/6127)), and the privileged command's stderr now reaches `app.log` instead of only a bare exit code hitting the toast ([#6198](https://github.com/diegosouzapw/OmniRoute/issues/6198)). The default DNS step (`addDNSEntry`) was called unguarded while cert install and the two sibling DNS steps were each best-effort — in the runtime Docker image (`USER node`, no `sudo`, read-only `/etc/hosts`) it threw `Command failed with code 1` out of `startMitmInternal` and killed the whole start, discarding the stderr. The three DNS steps are extracted into a best-effort `provisionDnsEntries()` where each failure is logged with the full `err` (stderr included, folded in by `systemCommands.ts`) and never aborts the start. Regression guard: `tests/unit/mitm-dns-graceful-degrade-6127.test.ts` (4).
+
+- **fix(providers):** copilot-m365-web now supports the M365 Education "Starter / OfficeWebIncludedCopilot" tier and no longer returns an empty `content:null` stream ([#6210](https://github.com/diegosouzapw/OmniRoute/issues/6210)). Two gaps: (1) `buildWsUrl()` hardcoded the individual-consumer scenario (`OfficeWebPaidConsumerCopilot`, `isEdu=false`) — the EDU tier is now opt-in via `providerSpecificData.tier="edu"`, emitting `scenario=OfficeWebIncludedCopilot`/`isEdu=true` (the individual path is unchanged); (2) the EDU/GPT-5.5 path streams deltas via `arguments[0].writeAtCursor` (incremental) instead of only `messages[].text` (accumulated snapshots), which the parser dropped — a new `accumulateBotContent()` folds both formats, with `type:2 item.result.message` as a last-resort fallback. Regression guard: `tests/unit/copilot-m365-edu-writeatcursor-6210.test.ts` (10). (thanks @qpeyba)
+
+- **fix(providers):** GitLab Duo executor now feeds tool results back into the prompt instead of looping ([#6220](https://github.com/diegosouzapw/OmniRoute/issues/6220)) — `buildPrompt()` branched only on `system`/`user` and took `userParts.at(-1)`, silently dropping the `assistant{tool_calls}` + `tool{result}` turns the client appended, so the reconstructed prompt was byte-identical to turn 1 and the model re-emitted the same `<tool>` call forever. When a tool exchange is present the full conversation is now serialized, folding each tool result back keyed by its `tool_call_id`; simple conversations keep the legacy shape. Complements the tool_call emission from [#6051](https://github.com/diegosouzapw/OmniRoute/issues/6051) (the `kilo-duplicate` label was a false positive — different, sequential defect). Regression guard: `tests/unit/gitlab-tool-result-feedback-6220.test.ts` (4).
+
+- **fix(providers):** opencode-go/opencode-zen can now synthesize the OpenCode CLI identity headers Cloudflare requires on VPS egress ([#5997](https://github.com/diegosouzapw/OmniRoute/issues/5997)) — on a datacenter VPS, `opencode.ai/zen/go/v1/chat/completions` 403s (HTML challenge) requests lacking CLI identity, while the reporter's control curl proved that `User-Agent: opencode-cli/1.0.0` + `x-opencode-client: cli` + `x-opencode-project: default` + fresh request/session UUIDs succeed. Opt-in via `OPENCODE_SYNTHESIZE_CLI_HEADERS=true` (values overridable via `OPENCODE_GO_USER_AGENT`/`OPENCODE_USER_AGENT`/`OPENCODE_CLIENT`/`OPENCODE_PROJECT`); it fills only headers the client did not already send. Kept **off by default** — the forward-only path is deliberate (fabricating a wrong value risks upstream rejection; a prior dedup regressed with `opencode/local`), so this replaces the fragile local header-injection shim without changing default behavior. Regression guard: `tests/unit/opencode-cli-headers-synthesis-5997.test.ts` (6). (thanks @aleksesipenko)
+- fix(resilience): sticky session affinity now evicts and fails over to another account when the pinned account is exhausted/unavailable (#6219)
+
+- fix(sse): Responses API passthrough now drops internal commentary-phase output before forwarding to clients (gated by RESPONSES_PASSTHROUGH_DROP_COMMENTARY, default on) (#6199)
+
+- **fix(sse):** tool-call function schemas with a root `type: null` are now coerced to `type: "object"` before dispatch ([#6359](https://github.com/diegosouzapw/OmniRoute/issues/6359)) — clients like the Codex app emit `parameters: { type: null, ... }` for some tools, which OpenAI-compatible upstreams reject with `400 Invalid schema for function '...': schema must be a JSON Schema of 'type: "object"', got 'type: null'`, failing the whole request. `toolSchemaSanitizer` already stripped the null; it now re-adds the mandatory root `"object"` type (and empty `properties`/open `additionalProperties` when absent). Combinator roots (`anyOf`/`oneOf`/`allOf`) and explicit root types are left untouched. Regression guard: 5 new cases in `tests/unit/tool-schema-sanitizer.test.mjs`.
+
+- **fix(docker):** AgentBridge no longer fails to start on npm/Electron/VPS installs with "MITM manager stub reached at runtime" ([#6344](https://github.com/diegosouzapw/OmniRoute/issues/6344)) — v3.8.45 flipped the production bundler default to Turbopack, but `next.config.mjs` aliased `@/mitm/manager` to its Docker-only degraded stub **unconditionally**. That was harmless while Docker (which sets the alias intentionally for #3390 graceful degradation) was the sole Turbopack consumer, but once every artifact built with Turbopack the stub shipped to all non-Docker users and `startMitm` threw on the first Agent-Bridge start. The alias is now opt-in via `OMNIROUTE_MITM_STUB=1` (set only by the Dockerfile) through the shared `scripts/build/mitm-stub-flag.mjs` helper; default builds bundle the real manager. Regression guard: `tests/unit/mitm-stub-alias-6344.test.mjs` (4).
+
+- **fix(proxy):** stop the v3.8.44 proxy regression that leaked the real IP and disabled healthy proxies ([#6246](https://github.com/diegosouzapw/OmniRoute/issues/6246)). Two coupled defects from the new health scheduler: (1) **IP leak** — when a proxy assigned to a connection was marked `inactive`, resolution fell through to a **direct** egress instead of blocking, exposing the operator's real IP; (2) **over-deactivation** — the sweep flipped a proxy to `inactive` on the **first** failed probe and counted our own 5s timeout / a probe-target `5xx` as the proxy's fault, so healthy paid proxies vanished from egress selection ("my proxies are not being used anymore"). Fix: the sweep decision is extracted into a pure, network-free `decideProxyHealthAction` (`src/lib/proxyHealth/decision.ts`) — by default the health check now **only counts/logs and never downgrades status** (a proxy is downgraded/removed only with `PROXY_AUTO_REMOVE=true`, after `PROXY_AUTO_REMOVE_AFTER` **consecutive** conclusive failures); probes are classified tri-state so an inconclusive result (our timeout, or a `5xx` from the probe target) never penalizes the proxy, and the probe timeout is raised 5s→15s. Separately, `safeResolveProxy` now **fails closed** via the existing policy: a connection whose assigned proxy is dead is blocked instead of leaking direct (`hasBlockingProxyAssignment`), honoring the explicit `proxy off` toggles and the `PROXY_FAIL_OPEN=true` opt-out. Existing proxies stuck `inactive` by the old behavior need a one-time manual re-activate (the operator owns proxy status). Regression guards: `tests/unit/proxy-health-decide-action-6246.test.ts`, `tests/unit/proxy-assigned-unavailable-6246.test.ts`.
+
+- **fix(proxy):** make "Test All" read-only and add bulk enable/disable ([#6246](https://github.com/diegosouzapw/OmniRoute/issues/6246)). Complements the core fail-closed / scheduler fix (#6296) with the two remaining reporter asks. (1) The **"Test All" button** (`POST /api/settings/proxies/auto-test`) used to flip a proxy to `inactive` on a failed reachability probe; since the egress selector excludes `inactive` proxies, a flaky probe (an unreachable `httpbin.org`, a proxy that blocks `HEAD`, or a slow paid proxy) silently disabled every proxy that failed — "Test All" is now **read-only by default** (only the operator sets a proxy active/inactive; opt back into the legacy test-and-set with `PROXY_HEALTH_AUTO_DEACTIVATE=true`). (2) Adds a **bulk enable/disable** proxies endpoint + toolbar action (`POST /api/settings/proxies/batch-activate`) so an operator can re-activate proxies in one click. Regression guard: `tests/unit/proxy-health-6246.test.ts`. (thanks @tenshiak)
+
+- **chatcore (tools): stop the default 128-tool cap from silently dropping opencode's `task`/MCP tools.** opencode (used as an MCP/agent host) sends a large tool list; when it exceeds the speculative `MAX_TOOLS_LIMIT` (128) default, `truncateToolList` did a blind `tools.slice(0, 128)`, dropping every tool past index 128 — including opencode's built-in `task` tool (subagent launch) and many MCP tools, so models routed through OmniRoute could no longer spawn subagents or reach part of their tools. The cap exists to avoid upstream `400`s for providers with real hard limits (e.g. grok-cli 200), so it is kept for those: detection of the opencode client (`isOpencodeClient` — any `x-opencode-*` header, or `opencode` in the user-agent) now only bypasses the **speculative 128 default**, never a known provider ceiling. Precedence is explicit — a proactive/detected provider limit always truncates (even for opencode); otherwise opencode forwards its full tool list; otherwise the unchanged 128 default applies to every other client. Refactors `getEffectiveToolLimit` into `getKnownToolLimit(provider) ?? DEFAULT_LIMIT` (byte-identical for existing callers) and fixes a cosmetic debug-log that reported the truncated count instead of the original. Regression guard: `tests/unit/tool-limit-detector.test.ts`.
+
+- **fix(mitm):** the macOS MITM-cert install check now matches the system keychain again. `security find-certificate -a -Z` prints the SHA-1 as a colon-less hex string, but the installed-check compared it against `getCertFingerprint()`'s colon-separated form, so the substring match never hit — the cert was reported as not-installed and re-prompted for the sudo install on every run. Fingerprints are now normalized (colons stripped, upper-cased) on both sides via the extracted `macCertOutputHasFingerprint` helper. Regression guard: `tests/unit/mitm-cert-mac-fingerprint.test.ts`. ([#6204](https://github.com/diegosouzapw/OmniRoute/pull/6204), closes [#6134](https://github.com/diegosouzapw/OmniRoute/issues/6134) — thanks @rianonehub)
+
+- **fix(api):** `/v1/messages/count_tokens` now counts `tool_use`, `tool_result` and `thinking` content blocks (and array-form `system` prompts) in the local-estimation path, instead of only `text`. Real agentic conversations keep ~95% of their tokens inside tool results; the previous estimate returned near-zero for them, which silently broke Claude Code's auto-compaction (context grew past the window with no compaction until the upstream API rejected the request). The real provider-side count path is unchanged. Regression guard: `tests/unit/messages-count-tokens-route.test.ts`. ([#6221](https://github.com/diegosouzapw/OmniRoute/pull/6221) — thanks @luweiCN)
+
+- **fix(antigravity):** strip a trailing assistant prefill turn for Vertex Claude models to avoid upstream 400s ([#6114](https://github.com/diegosouzapw/OmniRoute/pull/6114)). Regression guard: `tests/unit/antigravity-claude-prefill-strip.test.ts`. (thanks @anki1kr)
+
+- **fix(security):** the mutable cloud-agent routes (`/api/cloud/credentials/update`, `/api/cloud/models/alias`) now require management auth instead of being treated as public. They were classified as public API routes, so a request without management credentials could update stored cloud-agent credentials and model aliases. They are removed from the public-route set, classified as management routes in the authz pipeline, and gated by `requireManagementAuth`; cloud **read**/auth routes stay public. Regression guards: `tests/unit/cloud-write-auth.test.ts`, `tests/unit/authz/classify.test.ts`, `tests/unit/public-api-routes.test.ts`. ([#6233](https://github.com/diegosouzapw/OmniRoute/pull/6233) — thanks @vittoroliveira-dev)
+
+- **refactor(dashboard):** extract the onboarding-wizard "Open provider details" link target into a pure, unit-tested `buildProviderDetailsHref(connection)` helper. The wizard already routes by `connection.id` (the node UUID) rather than the provider category slug (#6144/#6145); this hardens that behavior behind a tested helper that guards a missing id/connection. Regression guard: `tests/unit/provider-onboarding-href.test.ts`. ([#6166](https://github.com/diegosouzapw/OmniRoute/pull/6166) — thanks @KooshaPari)
+
+- **fix(security):** the doubao synthetic device-id generator now derives its digits via an unbiased crypto-random draw (rejection sampling over `crypto.randomBytes()`) instead of a `% 10` reduction, closing a CodeQL `js/biased-cryptographic-random` finding.
+- **fix(agentSkills):** the GitHub-skills generator now resolves `outputDir` to an absolute path before writing, fixing a regression introduced by #6366 (relative-to-cwd base path) that could write generated skill files to the wrong directory.
+- **fix(security):** `/api/keys/{id}/devices` now checks the HTTP method before auth/validation, returning a `405` for non-GET/DELETE verbs instead of a misleading `401`/`500` (closes a `dast-smoke` QUERY-method finding).
+- **fix(quality):** clear the last 2 heavy quality-gate reds on the release tip (cycle pre-flight).
+- **fix(mitm):** the test suite and CI can never mutate the OS trust store — `OMNIROUTE_SKIP_SYSTEM_TRUST=1` is set globally for tests/CI so `installCert`/`uninstallCert`/`installTproxyCa` skip the privileged OS dispatch ([#6310](https://github.com/diegosouzapw/OmniRoute/pull/6310); full detail is under the [3.8.45] section below — this branch received it via the parallel-cycle sync-back).
+- **fix(api):** `POST /api/github-skills` now Zod-validates its request body; documented the new quality-gate env vars and pinned the merge-integrity GitHub Actions to a commit SHA.
+- **fix(skills):** generate the missing `omni-github-skills` registry entry and align the agent-skills catalog-count tests (follow-up to #6186).
+- **fix(quality):** clear the cycle's 11 net-new ESLint errors and make `validate-release-green` suppressions-aware.
+
+### 📝 Maintenance
+
+- **i18n(it):** add 118 missing Italian (`it`) translations (net-additive — no existing keys dropped, valid JSON), improving Italian UI coverage. ([#6212](https://github.com/diegosouzapw/OmniRoute/pull/6212) — thanks @serverless83)
+- **chore(providers):** remove deprecated **MiMo V2** model entries from the catalogs (xiaomi-mimo, opencode-go, zenmux-free, audio TTS) — the upstream V2 line is superseded by MiMo V2.5; drops `mimo-v2-tts`, `mimo-v2-pro`, `mimo-v2-omni`, `mimo-v2-flash`, `mimo-v2-flash-free` and realigns the provider-catalog tests. ([#6248](https://github.com/diegosouzapw/OmniRoute/pull/6248) — thanks @backryun)
+- **chore(release):** ~50 commits on this branch are v3.8.45 pre-flight/hardening fixes and CI-perf work that landed here via the parallel-cycle sync-back (`sync-next-cycle.mjs`, Hard Rule #21) **after** the `v3.8.45` git tag was cut, and are already fully documented under the **[3.8.45]** section below — listed here only so the per-cycle commit-coverage check (`npm run release:uncovered`) doesn't flag them as gaps. Provider/catalog/UX/backend fixes: #6041, #6078, #6108, #6135, #6148, #6149, #6154, #6158, #6161, #6162, #6163, #6164, #6165, #6170, #6177, #6178, #6181, #6186, #6187, #6191, #6193, #6194, #6195, #6200, #6205, #6208, #6209, #6211, #6213, #6223, #6224, #6225, #6226, #6227, #6228, #6229, #6230, #6235, #6291, #6292. CI/release-pipeline work: #6167, #6203, #6214, #6215, #6218, #6273, #6275, #6283, #6284, #6285, #6300, #6305.
+- **chore(release):** additional zero-ref release-cycle plumbing on this branch, kept out of `release:uncovered` on purpose (no `#N` in the commit subject to cite): opening the v3.8.46 cycle, opening/closing the v3.8.45 cycle, the finalized [3.8.45] CHANGELOG i18n sync-back to 42 mirrors, the v3.8.45 cognitive/cyclomatic and file-size drift rebaselines, ESLint stale-suppression pruning (4,273 → 4,233), and clearing test-masking/docs-all pre-flight reds for v3.8.45.
+
+### ⚡ Performance & Infrastructure
+
+- **perf(release-green):** the pre-flight validator (`scripts/quality/validate-release-green.mjs`) now runs its 4 slow suites (unit / vitest / integration / pack-artifact) **concurrently** via `Promise.all` — pre-flight wall time drops from ~the sum of the suites to ~the slowest one (~30min saved per round; Phase 0 was the nº1 bottleneck of the v3.8.45 release benchmark, 2h54 of 6h34 e2e). Guard: `tests/unit/validate-release-green.test.ts` ("runs the slow suites CONCURRENTLY"). ([#6319](https://github.com/diegosouzapw/OmniRoute/pull/6319))
+- **fix(ci):** `scripts/release/sync-next-cycle.mjs` — two defects found live in its first production run (v3.8.45 Phase 5): (1) the `git()` helper's default 1 MiB `maxBuffer` crashed with `ENOBUFS` on `git show origin/main:CHANGELOG.md` (the CHANGELOG alone is >1 MiB) — widened to 64 MiB; (2) the i18n resync only propagated the `[NEXT]` (TBD) section, leaving the just-shipped finalized section as "— TBD" in all 42 mirrors — it now also syncs `[prevVersion]` bounded by the heading below it (new exported pure helper `versionAfter`). Guards: +5 tests in `tests/unit/sync-next-cycle.test.ts` (8/8). ([#6327](https://github.com/diegosouzapw/OmniRoute/pull/6327))
+- **test(ci):** concurrency-sensitive flaky tests are **quarantined into a serial pass** (`tests/unit/serial/`, `--test-concurrency=1`, appended to every unit runner incl. sharded variants — the serial pass is sharded too so concurrent shard jobs never self-collide). Initial set: `glm-coding-plan-monthly-3580`, `quota-division-blocks`, `provider-health-autopilot`, `combo-health-autopilot` — the class behind the ~28min CI wedges/re-runs (two live 1h+ wedges cancelled during this PR's own validation). Discovery + TIA gates track the new glob; systemic root cause (async logger writing after teardown) tracked in [#6360](https://github.com/diegosouzapw/OmniRoute/issues/6360). Guard: `tests/unit/test-serial-quarantine.test.ts` (4). ([#6347](https://github.com/diegosouzapw/OmniRoute/pull/6347))
+
+### 🙌 Contributors
+
+Thanks to everyone whose work landed in v3.8.46:
+
+| Contributor                                                    | PRs / Issues           |
+| -------------------------------------------------------------- | ---------------------- |
+| [@2220258345](https://github.com/2220258345)                   | direct commit / report |
+| [@abdofallah](https://github.com/abdofallah)                   | direct commit / report |
+| [@adentdk](https://github.com/adentdk)                         | direct commit / report |
+| [@aleksesipenko](https://github.com/aleksesipenko)             | direct commit / report |
+| [@anki1kr](https://github.com/anki1kr)                         | direct commit / report |
+| [@anungma](https://github.com/anungma)                         | direct commit / report |
+| [@arpicato](https://github.com/arpicato)                       | direct commit / report |
+| [@backryun](https://github.com/backryun)                       | #6248                  |
+| [@binsarjr](https://github.com/binsarjr)                       | direct commit / report |
+| [@brightfiscalband](https://github.com/brightfiscalband)       | direct commit / report |
+| [@chirag127](https://github.com/chirag127)                     | #6501, #6506           |
+| [@developerjillur](https://github.com/developerjillur)         | direct commit / report |
+| [@dilneiss](https://github.com/dilneiss)                       | #6499                  |
+| [@dtybnrj](https://github.com/dtybnrj)                         | direct commit / report |
+| [@Forcerecon](https://github.com/Forcerecon)                   | direct commit / report |
+| [@hao3039032](https://github.com/hao3039032)                   | direct commit / report |
+| [@Iammilansoni](https://github.com/Iammilansoni)               | #6150, #6245           |
+| [@jmengit](https://github.com/jmengit)                         | direct commit / report |
+| [@jordansilly77-stack](https://github.com/jordansilly77-stack) | direct commit / report |
+| [@JxnLexn](https://github.com/JxnLexn)                         | direct commit / report |
+| [@KooshaPari](https://github.com/KooshaPari)                   | #6166                  |
+| [@loopyd](https://github.com/loopyd)                           | direct commit / report |
+| [@luweiCN](https://github.com/luweiCN)                         | #6221                  |
+| [@makcimbx](https://github.com/makcimbx)                       | direct commit / report |
+| [@muflifadla38](https://github.com/muflifadla38)               | direct commit / report |
+| [@newnol](https://github.com/newnol)                           | direct commit / report |
+| [@ofekbetzalel](https://github.com/ofekbetzalel)               | direct commit / report |
+| [@ohahe52-dot](https://github.com/ohahe52-dot)                 | direct commit / report |
+| [@phidinhmanh](https://github.com/phidinhmanh)                 | direct commit / report |
+| [@powellnorma](https://github.com/powellnorma)                 | direct commit / report |
+| [@qpeyba](https://github.com/qpeyba)                           | direct commit / report |
+| [@RaviTharuma](https://github.com/RaviTharuma)                 | direct commit / report |
+| [@RCrushMe](https://github.com/RCrushMe)                       | direct commit / report |
+| [@rianonehub](https://github.com/rianonehub)                   | #6134, #6204           |
+| [@serverless83](https://github.com/serverless83)               | #6212                  |
+| [@swingtempo](https://github.com/swingtempo)                   | direct commit / report |
+| [@tenshiak](https://github.com/tenshiak)                       | direct commit / report |
+| [@ThongAccount](https://github.com/ThongAccount)               | direct commit / report |
+| [@UnrealAryan](https://github.com/UnrealAryan)                 | direct commit / report |
+| [@vinayakkulkarni](https://github.com/vinayakkulkarni)         | direct commit / report |
+| [@vittoroliveira-dev](https://github.com/vittoroliveira-dev)   | #6233                  |
+| [@warelik](https://github.com/warelik)                         | direct commit / report |
+| [@xxy9468615](https://github.com/xxy9468615)                   | direct commit / report |
+| [@xz-dev](https://github.com/xz-dev)                           | direct commit / report |
+| [@yanpaing007](https://github.com/yanpaing007)                 | direct commit / report |
+| [@diegosouzapw](https://github.com/diegosouzapw)               | maintainer             |
+
+---
+
 ## [3.8.45] — 2026-07-06
 
 ### ✨ New Features

docs/i18n/sk/CHANGELOG.md

@@ -6,6 +6,213 @@
 
 ## [3.8.31] — 2026-06-20
 
+## [3.8.46] — TBD
+
+### ✨ New Features
+
+- **feat(sse):** **hide paid-only models from `auto/*` routing** when `hidePaidModels` is on ([#6512](https://github.com/diegosouzapw/OmniRoute/issues/6512)) — follow-up to #6328/#6495. PR #6495 hid paid-only models from the `GET /v1/models` listing, but `auto/*` combos (`auto/best-coding`, `auto/glm`, …) could still pick a paid-only backend into their candidate pool → a 402/403 at request time. `createVirtualAutoCombo` now filters the candidate pool through the new pure `open-sse/services/autoCombo/paidModelFilter.ts` (`filterPaidOnlyCandidates`), applying the same free-model predicate #6495 uses in `catalog.ts` (`providerHasFreeModels(provider) && isFreeModel(provider, {id})`) whenever `settings.hidePaidModels === true`. Applied before the category/tier/family narrowing, so it covers every `auto/*` combo; an all-paid pool degrades to the existing graceful empty-pool path. **Opt-in — default OFF leaves the pool unchanged** (identity). Regression guard: `tests/unit/autoCombo/paid-model-filter-6512.test.ts` (4, incl. the default-off identity guard).
+- **feat(sse):** **provider-family auto combos** — `auto/glm`, `auto/minimax`, `auto/mimo`, `auto/zai`, `auto/gemma`, `auto/llama`, `auto/gemini` ([#6453](https://github.com/diegosouzapw/OmniRoute/issues/6453)) — new routable ids that materialize an on-demand virtual combo spanning whatever installed backends currently expose that model family, degrading gracefully as backends rotate. A new pure `open-sse/services/autoCombo/modelFamily.ts` (`detectModelFamily`) classifies by model-id prefix for six families; `zai` is instead resolved by provider id (z.ai's hosted API serves the same `glm-*` model ids as every other GLM backend, so `auto/zai` means "route to my z.ai backend specifically" vs `auto/glm`'s "any connected GLM backend"). Reuses the existing `createVirtualAutoCombo` on-demand materialization path (no DB writes) and the `/v1/models` catalog advertising loop. Regression guard: `tests/unit/autoCombo/provider-family-combos.test.ts` (11).
+- **feat(proxy):** native **proxy-pool round-robin / egress IP rotation** ([#6365](https://github.com/diegosouzapw/OmniRoute/issues/6365)) — a scope (global / provider / account) can now hold **multiple** proxies as a pool with a rotation strategy, so outbound requests cycle their egress IP instead of pinning one proxy per scope. Migration `117_proxy_pool_rotation.sql` lifts the `UNIQUE(scope, scope_id)` constraint (rebuild via the canonical rename/copy/drop; existing single assignments become 1-element pools) and adds a `proxy_scope_rotation` companion table holding the per-scope strategy + a persisted monotonic round-robin cursor. Strategies: `round-robin` (default, monotonic cursor — never `Math.random`), `random`, and `sticky-per-N-min`. Resolution (`resolveProxyForScopeFromRegistry` / `resolveProxyForConnectionFromRegistry`) now fetches the alive, position-ordered candidate set (unchanged `PROXY_ALIVE_PREDICATE`) and applies the strategy; an empty / all-dead pool still returns `null` — the #6246 fail-closed guard is **untouched** (never falls through to direct egress). Backend + DB only; dashboard pool-builder UI is a follow-up. Regression guard: `tests/unit/proxy-pool-rotation-6365.test.ts` (8, incl. fail-closed + backward-compat).
+- **feat(providers):** end-to-end **tool/function calling on the native Gemini `/v1beta` endpoint** ([#6222](https://github.com/diegosouzapw/OmniRoute/issues/6222)) — both directions of the Gemini↔OpenAI conversion now preserve tool data (previously silently dropped). Request side: `convertGeminiToInternal` (extracted to its own testable module) maps `tools[].functionDeclarations` → OpenAI `tools`, prior `functionCall` parts → assistant `tool_calls`, and `functionResponse` parts → `tool`-role messages. Response side: `convertOpenAIResponseToGemini` emits `parts[].functionCall {name,args}` from `message.tool_calls`, and the streaming `openAIChunkToGeminiChunk` accumulates fragmented `tool_calls` deltas by index into complete `functionCall` parts. The non-Gemini client paths (Claude, OpenAI-Responses) already preserved tool calls — this closes the gap specific to the native Gemini surface. Regression guard: `tests/unit/v1beta-gemini-tool-calling-6222.test.ts` (6, incl. a streaming SSE round-trip).
+- **feat(providers):** copilot-m365-web **enterprise / work tier** support ([#6334](https://github.com/diegosouzapw/OmniRoute/issues/6334)) — mirrors the EDU-tier pattern (#6210): `M365ConnectionParams` gains an `agent` field, a new opt-in `M365_ENTERPRISE_OVERRIDES` preset (`agent=work`, `scenario=officeweb`, `licenseType=Premium`) applies via `providerSpecificData.tier="enterprise"` (alias `"work"`), and `agent` is also overridable directly via `providerSpecificData.agent`. `buildWsUrl` was hardcoding `agent="web"` (the one enterprise-distinguishing param with no override path), so a Premium work account handshook then returned an empty stream. The individual and EDU paths are untouched. Kilo's dup flag vs #6210 (EDU tier) was a false positive — different tier. Regression guard: `tests/unit/copilot-m365-enterprise-6334.test.ts` (7). End-to-end confirmation on a real Premium work account is a live-VPS validation follow-up (Hard Rule #18). (thanks @Forcerecon)
+- **feat(api):** standardized, provider-agnostic **`effort` + `thinking` request params** ([#6241](https://github.com/diegosouzapw/OmniRoute/issues/6241)) — a thin standardization layer over the existing mature per-provider reasoning plumbing (no provider mapper touched). `providerChatCompletionSchema` gains a canonical `effort` (reusing the shared `none/low/medium/high/xhigh` vocabulary — the UI tiers `extra`/`max` collapse onto `xhigh`) and a boolean `thinking`. A pure `normalizeReasoningRequest` (wired once in `src/sse/handlers/chat.ts`, before any reasoning field is read) folds them onto the fields the translators already consume (`reasoning_effort` / `reasoning.effort` / `thinking`), so they fan out to Anthropic / Gemini / xAI / Responses — an explicit client `reasoning_effort` / object-shaped `thinking` always wins (backward-compatible). `/models` additively exposes `supportsThinking` + `effort_tiers` so the frontend can render the toggles (UI component is a follow-up). Regression guard: `tests/unit/effort-thinking-standardization-6241.test.ts` (12). (thanks @Iammilansoni, @shabeer)
+- **feat(combo):** new **`pipeline` (sequential) combo strategy** ([#6297](https://github.com/diegosouzapw/OmniRoute/issues/6297)) — the 18th routing strategy runs targets **in order**, threading each step's output into the next step's input, with an optional per-step `prompt` (system instruction); only the final step's response is returned. Distinct from `fusion` (parallel fan-out + judge). Implemented as a self-contained `open-sse/services/pipeline.ts` (sibling to `fusion.ts`), dispatched from `combo.ts`; the step list reuses `combo.models` order and reads an optional `prompt` off each target (backward-compatible — ignored by every other strategy). Intermediate steps run non-streaming with tools stripped (complete prose to thread forward); the final step keeps the client's `stream` flag + tools. A failing/empty/unparseable intermediate step fails the whole pipeline explicitly via a sanitized error (never silently swallowed). Kilo's dup flag vs #563 was a false positive (that's model→chain selection; this is a sequential chain). Regression guard: `tests/unit/combo-pipeline-strategy.test.ts` (5). (thanks @ofekbetzalel)
+- **feat(ci):** `check:test-masking` now flags **inline-reimplemented prod conditions** ([#6348](https://github.com/diegosouzapw/OmniRoute/issues/6348)) — a new **report-only** subcheck (v2, 6A.10 family) catches the wrong-shape contract test: a test that recomputes the condition under test inline instead of importing/exercising the real function (the #6216 class, where `=== 500` → `>= 500` stayed green because the test re-implemented the branch). For each added/modified test file it warns when the file textually duplicates a ≥3-token conditional from a production file touched in the same PR **and** does not import the symbol/module owning it, via a pure, fixture-tested `findReimplementedConditions()` with an allowlist mirroring `assertReductionAllowlist`. Report-only for now (does not fail the gate) — to be promoted to blocking after a triage cycle. Regression guard: `tests/unit/check-test-masking.test.ts` (45).
+- **feat(sse):** per-connection routing override (native vs CLIProxyAPI) ([#6339](https://github.com/diegosouzapw/OmniRoute/issues/6339)) — the previously-dead `isCliproxyapiDeepModeEnabled` helper is now wired into `resolveExecutorWithProxy`: a single connection can opt itself into the CLIProxyAPI passthrough executor via `providerSpecificData.cliproxyapiMode="claude-native"`, with precedence **connection override > provider `upstream_proxy_config` mode > default**. `resolveExecutorWithProxy` now receives the resolved connection's `providerSpecificData` (threaded from `chatCore.ts`), so one connection can deep-route while the provider's other connections stay native — no DB schema change (the toggle rides in `providerSpecificData`). Also resolves the same-provider mixing ask in #6340. Regression guard: `tests/unit/chatcore-executor-proxy.test.ts` (9). (thanks @RaviTharuma)
+- **feat(dashboard):** "Add session cookie" modal now shows a prominent **"Open ‹host› →"** link to the provider's own site ([#6268](https://github.com/diegosouzapw/OmniRoute/issues/6268)) — every `-web` cookie-session provider (chatgpt-web, claude-web, gemini-web, kimi-web, lmarena, qwen-web, m365-copilot-web, …) renders a one-click external link (opening the provider's login/home page in a new tab) so operators no longer tab away to retype the URL mid-setup. The host resolves from a pure, unit-tested `resolveWebProviderHost()` (prefers `WEB_COOKIE_PROVIDERS[id].website`, falls back to the registry `baseUrl` origin); non-web providers render exactly as before. Kilo's dup flag vs #6265 (modal-too-small-on-1080p) was a false positive — distinct concern. Regression guard: `tests/unit/resolve-web-provider-host.test.ts` (5). (thanks @chirag127)
+- **feat(providers):** add **DigitalOcean** AI (serverless inference) as an OpenAI-compatible API-key provider ([#6373](https://github.com/diegosouzapw/OmniRoute/pull/6373)) — base `https://inference.do-ai.run/v1`, wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/digitalocean/`, `src/shared/constants/providers/apikey/inference-hosts.ts`). Regression guard: `tests/unit/digitalocean-provider.test.ts`. (thanks @newnol)
+- **feat(providers):** add **Huancheng Public API** (`hcnsec`) as an OpenAI-compatible regional provider ([#6410](https://github.com/diegosouzapw/OmniRoute/pull/6410)) — Xinjiang Huancheng Cybersecurity's public LLM platform (base `https://api.hcnsec.cn/v1`, free credits via daily check-ins), wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/hcnsec/`, `src/shared/constants/providers/apikey/regional.ts`). Regression guard: `tests/unit/hcnsec-provider.test.ts`. (thanks @UnrealAryan)
+- **feat(dashboard):** the web-session credential guide now shows an **"Open {host}" link** ([#6316](https://github.com/diegosouzapw/OmniRoute/pull/6316)) to the provider's sign-in site (derived from the provider `website` via `getProviderWebsiteHost`), so you can jump straight to the page where the cookie/session must be captured. Regression guard: `tests/unit/web-session-provider-link-6316.test.ts`. (thanks @jordansilly77-stack)
+- **feat(cerebras):** add the **Gemma 4 31B** model (`gemma-4-31b`) to the Cerebras registry + pricing table ([#6331](https://github.com/diegosouzapw/OmniRoute/pull/6331)). Regression guard extends `tests/unit/t28-model-catalog-updates.test.ts`. (thanks @backryun)
+- **feat(providers):** add **Yuanbao (web)** as a cookie-session provider ([#6196](https://github.com/diegosouzapw/OmniRoute/issues/6196)) — `yuanbao-web` (Tencent Yuanbao, `yuanbao.tencent.com`) with cookie-only auth (`hy_user`/`hy_token` + public agent id), SSE→OpenAI translation incl. `reasoning_content`, exposing DeepSeek V3/R1 + Hunyuan / Hunyuan-T1. Regression guard: `tests/unit/providers-yuanbao-web.test.ts`. `together-web` was **deferred** (no verifiable web-session endpoint — needs a captured request) and `huggingchat-web` **dropped** (the existing `huggingchat` already is a web-cookie provider). (thanks @chirag127)
+- **feat(providers):** route the built-in **agentrouter** through the dynamic Claude-Code wire image ([#6056](https://github.com/diegosouzapw/OmniRoute/issues/6056)) — a small static allow-set (`CC_WIRE_IMAGE_BUILTINS` in `open-sse/services/ccWireImageBuiltins.ts`), consulted by `isClaudeCodeCompatible` / `isClaudeCodeCompatibleProvider` / `applyFingerprint`, makes agentrouter adopt the CC wire-image headers + fingerprint **while guarding the CC baseUrl/auth branches** so it keeps its own registry `baseUrl` and `x-api-key` auth. Regression guard: `tests/unit/agentrouter-cc-wire-image.test.ts` (asserts the wire image is applied AND agentrouter's baseUrl/auth are preserved). Live WAF-acceptance against agentrouter.org is a VPS validation follow-up (Hard Rule #18).
+- **feat(providers):** **bulk-add API keys for Cloudflare Workers AI** ([#6174](https://github.com/diegosouzapw/OmniRoute/issues/6174)) — `cloudflare-ai` is removed from the bulk-add exclusion list and the bulk parser gains a 3-field `name|accountId|apiKey` mode; the bulk route now builds a **per-entry** `providerSpecificData` so each key carries its own `accountId` (fixing the previous shared-object reuse), and both the create + key-validation paths receive it. Regression guard: `tests/unit/bulk-api-key-parser-cloudflare.test.ts`. (thanks @muflifadla38)
+- **feat(dashboard):** routing/settings UX clarity ([#6147](https://github.com/diegosouzapw/OmniRoute/issues/6147)) — (1) weighted combos show the **effective routing share %** next to each weight when weights don't sum to 100 (`WeightTotalBar.tsx`); (2) the status widget's user-facing **"Cloud Sync" label is renamed** to "Remote Settings Sync" (`CloudSyncStatus.tsx`; internal ids/state untouched); (3) built-in providers gain an **opt-in advanced base-URL override** (`isBaseUrlOverrideEligibleProvider`, hidden behind an "Advanced" toggle, reusing the existing `providerSpecificData.baseUrl` persistence — not globally widened). Regression guard: `tests/unit/routing-settings-ux-6147.test.ts`.
+- **feat(combo):** add an option to **disable session stickiness**, per-combo or globally — round-robin / random combos can rotate to a different connection on every request instead of pinning a whole conversation to one connection by its first-message hash. Resolution precedence per-combo `config.disableSessionStickiness` → global `settings.disableSessionStickiness` → default `false` (preserves the #3825 prompt-cache/504 fix); gates **both** stickiness call sites in `open-sse/services/combo.ts`. Exposed as a global toggle (Combo Defaults) and a per-combo Inherit/on/off control. ([#6168](https://github.com/diegosouzapw/OmniRoute/issues/6168)) Regression guard: `tests/unit/combo-disable-session-stickiness.test.ts`. (thanks @RCrushMe)
+- **feat(docker):** add the `OMNIROUTE_NO_SUDO` env flag for root-less / user-namespaced deployments — the MITM cert-trust command path (`resolveSudoSpawn` in `src/mitm/systemCommands.ts`) now strips the leading `sudo` when the flag is truthy, in addition to the existing root / sudo-missing cases, so the Proxy Agent runs without `sudo` (the operator trusts the CA manually, e.g. via `NODE_EXTRA_CA_CERTS`). Argv-array `spawn` preserved — no shell interpolation (Hard Rule #13). ([#6122](https://github.com/diegosouzapw/OmniRoute/issues/6122)) Regression guard: `tests/unit/mitm-systemCommands-no-sudo.test.ts`. (thanks @powellnorma)
+- **feat(providers):** add **Requesty** as an OpenAI-compatible gateway provider (BYOK, base `https://router.requesty.ai/v1`, ~200 free requests/day) — wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/requesty/`, `src/shared/constants/providers/apikey/gateways.ts`). ([#6120](https://github.com/diegosouzapw/OmniRoute/issues/6120)) Regression guard: `tests/unit/requesty-provider.test.ts`. (thanks @chirag127)
+- **feat(dashboard):** add **configured-only / available-only filters** to the Free Provider Rankings page ([#6150](https://github.com/diegosouzapw/OmniRoute/issues/6150)) — hide providers you haven't configured, or whose connections are all rate-limited / out of quota, via server-side query params (`?configuredOnly` / `?availableOnly` on `GET /api/free-provider-rankings`) backed by a testable lib helper reusing the in-process connection state (no Redis). Both filters default off, so the default view is unchanged; this supersedes the earlier client-side "Configured Only" toggle (#6245) with an available-only dimension and unit-tested logic. Regression guard: `tests/unit/freeProviderRankings-filters.test.ts`.
+
+### 🔧 Bug Fixes
+
+- **fix(dashboard):** adding a second API key connection for the same provider no longer silently overwrites the first — the Add-API-key modal now derives a unique default connection name (`main`, then `main-2`, `main-3`, …) so the backend name-based upsert can't collide (#6499 — thanks @dilneiss).
+- **fix(compression):** the session-dedup engine now also deduplicates a large multi-line block repeated **within a single message** (intra-message dedup), not just across turns; the compression-preview API surfaces a `fallbackReason`, and the fusion panel reports how many models were rate-limited vs failed on total-panel failure (#6501 — thanks @chirag127).
+- **fix(compression):** a stacked-pipeline step naming an unregistered engine now surfaces a `validationErrors` entry instead of silently no-op'ing, so misconfigured pipelines are visible in the preview API (#6506 — thanks @chirag127).
+- **feat(usage):** add a **Codex reset-credit redemption** flow to the Provider Limits UI ([#6361](https://github.com/diegosouzapw/OmniRoute/pull/6361)) — a `useCodexResetCreditRedemption` hook + `/api/usage/codex-reset-credit` route + `codexResetCredits` lib let you redeem banked Codex reset credits from the quota card. Regression guard: `tests/unit/codex-reset-credits.test.ts`. (thanks @JxnLexn)
+- **feat(glm):** add **team-plan quota settings** for `glm-cn` connections ([#6351](https://github.com/diegosouzapw/OmniRoute/pull/6351)) — a dedicated `GlmTeamQuotaFields` form section (team quota id / limits) threaded through the Add/Edit connection modals, persisted via `providerSpecificData`, with the GLM usage service reading the team quota. Regression guards: `tests/unit/glm-team-quota.test.ts`, `provider-specific-data-schema.test.ts`. (thanks @hao3039032)
+- **feat(providers):** add **TinyFish** web-fetch/search support ([#6349](https://github.com/diegosouzapw/OmniRoute/pull/6349)) — a `tinyfish-fetch` executor + `/v1/web/fetch` route + MCP web-fetch tool, registered as a specialty-media provider with request-validation and a search-provider catalog entry. Regression guards: `tests/unit/executor-tinyfish-fetch.test.ts`, `web-fetch-handler.test.ts`, `mcp-web-fetch-tool.test.ts`, `provider-validation-tinyfish.test.ts`. (thanks @dtybnrj)
+- **fix(cli):** `omniroute launch-codex` now spawns `codex.cmd` through a shell on Windows (the npm `.cmd` shim is unresolvable by bare `spawn` → ENOENT), mirroring the qodercli Windows fix (#6263) ([#6312](https://github.com/diegosouzapw/OmniRoute/pull/6312)). Regression guard: `tests/unit/launch-codex-windows-spawn-6312.test.ts`. (thanks @swingtempo)
+- **fix(codex):** isolate the **Spark** quota from the shared Codex quota and stabilize the quota UI ordering / hydration so per-scope limits render consistently ([#6336](https://github.com/diegosouzapw/OmniRoute/pull/6336)). Regression guards: `tests/unit/codex-quota-selection-hydration.test.ts`, `provider-limits-ui.test.ts` + 3 more. (thanks @xz-dev)
+- **feat(api):** add a `hidePaidModels` setting that filters paid-only models out of the `/v1/models` catalog. Regression guard: `tests/unit/models-catalog-hide-paid.test.ts`. (thanks @chirag127)
+- **fix(api-manager):** the fallback model picker now preserves combos instead of dropping them when a primary model is unavailable ([#6443](https://github.com/diegosouzapw/OmniRoute/pull/6443)). Regression guard: `tests/unit/api-manager-page-static.test.ts`. (thanks @jmengit)
+- **fix(providers):** recoverable Antigravity / Cloud-Code (Gemini Code Assist) `403` responses ([#6452](https://github.com/diegosouzapw/OmniRoute/pull/6452)) are now classified as a retryable project-config error instead of a terminal account ban, so a fixable project/API-disabled 403 no longer forces a ~1-year cooldown / full OAuth reconnect. Regression guard: `tests/unit/errorclassifier-antigravity-403.test.ts`. (thanks @developerjillur)
+- **fix(mitm):** `sanitizeHeaders` now redacts `Set-Cookie` response headers so upstream session cookies never leak into logs / diagnostics ([#6451](https://github.com/diegosouzapw/OmniRoute/pull/6451)). Regression guard: `tests/unit/mitm-sanitize-headers.test.ts`. (thanks @developerjillur)
+- **fix(api):** `/api/compression/preview` now accepts `mode: "caveman"` and correctly handles stacked / zero-compression previews ([#6425](https://github.com/diegosouzapw/OmniRoute/issues/6425)). Regression guard: `tests/unit/api/compression-preview-caveman-and-stacked-6425.test.ts`. (thanks @chirag127)
+- **feat(providers):** add **Zed** hosted LLM aggregator as a native-app provider ([#6118](https://github.com/diegosouzapw/OmniRoute/pull/6118)) — OAuth sign-in via the Zed hosted flow, registered through the shared provider registry + executor. Regression guards: `tests/unit/zed-oauth-provider.test.ts`, `zed-import-utils.test.ts`, `zed-docker-detect.test.ts`, `mitm-handler-zed.test.ts`. VPS-validated via live operator login (Hard Rule #18).
+- **fix(oauth):** the Kiro SSO-cache auto-import now **preserves the IDC region** — cross-region Amazon Q / Kiro profiles imported from the SSO cache are no longer collapsed to the default region ([#6113](https://github.com/diegosouzapw/OmniRoute/pull/6113)). Regression guard: `tests/unit/kiro-auto-import-idc-2059.test.ts`. VPS-validated via live operator login (Hard Rule #18).
+- **fix(dashboard):** passthrough model aliases no longer collide when two namespaced model ids share a last segment (port from 9router#1850, [#6431](https://github.com/diegosouzapw/OmniRoute/pull/6431)). `enx/gpt-5.5` and `enx/codebuddy/gpt-5.5` both auto-generated the alias `gpt-5.5`, so the second model could never be added (the UI just alerted "alias already exists"). Aliases are now disambiguated deterministically — bare last segment when free, then parent-qualified (`codebuddy-gpt-5.5`), then a numeric suffix — while re-adding the exact same model id is still blocked. Regression guard: `tests/unit/passthrough-alias-1850.test.ts`. (thanks @arpicato)
+- **fix(translator):** preserve a Gemini `functionResponse` co-located with other parts (another `functionCall`, or trailing `text`) in the same content when translating **Gemini → OpenAI** ([#6376](https://github.com/diegosouzapw/OmniRoute/pull/6376)). `convertGeminiContent()` early-returned the tool message on the first `functionResponse` part, dropping any co-located parts; such contents are now pre-split (one tool message per `functionResponse`, emitted first, plus one message for the remaining parts). Regression guard: `tests/unit/gemini-to-openai-function-response.test.ts`. (thanks @warelik)
+- **fix(headroom):** detect a python interpreter managed by **mise / pyenv / asdf / conda** (port from 9router#2353, [#6382](https://github.com/diegosouzapw/OmniRoute/pull/6382)). Headroom's python probe (`src/lib/headroom/detect.ts`) searched a hardcoded `PATH`, but version managers expose their interpreters via shim dirs that only join `PATH` through interactive-shell activation — which the non-interactive server never runs, so a managed python (≥3.10) was invisible and Headroom reported it missing. The search path now prepends the well-known shim/bin dirs (`~/.local/share/mise/shims`, `~/.pyenv/shims`, `~/.asdf/shims`, `$CONDA_PREFIX/bin`, `~/.local/bin`, respecting `MISE_DATA_DIR`/`PYENV_ROOT`/`ASDF_DATA_DIR` when set), and a new `HEADROOM_PYTHON` env override lets operators point straight at their interpreter (mirroring `HEADROOM_URL`). Still shell-free (`execFileSync`). Regression guard: `tests/unit/headroom-detect.test.ts` (5). (thanks @loopyd)
+- **fix(executors):** strip the OpenAI-Codex/Claude-CLI `client_metadata` passthrough field for **NVIDIA** requests (port from 9router#1887, [#6411](https://github.com/diegosouzapw/OmniRoute/pull/6411)). NVIDIA's OpenAI-compatible wrapper rejects it with `400 Unsupported parameter`, the same class already handled for `cerebras`/`mistral`; `nvidia` (executor `default`) was missing from the strip allowlist so Codex/Claude-Code passthrough requests 400'd. Regression guard: `tests/unit/executor-default-strip-client-metadata.test.ts` (+nvidia case). (thanks @phidinhmanh)
+- **fix(translator):** strip the Claude-style `thinking` field for **NVIDIA `z-ai/glm-5.2`** (port from 9router#2023, [#6413](https://github.com/diegosouzapw/OmniRoute/pull/6413)). NVIDIA's OpenAI-compatible wrapper 400s on `thinking` (a Claude-format client routed here leaves a `thinking:{type:"adaptive"}`); the existing strip rule only dropped `reasoning`. Same class already handled for `minimax-m2.7`. Regression guard: `tests/unit/nvidia-minimax-thinking-strip.test.ts` (+glm-5.2 case). (thanks @phidinhmanh)
+- **fix(translator):** suppress the streamed `</think>` close marker for the **Antigravity IDE** client (port from 9router#1061, [#6415](https://github.com/diegosouzapw/OmniRoute/pull/6415)). On thinking-only turns Antigravity rendered a bare `</think>` as the sole visible content, tripping its loop-detection and wasting requests. Antigravity's UA (`vscode/<v> (Antigravity/<v>)`) is added to the marker-suppress allowlist (alongside OpenCode); Claude Code / Cursor still get the marker, and `x-omniroute-thinking-marker: on` force-restores it. Regression guard: `tests/unit/think-close-marker-suppress-5245.test.ts`. (thanks @abdofallah)
+- **fix(executors):** strip nested `reasoning_content` from messages for **Mistral** (port from 9router#1649, [#6417](https://github.com/diegosouzapw/OmniRoute/pull/6417)). Mistral's API returns `422 extra_forbidden` when an assistant message carries `reasoning_content` (replayed thinking from a prior turn, e.g. via the Codex `/responses` path); the generic top-level 400 field-downgrade retry never covered the nested per-message field. `DefaultExecutor` now strips it for provider `mistral` only, so DeepSeek (which requires replayed `reasoning_content`) is unaffected. Regression guard: `tests/unit/mistral-strip-reasoning-content-1649.test.ts`. (thanks @xxy9468615)
+- **fix(executors):** strip the `client_metadata` passthrough field on the **OpenCode** path (port from 9router#1442, [#6418](https://github.com/diegosouzapw/OmniRoute/pull/6418)). OpenCode upstreams (e.g. `kimi-k2.6` via opencode-go) reject it with `400 "Extra inputs are not permitted, field: 'client_metadata'"`; the DefaultExecutor strip only covered cerebras/mistral and `OpencodeExecutor` extends `BaseExecutor` directly, so nothing removed it there. Regression guard: `tests/unit/opencode-strip-client-metadata-1442.test.ts`. (thanks @yanpaing007)
+- **fix(executors):** inject the `reasoning_content` echo for the native **Moonshot Kimi** provider (port from 9router#1480, [#6419](https://github.com/diegosouzapw/OmniRoute/pull/6419)). Kimi (executor `default`) is a thinking-mode upstream that 400s with "reasoning_content must be passed back" when a prior assistant turn lacks it; the placeholder injection was only wired into the OpenCode meta-provider, so direct multi-turn Kimi conversations failed. Scoped to `kimi` (gateway-served models matching the thinking-model name pattern are unaffected). Regression guard: `tests/unit/kimi-native-reasoning-injected-1480.test.ts`. (thanks @2220258345)
+- **fix(executors):** recover from a strict gateway's `context_management: Extra inputs are not permitted` 400 (port from 9router#1468, [#6420](https://github.com/diegosouzapw/OmniRoute/pull/6420)). **Claude Code** always sends a top-level `context_management` field; strict anthropic-compatible gateways reject it. The dedicated context-editing 400-fallback only fired when OmniRoute's own `contextEditing` feature was enabled (default off), so a client-sent field passed through untouched and 400'd. `context_management` is now in the generic reactive field-strip list, so it's stripped-and-retried once regardless of the feature flag (with correct request re-signing for claude-compatible relays). Regression guard: `tests/unit/provider-field-strips.test.ts`. (thanks @ohahe52-dot)
+- **fix(network):** enable **RFC 8305 Happy Eyeballs** (`autoSelectFamily`) on the direct-egress undici dispatcher (port from 9router#1237, [#6423](https://github.com/diegosouzapw/OmniRoute/pull/6423)). When DNS returns both IPv6 (AAAA) and IPv4 (A) and the IPv6 route is broken (e.g. a NAT64 `64:ff9b::` prefix without routing), undici tried IPv6 first and hung until `ETIMEDOUT` (then a 502 + account lockout), even though `curl` reached the same host. The direct dispatcher now races both families and uses whichever connects first. Proxy paths pin family via `proxyTls` and are unaffected. Regression guard: `tests/unit/direct-dispatcher-pipelining-4580.test.ts`. (thanks @adentdk)
+- **fix(combo):** round-robin now advances the rotation pointer past the model that **actually served**, not the eagerly-scheduled one (port from 9router#948, [#6428](https://github.com/diegosouzapw/OmniRoute/pull/6428)). With `stickyLimit: 1` (true round-robin), when the scheduled model failed and a _different_ model served via fallback, the counter had already advanced +1 from the scheduled index — so the next request reused the fallback-served model, degrading round-robin into hot-spotting on whichever model was healthy. The pointer now advances to the served index + 1 (mirroring the sticky-limit>1 path). Session-stickiness (#3825) and distribution are preserved. Regression guard: `tests/unit/combo-rr-fallback-advance-948.test.ts`. (thanks @binsarjr)
+- **fix(sse):** a non-string `model` field is now rejected with a `400` before the resolver, instead of crashing downstream `.toLowerCase()`/`.split()` calls into an empty-body `500` that escapes the error sanitizer ([#6407](https://github.com/diegosouzapw/OmniRoute/issues/6407)). Regression guard: `tests/unit/chat-non-string-model-6407.test.ts`. (thanks @chirag127)
+- **fix(api):** unknown `/api/*` routes now return a JSON `404` (instead of the dashboard HTML shell) and scalar chat params (`model`/`temperature`/etc.) are validated **before** the provider lookup so malformed requests fail fast with a clear `400` ([#6424](https://github.com/diegosouzapw/OmniRoute/issues/6424), [#6412](https://github.com/diegosouzapw/OmniRoute/issues/6412)). Regression guards: `tests/unit/api/api-catchall-json-404.test.ts`, `tests/unit/chat-early-schema-validation-6412.test.ts`. (thanks @chirag127)
+- **fix(api):** `/v1/chat/completions` now rejects a non-JSON `Content-Type` with a `400` before parsing the body ([#6414](https://github.com/diegosouzapw/OmniRoute/issues/6414)). Regression guard: `tests/unit/v1-chat-completions-content-type-6414.test.ts`. (thanks @chirag127)
+- **fix(api):** the `X-OmniRoute-Compression` response header is now echoed on `/v1/chat/completions` and `/v1/completions` ([#6422](https://github.com/diegosouzapw/OmniRoute/issues/6422)). Regression guard: `tests/unit/compression-header-echo-6422.test.ts`. (thanks @chirag127)
+- **fix(api):** concurrent `GET /v1/models` requests are coalesced into a single catalog build ([#6408](https://github.com/diegosouzapw/OmniRoute/issues/6408)). Regression guard: `tests/unit/v1-models-concurrent-6408.test.ts`. (thanks @chirag127)
+- **fix(api):** `/v1/completions` now echoes the requested `body.model` in its JSON + streamed responses ([#6429](https://github.com/diegosouzapw/OmniRoute/pull/6429)). Regression guard: `tests/unit/completions-body-model-echo.test.ts`. (thanks @chirag127)
+- **fix(api):** env-var master keys now see the full `/v1/models` catalog ([#6406](https://github.com/diegosouzapw/OmniRoute/issues/6406)). Regression guard: `tests/unit/models-catalog-envkey-6406.test.ts`. (thanks @chirag127)
+- **fix(api):** non-streaming `/v1/completions` responses now echo `body.model` aligned with the `X-OmniRoute-Model` header ([#6426](https://github.com/diegosouzapw/OmniRoute/issues/6426)). Regression guard: `tests/unit/v1-completions-model-header-match-6426.test.ts`. (thanks @chirag127)
+- **fix(api):** unknown `/v1/*` routes now return a JSON `404 not_found` instead of the Next.js dashboard HTML shell ([#6405](https://github.com/diegosouzapw/OmniRoute/issues/6405)). Regression guard: `tests/unit/api/v1-catchall-json-404.test.ts`. (thanks @chirag127)
+- **fix(api):** the per-connection provider models route now degrades to the shipped catalog when a provider's `/models` endpoint answers with a redirect ([#6267](https://github.com/diegosouzapw/OmniRoute/issues/6267)) — a `qwen-web` import failed with a raw `Redirect blocked … (307)` 503. `safeOutboundFetch` throws `REDIRECT_BLOCKED` on the 307, `getSafeOutboundFetchErrorStatus` maps it to 503, and `buildDiscoveryErrorFallbackResponse` treated every 503 as a hard error — so the non-empty `getModelsByProviderId("qwen-web")` catalog was never surfaced. A models-endpoint redirect is not a fixable-config error (unlike `URL_GUARD_BLOCKED`/`INVALID_URL`, which stay hard errors), so it now falls back to the local/cached catalog before the 503 short-circuit. General fix — covers any config-driven provider that 307s. Regression guard: `tests/unit/provider-models-qwen-web-redirect-6267.test.ts`. (thanks @chirag127)
+- **fix(api):** the per-connection provider models route (MCP `list_models_catalog` + the dashboard import view) now merges USER-ADDED custom models into its response ([#6247](https://github.com/diegosouzapw/OmniRoute/issues/6247)) — custom models live in the `key_value` namespace `customModels`, which the live REST `/api/v1/models` already merges, but `src/app/api/providers/[id]/models/route.ts` never read `getCustomModels`, so custom models were dropped on both the discovery-success and local_catalog paths. They are now folded into the returned model list (deduped by id, stamped `owned_by: provider`), fixing MCP + the dashboard import view in one place. Regression guard: `tests/unit/provider-models-custom-merge-6247.test.ts`. (thanks @RCrushMe)
+
+- **fix(providers):** GitLab Duo tool-calling follow-up turns no longer fail upstream with `422 {"detail":"Validation error"}` (tokens 0/0, rejected pre-inference). The #6234 tool-result-feedback fix serialized the **entire** multi-turn conversation into GitLab's single-file `code_suggestions` (`small_file`) generation endpoint — folded history that turn-N sent as an oversized `current_file.content_above_cursor` **and** duplicated verbatim into `user_instruction`, tripping the AI-Gateway's `small_file` validation guard. The executor now **bounds** that prompt: it keeps system + latest user message + the most-recent tool round (dropping older turns), caps oversized tool results, and stops duplicating the full prompt into `user_instruction` (which now carries only the short latest user message) — while still feeding the most-recent tool result back so the agent continues (`open-sse/executors/gitlab.ts`, [#6220](https://github.com/diegosouzapw/OmniRoute/issues/6220)). The unit test covers the bounding logic; the upstream 422→200 clearing is VPS-only (Hard Rule #18). Regression guard: `tests/unit/gitlab-tool-exchange-bounded-6220.test.ts`.
+
+- **fix(i18n):** the provider-detail (`/dashboard/providers/[id]`) connection-status filter labels no longer render as `__MISSING__:All` / `__MISSING__:Active` / `__MISSING__:Error` / `__MISSING__:Banned` / `__MISSING__:CreditsExhausted` in non-English locales (notably pt-BR) ([#6290](https://github.com/diegosouzapw/OmniRoute/issues/6290)). Root cause was **not** the namespace mismatch the issue guessed — the `providers.filter*` keys resolve correctly in `en.json`; the debt lived in the locale mirrors (`src/i18n/messages/*.json`), where these five keys carried the `__MISSING__:` sync sentinel in ~15 locales and were absent entirely in ~26 others, so next-intl found the key and echoed the sentinel verbatim. All 40 non-English/-Chinese mirrors now ship real translations for the five `providers.filter*` labels. Regression guard: `tests/unit/i18n-provider-filter-keys-6290.test.ts`. (thanks @diegosouzapw)
+
+- **fix(providers):** the `copilot-m365-web` streaming executor now emits `debug`-level WebSocket diagnostics ([#6210](https://github.com/diegosouzapw/OmniRoute/issues/6210)) — the outbound WS URL (with the `access_token` **redacted** via `redactWsUrl()`), handshake success/failure, and each received SignalR frame's `type`/`target`. Previously the streaming path logged nothing, so an empty `content:null` response (the M365 Education / Starter tier symptom fixed in #6234) was undiagnosable even at `APP_LOG_LEVEL=debug`. The change is debug-level and side-effect-free — it does not alter streaming behavior or the frame parser, and the token never reaches the logs. Regression guard: `tests/unit/copilot-m365-web-logging-6210.test.ts` (thanks @qpeyba)
+
+- **fix(resilience):** a round-robin combo no longer returns `503 all upstream accounts are unavailable` when a compatibility-rejected target is actually healthy ([#6238](https://github.com/diegosouzapw/OmniRoute/issues/6238)). `filterTargetsByRequestCompatibility` drops request-incompatible targets (tool/vision/structured-output unsupported, or below the required context window) **before** any availability check runs, and its `compatible.length === 0` safety net only fired when _all_ targets were filtered — not when the kept targets later all turned out runtime-unavailable (circuit-open / cooldown / no credentials). So a combo could 503 while a compat-rejected-but-healthy provider sat unused. `handleRoundRobinCombo` now keeps the compat-rejected set and, when every compat-kept target was skipped without a single real attempt, probes those rejected targets as a **last-resort fallback tier** (via the new pure `open-sse/services/combo/comboCompatFallback.ts`) before crystallizing the 503. Regression guard: `tests/unit/combo-roundrobin-compat-fallback-6238.test.ts`. (thanks @ThongAccount)
+
+- **fix(startup):** best-effort self-heal for a corrupted Turbopack dev cache on Windows ([#6289](https://github.com/diegosouzapw/OmniRoute/issues/6289)). On Windows, `pnpm dev` can fail at startup when Turbopack `mmap`s a persistent-cache SST file and the OS refuses the mapping (`os error 1455` — "paging file too small"), which Turbopack surfaces as a misleading `Module not found: Can't resolve '@/shared/utils/machine'`. This is a **known upstream Turbopack cache-corruption bug — not our code**. The dev launcher (`scripts/dev/run-next.mjs`) now wraps `nextApp.prepare()` and, when it rejects with that signature (`isTurbopackCacheCorruption` in the new `scripts/dev/turbopackCacheHeal.mjs`), purges `.build/next/**/cache/turbopack` and retries **once** with a clear log. **Caveat — best-effort only:** the corruption often surfaces as a runtime overlay rather than a `prepare()` rejection, so this cannot always intercept it; the reliable remedy remains manually deleting the Turbopack cache dir. Regression guard: `tests/unit/turbopack-cache-heal-6289.test.ts`. (thanks @chirag127)
+
+- **fix(providers):** qodercli PAT auth no longer fails with `spawn qodercli ENOENT` on Windows ([#6263](https://github.com/diegosouzapw/OmniRoute/issues/6263)) — `spawnQoderCli` spawned the bare `qodercli` name with `shell:false` and an unenriched env, so the npm `.cmd` wrapper under `%APPDATA%\npm` (a user-PATH directory) was never resolved. It now resolves the absolute `.cmd`/`.exe` path through the existing `getCliRuntimeStatus("qoder")` resolver in `src/shared/services/cliRuntime.ts` (memoized), spawns with `shell` when the target is a `.cmd`/`.bat`, and uses the cliRuntime-enriched env (PATH + PATHEXT + APPDATA); the ENOENT error now lists the searched paths plus the `CLI_QODER_BIN` override. End-to-end spawn on a real Windows host is host-only (Hard Rule #18); the path-resolution logic is unit-tested. Regression guard: `tests/unit/qodercli-windows-resolve-6263.test.ts`. (thanks @chirag127)
+
+- **fix(sse):** the reasoning-token buffer no longer inflates **probe-sized `max_tokens`** ([#6274](https://github.com/diegosouzapw/OmniRoute/issues/6274)) — Claude Code's `/model` capability check sends `max_tokens: 1`, but for a thinking-capable model with a large output cap (e.g. `glm-5.2`) the #3587 headroom heuristic (`max(current + 1000, ceil(current * 1.5))`) rewrote it to `1001` and forwarded that upstream, wasting tokens on a request that was never a genuine reasoning budget. `resolveReasoningBufferedMaxTokens()` (`open-sse/services/reasoningTokenBuffer.ts`) now short-circuits and returns the caller's value verbatim when it is below the new `REASONING_BUFFER_MIN_TRIGGER` (256) threshold — a tiny explicit limit is a probe, not a reasoning request. Real budgets still receive the #3587 headroom unchanged, and the guard runs after the existing capability checks so unknown / non-reasoning models keep returning `null`. Regression guard: `tests/unit/reasoning-token-buffer-6274.test.ts`. (thanks @brightfiscalband)
+
+- **fix(cli):** `omniroute reset-password` now works as a real subcommand, and password resets over piped (non-TTY) stdin actually apply ([#6261](https://github.com/diegosouzapw/OmniRoute/issues/6261), [#6258](https://github.com/diegosouzapw/OmniRoute/issues/6258)). Two coupled defects: (1) **#6261** — `bin/omniroute.mjs` routed everything through Commander with only two pre-Commander bypasses (`--mcp`, `reset-encrypted-columns`), so `omniroute reset-password` was rejected as an unknown command; only the separate `omniroute-reset-password` bin worked, while the docs falsely advertised the subcommand (incl. a bogus "legacy alias still works"). A pre-Commander bypass mirroring `reset-encrypted-columns` now dynamically imports `bin/reset-password.mjs` (which self-executes) before Commander parses; the three doc lines were corrected. (2) **#6258** — `bin/reset-password.mjs` issued two sequential `rl.question` prompts; under piped stdin the second read never settled at EOF, so `main()` never reached `resetManagementPassword` and the reset was a silent no-op (both prompts printed, no success, password unchanged). The CLI now detects non-TTY stdin and reads it once (first line = password, second line = confirm if present, else reused), adds a `--password-stdin` flag (entire stdin is the password, no confirmation), and exits `0` explicitly so the success line always flushes; interactive TTY behavior is unchanged. Regression guard: `tests/unit/reset-password-cli-6261-6258.test.ts` (3). (thanks @chirag127)
+
+- **fix(db):** the mass-migration **safety abort** now tells the operator how to bypass it and stops flooding the log ([#6260](https://github.com/diegosouzapw/OmniRoute/issues/6260)) — after restoring a backup that wiped the migration tracking table, `runMigrations()` threw the abort on every downstream `ensureDbInitialized()`, re-logging the full banner 11+ times, and the message never mentioned the existing `OMNIROUTE_MAX_PENDING_MIGRATIONS` escape hatch. The abort text now appends a bypass hint (set `OMNIROUTE_MAX_PENDING_MIGRATIONS=0` in `server.env` / `DATA_DIR/.env`), and a new `MigrationSafetyAbortError` is memoized so repeated calls in the same process throw the same instance and emit a single concise line instead of the full cascade. Regression guard: `tests/unit/migration-safety-abort-6260.test.ts`. (thanks @chirag127)
+
+- **fix(auth):** importing a **distinct** Codex/ChatGPT OAuth `auth.json` is no longer falsely rejected as "already exists" when it belongs to a different user in the same workspace ([#6301](https://github.com/diegosouzapw/OmniRoute/issues/6301)). `findExistingCodexConnection` (in `src/lib/oauth/utils/codexAuthImport.ts`) deduped **only** on `providerSpecificData.workspaceId === accountId`, where `accountId` is the shared `chatgpt_account_id`/`tokens.account_id` — so two members of the same ChatGPT Team collapsed onto a single connection (409 `duplicate_account`). The id_token's `https://api.openai.com/auth` claim carries a per-user `chatgpt_user_id` alongside the workspace id (the device-flow path already persisted it as `chatgptUserId`, but the import path did not). Now `parseAndValidateCodexAuth` extracts `userId` (`chatgpt_user_id` → `user_id` → JWT `sub`) into `ParsedCodexAuth`, the create/update paths persist `chatgptUserId` in `providerSpecificData` (mirroring `codex.ts`), and dedup keys on `workspaceId` **AND** `chatgptUserId` — with a backward-compat fallback to legacy accountId-only matching when no stored connection for that workspace records a `chatgptUserId`, so genuinely-same accounts still dedup. Regression guard: `tests/unit/codex-auth-import-userid-dedup-6301.test.ts` (4). (thanks @anungma)
+
+- **fix(providers):** importing models for the **venice-web** provider no longer fails with a red "Provider venice-web does not support models listing" ([#6269](https://github.com/diegosouzapw/OmniRoute/issues/6269)). `venice-web` is a web-cookie provider with an executor but no upstream `/v1/models` endpoint and no registry `models`, so the models route fell through to the tail `400`. Mirroring the `jules`/`linkup-search`/`ollama-search` fix (#5569), it now ships a static local catalog entry in `src/lib/providers/staticModels.ts` — seeding the current Venice lineup (`venice-uncensored`, `llama-3.3-70b`, `qwen3-235b`, `qwen3-4b`, `deepseek-r1-671b`; Venice rotates its catalog, see docs.venice.ai/models/overview) — so the route returns `200` with `source:"local_catalog"`, `intentional:true`. Regression guard: `tests/unit/static-models-venice-web-6269.test.ts`. (thanks @chirag127)
+
+- **fix(api):** the specialty model catalogs (`/v1/embeddings`, `/v1/images`, `/v1/music`, `/v1/videos` model lists) are now derived from the **unified catalog filtered by a predicate** (`getSpecialtyModelsResponse`) instead of ad-hoc per-route logic, so they consistently respect active-credential visibility and stay in sync with the main catalog ([#6303](https://github.com/diegosouzapw/OmniRoute/pull/6303)). Regression guard: `tests/unit/specialty-model-catalog-routes.test.ts`. (thanks @makcimbx)
+- **fix(api):** the agent-bridge server route now resolves the MITM manager via a **dynamic `import("@/mitm/manager.runtime")`** so Turbopack does not statically pull the stub (or over-bundle the manager), and the agent-skills generator anchors its output base path with `path.join(process.cwd(), …)` so Turbopack's static analyzer stops tracing the whole project root ([#6329](https://github.com/diegosouzapw/OmniRoute/issues/6329), [#6366](https://github.com/diegosouzapw/OmniRoute/pull/6366)). Regression guard: `tests/unit/agent-bridge-server-route-dynamic-import.test.ts`. (thanks @Iammilansoni)
+- **fix(api):** internal probes (combo-test, cloud-sync verify) now pick a **management-scoped / allow-all API key** instead of naively grabbing `getApiKeys()[0]` — a restricted `self:usage` first row made the probe fail with "Model X is not allowed for this API key" even when the combo path was healthy (`pickApiKeyForInternalUse` in `src/lib/db/apiKeys.ts`). The API-manager model editor also falls back to `/api/models?all=true` when `/v1/models` is catalog-protected ([#6372](https://github.com/diegosouzapw/OmniRoute/pull/6372)). Regression guard: `tests/unit/pick-internal-api-key-6372.test.ts`. (thanks @jmengit)
+- **fix(live-ws):** the Live Dashboard WebSocket server now **rejects on bind failure** (e.g. `EADDRINUSE` when the API bridge already holds the port) instead of letting the error surface as an unhandled `error` event that crash-loops the process — the `error` listener is attached to `wss` (not `server`) and releases the EventBus subscription on a failed start ([#6324](https://github.com/diegosouzapw/OmniRoute/issues/6324)). Regression guard: `tests/unit/live-ws-eaddrinuse-6324.test.ts`. (thanks @vinayakkulkarni)
+- **fix(dashboard):** the Home provider-topology widget now trusts the live provider-metrics snapshot — it uses `topology.errorProvider` and live `activeRequests` directly instead of re-deriving state from a stale `lastErrorAt` or applying a frontend timeout filter, so the topology reflects real-time provider health ([#6322](https://github.com/diegosouzapw/OmniRoute/pull/6322)). Regression guard: `tests/unit/home-provider-topology-live-state.test.ts`. (thanks @xz-dev)
+- **fix(sse):** strip zero-width markers from streamed **tool-call arguments** — a follow-up to [#5857](https://github.com/diegosouzapw/OmniRoute/pull/5857). That PR removed injected zero-width joiners (U+200D) from streamed assistant text/reasoning but deliberately left tool-call argument JSON byte-exact. The request-side obfuscation (`open-sse/services/claudeCodeObfuscation.ts`) injects ZWJ into agent words — including the temp path inside the Bash tool description — and Claude models copy that verbatim into generated commands, which are delivered as tool-call arguments rather than assistant text. As a result the ZWJ survived and corrupted code blocks (e.g. a temp path rendered with an invisible joiner). Now `open-sse/handlers/responseSanitizer.ts` strips zero-width code points from tool-call argument strings at every emit site (OpenAI non-stream/stream chat `tool_calls` + legacy `function_call`, native Responses `function_call` items, the OpenAI→Responses conversion, and the native Responses streaming `response.function_call_arguments.delta/.done` events). Only zero-width code points are removed; JSON structure and all other bytes stay identical (no parse/restringify), so normal arguments remain byte-exact. Regression guard: 6 new cases in `tests/unit/response-sanitizer.test.ts` (suite 50/50).
+
+- **fix(nodejs):** the default app log path now resolves under `DATA_DIR` (`~/.omniroute/logs/application/app.log`) instead of `process.cwd()` ([#6197](https://github.com/diegosouzapw/OmniRoute/issues/6197)) — the globally-installed CLI runs from an arbitrary working directory, so anchoring the default to cwd made file logging silently write to (or no-op under) an unrelated directory, contradicting the documented `.env.example` default. `getAppLogFilePath()` now computes the default lazily via the pure `resolveDataDir()` resolver (honours a per-process `DATA_DIR`, no directory-creation side effect); an explicit `APP_LOG_FILE_PATH` still wins. Regression guard: `tests/unit/logenv-datadir-path-6197.test.ts` (3).
+
+- **fix(docker):** AgentBridge/`startMitm` no longer aborts in containers/headless when the Antigravity-default DNS step can't write `/etc/hosts` ([#6127](https://github.com/diegosouzapw/OmniRoute/issues/6127)), and the privileged command's stderr now reaches `app.log` instead of only a bare exit code hitting the toast ([#6198](https://github.com/diegosouzapw/OmniRoute/issues/6198)). The default DNS step (`addDNSEntry`) was called unguarded while cert install and the two sibling DNS steps were each best-effort — in the runtime Docker image (`USER node`, no `sudo`, read-only `/etc/hosts`) it threw `Command failed with code 1` out of `startMitmInternal` and killed the whole start, discarding the stderr. The three DNS steps are extracted into a best-effort `provisionDnsEntries()` where each failure is logged with the full `err` (stderr included, folded in by `systemCommands.ts`) and never aborts the start. Regression guard: `tests/unit/mitm-dns-graceful-degrade-6127.test.ts` (4).
+
+- **fix(providers):** copilot-m365-web now supports the M365 Education "Starter / OfficeWebIncludedCopilot" tier and no longer returns an empty `content:null` stream ([#6210](https://github.com/diegosouzapw/OmniRoute/issues/6210)). Two gaps: (1) `buildWsUrl()` hardcoded the individual-consumer scenario (`OfficeWebPaidConsumerCopilot`, `isEdu=false`) — the EDU tier is now opt-in via `providerSpecificData.tier="edu"`, emitting `scenario=OfficeWebIncludedCopilot`/`isEdu=true` (the individual path is unchanged); (2) the EDU/GPT-5.5 path streams deltas via `arguments[0].writeAtCursor` (incremental) instead of only `messages[].text` (accumulated snapshots), which the parser dropped — a new `accumulateBotContent()` folds both formats, with `type:2 item.result.message` as a last-resort fallback. Regression guard: `tests/unit/copilot-m365-edu-writeatcursor-6210.test.ts` (10). (thanks @qpeyba)
+
+- **fix(providers):** GitLab Duo executor now feeds tool results back into the prompt instead of looping ([#6220](https://github.com/diegosouzapw/OmniRoute/issues/6220)) — `buildPrompt()` branched only on `system`/`user` and took `userParts.at(-1)`, silently dropping the `assistant{tool_calls}` + `tool{result}` turns the client appended, so the reconstructed prompt was byte-identical to turn 1 and the model re-emitted the same `<tool>` call forever. When a tool exchange is present the full conversation is now serialized, folding each tool result back keyed by its `tool_call_id`; simple conversations keep the legacy shape. Complements the tool_call emission from [#6051](https://github.com/diegosouzapw/OmniRoute/issues/6051) (the `kilo-duplicate` label was a false positive — different, sequential defect). Regression guard: `tests/unit/gitlab-tool-result-feedback-6220.test.ts` (4).
+
+- **fix(providers):** opencode-go/opencode-zen can now synthesize the OpenCode CLI identity headers Cloudflare requires on VPS egress ([#5997](https://github.com/diegosouzapw/OmniRoute/issues/5997)) — on a datacenter VPS, `opencode.ai/zen/go/v1/chat/completions` 403s (HTML challenge) requests lacking CLI identity, while the reporter's control curl proved that `User-Agent: opencode-cli/1.0.0` + `x-opencode-client: cli` + `x-opencode-project: default` + fresh request/session UUIDs succeed. Opt-in via `OPENCODE_SYNTHESIZE_CLI_HEADERS=true` (values overridable via `OPENCODE_GO_USER_AGENT`/`OPENCODE_USER_AGENT`/`OPENCODE_CLIENT`/`OPENCODE_PROJECT`); it fills only headers the client did not already send. Kept **off by default** — the forward-only path is deliberate (fabricating a wrong value risks upstream rejection; a prior dedup regressed with `opencode/local`), so this replaces the fragile local header-injection shim without changing default behavior. Regression guard: `tests/unit/opencode-cli-headers-synthesis-5997.test.ts` (6). (thanks @aleksesipenko)
+- fix(resilience): sticky session affinity now evicts and fails over to another account when the pinned account is exhausted/unavailable (#6219)
+
+- fix(sse): Responses API passthrough now drops internal commentary-phase output before forwarding to clients (gated by RESPONSES_PASSTHROUGH_DROP_COMMENTARY, default on) (#6199)
+
+- **fix(sse):** tool-call function schemas with a root `type: null` are now coerced to `type: "object"` before dispatch ([#6359](https://github.com/diegosouzapw/OmniRoute/issues/6359)) — clients like the Codex app emit `parameters: { type: null, ... }` for some tools, which OpenAI-compatible upstreams reject with `400 Invalid schema for function '...': schema must be a JSON Schema of 'type: "object"', got 'type: null'`, failing the whole request. `toolSchemaSanitizer` already stripped the null; it now re-adds the mandatory root `"object"` type (and empty `properties`/open `additionalProperties` when absent). Combinator roots (`anyOf`/`oneOf`/`allOf`) and explicit root types are left untouched. Regression guard: 5 new cases in `tests/unit/tool-schema-sanitizer.test.mjs`.
+
+- **fix(docker):** AgentBridge no longer fails to start on npm/Electron/VPS installs with "MITM manager stub reached at runtime" ([#6344](https://github.com/diegosouzapw/OmniRoute/issues/6344)) — v3.8.45 flipped the production bundler default to Turbopack, but `next.config.mjs` aliased `@/mitm/manager` to its Docker-only degraded stub **unconditionally**. That was harmless while Docker (which sets the alias intentionally for #3390 graceful degradation) was the sole Turbopack consumer, but once every artifact built with Turbopack the stub shipped to all non-Docker users and `startMitm` threw on the first Agent-Bridge start. The alias is now opt-in via `OMNIROUTE_MITM_STUB=1` (set only by the Dockerfile) through the shared `scripts/build/mitm-stub-flag.mjs` helper; default builds bundle the real manager. Regression guard: `tests/unit/mitm-stub-alias-6344.test.mjs` (4).
+
+- **fix(proxy):** stop the v3.8.44 proxy regression that leaked the real IP and disabled healthy proxies ([#6246](https://github.com/diegosouzapw/OmniRoute/issues/6246)). Two coupled defects from the new health scheduler: (1) **IP leak** — when a proxy assigned to a connection was marked `inactive`, resolution fell through to a **direct** egress instead of blocking, exposing the operator's real IP; (2) **over-deactivation** — the sweep flipped a proxy to `inactive` on the **first** failed probe and counted our own 5s timeout / a probe-target `5xx` as the proxy's fault, so healthy paid proxies vanished from egress selection ("my proxies are not being used anymore"). Fix: the sweep decision is extracted into a pure, network-free `decideProxyHealthAction` (`src/lib/proxyHealth/decision.ts`) — by default the health check now **only counts/logs and never downgrades status** (a proxy is downgraded/removed only with `PROXY_AUTO_REMOVE=true`, after `PROXY_AUTO_REMOVE_AFTER` **consecutive** conclusive failures); probes are classified tri-state so an inconclusive result (our timeout, or a `5xx` from the probe target) never penalizes the proxy, and the probe timeout is raised 5s→15s. Separately, `safeResolveProxy` now **fails closed** via the existing policy: a connection whose assigned proxy is dead is blocked instead of leaking direct (`hasBlockingProxyAssignment`), honoring the explicit `proxy off` toggles and the `PROXY_FAIL_OPEN=true` opt-out. Existing proxies stuck `inactive` by the old behavior need a one-time manual re-activate (the operator owns proxy status). Regression guards: `tests/unit/proxy-health-decide-action-6246.test.ts`, `tests/unit/proxy-assigned-unavailable-6246.test.ts`.
+
+- **fix(proxy):** make "Test All" read-only and add bulk enable/disable ([#6246](https://github.com/diegosouzapw/OmniRoute/issues/6246)). Complements the core fail-closed / scheduler fix (#6296) with the two remaining reporter asks. (1) The **"Test All" button** (`POST /api/settings/proxies/auto-test`) used to flip a proxy to `inactive` on a failed reachability probe; since the egress selector excludes `inactive` proxies, a flaky probe (an unreachable `httpbin.org`, a proxy that blocks `HEAD`, or a slow paid proxy) silently disabled every proxy that failed — "Test All" is now **read-only by default** (only the operator sets a proxy active/inactive; opt back into the legacy test-and-set with `PROXY_HEALTH_AUTO_DEACTIVATE=true`). (2) Adds a **bulk enable/disable** proxies endpoint + toolbar action (`POST /api/settings/proxies/batch-activate`) so an operator can re-activate proxies in one click. Regression guard: `tests/unit/proxy-health-6246.test.ts`. (thanks @tenshiak)
+
+- **chatcore (tools): stop the default 128-tool cap from silently dropping opencode's `task`/MCP tools.** opencode (used as an MCP/agent host) sends a large tool list; when it exceeds the speculative `MAX_TOOLS_LIMIT` (128) default, `truncateToolList` did a blind `tools.slice(0, 128)`, dropping every tool past index 128 — including opencode's built-in `task` tool (subagent launch) and many MCP tools, so models routed through OmniRoute could no longer spawn subagents or reach part of their tools. The cap exists to avoid upstream `400`s for providers with real hard limits (e.g. grok-cli 200), so it is kept for those: detection of the opencode client (`isOpencodeClient` — any `x-opencode-*` header, or `opencode` in the user-agent) now only bypasses the **speculative 128 default**, never a known provider ceiling. Precedence is explicit — a proactive/detected provider limit always truncates (even for opencode); otherwise opencode forwards its full tool list; otherwise the unchanged 128 default applies to every other client. Refactors `getEffectiveToolLimit` into `getKnownToolLimit(provider) ?? DEFAULT_LIMIT` (byte-identical for existing callers) and fixes a cosmetic debug-log that reported the truncated count instead of the original. Regression guard: `tests/unit/tool-limit-detector.test.ts`.
+
+- **fix(mitm):** the macOS MITM-cert install check now matches the system keychain again. `security find-certificate -a -Z` prints the SHA-1 as a colon-less hex string, but the installed-check compared it against `getCertFingerprint()`'s colon-separated form, so the substring match never hit — the cert was reported as not-installed and re-prompted for the sudo install on every run. Fingerprints are now normalized (colons stripped, upper-cased) on both sides via the extracted `macCertOutputHasFingerprint` helper. Regression guard: `tests/unit/mitm-cert-mac-fingerprint.test.ts`. ([#6204](https://github.com/diegosouzapw/OmniRoute/pull/6204), closes [#6134](https://github.com/diegosouzapw/OmniRoute/issues/6134) — thanks @rianonehub)
+
+- **fix(api):** `/v1/messages/count_tokens` now counts `tool_use`, `tool_result` and `thinking` content blocks (and array-form `system` prompts) in the local-estimation path, instead of only `text`. Real agentic conversations keep ~95% of their tokens inside tool results; the previous estimate returned near-zero for them, which silently broke Claude Code's auto-compaction (context grew past the window with no compaction until the upstream API rejected the request). The real provider-side count path is unchanged. Regression guard: `tests/unit/messages-count-tokens-route.test.ts`. ([#6221](https://github.com/diegosouzapw/OmniRoute/pull/6221) — thanks @luweiCN)
+
+- **fix(antigravity):** strip a trailing assistant prefill turn for Vertex Claude models to avoid upstream 400s ([#6114](https://github.com/diegosouzapw/OmniRoute/pull/6114)). Regression guard: `tests/unit/antigravity-claude-prefill-strip.test.ts`. (thanks @anki1kr)
+
+- **fix(security):** the mutable cloud-agent routes (`/api/cloud/credentials/update`, `/api/cloud/models/alias`) now require management auth instead of being treated as public. They were classified as public API routes, so a request without management credentials could update stored cloud-agent credentials and model aliases. They are removed from the public-route set, classified as management routes in the authz pipeline, and gated by `requireManagementAuth`; cloud **read**/auth routes stay public. Regression guards: `tests/unit/cloud-write-auth.test.ts`, `tests/unit/authz/classify.test.ts`, `tests/unit/public-api-routes.test.ts`. ([#6233](https://github.com/diegosouzapw/OmniRoute/pull/6233) — thanks @vittoroliveira-dev)
+
+- **refactor(dashboard):** extract the onboarding-wizard "Open provider details" link target into a pure, unit-tested `buildProviderDetailsHref(connection)` helper. The wizard already routes by `connection.id` (the node UUID) rather than the provider category slug (#6144/#6145); this hardens that behavior behind a tested helper that guards a missing id/connection. Regression guard: `tests/unit/provider-onboarding-href.test.ts`. ([#6166](https://github.com/diegosouzapw/OmniRoute/pull/6166) — thanks @KooshaPari)
+
+- **fix(security):** the doubao synthetic device-id generator now derives its digits via an unbiased crypto-random draw (rejection sampling over `crypto.randomBytes()`) instead of a `% 10` reduction, closing a CodeQL `js/biased-cryptographic-random` finding.
+- **fix(agentSkills):** the GitHub-skills generator now resolves `outputDir` to an absolute path before writing, fixing a regression introduced by #6366 (relative-to-cwd base path) that could write generated skill files to the wrong directory.
+- **fix(security):** `/api/keys/{id}/devices` now checks the HTTP method before auth/validation, returning a `405` for non-GET/DELETE verbs instead of a misleading `401`/`500` (closes a `dast-smoke` QUERY-method finding).
+- **fix(quality):** clear the last 2 heavy quality-gate reds on the release tip (cycle pre-flight).
+- **fix(mitm):** the test suite and CI can never mutate the OS trust store — `OMNIROUTE_SKIP_SYSTEM_TRUST=1` is set globally for tests/CI so `installCert`/`uninstallCert`/`installTproxyCa` skip the privileged OS dispatch ([#6310](https://github.com/diegosouzapw/OmniRoute/pull/6310); full detail is under the [3.8.45] section below — this branch received it via the parallel-cycle sync-back).
+- **fix(api):** `POST /api/github-skills` now Zod-validates its request body; documented the new quality-gate env vars and pinned the merge-integrity GitHub Actions to a commit SHA.
+- **fix(skills):** generate the missing `omni-github-skills` registry entry and align the agent-skills catalog-count tests (follow-up to #6186).
+- **fix(quality):** clear the cycle's 11 net-new ESLint errors and make `validate-release-green` suppressions-aware.
+
+### 📝 Maintenance
+
+- **i18n(it):** add 118 missing Italian (`it`) translations (net-additive — no existing keys dropped, valid JSON), improving Italian UI coverage. ([#6212](https://github.com/diegosouzapw/OmniRoute/pull/6212) — thanks @serverless83)
+- **chore(providers):** remove deprecated **MiMo V2** model entries from the catalogs (xiaomi-mimo, opencode-go, zenmux-free, audio TTS) — the upstream V2 line is superseded by MiMo V2.5; drops `mimo-v2-tts`, `mimo-v2-pro`, `mimo-v2-omni`, `mimo-v2-flash`, `mimo-v2-flash-free` and realigns the provider-catalog tests. ([#6248](https://github.com/diegosouzapw/OmniRoute/pull/6248) — thanks @backryun)
+- **chore(release):** ~50 commits on this branch are v3.8.45 pre-flight/hardening fixes and CI-perf work that landed here via the parallel-cycle sync-back (`sync-next-cycle.mjs`, Hard Rule #21) **after** the `v3.8.45` git tag was cut, and are already fully documented under the **[3.8.45]** section below — listed here only so the per-cycle commit-coverage check (`npm run release:uncovered`) doesn't flag them as gaps. Provider/catalog/UX/backend fixes: #6041, #6078, #6108, #6135, #6148, #6149, #6154, #6158, #6161, #6162, #6163, #6164, #6165, #6170, #6177, #6178, #6181, #6186, #6187, #6191, #6193, #6194, #6195, #6200, #6205, #6208, #6209, #6211, #6213, #6223, #6224, #6225, #6226, #6227, #6228, #6229, #6230, #6235, #6291, #6292. CI/release-pipeline work: #6167, #6203, #6214, #6215, #6218, #6273, #6275, #6283, #6284, #6285, #6300, #6305.
+- **chore(release):** additional zero-ref release-cycle plumbing on this branch, kept out of `release:uncovered` on purpose (no `#N` in the commit subject to cite): opening the v3.8.46 cycle, opening/closing the v3.8.45 cycle, the finalized [3.8.45] CHANGELOG i18n sync-back to 42 mirrors, the v3.8.45 cognitive/cyclomatic and file-size drift rebaselines, ESLint stale-suppression pruning (4,273 → 4,233), and clearing test-masking/docs-all pre-flight reds for v3.8.45.
+
+### ⚡ Performance & Infrastructure
+
+- **perf(release-green):** the pre-flight validator (`scripts/quality/validate-release-green.mjs`) now runs its 4 slow suites (unit / vitest / integration / pack-artifact) **concurrently** via `Promise.all` — pre-flight wall time drops from ~the sum of the suites to ~the slowest one (~30min saved per round; Phase 0 was the nº1 bottleneck of the v3.8.45 release benchmark, 2h54 of 6h34 e2e). Guard: `tests/unit/validate-release-green.test.ts` ("runs the slow suites CONCURRENTLY"). ([#6319](https://github.com/diegosouzapw/OmniRoute/pull/6319))
+- **fix(ci):** `scripts/release/sync-next-cycle.mjs` — two defects found live in its first production run (v3.8.45 Phase 5): (1) the `git()` helper's default 1 MiB `maxBuffer` crashed with `ENOBUFS` on `git show origin/main:CHANGELOG.md` (the CHANGELOG alone is >1 MiB) — widened to 64 MiB; (2) the i18n resync only propagated the `[NEXT]` (TBD) section, leaving the just-shipped finalized section as "— TBD" in all 42 mirrors — it now also syncs `[prevVersion]` bounded by the heading below it (new exported pure helper `versionAfter`). Guards: +5 tests in `tests/unit/sync-next-cycle.test.ts` (8/8). ([#6327](https://github.com/diegosouzapw/OmniRoute/pull/6327))
+- **test(ci):** concurrency-sensitive flaky tests are **quarantined into a serial pass** (`tests/unit/serial/`, `--test-concurrency=1`, appended to every unit runner incl. sharded variants — the serial pass is sharded too so concurrent shard jobs never self-collide). Initial set: `glm-coding-plan-monthly-3580`, `quota-division-blocks`, `provider-health-autopilot`, `combo-health-autopilot` — the class behind the ~28min CI wedges/re-runs (two live 1h+ wedges cancelled during this PR's own validation). Discovery + TIA gates track the new glob; systemic root cause (async logger writing after teardown) tracked in [#6360](https://github.com/diegosouzapw/OmniRoute/issues/6360). Guard: `tests/unit/test-serial-quarantine.test.ts` (4). ([#6347](https://github.com/diegosouzapw/OmniRoute/pull/6347))
+
+### 🙌 Contributors
+
+Thanks to everyone whose work landed in v3.8.46:
+
+| Contributor                                                    | PRs / Issues           |
+| -------------------------------------------------------------- | ---------------------- |
+| [@2220258345](https://github.com/2220258345)                   | direct commit / report |
+| [@abdofallah](https://github.com/abdofallah)                   | direct commit / report |
+| [@adentdk](https://github.com/adentdk)                         | direct commit / report |
+| [@aleksesipenko](https://github.com/aleksesipenko)             | direct commit / report |
+| [@anki1kr](https://github.com/anki1kr)                         | direct commit / report |
+| [@anungma](https://github.com/anungma)                         | direct commit / report |
+| [@arpicato](https://github.com/arpicato)                       | direct commit / report |
+| [@backryun](https://github.com/backryun)                       | #6248                  |
+| [@binsarjr](https://github.com/binsarjr)                       | direct commit / report |
+| [@brightfiscalband](https://github.com/brightfiscalband)       | direct commit / report |
+| [@chirag127](https://github.com/chirag127)                     | #6501, #6506           |
+| [@developerjillur](https://github.com/developerjillur)         | direct commit / report |
+| [@dilneiss](https://github.com/dilneiss)                       | #6499                  |
+| [@dtybnrj](https://github.com/dtybnrj)                         | direct commit / report |
+| [@Forcerecon](https://github.com/Forcerecon)                   | direct commit / report |
+| [@hao3039032](https://github.com/hao3039032)                   | direct commit / report |
+| [@Iammilansoni](https://github.com/Iammilansoni)               | #6150, #6245           |
+| [@jmengit](https://github.com/jmengit)                         | direct commit / report |
+| [@jordansilly77-stack](https://github.com/jordansilly77-stack) | direct commit / report |
+| [@JxnLexn](https://github.com/JxnLexn)                         | direct commit / report |
+| [@KooshaPari](https://github.com/KooshaPari)                   | #6166                  |
+| [@loopyd](https://github.com/loopyd)                           | direct commit / report |
+| [@luweiCN](https://github.com/luweiCN)                         | #6221                  |
+| [@makcimbx](https://github.com/makcimbx)                       | direct commit / report |
+| [@muflifadla38](https://github.com/muflifadla38)               | direct commit / report |
+| [@newnol](https://github.com/newnol)                           | direct commit / report |
+| [@ofekbetzalel](https://github.com/ofekbetzalel)               | direct commit / report |
+| [@ohahe52-dot](https://github.com/ohahe52-dot)                 | direct commit / report |
+| [@phidinhmanh](https://github.com/phidinhmanh)                 | direct commit / report |
+| [@powellnorma](https://github.com/powellnorma)                 | direct commit / report |
+| [@qpeyba](https://github.com/qpeyba)                           | direct commit / report |
+| [@RaviTharuma](https://github.com/RaviTharuma)                 | direct commit / report |
+| [@RCrushMe](https://github.com/RCrushMe)                       | direct commit / report |
+| [@rianonehub](https://github.com/rianonehub)                   | #6134, #6204           |
+| [@serverless83](https://github.com/serverless83)               | #6212                  |
+| [@swingtempo](https://github.com/swingtempo)                   | direct commit / report |
+| [@tenshiak](https://github.com/tenshiak)                       | direct commit / report |
+| [@ThongAccount](https://github.com/ThongAccount)               | direct commit / report |
+| [@UnrealAryan](https://github.com/UnrealAryan)                 | direct commit / report |
+| [@vinayakkulkarni](https://github.com/vinayakkulkarni)         | direct commit / report |
+| [@vittoroliveira-dev](https://github.com/vittoroliveira-dev)   | #6233                  |
+| [@warelik](https://github.com/warelik)                         | direct commit / report |
+| [@xxy9468615](https://github.com/xxy9468615)                   | direct commit / report |
+| [@xz-dev](https://github.com/xz-dev)                           | direct commit / report |
+| [@yanpaing007](https://github.com/yanpaing007)                 | direct commit / report |
+| [@diegosouzapw](https://github.com/diegosouzapw)               | maintainer             |
+
+---
+
 ## [3.8.45] — 2026-07-06
 
 ### ✨ New Features

docs/i18n/sv/CHANGELOG.md

@@ -6,6 +6,213 @@
 
 ## [3.8.31] — 2026-06-20
 
+## [3.8.46] — TBD
+
+### ✨ New Features
+
+- **feat(sse):** **hide paid-only models from `auto/*` routing** when `hidePaidModels` is on ([#6512](https://github.com/diegosouzapw/OmniRoute/issues/6512)) — follow-up to #6328/#6495. PR #6495 hid paid-only models from the `GET /v1/models` listing, but `auto/*` combos (`auto/best-coding`, `auto/glm`, …) could still pick a paid-only backend into their candidate pool → a 402/403 at request time. `createVirtualAutoCombo` now filters the candidate pool through the new pure `open-sse/services/autoCombo/paidModelFilter.ts` (`filterPaidOnlyCandidates`), applying the same free-model predicate #6495 uses in `catalog.ts` (`providerHasFreeModels(provider) && isFreeModel(provider, {id})`) whenever `settings.hidePaidModels === true`. Applied before the category/tier/family narrowing, so it covers every `auto/*` combo; an all-paid pool degrades to the existing graceful empty-pool path. **Opt-in — default OFF leaves the pool unchanged** (identity). Regression guard: `tests/unit/autoCombo/paid-model-filter-6512.test.ts` (4, incl. the default-off identity guard).
+- **feat(sse):** **provider-family auto combos** — `auto/glm`, `auto/minimax`, `auto/mimo`, `auto/zai`, `auto/gemma`, `auto/llama`, `auto/gemini` ([#6453](https://github.com/diegosouzapw/OmniRoute/issues/6453)) — new routable ids that materialize an on-demand virtual combo spanning whatever installed backends currently expose that model family, degrading gracefully as backends rotate. A new pure `open-sse/services/autoCombo/modelFamily.ts` (`detectModelFamily`) classifies by model-id prefix for six families; `zai` is instead resolved by provider id (z.ai's hosted API serves the same `glm-*` model ids as every other GLM backend, so `auto/zai` means "route to my z.ai backend specifically" vs `auto/glm`'s "any connected GLM backend"). Reuses the existing `createVirtualAutoCombo` on-demand materialization path (no DB writes) and the `/v1/models` catalog advertising loop. Regression guard: `tests/unit/autoCombo/provider-family-combos.test.ts` (11).
+- **feat(proxy):** native **proxy-pool round-robin / egress IP rotation** ([#6365](https://github.com/diegosouzapw/OmniRoute/issues/6365)) — a scope (global / provider / account) can now hold **multiple** proxies as a pool with a rotation strategy, so outbound requests cycle their egress IP instead of pinning one proxy per scope. Migration `117_proxy_pool_rotation.sql` lifts the `UNIQUE(scope, scope_id)` constraint (rebuild via the canonical rename/copy/drop; existing single assignments become 1-element pools) and adds a `proxy_scope_rotation` companion table holding the per-scope strategy + a persisted monotonic round-robin cursor. Strategies: `round-robin` (default, monotonic cursor — never `Math.random`), `random`, and `sticky-per-N-min`. Resolution (`resolveProxyForScopeFromRegistry` / `resolveProxyForConnectionFromRegistry`) now fetches the alive, position-ordered candidate set (unchanged `PROXY_ALIVE_PREDICATE`) and applies the strategy; an empty / all-dead pool still returns `null` — the #6246 fail-closed guard is **untouched** (never falls through to direct egress). Backend + DB only; dashboard pool-builder UI is a follow-up. Regression guard: `tests/unit/proxy-pool-rotation-6365.test.ts` (8, incl. fail-closed + backward-compat).
+- **feat(providers):** end-to-end **tool/function calling on the native Gemini `/v1beta` endpoint** ([#6222](https://github.com/diegosouzapw/OmniRoute/issues/6222)) — both directions of the Gemini↔OpenAI conversion now preserve tool data (previously silently dropped). Request side: `convertGeminiToInternal` (extracted to its own testable module) maps `tools[].functionDeclarations` → OpenAI `tools`, prior `functionCall` parts → assistant `tool_calls`, and `functionResponse` parts → `tool`-role messages. Response side: `convertOpenAIResponseToGemini` emits `parts[].functionCall {name,args}` from `message.tool_calls`, and the streaming `openAIChunkToGeminiChunk` accumulates fragmented `tool_calls` deltas by index into complete `functionCall` parts. The non-Gemini client paths (Claude, OpenAI-Responses) already preserved tool calls — this closes the gap specific to the native Gemini surface. Regression guard: `tests/unit/v1beta-gemini-tool-calling-6222.test.ts` (6, incl. a streaming SSE round-trip).
+- **feat(providers):** copilot-m365-web **enterprise / work tier** support ([#6334](https://github.com/diegosouzapw/OmniRoute/issues/6334)) — mirrors the EDU-tier pattern (#6210): `M365ConnectionParams` gains an `agent` field, a new opt-in `M365_ENTERPRISE_OVERRIDES` preset (`agent=work`, `scenario=officeweb`, `licenseType=Premium`) applies via `providerSpecificData.tier="enterprise"` (alias `"work"`), and `agent` is also overridable directly via `providerSpecificData.agent`. `buildWsUrl` was hardcoding `agent="web"` (the one enterprise-distinguishing param with no override path), so a Premium work account handshook then returned an empty stream. The individual and EDU paths are untouched. Kilo's dup flag vs #6210 (EDU tier) was a false positive — different tier. Regression guard: `tests/unit/copilot-m365-enterprise-6334.test.ts` (7). End-to-end confirmation on a real Premium work account is a live-VPS validation follow-up (Hard Rule #18). (thanks @Forcerecon)
+- **feat(api):** standardized, provider-agnostic **`effort` + `thinking` request params** ([#6241](https://github.com/diegosouzapw/OmniRoute/issues/6241)) — a thin standardization layer over the existing mature per-provider reasoning plumbing (no provider mapper touched). `providerChatCompletionSchema` gains a canonical `effort` (reusing the shared `none/low/medium/high/xhigh` vocabulary — the UI tiers `extra`/`max` collapse onto `xhigh`) and a boolean `thinking`. A pure `normalizeReasoningRequest` (wired once in `src/sse/handlers/chat.ts`, before any reasoning field is read) folds them onto the fields the translators already consume (`reasoning_effort` / `reasoning.effort` / `thinking`), so they fan out to Anthropic / Gemini / xAI / Responses — an explicit client `reasoning_effort` / object-shaped `thinking` always wins (backward-compatible). `/models` additively exposes `supportsThinking` + `effort_tiers` so the frontend can render the toggles (UI component is a follow-up). Regression guard: `tests/unit/effort-thinking-standardization-6241.test.ts` (12). (thanks @Iammilansoni, @shabeer)
+- **feat(combo):** new **`pipeline` (sequential) combo strategy** ([#6297](https://github.com/diegosouzapw/OmniRoute/issues/6297)) — the 18th routing strategy runs targets **in order**, threading each step's output into the next step's input, with an optional per-step `prompt` (system instruction); only the final step's response is returned. Distinct from `fusion` (parallel fan-out + judge). Implemented as a self-contained `open-sse/services/pipeline.ts` (sibling to `fusion.ts`), dispatched from `combo.ts`; the step list reuses `combo.models` order and reads an optional `prompt` off each target (backward-compatible — ignored by every other strategy). Intermediate steps run non-streaming with tools stripped (complete prose to thread forward); the final step keeps the client's `stream` flag + tools. A failing/empty/unparseable intermediate step fails the whole pipeline explicitly via a sanitized error (never silently swallowed). Kilo's dup flag vs #563 was a false positive (that's model→chain selection; this is a sequential chain). Regression guard: `tests/unit/combo-pipeline-strategy.test.ts` (5). (thanks @ofekbetzalel)
+- **feat(ci):** `check:test-masking` now flags **inline-reimplemented prod conditions** ([#6348](https://github.com/diegosouzapw/OmniRoute/issues/6348)) — a new **report-only** subcheck (v2, 6A.10 family) catches the wrong-shape contract test: a test that recomputes the condition under test inline instead of importing/exercising the real function (the #6216 class, where `=== 500` → `>= 500` stayed green because the test re-implemented the branch). For each added/modified test file it warns when the file textually duplicates a ≥3-token conditional from a production file touched in the same PR **and** does not import the symbol/module owning it, via a pure, fixture-tested `findReimplementedConditions()` with an allowlist mirroring `assertReductionAllowlist`. Report-only for now (does not fail the gate) — to be promoted to blocking after a triage cycle. Regression guard: `tests/unit/check-test-masking.test.ts` (45).
+- **feat(sse):** per-connection routing override (native vs CLIProxyAPI) ([#6339](https://github.com/diegosouzapw/OmniRoute/issues/6339)) — the previously-dead `isCliproxyapiDeepModeEnabled` helper is now wired into `resolveExecutorWithProxy`: a single connection can opt itself into the CLIProxyAPI passthrough executor via `providerSpecificData.cliproxyapiMode="claude-native"`, with precedence **connection override > provider `upstream_proxy_config` mode > default**. `resolveExecutorWithProxy` now receives the resolved connection's `providerSpecificData` (threaded from `chatCore.ts`), so one connection can deep-route while the provider's other connections stay native — no DB schema change (the toggle rides in `providerSpecificData`). Also resolves the same-provider mixing ask in #6340. Regression guard: `tests/unit/chatcore-executor-proxy.test.ts` (9). (thanks @RaviTharuma)
+- **feat(dashboard):** "Add session cookie" modal now shows a prominent **"Open ‹host› →"** link to the provider's own site ([#6268](https://github.com/diegosouzapw/OmniRoute/issues/6268)) — every `-web` cookie-session provider (chatgpt-web, claude-web, gemini-web, kimi-web, lmarena, qwen-web, m365-copilot-web, …) renders a one-click external link (opening the provider's login/home page in a new tab) so operators no longer tab away to retype the URL mid-setup. The host resolves from a pure, unit-tested `resolveWebProviderHost()` (prefers `WEB_COOKIE_PROVIDERS[id].website`, falls back to the registry `baseUrl` origin); non-web providers render exactly as before. Kilo's dup flag vs #6265 (modal-too-small-on-1080p) was a false positive — distinct concern. Regression guard: `tests/unit/resolve-web-provider-host.test.ts` (5). (thanks @chirag127)
+- **feat(providers):** add **DigitalOcean** AI (serverless inference) as an OpenAI-compatible API-key provider ([#6373](https://github.com/diegosouzapw/OmniRoute/pull/6373)) — base `https://inference.do-ai.run/v1`, wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/digitalocean/`, `src/shared/constants/providers/apikey/inference-hosts.ts`). Regression guard: `tests/unit/digitalocean-provider.test.ts`. (thanks @newnol)
+- **feat(providers):** add **Huancheng Public API** (`hcnsec`) as an OpenAI-compatible regional provider ([#6410](https://github.com/diegosouzapw/OmniRoute/pull/6410)) — Xinjiang Huancheng Cybersecurity's public LLM platform (base `https://api.hcnsec.cn/v1`, free credits via daily check-ins), wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/hcnsec/`, `src/shared/constants/providers/apikey/regional.ts`). Regression guard: `tests/unit/hcnsec-provider.test.ts`. (thanks @UnrealAryan)
+- **feat(dashboard):** the web-session credential guide now shows an **"Open {host}" link** ([#6316](https://github.com/diegosouzapw/OmniRoute/pull/6316)) to the provider's sign-in site (derived from the provider `website` via `getProviderWebsiteHost`), so you can jump straight to the page where the cookie/session must be captured. Regression guard: `tests/unit/web-session-provider-link-6316.test.ts`. (thanks @jordansilly77-stack)
+- **feat(cerebras):** add the **Gemma 4 31B** model (`gemma-4-31b`) to the Cerebras registry + pricing table ([#6331](https://github.com/diegosouzapw/OmniRoute/pull/6331)). Regression guard extends `tests/unit/t28-model-catalog-updates.test.ts`. (thanks @backryun)
+- **feat(providers):** add **Yuanbao (web)** as a cookie-session provider ([#6196](https://github.com/diegosouzapw/OmniRoute/issues/6196)) — `yuanbao-web` (Tencent Yuanbao, `yuanbao.tencent.com`) with cookie-only auth (`hy_user`/`hy_token` + public agent id), SSE→OpenAI translation incl. `reasoning_content`, exposing DeepSeek V3/R1 + Hunyuan / Hunyuan-T1. Regression guard: `tests/unit/providers-yuanbao-web.test.ts`. `together-web` was **deferred** (no verifiable web-session endpoint — needs a captured request) and `huggingchat-web` **dropped** (the existing `huggingchat` already is a web-cookie provider). (thanks @chirag127)
+- **feat(providers):** route the built-in **agentrouter** through the dynamic Claude-Code wire image ([#6056](https://github.com/diegosouzapw/OmniRoute/issues/6056)) — a small static allow-set (`CC_WIRE_IMAGE_BUILTINS` in `open-sse/services/ccWireImageBuiltins.ts`), consulted by `isClaudeCodeCompatible` / `isClaudeCodeCompatibleProvider` / `applyFingerprint`, makes agentrouter adopt the CC wire-image headers + fingerprint **while guarding the CC baseUrl/auth branches** so it keeps its own registry `baseUrl` and `x-api-key` auth. Regression guard: `tests/unit/agentrouter-cc-wire-image.test.ts` (asserts the wire image is applied AND agentrouter's baseUrl/auth are preserved). Live WAF-acceptance against agentrouter.org is a VPS validation follow-up (Hard Rule #18).
+- **feat(providers):** **bulk-add API keys for Cloudflare Workers AI** ([#6174](https://github.com/diegosouzapw/OmniRoute/issues/6174)) — `cloudflare-ai` is removed from the bulk-add exclusion list and the bulk parser gains a 3-field `name|accountId|apiKey` mode; the bulk route now builds a **per-entry** `providerSpecificData` so each key carries its own `accountId` (fixing the previous shared-object reuse), and both the create + key-validation paths receive it. Regression guard: `tests/unit/bulk-api-key-parser-cloudflare.test.ts`. (thanks @muflifadla38)
+- **feat(dashboard):** routing/settings UX clarity ([#6147](https://github.com/diegosouzapw/OmniRoute/issues/6147)) — (1) weighted combos show the **effective routing share %** next to each weight when weights don't sum to 100 (`WeightTotalBar.tsx`); (2) the status widget's user-facing **"Cloud Sync" label is renamed** to "Remote Settings Sync" (`CloudSyncStatus.tsx`; internal ids/state untouched); (3) built-in providers gain an **opt-in advanced base-URL override** (`isBaseUrlOverrideEligibleProvider`, hidden behind an "Advanced" toggle, reusing the existing `providerSpecificData.baseUrl` persistence — not globally widened). Regression guard: `tests/unit/routing-settings-ux-6147.test.ts`.
+- **feat(combo):** add an option to **disable session stickiness**, per-combo or globally — round-robin / random combos can rotate to a different connection on every request instead of pinning a whole conversation to one connection by its first-message hash. Resolution precedence per-combo `config.disableSessionStickiness` → global `settings.disableSessionStickiness` → default `false` (preserves the #3825 prompt-cache/504 fix); gates **both** stickiness call sites in `open-sse/services/combo.ts`. Exposed as a global toggle (Combo Defaults) and a per-combo Inherit/on/off control. ([#6168](https://github.com/diegosouzapw/OmniRoute/issues/6168)) Regression guard: `tests/unit/combo-disable-session-stickiness.test.ts`. (thanks @RCrushMe)
+- **feat(docker):** add the `OMNIROUTE_NO_SUDO` env flag for root-less / user-namespaced deployments — the MITM cert-trust command path (`resolveSudoSpawn` in `src/mitm/systemCommands.ts`) now strips the leading `sudo` when the flag is truthy, in addition to the existing root / sudo-missing cases, so the Proxy Agent runs without `sudo` (the operator trusts the CA manually, e.g. via `NODE_EXTRA_CA_CERTS`). Argv-array `spawn` preserved — no shell interpolation (Hard Rule #13). ([#6122](https://github.com/diegosouzapw/OmniRoute/issues/6122)) Regression guard: `tests/unit/mitm-systemCommands-no-sudo.test.ts`. (thanks @powellnorma)
+- **feat(providers):** add **Requesty** as an OpenAI-compatible gateway provider (BYOK, base `https://router.requesty.ai/v1`, ~200 free requests/day) — wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/requesty/`, `src/shared/constants/providers/apikey/gateways.ts`). ([#6120](https://github.com/diegosouzapw/OmniRoute/issues/6120)) Regression guard: `tests/unit/requesty-provider.test.ts`. (thanks @chirag127)
+- **feat(dashboard):** add **configured-only / available-only filters** to the Free Provider Rankings page ([#6150](https://github.com/diegosouzapw/OmniRoute/issues/6150)) — hide providers you haven't configured, or whose connections are all rate-limited / out of quota, via server-side query params (`?configuredOnly` / `?availableOnly` on `GET /api/free-provider-rankings`) backed by a testable lib helper reusing the in-process connection state (no Redis). Both filters default off, so the default view is unchanged; this supersedes the earlier client-side "Configured Only" toggle (#6245) with an available-only dimension and unit-tested logic. Regression guard: `tests/unit/freeProviderRankings-filters.test.ts`.
+
+### 🔧 Bug Fixes
+
+- **fix(dashboard):** adding a second API key connection for the same provider no longer silently overwrites the first — the Add-API-key modal now derives a unique default connection name (`main`, then `main-2`, `main-3`, …) so the backend name-based upsert can't collide (#6499 — thanks @dilneiss).
+- **fix(compression):** the session-dedup engine now also deduplicates a large multi-line block repeated **within a single message** (intra-message dedup), not just across turns; the compression-preview API surfaces a `fallbackReason`, and the fusion panel reports how many models were rate-limited vs failed on total-panel failure (#6501 — thanks @chirag127).
+- **fix(compression):** a stacked-pipeline step naming an unregistered engine now surfaces a `validationErrors` entry instead of silently no-op'ing, so misconfigured pipelines are visible in the preview API (#6506 — thanks @chirag127).
+- **feat(usage):** add a **Codex reset-credit redemption** flow to the Provider Limits UI ([#6361](https://github.com/diegosouzapw/OmniRoute/pull/6361)) — a `useCodexResetCreditRedemption` hook + `/api/usage/codex-reset-credit` route + `codexResetCredits` lib let you redeem banked Codex reset credits from the quota card. Regression guard: `tests/unit/codex-reset-credits.test.ts`. (thanks @JxnLexn)
+- **feat(glm):** add **team-plan quota settings** for `glm-cn` connections ([#6351](https://github.com/diegosouzapw/OmniRoute/pull/6351)) — a dedicated `GlmTeamQuotaFields` form section (team quota id / limits) threaded through the Add/Edit connection modals, persisted via `providerSpecificData`, with the GLM usage service reading the team quota. Regression guards: `tests/unit/glm-team-quota.test.ts`, `provider-specific-data-schema.test.ts`. (thanks @hao3039032)
+- **feat(providers):** add **TinyFish** web-fetch/search support ([#6349](https://github.com/diegosouzapw/OmniRoute/pull/6349)) — a `tinyfish-fetch` executor + `/v1/web/fetch` route + MCP web-fetch tool, registered as a specialty-media provider with request-validation and a search-provider catalog entry. Regression guards: `tests/unit/executor-tinyfish-fetch.test.ts`, `web-fetch-handler.test.ts`, `mcp-web-fetch-tool.test.ts`, `provider-validation-tinyfish.test.ts`. (thanks @dtybnrj)
+- **fix(cli):** `omniroute launch-codex` now spawns `codex.cmd` through a shell on Windows (the npm `.cmd` shim is unresolvable by bare `spawn` → ENOENT), mirroring the qodercli Windows fix (#6263) ([#6312](https://github.com/diegosouzapw/OmniRoute/pull/6312)). Regression guard: `tests/unit/launch-codex-windows-spawn-6312.test.ts`. (thanks @swingtempo)
+- **fix(codex):** isolate the **Spark** quota from the shared Codex quota and stabilize the quota UI ordering / hydration so per-scope limits render consistently ([#6336](https://github.com/diegosouzapw/OmniRoute/pull/6336)). Regression guards: `tests/unit/codex-quota-selection-hydration.test.ts`, `provider-limits-ui.test.ts` + 3 more. (thanks @xz-dev)
+- **feat(api):** add a `hidePaidModels` setting that filters paid-only models out of the `/v1/models` catalog. Regression guard: `tests/unit/models-catalog-hide-paid.test.ts`. (thanks @chirag127)
+- **fix(api-manager):** the fallback model picker now preserves combos instead of dropping them when a primary model is unavailable ([#6443](https://github.com/diegosouzapw/OmniRoute/pull/6443)). Regression guard: `tests/unit/api-manager-page-static.test.ts`. (thanks @jmengit)
+- **fix(providers):** recoverable Antigravity / Cloud-Code (Gemini Code Assist) `403` responses ([#6452](https://github.com/diegosouzapw/OmniRoute/pull/6452)) are now classified as a retryable project-config error instead of a terminal account ban, so a fixable project/API-disabled 403 no longer forces a ~1-year cooldown / full OAuth reconnect. Regression guard: `tests/unit/errorclassifier-antigravity-403.test.ts`. (thanks @developerjillur)
+- **fix(mitm):** `sanitizeHeaders` now redacts `Set-Cookie` response headers so upstream session cookies never leak into logs / diagnostics ([#6451](https://github.com/diegosouzapw/OmniRoute/pull/6451)). Regression guard: `tests/unit/mitm-sanitize-headers.test.ts`. (thanks @developerjillur)
+- **fix(api):** `/api/compression/preview` now accepts `mode: "caveman"` and correctly handles stacked / zero-compression previews ([#6425](https://github.com/diegosouzapw/OmniRoute/issues/6425)). Regression guard: `tests/unit/api/compression-preview-caveman-and-stacked-6425.test.ts`. (thanks @chirag127)
+- **feat(providers):** add **Zed** hosted LLM aggregator as a native-app provider ([#6118](https://github.com/diegosouzapw/OmniRoute/pull/6118)) — OAuth sign-in via the Zed hosted flow, registered through the shared provider registry + executor. Regression guards: `tests/unit/zed-oauth-provider.test.ts`, `zed-import-utils.test.ts`, `zed-docker-detect.test.ts`, `mitm-handler-zed.test.ts`. VPS-validated via live operator login (Hard Rule #18).
+- **fix(oauth):** the Kiro SSO-cache auto-import now **preserves the IDC region** — cross-region Amazon Q / Kiro profiles imported from the SSO cache are no longer collapsed to the default region ([#6113](https://github.com/diegosouzapw/OmniRoute/pull/6113)). Regression guard: `tests/unit/kiro-auto-import-idc-2059.test.ts`. VPS-validated via live operator login (Hard Rule #18).
+- **fix(dashboard):** passthrough model aliases no longer collide when two namespaced model ids share a last segment (port from 9router#1850, [#6431](https://github.com/diegosouzapw/OmniRoute/pull/6431)). `enx/gpt-5.5` and `enx/codebuddy/gpt-5.5` both auto-generated the alias `gpt-5.5`, so the second model could never be added (the UI just alerted "alias already exists"). Aliases are now disambiguated deterministically — bare last segment when free, then parent-qualified (`codebuddy-gpt-5.5`), then a numeric suffix — while re-adding the exact same model id is still blocked. Regression guard: `tests/unit/passthrough-alias-1850.test.ts`. (thanks @arpicato)
+- **fix(translator):** preserve a Gemini `functionResponse` co-located with other parts (another `functionCall`, or trailing `text`) in the same content when translating **Gemini → OpenAI** ([#6376](https://github.com/diegosouzapw/OmniRoute/pull/6376)). `convertGeminiContent()` early-returned the tool message on the first `functionResponse` part, dropping any co-located parts; such contents are now pre-split (one tool message per `functionResponse`, emitted first, plus one message for the remaining parts). Regression guard: `tests/unit/gemini-to-openai-function-response.test.ts`. (thanks @warelik)
+- **fix(headroom):** detect a python interpreter managed by **mise / pyenv / asdf / conda** (port from 9router#2353, [#6382](https://github.com/diegosouzapw/OmniRoute/pull/6382)). Headroom's python probe (`src/lib/headroom/detect.ts`) searched a hardcoded `PATH`, but version managers expose their interpreters via shim dirs that only join `PATH` through interactive-shell activation — which the non-interactive server never runs, so a managed python (≥3.10) was invisible and Headroom reported it missing. The search path now prepends the well-known shim/bin dirs (`~/.local/share/mise/shims`, `~/.pyenv/shims`, `~/.asdf/shims`, `$CONDA_PREFIX/bin`, `~/.local/bin`, respecting `MISE_DATA_DIR`/`PYENV_ROOT`/`ASDF_DATA_DIR` when set), and a new `HEADROOM_PYTHON` env override lets operators point straight at their interpreter (mirroring `HEADROOM_URL`). Still shell-free (`execFileSync`). Regression guard: `tests/unit/headroom-detect.test.ts` (5). (thanks @loopyd)
+- **fix(executors):** strip the OpenAI-Codex/Claude-CLI `client_metadata` passthrough field for **NVIDIA** requests (port from 9router#1887, [#6411](https://github.com/diegosouzapw/OmniRoute/pull/6411)). NVIDIA's OpenAI-compatible wrapper rejects it with `400 Unsupported parameter`, the same class already handled for `cerebras`/`mistral`; `nvidia` (executor `default`) was missing from the strip allowlist so Codex/Claude-Code passthrough requests 400'd. Regression guard: `tests/unit/executor-default-strip-client-metadata.test.ts` (+nvidia case). (thanks @phidinhmanh)
+- **fix(translator):** strip the Claude-style `thinking` field for **NVIDIA `z-ai/glm-5.2`** (port from 9router#2023, [#6413](https://github.com/diegosouzapw/OmniRoute/pull/6413)). NVIDIA's OpenAI-compatible wrapper 400s on `thinking` (a Claude-format client routed here leaves a `thinking:{type:"adaptive"}`); the existing strip rule only dropped `reasoning`. Same class already handled for `minimax-m2.7`. Regression guard: `tests/unit/nvidia-minimax-thinking-strip.test.ts` (+glm-5.2 case). (thanks @phidinhmanh)
+- **fix(translator):** suppress the streamed `</think>` close marker for the **Antigravity IDE** client (port from 9router#1061, [#6415](https://github.com/diegosouzapw/OmniRoute/pull/6415)). On thinking-only turns Antigravity rendered a bare `</think>` as the sole visible content, tripping its loop-detection and wasting requests. Antigravity's UA (`vscode/<v> (Antigravity/<v>)`) is added to the marker-suppress allowlist (alongside OpenCode); Claude Code / Cursor still get the marker, and `x-omniroute-thinking-marker: on` force-restores it. Regression guard: `tests/unit/think-close-marker-suppress-5245.test.ts`. (thanks @abdofallah)
+- **fix(executors):** strip nested `reasoning_content` from messages for **Mistral** (port from 9router#1649, [#6417](https://github.com/diegosouzapw/OmniRoute/pull/6417)). Mistral's API returns `422 extra_forbidden` when an assistant message carries `reasoning_content` (replayed thinking from a prior turn, e.g. via the Codex `/responses` path); the generic top-level 400 field-downgrade retry never covered the nested per-message field. `DefaultExecutor` now strips it for provider `mistral` only, so DeepSeek (which requires replayed `reasoning_content`) is unaffected. Regression guard: `tests/unit/mistral-strip-reasoning-content-1649.test.ts`. (thanks @xxy9468615)
+- **fix(executors):** strip the `client_metadata` passthrough field on the **OpenCode** path (port from 9router#1442, [#6418](https://github.com/diegosouzapw/OmniRoute/pull/6418)). OpenCode upstreams (e.g. `kimi-k2.6` via opencode-go) reject it with `400 "Extra inputs are not permitted, field: 'client_metadata'"`; the DefaultExecutor strip only covered cerebras/mistral and `OpencodeExecutor` extends `BaseExecutor` directly, so nothing removed it there. Regression guard: `tests/unit/opencode-strip-client-metadata-1442.test.ts`. (thanks @yanpaing007)
+- **fix(executors):** inject the `reasoning_content` echo for the native **Moonshot Kimi** provider (port from 9router#1480, [#6419](https://github.com/diegosouzapw/OmniRoute/pull/6419)). Kimi (executor `default`) is a thinking-mode upstream that 400s with "reasoning_content must be passed back" when a prior assistant turn lacks it; the placeholder injection was only wired into the OpenCode meta-provider, so direct multi-turn Kimi conversations failed. Scoped to `kimi` (gateway-served models matching the thinking-model name pattern are unaffected). Regression guard: `tests/unit/kimi-native-reasoning-injected-1480.test.ts`. (thanks @2220258345)
+- **fix(executors):** recover from a strict gateway's `context_management: Extra inputs are not permitted` 400 (port from 9router#1468, [#6420](https://github.com/diegosouzapw/OmniRoute/pull/6420)). **Claude Code** always sends a top-level `context_management` field; strict anthropic-compatible gateways reject it. The dedicated context-editing 400-fallback only fired when OmniRoute's own `contextEditing` feature was enabled (default off), so a client-sent field passed through untouched and 400'd. `context_management` is now in the generic reactive field-strip list, so it's stripped-and-retried once regardless of the feature flag (with correct request re-signing for claude-compatible relays). Regression guard: `tests/unit/provider-field-strips.test.ts`. (thanks @ohahe52-dot)
+- **fix(network):** enable **RFC 8305 Happy Eyeballs** (`autoSelectFamily`) on the direct-egress undici dispatcher (port from 9router#1237, [#6423](https://github.com/diegosouzapw/OmniRoute/pull/6423)). When DNS returns both IPv6 (AAAA) and IPv4 (A) and the IPv6 route is broken (e.g. a NAT64 `64:ff9b::` prefix without routing), undici tried IPv6 first and hung until `ETIMEDOUT` (then a 502 + account lockout), even though `curl` reached the same host. The direct dispatcher now races both families and uses whichever connects first. Proxy paths pin family via `proxyTls` and are unaffected. Regression guard: `tests/unit/direct-dispatcher-pipelining-4580.test.ts`. (thanks @adentdk)
+- **fix(combo):** round-robin now advances the rotation pointer past the model that **actually served**, not the eagerly-scheduled one (port from 9router#948, [#6428](https://github.com/diegosouzapw/OmniRoute/pull/6428)). With `stickyLimit: 1` (true round-robin), when the scheduled model failed and a _different_ model served via fallback, the counter had already advanced +1 from the scheduled index — so the next request reused the fallback-served model, degrading round-robin into hot-spotting on whichever model was healthy. The pointer now advances to the served index + 1 (mirroring the sticky-limit>1 path). Session-stickiness (#3825) and distribution are preserved. Regression guard: `tests/unit/combo-rr-fallback-advance-948.test.ts`. (thanks @binsarjr)
+- **fix(sse):** a non-string `model` field is now rejected with a `400` before the resolver, instead of crashing downstream `.toLowerCase()`/`.split()` calls into an empty-body `500` that escapes the error sanitizer ([#6407](https://github.com/diegosouzapw/OmniRoute/issues/6407)). Regression guard: `tests/unit/chat-non-string-model-6407.test.ts`. (thanks @chirag127)
+- **fix(api):** unknown `/api/*` routes now return a JSON `404` (instead of the dashboard HTML shell) and scalar chat params (`model`/`temperature`/etc.) are validated **before** the provider lookup so malformed requests fail fast with a clear `400` ([#6424](https://github.com/diegosouzapw/OmniRoute/issues/6424), [#6412](https://github.com/diegosouzapw/OmniRoute/issues/6412)). Regression guards: `tests/unit/api/api-catchall-json-404.test.ts`, `tests/unit/chat-early-schema-validation-6412.test.ts`. (thanks @chirag127)
+- **fix(api):** `/v1/chat/completions` now rejects a non-JSON `Content-Type` with a `400` before parsing the body ([#6414](https://github.com/diegosouzapw/OmniRoute/issues/6414)). Regression guard: `tests/unit/v1-chat-completions-content-type-6414.test.ts`. (thanks @chirag127)
+- **fix(api):** the `X-OmniRoute-Compression` response header is now echoed on `/v1/chat/completions` and `/v1/completions` ([#6422](https://github.com/diegosouzapw/OmniRoute/issues/6422)). Regression guard: `tests/unit/compression-header-echo-6422.test.ts`. (thanks @chirag127)
+- **fix(api):** concurrent `GET /v1/models` requests are coalesced into a single catalog build ([#6408](https://github.com/diegosouzapw/OmniRoute/issues/6408)). Regression guard: `tests/unit/v1-models-concurrent-6408.test.ts`. (thanks @chirag127)
+- **fix(api):** `/v1/completions` now echoes the requested `body.model` in its JSON + streamed responses ([#6429](https://github.com/diegosouzapw/OmniRoute/pull/6429)). Regression guard: `tests/unit/completions-body-model-echo.test.ts`. (thanks @chirag127)
+- **fix(api):** env-var master keys now see the full `/v1/models` catalog ([#6406](https://github.com/diegosouzapw/OmniRoute/issues/6406)). Regression guard: `tests/unit/models-catalog-envkey-6406.test.ts`. (thanks @chirag127)
+- **fix(api):** non-streaming `/v1/completions` responses now echo `body.model` aligned with the `X-OmniRoute-Model` header ([#6426](https://github.com/diegosouzapw/OmniRoute/issues/6426)). Regression guard: `tests/unit/v1-completions-model-header-match-6426.test.ts`. (thanks @chirag127)
+- **fix(api):** unknown `/v1/*` routes now return a JSON `404 not_found` instead of the Next.js dashboard HTML shell ([#6405](https://github.com/diegosouzapw/OmniRoute/issues/6405)). Regression guard: `tests/unit/api/v1-catchall-json-404.test.ts`. (thanks @chirag127)
+- **fix(api):** the per-connection provider models route now degrades to the shipped catalog when a provider's `/models` endpoint answers with a redirect ([#6267](https://github.com/diegosouzapw/OmniRoute/issues/6267)) — a `qwen-web` import failed with a raw `Redirect blocked … (307)` 503. `safeOutboundFetch` throws `REDIRECT_BLOCKED` on the 307, `getSafeOutboundFetchErrorStatus` maps it to 503, and `buildDiscoveryErrorFallbackResponse` treated every 503 as a hard error — so the non-empty `getModelsByProviderId("qwen-web")` catalog was never surfaced. A models-endpoint redirect is not a fixable-config error (unlike `URL_GUARD_BLOCKED`/`INVALID_URL`, which stay hard errors), so it now falls back to the local/cached catalog before the 503 short-circuit. General fix — covers any config-driven provider that 307s. Regression guard: `tests/unit/provider-models-qwen-web-redirect-6267.test.ts`. (thanks @chirag127)
+- **fix(api):** the per-connection provider models route (MCP `list_models_catalog` + the dashboard import view) now merges USER-ADDED custom models into its response ([#6247](https://github.com/diegosouzapw/OmniRoute/issues/6247)) — custom models live in the `key_value` namespace `customModels`, which the live REST `/api/v1/models` already merges, but `src/app/api/providers/[id]/models/route.ts` never read `getCustomModels`, so custom models were dropped on both the discovery-success and local_catalog paths. They are now folded into the returned model list (deduped by id, stamped `owned_by: provider`), fixing MCP + the dashboard import view in one place. Regression guard: `tests/unit/provider-models-custom-merge-6247.test.ts`. (thanks @RCrushMe)
+
+- **fix(providers):** GitLab Duo tool-calling follow-up turns no longer fail upstream with `422 {"detail":"Validation error"}` (tokens 0/0, rejected pre-inference). The #6234 tool-result-feedback fix serialized the **entire** multi-turn conversation into GitLab's single-file `code_suggestions` (`small_file`) generation endpoint — folded history that turn-N sent as an oversized `current_file.content_above_cursor` **and** duplicated verbatim into `user_instruction`, tripping the AI-Gateway's `small_file` validation guard. The executor now **bounds** that prompt: it keeps system + latest user message + the most-recent tool round (dropping older turns), caps oversized tool results, and stops duplicating the full prompt into `user_instruction` (which now carries only the short latest user message) — while still feeding the most-recent tool result back so the agent continues (`open-sse/executors/gitlab.ts`, [#6220](https://github.com/diegosouzapw/OmniRoute/issues/6220)). The unit test covers the bounding logic; the upstream 422→200 clearing is VPS-only (Hard Rule #18). Regression guard: `tests/unit/gitlab-tool-exchange-bounded-6220.test.ts`.
+
+- **fix(i18n):** the provider-detail (`/dashboard/providers/[id]`) connection-status filter labels no longer render as `__MISSING__:All` / `__MISSING__:Active` / `__MISSING__:Error` / `__MISSING__:Banned` / `__MISSING__:CreditsExhausted` in non-English locales (notably pt-BR) ([#6290](https://github.com/diegosouzapw/OmniRoute/issues/6290)). Root cause was **not** the namespace mismatch the issue guessed — the `providers.filter*` keys resolve correctly in `en.json`; the debt lived in the locale mirrors (`src/i18n/messages/*.json`), where these five keys carried the `__MISSING__:` sync sentinel in ~15 locales and were absent entirely in ~26 others, so next-intl found the key and echoed the sentinel verbatim. All 40 non-English/-Chinese mirrors now ship real translations for the five `providers.filter*` labels. Regression guard: `tests/unit/i18n-provider-filter-keys-6290.test.ts`. (thanks @diegosouzapw)
+
+- **fix(providers):** the `copilot-m365-web` streaming executor now emits `debug`-level WebSocket diagnostics ([#6210](https://github.com/diegosouzapw/OmniRoute/issues/6210)) — the outbound WS URL (with the `access_token` **redacted** via `redactWsUrl()`), handshake success/failure, and each received SignalR frame's `type`/`target`. Previously the streaming path logged nothing, so an empty `content:null` response (the M365 Education / Starter tier symptom fixed in #6234) was undiagnosable even at `APP_LOG_LEVEL=debug`. The change is debug-level and side-effect-free — it does not alter streaming behavior or the frame parser, and the token never reaches the logs. Regression guard: `tests/unit/copilot-m365-web-logging-6210.test.ts` (thanks @qpeyba)
+
+- **fix(resilience):** a round-robin combo no longer returns `503 all upstream accounts are unavailable` when a compatibility-rejected target is actually healthy ([#6238](https://github.com/diegosouzapw/OmniRoute/issues/6238)). `filterTargetsByRequestCompatibility` drops request-incompatible targets (tool/vision/structured-output unsupported, or below the required context window) **before** any availability check runs, and its `compatible.length === 0` safety net only fired when _all_ targets were filtered — not when the kept targets later all turned out runtime-unavailable (circuit-open / cooldown / no credentials). So a combo could 503 while a compat-rejected-but-healthy provider sat unused. `handleRoundRobinCombo` now keeps the compat-rejected set and, when every compat-kept target was skipped without a single real attempt, probes those rejected targets as a **last-resort fallback tier** (via the new pure `open-sse/services/combo/comboCompatFallback.ts`) before crystallizing the 503. Regression guard: `tests/unit/combo-roundrobin-compat-fallback-6238.test.ts`. (thanks @ThongAccount)
+
+- **fix(startup):** best-effort self-heal for a corrupted Turbopack dev cache on Windows ([#6289](https://github.com/diegosouzapw/OmniRoute/issues/6289)). On Windows, `pnpm dev` can fail at startup when Turbopack `mmap`s a persistent-cache SST file and the OS refuses the mapping (`os error 1455` — "paging file too small"), which Turbopack surfaces as a misleading `Module not found: Can't resolve '@/shared/utils/machine'`. This is a **known upstream Turbopack cache-corruption bug — not our code**. The dev launcher (`scripts/dev/run-next.mjs`) now wraps `nextApp.prepare()` and, when it rejects with that signature (`isTurbopackCacheCorruption` in the new `scripts/dev/turbopackCacheHeal.mjs`), purges `.build/next/**/cache/turbopack` and retries **once** with a clear log. **Caveat — best-effort only:** the corruption often surfaces as a runtime overlay rather than a `prepare()` rejection, so this cannot always intercept it; the reliable remedy remains manually deleting the Turbopack cache dir. Regression guard: `tests/unit/turbopack-cache-heal-6289.test.ts`. (thanks @chirag127)
+
+- **fix(providers):** qodercli PAT auth no longer fails with `spawn qodercli ENOENT` on Windows ([#6263](https://github.com/diegosouzapw/OmniRoute/issues/6263)) — `spawnQoderCli` spawned the bare `qodercli` name with `shell:false` and an unenriched env, so the npm `.cmd` wrapper under `%APPDATA%\npm` (a user-PATH directory) was never resolved. It now resolves the absolute `.cmd`/`.exe` path through the existing `getCliRuntimeStatus("qoder")` resolver in `src/shared/services/cliRuntime.ts` (memoized), spawns with `shell` when the target is a `.cmd`/`.bat`, and uses the cliRuntime-enriched env (PATH + PATHEXT + APPDATA); the ENOENT error now lists the searched paths plus the `CLI_QODER_BIN` override. End-to-end spawn on a real Windows host is host-only (Hard Rule #18); the path-resolution logic is unit-tested. Regression guard: `tests/unit/qodercli-windows-resolve-6263.test.ts`. (thanks @chirag127)
+
+- **fix(sse):** the reasoning-token buffer no longer inflates **probe-sized `max_tokens`** ([#6274](https://github.com/diegosouzapw/OmniRoute/issues/6274)) — Claude Code's `/model` capability check sends `max_tokens: 1`, but for a thinking-capable model with a large output cap (e.g. `glm-5.2`) the #3587 headroom heuristic (`max(current + 1000, ceil(current * 1.5))`) rewrote it to `1001` and forwarded that upstream, wasting tokens on a request that was never a genuine reasoning budget. `resolveReasoningBufferedMaxTokens()` (`open-sse/services/reasoningTokenBuffer.ts`) now short-circuits and returns the caller's value verbatim when it is below the new `REASONING_BUFFER_MIN_TRIGGER` (256) threshold — a tiny explicit limit is a probe, not a reasoning request. Real budgets still receive the #3587 headroom unchanged, and the guard runs after the existing capability checks so unknown / non-reasoning models keep returning `null`. Regression guard: `tests/unit/reasoning-token-buffer-6274.test.ts`. (thanks @brightfiscalband)
+
+- **fix(cli):** `omniroute reset-password` now works as a real subcommand, and password resets over piped (non-TTY) stdin actually apply ([#6261](https://github.com/diegosouzapw/OmniRoute/issues/6261), [#6258](https://github.com/diegosouzapw/OmniRoute/issues/6258)). Two coupled defects: (1) **#6261** — `bin/omniroute.mjs` routed everything through Commander with only two pre-Commander bypasses (`--mcp`, `reset-encrypted-columns`), so `omniroute reset-password` was rejected as an unknown command; only the separate `omniroute-reset-password` bin worked, while the docs falsely advertised the subcommand (incl. a bogus "legacy alias still works"). A pre-Commander bypass mirroring `reset-encrypted-columns` now dynamically imports `bin/reset-password.mjs` (which self-executes) before Commander parses; the three doc lines were corrected. (2) **#6258** — `bin/reset-password.mjs` issued two sequential `rl.question` prompts; under piped stdin the second read never settled at EOF, so `main()` never reached `resetManagementPassword` and the reset was a silent no-op (both prompts printed, no success, password unchanged). The CLI now detects non-TTY stdin and reads it once (first line = password, second line = confirm if present, else reused), adds a `--password-stdin` flag (entire stdin is the password, no confirmation), and exits `0` explicitly so the success line always flushes; interactive TTY behavior is unchanged. Regression guard: `tests/unit/reset-password-cli-6261-6258.test.ts` (3). (thanks @chirag127)
+
+- **fix(db):** the mass-migration **safety abort** now tells the operator how to bypass it and stops flooding the log ([#6260](https://github.com/diegosouzapw/OmniRoute/issues/6260)) — after restoring a backup that wiped the migration tracking table, `runMigrations()` threw the abort on every downstream `ensureDbInitialized()`, re-logging the full banner 11+ times, and the message never mentioned the existing `OMNIROUTE_MAX_PENDING_MIGRATIONS` escape hatch. The abort text now appends a bypass hint (set `OMNIROUTE_MAX_PENDING_MIGRATIONS=0` in `server.env` / `DATA_DIR/.env`), and a new `MigrationSafetyAbortError` is memoized so repeated calls in the same process throw the same instance and emit a single concise line instead of the full cascade. Regression guard: `tests/unit/migration-safety-abort-6260.test.ts`. (thanks @chirag127)
+
+- **fix(auth):** importing a **distinct** Codex/ChatGPT OAuth `auth.json` is no longer falsely rejected as "already exists" when it belongs to a different user in the same workspace ([#6301](https://github.com/diegosouzapw/OmniRoute/issues/6301)). `findExistingCodexConnection` (in `src/lib/oauth/utils/codexAuthImport.ts`) deduped **only** on `providerSpecificData.workspaceId === accountId`, where `accountId` is the shared `chatgpt_account_id`/`tokens.account_id` — so two members of the same ChatGPT Team collapsed onto a single connection (409 `duplicate_account`). The id_token's `https://api.openai.com/auth` claim carries a per-user `chatgpt_user_id` alongside the workspace id (the device-flow path already persisted it as `chatgptUserId`, but the import path did not). Now `parseAndValidateCodexAuth` extracts `userId` (`chatgpt_user_id` → `user_id` → JWT `sub`) into `ParsedCodexAuth`, the create/update paths persist `chatgptUserId` in `providerSpecificData` (mirroring `codex.ts`), and dedup keys on `workspaceId` **AND** `chatgptUserId` — with a backward-compat fallback to legacy accountId-only matching when no stored connection for that workspace records a `chatgptUserId`, so genuinely-same accounts still dedup. Regression guard: `tests/unit/codex-auth-import-userid-dedup-6301.test.ts` (4). (thanks @anungma)
+
+- **fix(providers):** importing models for the **venice-web** provider no longer fails with a red "Provider venice-web does not support models listing" ([#6269](https://github.com/diegosouzapw/OmniRoute/issues/6269)). `venice-web` is a web-cookie provider with an executor but no upstream `/v1/models` endpoint and no registry `models`, so the models route fell through to the tail `400`. Mirroring the `jules`/`linkup-search`/`ollama-search` fix (#5569), it now ships a static local catalog entry in `src/lib/providers/staticModels.ts` — seeding the current Venice lineup (`venice-uncensored`, `llama-3.3-70b`, `qwen3-235b`, `qwen3-4b`, `deepseek-r1-671b`; Venice rotates its catalog, see docs.venice.ai/models/overview) — so the route returns `200` with `source:"local_catalog"`, `intentional:true`. Regression guard: `tests/unit/static-models-venice-web-6269.test.ts`. (thanks @chirag127)
+
+- **fix(api):** the specialty model catalogs (`/v1/embeddings`, `/v1/images`, `/v1/music`, `/v1/videos` model lists) are now derived from the **unified catalog filtered by a predicate** (`getSpecialtyModelsResponse`) instead of ad-hoc per-route logic, so they consistently respect active-credential visibility and stay in sync with the main catalog ([#6303](https://github.com/diegosouzapw/OmniRoute/pull/6303)). Regression guard: `tests/unit/specialty-model-catalog-routes.test.ts`. (thanks @makcimbx)
+- **fix(api):** the agent-bridge server route now resolves the MITM manager via a **dynamic `import("@/mitm/manager.runtime")`** so Turbopack does not statically pull the stub (or over-bundle the manager), and the agent-skills generator anchors its output base path with `path.join(process.cwd(), …)` so Turbopack's static analyzer stops tracing the whole project root ([#6329](https://github.com/diegosouzapw/OmniRoute/issues/6329), [#6366](https://github.com/diegosouzapw/OmniRoute/pull/6366)). Regression guard: `tests/unit/agent-bridge-server-route-dynamic-import.test.ts`. (thanks @Iammilansoni)
+- **fix(api):** internal probes (combo-test, cloud-sync verify) now pick a **management-scoped / allow-all API key** instead of naively grabbing `getApiKeys()[0]` — a restricted `self:usage` first row made the probe fail with "Model X is not allowed for this API key" even when the combo path was healthy (`pickApiKeyForInternalUse` in `src/lib/db/apiKeys.ts`). The API-manager model editor also falls back to `/api/models?all=true` when `/v1/models` is catalog-protected ([#6372](https://github.com/diegosouzapw/OmniRoute/pull/6372)). Regression guard: `tests/unit/pick-internal-api-key-6372.test.ts`. (thanks @jmengit)
+- **fix(live-ws):** the Live Dashboard WebSocket server now **rejects on bind failure** (e.g. `EADDRINUSE` when the API bridge already holds the port) instead of letting the error surface as an unhandled `error` event that crash-loops the process — the `error` listener is attached to `wss` (not `server`) and releases the EventBus subscription on a failed start ([#6324](https://github.com/diegosouzapw/OmniRoute/issues/6324)). Regression guard: `tests/unit/live-ws-eaddrinuse-6324.test.ts`. (thanks @vinayakkulkarni)
+- **fix(dashboard):** the Home provider-topology widget now trusts the live provider-metrics snapshot — it uses `topology.errorProvider` and live `activeRequests` directly instead of re-deriving state from a stale `lastErrorAt` or applying a frontend timeout filter, so the topology reflects real-time provider health ([#6322](https://github.com/diegosouzapw/OmniRoute/pull/6322)). Regression guard: `tests/unit/home-provider-topology-live-state.test.ts`. (thanks @xz-dev)
+- **fix(sse):** strip zero-width markers from streamed **tool-call arguments** — a follow-up to [#5857](https://github.com/diegosouzapw/OmniRoute/pull/5857). That PR removed injected zero-width joiners (U+200D) from streamed assistant text/reasoning but deliberately left tool-call argument JSON byte-exact. The request-side obfuscation (`open-sse/services/claudeCodeObfuscation.ts`) injects ZWJ into agent words — including the temp path inside the Bash tool description — and Claude models copy that verbatim into generated commands, which are delivered as tool-call arguments rather than assistant text. As a result the ZWJ survived and corrupted code blocks (e.g. a temp path rendered with an invisible joiner). Now `open-sse/handlers/responseSanitizer.ts` strips zero-width code points from tool-call argument strings at every emit site (OpenAI non-stream/stream chat `tool_calls` + legacy `function_call`, native Responses `function_call` items, the OpenAI→Responses conversion, and the native Responses streaming `response.function_call_arguments.delta/.done` events). Only zero-width code points are removed; JSON structure and all other bytes stay identical (no parse/restringify), so normal arguments remain byte-exact. Regression guard: 6 new cases in `tests/unit/response-sanitizer.test.ts` (suite 50/50).
+
+- **fix(nodejs):** the default app log path now resolves under `DATA_DIR` (`~/.omniroute/logs/application/app.log`) instead of `process.cwd()` ([#6197](https://github.com/diegosouzapw/OmniRoute/issues/6197)) — the globally-installed CLI runs from an arbitrary working directory, so anchoring the default to cwd made file logging silently write to (or no-op under) an unrelated directory, contradicting the documented `.env.example` default. `getAppLogFilePath()` now computes the default lazily via the pure `resolveDataDir()` resolver (honours a per-process `DATA_DIR`, no directory-creation side effect); an explicit `APP_LOG_FILE_PATH` still wins. Regression guard: `tests/unit/logenv-datadir-path-6197.test.ts` (3).
+
+- **fix(docker):** AgentBridge/`startMitm` no longer aborts in containers/headless when the Antigravity-default DNS step can't write `/etc/hosts` ([#6127](https://github.com/diegosouzapw/OmniRoute/issues/6127)), and the privileged command's stderr now reaches `app.log` instead of only a bare exit code hitting the toast ([#6198](https://github.com/diegosouzapw/OmniRoute/issues/6198)). The default DNS step (`addDNSEntry`) was called unguarded while cert install and the two sibling DNS steps were each best-effort — in the runtime Docker image (`USER node`, no `sudo`, read-only `/etc/hosts`) it threw `Command failed with code 1` out of `startMitmInternal` and killed the whole start, discarding the stderr. The three DNS steps are extracted into a best-effort `provisionDnsEntries()` where each failure is logged with the full `err` (stderr included, folded in by `systemCommands.ts`) and never aborts the start. Regression guard: `tests/unit/mitm-dns-graceful-degrade-6127.test.ts` (4).
+
+- **fix(providers):** copilot-m365-web now supports the M365 Education "Starter / OfficeWebIncludedCopilot" tier and no longer returns an empty `content:null` stream ([#6210](https://github.com/diegosouzapw/OmniRoute/issues/6210)). Two gaps: (1) `buildWsUrl()` hardcoded the individual-consumer scenario (`OfficeWebPaidConsumerCopilot`, `isEdu=false`) — the EDU tier is now opt-in via `providerSpecificData.tier="edu"`, emitting `scenario=OfficeWebIncludedCopilot`/`isEdu=true` (the individual path is unchanged); (2) the EDU/GPT-5.5 path streams deltas via `arguments[0].writeAtCursor` (incremental) instead of only `messages[].text` (accumulated snapshots), which the parser dropped — a new `accumulateBotContent()` folds both formats, with `type:2 item.result.message` as a last-resort fallback. Regression guard: `tests/unit/copilot-m365-edu-writeatcursor-6210.test.ts` (10). (thanks @qpeyba)
+
+- **fix(providers):** GitLab Duo executor now feeds tool results back into the prompt instead of looping ([#6220](https://github.com/diegosouzapw/OmniRoute/issues/6220)) — `buildPrompt()` branched only on `system`/`user` and took `userParts.at(-1)`, silently dropping the `assistant{tool_calls}` + `tool{result}` turns the client appended, so the reconstructed prompt was byte-identical to turn 1 and the model re-emitted the same `<tool>` call forever. When a tool exchange is present the full conversation is now serialized, folding each tool result back keyed by its `tool_call_id`; simple conversations keep the legacy shape. Complements the tool_call emission from [#6051](https://github.com/diegosouzapw/OmniRoute/issues/6051) (the `kilo-duplicate` label was a false positive — different, sequential defect). Regression guard: `tests/unit/gitlab-tool-result-feedback-6220.test.ts` (4).
+
+- **fix(providers):** opencode-go/opencode-zen can now synthesize the OpenCode CLI identity headers Cloudflare requires on VPS egress ([#5997](https://github.com/diegosouzapw/OmniRoute/issues/5997)) — on a datacenter VPS, `opencode.ai/zen/go/v1/chat/completions` 403s (HTML challenge) requests lacking CLI identity, while the reporter's control curl proved that `User-Agent: opencode-cli/1.0.0` + `x-opencode-client: cli` + `x-opencode-project: default` + fresh request/session UUIDs succeed. Opt-in via `OPENCODE_SYNTHESIZE_CLI_HEADERS=true` (values overridable via `OPENCODE_GO_USER_AGENT`/`OPENCODE_USER_AGENT`/`OPENCODE_CLIENT`/`OPENCODE_PROJECT`); it fills only headers the client did not already send. Kept **off by default** — the forward-only path is deliberate (fabricating a wrong value risks upstream rejection; a prior dedup regressed with `opencode/local`), so this replaces the fragile local header-injection shim without changing default behavior. Regression guard: `tests/unit/opencode-cli-headers-synthesis-5997.test.ts` (6). (thanks @aleksesipenko)
+- fix(resilience): sticky session affinity now evicts and fails over to another account when the pinned account is exhausted/unavailable (#6219)
+
+- fix(sse): Responses API passthrough now drops internal commentary-phase output before forwarding to clients (gated by RESPONSES_PASSTHROUGH_DROP_COMMENTARY, default on) (#6199)
+
+- **fix(sse):** tool-call function schemas with a root `type: null` are now coerced to `type: "object"` before dispatch ([#6359](https://github.com/diegosouzapw/OmniRoute/issues/6359)) — clients like the Codex app emit `parameters: { type: null, ... }` for some tools, which OpenAI-compatible upstreams reject with `400 Invalid schema for function '...': schema must be a JSON Schema of 'type: "object"', got 'type: null'`, failing the whole request. `toolSchemaSanitizer` already stripped the null; it now re-adds the mandatory root `"object"` type (and empty `properties`/open `additionalProperties` when absent). Combinator roots (`anyOf`/`oneOf`/`allOf`) and explicit root types are left untouched. Regression guard: 5 new cases in `tests/unit/tool-schema-sanitizer.test.mjs`.
+
+- **fix(docker):** AgentBridge no longer fails to start on npm/Electron/VPS installs with "MITM manager stub reached at runtime" ([#6344](https://github.com/diegosouzapw/OmniRoute/issues/6344)) — v3.8.45 flipped the production bundler default to Turbopack, but `next.config.mjs` aliased `@/mitm/manager` to its Docker-only degraded stub **unconditionally**. That was harmless while Docker (which sets the alias intentionally for #3390 graceful degradation) was the sole Turbopack consumer, but once every artifact built with Turbopack the stub shipped to all non-Docker users and `startMitm` threw on the first Agent-Bridge start. The alias is now opt-in via `OMNIROUTE_MITM_STUB=1` (set only by the Dockerfile) through the shared `scripts/build/mitm-stub-flag.mjs` helper; default builds bundle the real manager. Regression guard: `tests/unit/mitm-stub-alias-6344.test.mjs` (4).
+
+- **fix(proxy):** stop the v3.8.44 proxy regression that leaked the real IP and disabled healthy proxies ([#6246](https://github.com/diegosouzapw/OmniRoute/issues/6246)). Two coupled defects from the new health scheduler: (1) **IP leak** — when a proxy assigned to a connection was marked `inactive`, resolution fell through to a **direct** egress instead of blocking, exposing the operator's real IP; (2) **over-deactivation** — the sweep flipped a proxy to `inactive` on the **first** failed probe and counted our own 5s timeout / a probe-target `5xx` as the proxy's fault, so healthy paid proxies vanished from egress selection ("my proxies are not being used anymore"). Fix: the sweep decision is extracted into a pure, network-free `decideProxyHealthAction` (`src/lib/proxyHealth/decision.ts`) — by default the health check now **only counts/logs and never downgrades status** (a proxy is downgraded/removed only with `PROXY_AUTO_REMOVE=true`, after `PROXY_AUTO_REMOVE_AFTER` **consecutive** conclusive failures); probes are classified tri-state so an inconclusive result (our timeout, or a `5xx` from the probe target) never penalizes the proxy, and the probe timeout is raised 5s→15s. Separately, `safeResolveProxy` now **fails closed** via the existing policy: a connection whose assigned proxy is dead is blocked instead of leaking direct (`hasBlockingProxyAssignment`), honoring the explicit `proxy off` toggles and the `PROXY_FAIL_OPEN=true` opt-out. Existing proxies stuck `inactive` by the old behavior need a one-time manual re-activate (the operator owns proxy status). Regression guards: `tests/unit/proxy-health-decide-action-6246.test.ts`, `tests/unit/proxy-assigned-unavailable-6246.test.ts`.
+
+- **fix(proxy):** make "Test All" read-only and add bulk enable/disable ([#6246](https://github.com/diegosouzapw/OmniRoute/issues/6246)). Complements the core fail-closed / scheduler fix (#6296) with the two remaining reporter asks. (1) The **"Test All" button** (`POST /api/settings/proxies/auto-test`) used to flip a proxy to `inactive` on a failed reachability probe; since the egress selector excludes `inactive` proxies, a flaky probe (an unreachable `httpbin.org`, a proxy that blocks `HEAD`, or a slow paid proxy) silently disabled every proxy that failed — "Test All" is now **read-only by default** (only the operator sets a proxy active/inactive; opt back into the legacy test-and-set with `PROXY_HEALTH_AUTO_DEACTIVATE=true`). (2) Adds a **bulk enable/disable** proxies endpoint + toolbar action (`POST /api/settings/proxies/batch-activate`) so an operator can re-activate proxies in one click. Regression guard: `tests/unit/proxy-health-6246.test.ts`. (thanks @tenshiak)
+
+- **chatcore (tools): stop the default 128-tool cap from silently dropping opencode's `task`/MCP tools.** opencode (used as an MCP/agent host) sends a large tool list; when it exceeds the speculative `MAX_TOOLS_LIMIT` (128) default, `truncateToolList` did a blind `tools.slice(0, 128)`, dropping every tool past index 128 — including opencode's built-in `task` tool (subagent launch) and many MCP tools, so models routed through OmniRoute could no longer spawn subagents or reach part of their tools. The cap exists to avoid upstream `400`s for providers with real hard limits (e.g. grok-cli 200), so it is kept for those: detection of the opencode client (`isOpencodeClient` — any `x-opencode-*` header, or `opencode` in the user-agent) now only bypasses the **speculative 128 default**, never a known provider ceiling. Precedence is explicit — a proactive/detected provider limit always truncates (even for opencode); otherwise opencode forwards its full tool list; otherwise the unchanged 128 default applies to every other client. Refactors `getEffectiveToolLimit` into `getKnownToolLimit(provider) ?? DEFAULT_LIMIT` (byte-identical for existing callers) and fixes a cosmetic debug-log that reported the truncated count instead of the original. Regression guard: `tests/unit/tool-limit-detector.test.ts`.
+
+- **fix(mitm):** the macOS MITM-cert install check now matches the system keychain again. `security find-certificate -a -Z` prints the SHA-1 as a colon-less hex string, but the installed-check compared it against `getCertFingerprint()`'s colon-separated form, so the substring match never hit — the cert was reported as not-installed and re-prompted for the sudo install on every run. Fingerprints are now normalized (colons stripped, upper-cased) on both sides via the extracted `macCertOutputHasFingerprint` helper. Regression guard: `tests/unit/mitm-cert-mac-fingerprint.test.ts`. ([#6204](https://github.com/diegosouzapw/OmniRoute/pull/6204), closes [#6134](https://github.com/diegosouzapw/OmniRoute/issues/6134) — thanks @rianonehub)
+
+- **fix(api):** `/v1/messages/count_tokens` now counts `tool_use`, `tool_result` and `thinking` content blocks (and array-form `system` prompts) in the local-estimation path, instead of only `text`. Real agentic conversations keep ~95% of their tokens inside tool results; the previous estimate returned near-zero for them, which silently broke Claude Code's auto-compaction (context grew past the window with no compaction until the upstream API rejected the request). The real provider-side count path is unchanged. Regression guard: `tests/unit/messages-count-tokens-route.test.ts`. ([#6221](https://github.com/diegosouzapw/OmniRoute/pull/6221) — thanks @luweiCN)
+
+- **fix(antigravity):** strip a trailing assistant prefill turn for Vertex Claude models to avoid upstream 400s ([#6114](https://github.com/diegosouzapw/OmniRoute/pull/6114)). Regression guard: `tests/unit/antigravity-claude-prefill-strip.test.ts`. (thanks @anki1kr)
+
+- **fix(security):** the mutable cloud-agent routes (`/api/cloud/credentials/update`, `/api/cloud/models/alias`) now require management auth instead of being treated as public. They were classified as public API routes, so a request without management credentials could update stored cloud-agent credentials and model aliases. They are removed from the public-route set, classified as management routes in the authz pipeline, and gated by `requireManagementAuth`; cloud **read**/auth routes stay public. Regression guards: `tests/unit/cloud-write-auth.test.ts`, `tests/unit/authz/classify.test.ts`, `tests/unit/public-api-routes.test.ts`. ([#6233](https://github.com/diegosouzapw/OmniRoute/pull/6233) — thanks @vittoroliveira-dev)
+
+- **refactor(dashboard):** extract the onboarding-wizard "Open provider details" link target into a pure, unit-tested `buildProviderDetailsHref(connection)` helper. The wizard already routes by `connection.id` (the node UUID) rather than the provider category slug (#6144/#6145); this hardens that behavior behind a tested helper that guards a missing id/connection. Regression guard: `tests/unit/provider-onboarding-href.test.ts`. ([#6166](https://github.com/diegosouzapw/OmniRoute/pull/6166) — thanks @KooshaPari)
+
+- **fix(security):** the doubao synthetic device-id generator now derives its digits via an unbiased crypto-random draw (rejection sampling over `crypto.randomBytes()`) instead of a `% 10` reduction, closing a CodeQL `js/biased-cryptographic-random` finding.
+- **fix(agentSkills):** the GitHub-skills generator now resolves `outputDir` to an absolute path before writing, fixing a regression introduced by #6366 (relative-to-cwd base path) that could write generated skill files to the wrong directory.
+- **fix(security):** `/api/keys/{id}/devices` now checks the HTTP method before auth/validation, returning a `405` for non-GET/DELETE verbs instead of a misleading `401`/`500` (closes a `dast-smoke` QUERY-method finding).
+- **fix(quality):** clear the last 2 heavy quality-gate reds on the release tip (cycle pre-flight).
+- **fix(mitm):** the test suite and CI can never mutate the OS trust store — `OMNIROUTE_SKIP_SYSTEM_TRUST=1` is set globally for tests/CI so `installCert`/`uninstallCert`/`installTproxyCa` skip the privileged OS dispatch ([#6310](https://github.com/diegosouzapw/OmniRoute/pull/6310); full detail is under the [3.8.45] section below — this branch received it via the parallel-cycle sync-back).
+- **fix(api):** `POST /api/github-skills` now Zod-validates its request body; documented the new quality-gate env vars and pinned the merge-integrity GitHub Actions to a commit SHA.
+- **fix(skills):** generate the missing `omni-github-skills` registry entry and align the agent-skills catalog-count tests (follow-up to #6186).
+- **fix(quality):** clear the cycle's 11 net-new ESLint errors and make `validate-release-green` suppressions-aware.
+
+### 📝 Maintenance
+
+- **i18n(it):** add 118 missing Italian (`it`) translations (net-additive — no existing keys dropped, valid JSON), improving Italian UI coverage. ([#6212](https://github.com/diegosouzapw/OmniRoute/pull/6212) — thanks @serverless83)
+- **chore(providers):** remove deprecated **MiMo V2** model entries from the catalogs (xiaomi-mimo, opencode-go, zenmux-free, audio TTS) — the upstream V2 line is superseded by MiMo V2.5; drops `mimo-v2-tts`, `mimo-v2-pro`, `mimo-v2-omni`, `mimo-v2-flash`, `mimo-v2-flash-free` and realigns the provider-catalog tests. ([#6248](https://github.com/diegosouzapw/OmniRoute/pull/6248) — thanks @backryun)
+- **chore(release):** ~50 commits on this branch are v3.8.45 pre-flight/hardening fixes and CI-perf work that landed here via the parallel-cycle sync-back (`sync-next-cycle.mjs`, Hard Rule #21) **after** the `v3.8.45` git tag was cut, and are already fully documented under the **[3.8.45]** section below — listed here only so the per-cycle commit-coverage check (`npm run release:uncovered`) doesn't flag them as gaps. Provider/catalog/UX/backend fixes: #6041, #6078, #6108, #6135, #6148, #6149, #6154, #6158, #6161, #6162, #6163, #6164, #6165, #6170, #6177, #6178, #6181, #6186, #6187, #6191, #6193, #6194, #6195, #6200, #6205, #6208, #6209, #6211, #6213, #6223, #6224, #6225, #6226, #6227, #6228, #6229, #6230, #6235, #6291, #6292. CI/release-pipeline work: #6167, #6203, #6214, #6215, #6218, #6273, #6275, #6283, #6284, #6285, #6300, #6305.
+- **chore(release):** additional zero-ref release-cycle plumbing on this branch, kept out of `release:uncovered` on purpose (no `#N` in the commit subject to cite): opening the v3.8.46 cycle, opening/closing the v3.8.45 cycle, the finalized [3.8.45] CHANGELOG i18n sync-back to 42 mirrors, the v3.8.45 cognitive/cyclomatic and file-size drift rebaselines, ESLint stale-suppression pruning (4,273 → 4,233), and clearing test-masking/docs-all pre-flight reds for v3.8.45.
+
+### ⚡ Performance & Infrastructure
+
+- **perf(release-green):** the pre-flight validator (`scripts/quality/validate-release-green.mjs`) now runs its 4 slow suites (unit / vitest / integration / pack-artifact) **concurrently** via `Promise.all` — pre-flight wall time drops from ~the sum of the suites to ~the slowest one (~30min saved per round; Phase 0 was the nº1 bottleneck of the v3.8.45 release benchmark, 2h54 of 6h34 e2e). Guard: `tests/unit/validate-release-green.test.ts` ("runs the slow suites CONCURRENTLY"). ([#6319](https://github.com/diegosouzapw/OmniRoute/pull/6319))
+- **fix(ci):** `scripts/release/sync-next-cycle.mjs` — two defects found live in its first production run (v3.8.45 Phase 5): (1) the `git()` helper's default 1 MiB `maxBuffer` crashed with `ENOBUFS` on `git show origin/main:CHANGELOG.md` (the CHANGELOG alone is >1 MiB) — widened to 64 MiB; (2) the i18n resync only propagated the `[NEXT]` (TBD) section, leaving the just-shipped finalized section as "— TBD" in all 42 mirrors — it now also syncs `[prevVersion]` bounded by the heading below it (new exported pure helper `versionAfter`). Guards: +5 tests in `tests/unit/sync-next-cycle.test.ts` (8/8). ([#6327](https://github.com/diegosouzapw/OmniRoute/pull/6327))
+- **test(ci):** concurrency-sensitive flaky tests are **quarantined into a serial pass** (`tests/unit/serial/`, `--test-concurrency=1`, appended to every unit runner incl. sharded variants — the serial pass is sharded too so concurrent shard jobs never self-collide). Initial set: `glm-coding-plan-monthly-3580`, `quota-division-blocks`, `provider-health-autopilot`, `combo-health-autopilot` — the class behind the ~28min CI wedges/re-runs (two live 1h+ wedges cancelled during this PR's own validation). Discovery + TIA gates track the new glob; systemic root cause (async logger writing after teardown) tracked in [#6360](https://github.com/diegosouzapw/OmniRoute/issues/6360). Guard: `tests/unit/test-serial-quarantine.test.ts` (4). ([#6347](https://github.com/diegosouzapw/OmniRoute/pull/6347))
+
+### 🙌 Contributors
+
+Thanks to everyone whose work landed in v3.8.46:
+
+| Contributor                                                    | PRs / Issues           |
+| -------------------------------------------------------------- | ---------------------- |
+| [@2220258345](https://github.com/2220258345)                   | direct commit / report |
+| [@abdofallah](https://github.com/abdofallah)                   | direct commit / report |
+| [@adentdk](https://github.com/adentdk)                         | direct commit / report |
+| [@aleksesipenko](https://github.com/aleksesipenko)             | direct commit / report |
+| [@anki1kr](https://github.com/anki1kr)                         | direct commit / report |
+| [@anungma](https://github.com/anungma)                         | direct commit / report |
+| [@arpicato](https://github.com/arpicato)                       | direct commit / report |
+| [@backryun](https://github.com/backryun)                       | #6248                  |
+| [@binsarjr](https://github.com/binsarjr)                       | direct commit / report |
+| [@brightfiscalband](https://github.com/brightfiscalband)       | direct commit / report |
+| [@chirag127](https://github.com/chirag127)                     | #6501, #6506           |
+| [@developerjillur](https://github.com/developerjillur)         | direct commit / report |
+| [@dilneiss](https://github.com/dilneiss)                       | #6499                  |
+| [@dtybnrj](https://github.com/dtybnrj)                         | direct commit / report |
+| [@Forcerecon](https://github.com/Forcerecon)                   | direct commit / report |
+| [@hao3039032](https://github.com/hao3039032)                   | direct commit / report |
+| [@Iammilansoni](https://github.com/Iammilansoni)               | #6150, #6245           |
+| [@jmengit](https://github.com/jmengit)                         | direct commit / report |
+| [@jordansilly77-stack](https://github.com/jordansilly77-stack) | direct commit / report |
+| [@JxnLexn](https://github.com/JxnLexn)                         | direct commit / report |
+| [@KooshaPari](https://github.com/KooshaPari)                   | #6166                  |
+| [@loopyd](https://github.com/loopyd)                           | direct commit / report |
+| [@luweiCN](https://github.com/luweiCN)                         | #6221                  |
+| [@makcimbx](https://github.com/makcimbx)                       | direct commit / report |
+| [@muflifadla38](https://github.com/muflifadla38)               | direct commit / report |
+| [@newnol](https://github.com/newnol)                           | direct commit / report |
+| [@ofekbetzalel](https://github.com/ofekbetzalel)               | direct commit / report |
+| [@ohahe52-dot](https://github.com/ohahe52-dot)                 | direct commit / report |
+| [@phidinhmanh](https://github.com/phidinhmanh)                 | direct commit / report |
+| [@powellnorma](https://github.com/powellnorma)                 | direct commit / report |
+| [@qpeyba](https://github.com/qpeyba)                           | direct commit / report |
+| [@RaviTharuma](https://github.com/RaviTharuma)                 | direct commit / report |
+| [@RCrushMe](https://github.com/RCrushMe)                       | direct commit / report |
+| [@rianonehub](https://github.com/rianonehub)                   | #6134, #6204           |
+| [@serverless83](https://github.com/serverless83)               | #6212                  |
+| [@swingtempo](https://github.com/swingtempo)                   | direct commit / report |
+| [@tenshiak](https://github.com/tenshiak)                       | direct commit / report |
+| [@ThongAccount](https://github.com/ThongAccount)               | direct commit / report |
+| [@UnrealAryan](https://github.com/UnrealAryan)                 | direct commit / report |
+| [@vinayakkulkarni](https://github.com/vinayakkulkarni)         | direct commit / report |
+| [@vittoroliveira-dev](https://github.com/vittoroliveira-dev)   | #6233                  |
+| [@warelik](https://github.com/warelik)                         | direct commit / report |
+| [@xxy9468615](https://github.com/xxy9468615)                   | direct commit / report |
+| [@xz-dev](https://github.com/xz-dev)                           | direct commit / report |
+| [@yanpaing007](https://github.com/yanpaing007)                 | direct commit / report |
+| [@diegosouzapw](https://github.com/diegosouzapw)               | maintainer             |
+
+---
+
 ## [3.8.45] — 2026-07-06
 
 ### ✨ New Features

docs/i18n/sw/CHANGELOG.md

@@ -6,6 +6,213 @@
 
 ## [3.8.31] — 2026-06-20
 
+## [3.8.46] — TBD
+
+### ✨ New Features
+
+- **feat(sse):** **hide paid-only models from `auto/*` routing** when `hidePaidModels` is on ([#6512](https://github.com/diegosouzapw/OmniRoute/issues/6512)) — follow-up to #6328/#6495. PR #6495 hid paid-only models from the `GET /v1/models` listing, but `auto/*` combos (`auto/best-coding`, `auto/glm`, …) could still pick a paid-only backend into their candidate pool → a 402/403 at request time. `createVirtualAutoCombo` now filters the candidate pool through the new pure `open-sse/services/autoCombo/paidModelFilter.ts` (`filterPaidOnlyCandidates`), applying the same free-model predicate #6495 uses in `catalog.ts` (`providerHasFreeModels(provider) && isFreeModel(provider, {id})`) whenever `settings.hidePaidModels === true`. Applied before the category/tier/family narrowing, so it covers every `auto/*` combo; an all-paid pool degrades to the existing graceful empty-pool path. **Opt-in — default OFF leaves the pool unchanged** (identity). Regression guard: `tests/unit/autoCombo/paid-model-filter-6512.test.ts` (4, incl. the default-off identity guard).
+- **feat(sse):** **provider-family auto combos** — `auto/glm`, `auto/minimax`, `auto/mimo`, `auto/zai`, `auto/gemma`, `auto/llama`, `auto/gemini` ([#6453](https://github.com/diegosouzapw/OmniRoute/issues/6453)) — new routable ids that materialize an on-demand virtual combo spanning whatever installed backends currently expose that model family, degrading gracefully as backends rotate. A new pure `open-sse/services/autoCombo/modelFamily.ts` (`detectModelFamily`) classifies by model-id prefix for six families; `zai` is instead resolved by provider id (z.ai's hosted API serves the same `glm-*` model ids as every other GLM backend, so `auto/zai` means "route to my z.ai backend specifically" vs `auto/glm`'s "any connected GLM backend"). Reuses the existing `createVirtualAutoCombo` on-demand materialization path (no DB writes) and the `/v1/models` catalog advertising loop. Regression guard: `tests/unit/autoCombo/provider-family-combos.test.ts` (11).
+- **feat(proxy):** native **proxy-pool round-robin / egress IP rotation** ([#6365](https://github.com/diegosouzapw/OmniRoute/issues/6365)) — a scope (global / provider / account) can now hold **multiple** proxies as a pool with a rotation strategy, so outbound requests cycle their egress IP instead of pinning one proxy per scope. Migration `117_proxy_pool_rotation.sql` lifts the `UNIQUE(scope, scope_id)` constraint (rebuild via the canonical rename/copy/drop; existing single assignments become 1-element pools) and adds a `proxy_scope_rotation` companion table holding the per-scope strategy + a persisted monotonic round-robin cursor. Strategies: `round-robin` (default, monotonic cursor — never `Math.random`), `random`, and `sticky-per-N-min`. Resolution (`resolveProxyForScopeFromRegistry` / `resolveProxyForConnectionFromRegistry`) now fetches the alive, position-ordered candidate set (unchanged `PROXY_ALIVE_PREDICATE`) and applies the strategy; an empty / all-dead pool still returns `null` — the #6246 fail-closed guard is **untouched** (never falls through to direct egress). Backend + DB only; dashboard pool-builder UI is a follow-up. Regression guard: `tests/unit/proxy-pool-rotation-6365.test.ts` (8, incl. fail-closed + backward-compat).
+- **feat(providers):** end-to-end **tool/function calling on the native Gemini `/v1beta` endpoint** ([#6222](https://github.com/diegosouzapw/OmniRoute/issues/6222)) — both directions of the Gemini↔OpenAI conversion now preserve tool data (previously silently dropped). Request side: `convertGeminiToInternal` (extracted to its own testable module) maps `tools[].functionDeclarations` → OpenAI `tools`, prior `functionCall` parts → assistant `tool_calls`, and `functionResponse` parts → `tool`-role messages. Response side: `convertOpenAIResponseToGemini` emits `parts[].functionCall {name,args}` from `message.tool_calls`, and the streaming `openAIChunkToGeminiChunk` accumulates fragmented `tool_calls` deltas by index into complete `functionCall` parts. The non-Gemini client paths (Claude, OpenAI-Responses) already preserved tool calls — this closes the gap specific to the native Gemini surface. Regression guard: `tests/unit/v1beta-gemini-tool-calling-6222.test.ts` (6, incl. a streaming SSE round-trip).
+- **feat(providers):** copilot-m365-web **enterprise / work tier** support ([#6334](https://github.com/diegosouzapw/OmniRoute/issues/6334)) — mirrors the EDU-tier pattern (#6210): `M365ConnectionParams` gains an `agent` field, a new opt-in `M365_ENTERPRISE_OVERRIDES` preset (`agent=work`, `scenario=officeweb`, `licenseType=Premium`) applies via `providerSpecificData.tier="enterprise"` (alias `"work"`), and `agent` is also overridable directly via `providerSpecificData.agent`. `buildWsUrl` was hardcoding `agent="web"` (the one enterprise-distinguishing param with no override path), so a Premium work account handshook then returned an empty stream. The individual and EDU paths are untouched. Kilo's dup flag vs #6210 (EDU tier) was a false positive — different tier. Regression guard: `tests/unit/copilot-m365-enterprise-6334.test.ts` (7). End-to-end confirmation on a real Premium work account is a live-VPS validation follow-up (Hard Rule #18). (thanks @Forcerecon)
+- **feat(api):** standardized, provider-agnostic **`effort` + `thinking` request params** ([#6241](https://github.com/diegosouzapw/OmniRoute/issues/6241)) — a thin standardization layer over the existing mature per-provider reasoning plumbing (no provider mapper touched). `providerChatCompletionSchema` gains a canonical `effort` (reusing the shared `none/low/medium/high/xhigh` vocabulary — the UI tiers `extra`/`max` collapse onto `xhigh`) and a boolean `thinking`. A pure `normalizeReasoningRequest` (wired once in `src/sse/handlers/chat.ts`, before any reasoning field is read) folds them onto the fields the translators already consume (`reasoning_effort` / `reasoning.effort` / `thinking`), so they fan out to Anthropic / Gemini / xAI / Responses — an explicit client `reasoning_effort` / object-shaped `thinking` always wins (backward-compatible). `/models` additively exposes `supportsThinking` + `effort_tiers` so the frontend can render the toggles (UI component is a follow-up). Regression guard: `tests/unit/effort-thinking-standardization-6241.test.ts` (12). (thanks @Iammilansoni, @shabeer)
+- **feat(combo):** new **`pipeline` (sequential) combo strategy** ([#6297](https://github.com/diegosouzapw/OmniRoute/issues/6297)) — the 18th routing strategy runs targets **in order**, threading each step's output into the next step's input, with an optional per-step `prompt` (system instruction); only the final step's response is returned. Distinct from `fusion` (parallel fan-out + judge). Implemented as a self-contained `open-sse/services/pipeline.ts` (sibling to `fusion.ts`), dispatched from `combo.ts`; the step list reuses `combo.models` order and reads an optional `prompt` off each target (backward-compatible — ignored by every other strategy). Intermediate steps run non-streaming with tools stripped (complete prose to thread forward); the final step keeps the client's `stream` flag + tools. A failing/empty/unparseable intermediate step fails the whole pipeline explicitly via a sanitized error (never silently swallowed). Kilo's dup flag vs #563 was a false positive (that's model→chain selection; this is a sequential chain). Regression guard: `tests/unit/combo-pipeline-strategy.test.ts` (5). (thanks @ofekbetzalel)
+- **feat(ci):** `check:test-masking` now flags **inline-reimplemented prod conditions** ([#6348](https://github.com/diegosouzapw/OmniRoute/issues/6348)) — a new **report-only** subcheck (v2, 6A.10 family) catches the wrong-shape contract test: a test that recomputes the condition under test inline instead of importing/exercising the real function (the #6216 class, where `=== 500` → `>= 500` stayed green because the test re-implemented the branch). For each added/modified test file it warns when the file textually duplicates a ≥3-token conditional from a production file touched in the same PR **and** does not import the symbol/module owning it, via a pure, fixture-tested `findReimplementedConditions()` with an allowlist mirroring `assertReductionAllowlist`. Report-only for now (does not fail the gate) — to be promoted to blocking after a triage cycle. Regression guard: `tests/unit/check-test-masking.test.ts` (45).
+- **feat(sse):** per-connection routing override (native vs CLIProxyAPI) ([#6339](https://github.com/diegosouzapw/OmniRoute/issues/6339)) — the previously-dead `isCliproxyapiDeepModeEnabled` helper is now wired into `resolveExecutorWithProxy`: a single connection can opt itself into the CLIProxyAPI passthrough executor via `providerSpecificData.cliproxyapiMode="claude-native"`, with precedence **connection override > provider `upstream_proxy_config` mode > default**. `resolveExecutorWithProxy` now receives the resolved connection's `providerSpecificData` (threaded from `chatCore.ts`), so one connection can deep-route while the provider's other connections stay native — no DB schema change (the toggle rides in `providerSpecificData`). Also resolves the same-provider mixing ask in #6340. Regression guard: `tests/unit/chatcore-executor-proxy.test.ts` (9). (thanks @RaviTharuma)
+- **feat(dashboard):** "Add session cookie" modal now shows a prominent **"Open ‹host› →"** link to the provider's own site ([#6268](https://github.com/diegosouzapw/OmniRoute/issues/6268)) — every `-web` cookie-session provider (chatgpt-web, claude-web, gemini-web, kimi-web, lmarena, qwen-web, m365-copilot-web, …) renders a one-click external link (opening the provider's login/home page in a new tab) so operators no longer tab away to retype the URL mid-setup. The host resolves from a pure, unit-tested `resolveWebProviderHost()` (prefers `WEB_COOKIE_PROVIDERS[id].website`, falls back to the registry `baseUrl` origin); non-web providers render exactly as before. Kilo's dup flag vs #6265 (modal-too-small-on-1080p) was a false positive — distinct concern. Regression guard: `tests/unit/resolve-web-provider-host.test.ts` (5). (thanks @chirag127)
+- **feat(providers):** add **DigitalOcean** AI (serverless inference) as an OpenAI-compatible API-key provider ([#6373](https://github.com/diegosouzapw/OmniRoute/pull/6373)) — base `https://inference.do-ai.run/v1`, wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/digitalocean/`, `src/shared/constants/providers/apikey/inference-hosts.ts`). Regression guard: `tests/unit/digitalocean-provider.test.ts`. (thanks @newnol)
+- **feat(providers):** add **Huancheng Public API** (`hcnsec`) as an OpenAI-compatible regional provider ([#6410](https://github.com/diegosouzapw/OmniRoute/pull/6410)) — Xinjiang Huancheng Cybersecurity's public LLM platform (base `https://api.hcnsec.cn/v1`, free credits via daily check-ins), wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/hcnsec/`, `src/shared/constants/providers/apikey/regional.ts`). Regression guard: `tests/unit/hcnsec-provider.test.ts`. (thanks @UnrealAryan)
+- **feat(dashboard):** the web-session credential guide now shows an **"Open {host}" link** ([#6316](https://github.com/diegosouzapw/OmniRoute/pull/6316)) to the provider's sign-in site (derived from the provider `website` via `getProviderWebsiteHost`), so you can jump straight to the page where the cookie/session must be captured. Regression guard: `tests/unit/web-session-provider-link-6316.test.ts`. (thanks @jordansilly77-stack)
+- **feat(cerebras):** add the **Gemma 4 31B** model (`gemma-4-31b`) to the Cerebras registry + pricing table ([#6331](https://github.com/diegosouzapw/OmniRoute/pull/6331)). Regression guard extends `tests/unit/t28-model-catalog-updates.test.ts`. (thanks @backryun)
+- **feat(providers):** add **Yuanbao (web)** as a cookie-session provider ([#6196](https://github.com/diegosouzapw/OmniRoute/issues/6196)) — `yuanbao-web` (Tencent Yuanbao, `yuanbao.tencent.com`) with cookie-only auth (`hy_user`/`hy_token` + public agent id), SSE→OpenAI translation incl. `reasoning_content`, exposing DeepSeek V3/R1 + Hunyuan / Hunyuan-T1. Regression guard: `tests/unit/providers-yuanbao-web.test.ts`. `together-web` was **deferred** (no verifiable web-session endpoint — needs a captured request) and `huggingchat-web` **dropped** (the existing `huggingchat` already is a web-cookie provider). (thanks @chirag127)
+- **feat(providers):** route the built-in **agentrouter** through the dynamic Claude-Code wire image ([#6056](https://github.com/diegosouzapw/OmniRoute/issues/6056)) — a small static allow-set (`CC_WIRE_IMAGE_BUILTINS` in `open-sse/services/ccWireImageBuiltins.ts`), consulted by `isClaudeCodeCompatible` / `isClaudeCodeCompatibleProvider` / `applyFingerprint`, makes agentrouter adopt the CC wire-image headers + fingerprint **while guarding the CC baseUrl/auth branches** so it keeps its own registry `baseUrl` and `x-api-key` auth. Regression guard: `tests/unit/agentrouter-cc-wire-image.test.ts` (asserts the wire image is applied AND agentrouter's baseUrl/auth are preserved). Live WAF-acceptance against agentrouter.org is a VPS validation follow-up (Hard Rule #18).
+- **feat(providers):** **bulk-add API keys for Cloudflare Workers AI** ([#6174](https://github.com/diegosouzapw/OmniRoute/issues/6174)) — `cloudflare-ai` is removed from the bulk-add exclusion list and the bulk parser gains a 3-field `name|accountId|apiKey` mode; the bulk route now builds a **per-entry** `providerSpecificData` so each key carries its own `accountId` (fixing the previous shared-object reuse), and both the create + key-validation paths receive it. Regression guard: `tests/unit/bulk-api-key-parser-cloudflare.test.ts`. (thanks @muflifadla38)
+- **feat(dashboard):** routing/settings UX clarity ([#6147](https://github.com/diegosouzapw/OmniRoute/issues/6147)) — (1) weighted combos show the **effective routing share %** next to each weight when weights don't sum to 100 (`WeightTotalBar.tsx`); (2) the status widget's user-facing **"Cloud Sync" label is renamed** to "Remote Settings Sync" (`CloudSyncStatus.tsx`; internal ids/state untouched); (3) built-in providers gain an **opt-in advanced base-URL override** (`isBaseUrlOverrideEligibleProvider`, hidden behind an "Advanced" toggle, reusing the existing `providerSpecificData.baseUrl` persistence — not globally widened). Regression guard: `tests/unit/routing-settings-ux-6147.test.ts`.
+- **feat(combo):** add an option to **disable session stickiness**, per-combo or globally — round-robin / random combos can rotate to a different connection on every request instead of pinning a whole conversation to one connection by its first-message hash. Resolution precedence per-combo `config.disableSessionStickiness` → global `settings.disableSessionStickiness` → default `false` (preserves the #3825 prompt-cache/504 fix); gates **both** stickiness call sites in `open-sse/services/combo.ts`. Exposed as a global toggle (Combo Defaults) and a per-combo Inherit/on/off control. ([#6168](https://github.com/diegosouzapw/OmniRoute/issues/6168)) Regression guard: `tests/unit/combo-disable-session-stickiness.test.ts`. (thanks @RCrushMe)
+- **feat(docker):** add the `OMNIROUTE_NO_SUDO` env flag for root-less / user-namespaced deployments — the MITM cert-trust command path (`resolveSudoSpawn` in `src/mitm/systemCommands.ts`) now strips the leading `sudo` when the flag is truthy, in addition to the existing root / sudo-missing cases, so the Proxy Agent runs without `sudo` (the operator trusts the CA manually, e.g. via `NODE_EXTRA_CA_CERTS`). Argv-array `spawn` preserved — no shell interpolation (Hard Rule #13). ([#6122](https://github.com/diegosouzapw/OmniRoute/issues/6122)) Regression guard: `tests/unit/mitm-systemCommands-no-sudo.test.ts`. (thanks @powellnorma)
+- **feat(providers):** add **Requesty** as an OpenAI-compatible gateway provider (BYOK, base `https://router.requesty.ai/v1`, ~200 free requests/day) — wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/requesty/`, `src/shared/constants/providers/apikey/gateways.ts`). ([#6120](https://github.com/diegosouzapw/OmniRoute/issues/6120)) Regression guard: `tests/unit/requesty-provider.test.ts`. (thanks @chirag127)
+- **feat(dashboard):** add **configured-only / available-only filters** to the Free Provider Rankings page ([#6150](https://github.com/diegosouzapw/OmniRoute/issues/6150)) — hide providers you haven't configured, or whose connections are all rate-limited / out of quota, via server-side query params (`?configuredOnly` / `?availableOnly` on `GET /api/free-provider-rankings`) backed by a testable lib helper reusing the in-process connection state (no Redis). Both filters default off, so the default view is unchanged; this supersedes the earlier client-side "Configured Only" toggle (#6245) with an available-only dimension and unit-tested logic. Regression guard: `tests/unit/freeProviderRankings-filters.test.ts`.
+
+### 🔧 Bug Fixes
+
+- **fix(dashboard):** adding a second API key connection for the same provider no longer silently overwrites the first — the Add-API-key modal now derives a unique default connection name (`main`, then `main-2`, `main-3`, …) so the backend name-based upsert can't collide (#6499 — thanks @dilneiss).
+- **fix(compression):** the session-dedup engine now also deduplicates a large multi-line block repeated **within a single message** (intra-message dedup), not just across turns; the compression-preview API surfaces a `fallbackReason`, and the fusion panel reports how many models were rate-limited vs failed on total-panel failure (#6501 — thanks @chirag127).
+- **fix(compression):** a stacked-pipeline step naming an unregistered engine now surfaces a `validationErrors` entry instead of silently no-op'ing, so misconfigured pipelines are visible in the preview API (#6506 — thanks @chirag127).
+- **feat(usage):** add a **Codex reset-credit redemption** flow to the Provider Limits UI ([#6361](https://github.com/diegosouzapw/OmniRoute/pull/6361)) — a `useCodexResetCreditRedemption` hook + `/api/usage/codex-reset-credit` route + `codexResetCredits` lib let you redeem banked Codex reset credits from the quota card. Regression guard: `tests/unit/codex-reset-credits.test.ts`. (thanks @JxnLexn)
+- **feat(glm):** add **team-plan quota settings** for `glm-cn` connections ([#6351](https://github.com/diegosouzapw/OmniRoute/pull/6351)) — a dedicated `GlmTeamQuotaFields` form section (team quota id / limits) threaded through the Add/Edit connection modals, persisted via `providerSpecificData`, with the GLM usage service reading the team quota. Regression guards: `tests/unit/glm-team-quota.test.ts`, `provider-specific-data-schema.test.ts`. (thanks @hao3039032)
+- **feat(providers):** add **TinyFish** web-fetch/search support ([#6349](https://github.com/diegosouzapw/OmniRoute/pull/6349)) — a `tinyfish-fetch` executor + `/v1/web/fetch` route + MCP web-fetch tool, registered as a specialty-media provider with request-validation and a search-provider catalog entry. Regression guards: `tests/unit/executor-tinyfish-fetch.test.ts`, `web-fetch-handler.test.ts`, `mcp-web-fetch-tool.test.ts`, `provider-validation-tinyfish.test.ts`. (thanks @dtybnrj)
+- **fix(cli):** `omniroute launch-codex` now spawns `codex.cmd` through a shell on Windows (the npm `.cmd` shim is unresolvable by bare `spawn` → ENOENT), mirroring the qodercli Windows fix (#6263) ([#6312](https://github.com/diegosouzapw/OmniRoute/pull/6312)). Regression guard: `tests/unit/launch-codex-windows-spawn-6312.test.ts`. (thanks @swingtempo)
+- **fix(codex):** isolate the **Spark** quota from the shared Codex quota and stabilize the quota UI ordering / hydration so per-scope limits render consistently ([#6336](https://github.com/diegosouzapw/OmniRoute/pull/6336)). Regression guards: `tests/unit/codex-quota-selection-hydration.test.ts`, `provider-limits-ui.test.ts` + 3 more. (thanks @xz-dev)
+- **feat(api):** add a `hidePaidModels` setting that filters paid-only models out of the `/v1/models` catalog. Regression guard: `tests/unit/models-catalog-hide-paid.test.ts`. (thanks @chirag127)
+- **fix(api-manager):** the fallback model picker now preserves combos instead of dropping them when a primary model is unavailable ([#6443](https://github.com/diegosouzapw/OmniRoute/pull/6443)). Regression guard: `tests/unit/api-manager-page-static.test.ts`. (thanks @jmengit)
+- **fix(providers):** recoverable Antigravity / Cloud-Code (Gemini Code Assist) `403` responses ([#6452](https://github.com/diegosouzapw/OmniRoute/pull/6452)) are now classified as a retryable project-config error instead of a terminal account ban, so a fixable project/API-disabled 403 no longer forces a ~1-year cooldown / full OAuth reconnect. Regression guard: `tests/unit/errorclassifier-antigravity-403.test.ts`. (thanks @developerjillur)
+- **fix(mitm):** `sanitizeHeaders` now redacts `Set-Cookie` response headers so upstream session cookies never leak into logs / diagnostics ([#6451](https://github.com/diegosouzapw/OmniRoute/pull/6451)). Regression guard: `tests/unit/mitm-sanitize-headers.test.ts`. (thanks @developerjillur)
+- **fix(api):** `/api/compression/preview` now accepts `mode: "caveman"` and correctly handles stacked / zero-compression previews ([#6425](https://github.com/diegosouzapw/OmniRoute/issues/6425)). Regression guard: `tests/unit/api/compression-preview-caveman-and-stacked-6425.test.ts`. (thanks @chirag127)
+- **feat(providers):** add **Zed** hosted LLM aggregator as a native-app provider ([#6118](https://github.com/diegosouzapw/OmniRoute/pull/6118)) — OAuth sign-in via the Zed hosted flow, registered through the shared provider registry + executor. Regression guards: `tests/unit/zed-oauth-provider.test.ts`, `zed-import-utils.test.ts`, `zed-docker-detect.test.ts`, `mitm-handler-zed.test.ts`. VPS-validated via live operator login (Hard Rule #18).
+- **fix(oauth):** the Kiro SSO-cache auto-import now **preserves the IDC region** — cross-region Amazon Q / Kiro profiles imported from the SSO cache are no longer collapsed to the default region ([#6113](https://github.com/diegosouzapw/OmniRoute/pull/6113)). Regression guard: `tests/unit/kiro-auto-import-idc-2059.test.ts`. VPS-validated via live operator login (Hard Rule #18).
+- **fix(dashboard):** passthrough model aliases no longer collide when two namespaced model ids share a last segment (port from 9router#1850, [#6431](https://github.com/diegosouzapw/OmniRoute/pull/6431)). `enx/gpt-5.5` and `enx/codebuddy/gpt-5.5` both auto-generated the alias `gpt-5.5`, so the second model could never be added (the UI just alerted "alias already exists"). Aliases are now disambiguated deterministically — bare last segment when free, then parent-qualified (`codebuddy-gpt-5.5`), then a numeric suffix — while re-adding the exact same model id is still blocked. Regression guard: `tests/unit/passthrough-alias-1850.test.ts`. (thanks @arpicato)
+- **fix(translator):** preserve a Gemini `functionResponse` co-located with other parts (another `functionCall`, or trailing `text`) in the same content when translating **Gemini → OpenAI** ([#6376](https://github.com/diegosouzapw/OmniRoute/pull/6376)). `convertGeminiContent()` early-returned the tool message on the first `functionResponse` part, dropping any co-located parts; such contents are now pre-split (one tool message per `functionResponse`, emitted first, plus one message for the remaining parts). Regression guard: `tests/unit/gemini-to-openai-function-response.test.ts`. (thanks @warelik)
+- **fix(headroom):** detect a python interpreter managed by **mise / pyenv / asdf / conda** (port from 9router#2353, [#6382](https://github.com/diegosouzapw/OmniRoute/pull/6382)). Headroom's python probe (`src/lib/headroom/detect.ts`) searched a hardcoded `PATH`, but version managers expose their interpreters via shim dirs that only join `PATH` through interactive-shell activation — which the non-interactive server never runs, so a managed python (≥3.10) was invisible and Headroom reported it missing. The search path now prepends the well-known shim/bin dirs (`~/.local/share/mise/shims`, `~/.pyenv/shims`, `~/.asdf/shims`, `$CONDA_PREFIX/bin`, `~/.local/bin`, respecting `MISE_DATA_DIR`/`PYENV_ROOT`/`ASDF_DATA_DIR` when set), and a new `HEADROOM_PYTHON` env override lets operators point straight at their interpreter (mirroring `HEADROOM_URL`). Still shell-free (`execFileSync`). Regression guard: `tests/unit/headroom-detect.test.ts` (5). (thanks @loopyd)
+- **fix(executors):** strip the OpenAI-Codex/Claude-CLI `client_metadata` passthrough field for **NVIDIA** requests (port from 9router#1887, [#6411](https://github.com/diegosouzapw/OmniRoute/pull/6411)). NVIDIA's OpenAI-compatible wrapper rejects it with `400 Unsupported parameter`, the same class already handled for `cerebras`/`mistral`; `nvidia` (executor `default`) was missing from the strip allowlist so Codex/Claude-Code passthrough requests 400'd. Regression guard: `tests/unit/executor-default-strip-client-metadata.test.ts` (+nvidia case). (thanks @phidinhmanh)
+- **fix(translator):** strip the Claude-style `thinking` field for **NVIDIA `z-ai/glm-5.2`** (port from 9router#2023, [#6413](https://github.com/diegosouzapw/OmniRoute/pull/6413)). NVIDIA's OpenAI-compatible wrapper 400s on `thinking` (a Claude-format client routed here leaves a `thinking:{type:"adaptive"}`); the existing strip rule only dropped `reasoning`. Same class already handled for `minimax-m2.7`. Regression guard: `tests/unit/nvidia-minimax-thinking-strip.test.ts` (+glm-5.2 case). (thanks @phidinhmanh)
+- **fix(translator):** suppress the streamed `</think>` close marker for the **Antigravity IDE** client (port from 9router#1061, [#6415](https://github.com/diegosouzapw/OmniRoute/pull/6415)). On thinking-only turns Antigravity rendered a bare `</think>` as the sole visible content, tripping its loop-detection and wasting requests. Antigravity's UA (`vscode/<v> (Antigravity/<v>)`) is added to the marker-suppress allowlist (alongside OpenCode); Claude Code / Cursor still get the marker, and `x-omniroute-thinking-marker: on` force-restores it. Regression guard: `tests/unit/think-close-marker-suppress-5245.test.ts`. (thanks @abdofallah)
+- **fix(executors):** strip nested `reasoning_content` from messages for **Mistral** (port from 9router#1649, [#6417](https://github.com/diegosouzapw/OmniRoute/pull/6417)). Mistral's API returns `422 extra_forbidden` when an assistant message carries `reasoning_content` (replayed thinking from a prior turn, e.g. via the Codex `/responses` path); the generic top-level 400 field-downgrade retry never covered the nested per-message field. `DefaultExecutor` now strips it for provider `mistral` only, so DeepSeek (which requires replayed `reasoning_content`) is unaffected. Regression guard: `tests/unit/mistral-strip-reasoning-content-1649.test.ts`. (thanks @xxy9468615)
+- **fix(executors):** strip the `client_metadata` passthrough field on the **OpenCode** path (port from 9router#1442, [#6418](https://github.com/diegosouzapw/OmniRoute/pull/6418)). OpenCode upstreams (e.g. `kimi-k2.6` via opencode-go) reject it with `400 "Extra inputs are not permitted, field: 'client_metadata'"`; the DefaultExecutor strip only covered cerebras/mistral and `OpencodeExecutor` extends `BaseExecutor` directly, so nothing removed it there. Regression guard: `tests/unit/opencode-strip-client-metadata-1442.test.ts`. (thanks @yanpaing007)
+- **fix(executors):** inject the `reasoning_content` echo for the native **Moonshot Kimi** provider (port from 9router#1480, [#6419](https://github.com/diegosouzapw/OmniRoute/pull/6419)). Kimi (executor `default`) is a thinking-mode upstream that 400s with "reasoning_content must be passed back" when a prior assistant turn lacks it; the placeholder injection was only wired into the OpenCode meta-provider, so direct multi-turn Kimi conversations failed. Scoped to `kimi` (gateway-served models matching the thinking-model name pattern are unaffected). Regression guard: `tests/unit/kimi-native-reasoning-injected-1480.test.ts`. (thanks @2220258345)
+- **fix(executors):** recover from a strict gateway's `context_management: Extra inputs are not permitted` 400 (port from 9router#1468, [#6420](https://github.com/diegosouzapw/OmniRoute/pull/6420)). **Claude Code** always sends a top-level `context_management` field; strict anthropic-compatible gateways reject it. The dedicated context-editing 400-fallback only fired when OmniRoute's own `contextEditing` feature was enabled (default off), so a client-sent field passed through untouched and 400'd. `context_management` is now in the generic reactive field-strip list, so it's stripped-and-retried once regardless of the feature flag (with correct request re-signing for claude-compatible relays). Regression guard: `tests/unit/provider-field-strips.test.ts`. (thanks @ohahe52-dot)
+- **fix(network):** enable **RFC 8305 Happy Eyeballs** (`autoSelectFamily`) on the direct-egress undici dispatcher (port from 9router#1237, [#6423](https://github.com/diegosouzapw/OmniRoute/pull/6423)). When DNS returns both IPv6 (AAAA) and IPv4 (A) and the IPv6 route is broken (e.g. a NAT64 `64:ff9b::` prefix without routing), undici tried IPv6 first and hung until `ETIMEDOUT` (then a 502 + account lockout), even though `curl` reached the same host. The direct dispatcher now races both families and uses whichever connects first. Proxy paths pin family via `proxyTls` and are unaffected. Regression guard: `tests/unit/direct-dispatcher-pipelining-4580.test.ts`. (thanks @adentdk)
+- **fix(combo):** round-robin now advances the rotation pointer past the model that **actually served**, not the eagerly-scheduled one (port from 9router#948, [#6428](https://github.com/diegosouzapw/OmniRoute/pull/6428)). With `stickyLimit: 1` (true round-robin), when the scheduled model failed and a _different_ model served via fallback, the counter had already advanced +1 from the scheduled index — so the next request reused the fallback-served model, degrading round-robin into hot-spotting on whichever model was healthy. The pointer now advances to the served index + 1 (mirroring the sticky-limit>1 path). Session-stickiness (#3825) and distribution are preserved. Regression guard: `tests/unit/combo-rr-fallback-advance-948.test.ts`. (thanks @binsarjr)
+- **fix(sse):** a non-string `model` field is now rejected with a `400` before the resolver, instead of crashing downstream `.toLowerCase()`/`.split()` calls into an empty-body `500` that escapes the error sanitizer ([#6407](https://github.com/diegosouzapw/OmniRoute/issues/6407)). Regression guard: `tests/unit/chat-non-string-model-6407.test.ts`. (thanks @chirag127)
+- **fix(api):** unknown `/api/*` routes now return a JSON `404` (instead of the dashboard HTML shell) and scalar chat params (`model`/`temperature`/etc.) are validated **before** the provider lookup so malformed requests fail fast with a clear `400` ([#6424](https://github.com/diegosouzapw/OmniRoute/issues/6424), [#6412](https://github.com/diegosouzapw/OmniRoute/issues/6412)). Regression guards: `tests/unit/api/api-catchall-json-404.test.ts`, `tests/unit/chat-early-schema-validation-6412.test.ts`. (thanks @chirag127)
+- **fix(api):** `/v1/chat/completions` now rejects a non-JSON `Content-Type` with a `400` before parsing the body ([#6414](https://github.com/diegosouzapw/OmniRoute/issues/6414)). Regression guard: `tests/unit/v1-chat-completions-content-type-6414.test.ts`. (thanks @chirag127)
+- **fix(api):** the `X-OmniRoute-Compression` response header is now echoed on `/v1/chat/completions` and `/v1/completions` ([#6422](https://github.com/diegosouzapw/OmniRoute/issues/6422)). Regression guard: `tests/unit/compression-header-echo-6422.test.ts`. (thanks @chirag127)
+- **fix(api):** concurrent `GET /v1/models` requests are coalesced into a single catalog build ([#6408](https://github.com/diegosouzapw/OmniRoute/issues/6408)). Regression guard: `tests/unit/v1-models-concurrent-6408.test.ts`. (thanks @chirag127)
+- **fix(api):** `/v1/completions` now echoes the requested `body.model` in its JSON + streamed responses ([#6429](https://github.com/diegosouzapw/OmniRoute/pull/6429)). Regression guard: `tests/unit/completions-body-model-echo.test.ts`. (thanks @chirag127)
+- **fix(api):** env-var master keys now see the full `/v1/models` catalog ([#6406](https://github.com/diegosouzapw/OmniRoute/issues/6406)). Regression guard: `tests/unit/models-catalog-envkey-6406.test.ts`. (thanks @chirag127)
+- **fix(api):** non-streaming `/v1/completions` responses now echo `body.model` aligned with the `X-OmniRoute-Model` header ([#6426](https://github.com/diegosouzapw/OmniRoute/issues/6426)). Regression guard: `tests/unit/v1-completions-model-header-match-6426.test.ts`. (thanks @chirag127)
+- **fix(api):** unknown `/v1/*` routes now return a JSON `404 not_found` instead of the Next.js dashboard HTML shell ([#6405](https://github.com/diegosouzapw/OmniRoute/issues/6405)). Regression guard: `tests/unit/api/v1-catchall-json-404.test.ts`. (thanks @chirag127)
+- **fix(api):** the per-connection provider models route now degrades to the shipped catalog when a provider's `/models` endpoint answers with a redirect ([#6267](https://github.com/diegosouzapw/OmniRoute/issues/6267)) — a `qwen-web` import failed with a raw `Redirect blocked … (307)` 503. `safeOutboundFetch` throws `REDIRECT_BLOCKED` on the 307, `getSafeOutboundFetchErrorStatus` maps it to 503, and `buildDiscoveryErrorFallbackResponse` treated every 503 as a hard error — so the non-empty `getModelsByProviderId("qwen-web")` catalog was never surfaced. A models-endpoint redirect is not a fixable-config error (unlike `URL_GUARD_BLOCKED`/`INVALID_URL`, which stay hard errors), so it now falls back to the local/cached catalog before the 503 short-circuit. General fix — covers any config-driven provider that 307s. Regression guard: `tests/unit/provider-models-qwen-web-redirect-6267.test.ts`. (thanks @chirag127)
+- **fix(api):** the per-connection provider models route (MCP `list_models_catalog` + the dashboard import view) now merges USER-ADDED custom models into its response ([#6247](https://github.com/diegosouzapw/OmniRoute/issues/6247)) — custom models live in the `key_value` namespace `customModels`, which the live REST `/api/v1/models` already merges, but `src/app/api/providers/[id]/models/route.ts` never read `getCustomModels`, so custom models were dropped on both the discovery-success and local_catalog paths. They are now folded into the returned model list (deduped by id, stamped `owned_by: provider`), fixing MCP + the dashboard import view in one place. Regression guard: `tests/unit/provider-models-custom-merge-6247.test.ts`. (thanks @RCrushMe)
+
+- **fix(providers):** GitLab Duo tool-calling follow-up turns no longer fail upstream with `422 {"detail":"Validation error"}` (tokens 0/0, rejected pre-inference). The #6234 tool-result-feedback fix serialized the **entire** multi-turn conversation into GitLab's single-file `code_suggestions` (`small_file`) generation endpoint — folded history that turn-N sent as an oversized `current_file.content_above_cursor` **and** duplicated verbatim into `user_instruction`, tripping the AI-Gateway's `small_file` validation guard. The executor now **bounds** that prompt: it keeps system + latest user message + the most-recent tool round (dropping older turns), caps oversized tool results, and stops duplicating the full prompt into `user_instruction` (which now carries only the short latest user message) — while still feeding the most-recent tool result back so the agent continues (`open-sse/executors/gitlab.ts`, [#6220](https://github.com/diegosouzapw/OmniRoute/issues/6220)). The unit test covers the bounding logic; the upstream 422→200 clearing is VPS-only (Hard Rule #18). Regression guard: `tests/unit/gitlab-tool-exchange-bounded-6220.test.ts`.
+
+- **fix(i18n):** the provider-detail (`/dashboard/providers/[id]`) connection-status filter labels no longer render as `__MISSING__:All` / `__MISSING__:Active` / `__MISSING__:Error` / `__MISSING__:Banned` / `__MISSING__:CreditsExhausted` in non-English locales (notably pt-BR) ([#6290](https://github.com/diegosouzapw/OmniRoute/issues/6290)). Root cause was **not** the namespace mismatch the issue guessed — the `providers.filter*` keys resolve correctly in `en.json`; the debt lived in the locale mirrors (`src/i18n/messages/*.json`), where these five keys carried the `__MISSING__:` sync sentinel in ~15 locales and were absent entirely in ~26 others, so next-intl found the key and echoed the sentinel verbatim. All 40 non-English/-Chinese mirrors now ship real translations for the five `providers.filter*` labels. Regression guard: `tests/unit/i18n-provider-filter-keys-6290.test.ts`. (thanks @diegosouzapw)
+
+- **fix(providers):** the `copilot-m365-web` streaming executor now emits `debug`-level WebSocket diagnostics ([#6210](https://github.com/diegosouzapw/OmniRoute/issues/6210)) — the outbound WS URL (with the `access_token` **redacted** via `redactWsUrl()`), handshake success/failure, and each received SignalR frame's `type`/`target`. Previously the streaming path logged nothing, so an empty `content:null` response (the M365 Education / Starter tier symptom fixed in #6234) was undiagnosable even at `APP_LOG_LEVEL=debug`. The change is debug-level and side-effect-free — it does not alter streaming behavior or the frame parser, and the token never reaches the logs. Regression guard: `tests/unit/copilot-m365-web-logging-6210.test.ts` (thanks @qpeyba)
+
+- **fix(resilience):** a round-robin combo no longer returns `503 all upstream accounts are unavailable` when a compatibility-rejected target is actually healthy ([#6238](https://github.com/diegosouzapw/OmniRoute/issues/6238)). `filterTargetsByRequestCompatibility` drops request-incompatible targets (tool/vision/structured-output unsupported, or below the required context window) **before** any availability check runs, and its `compatible.length === 0` safety net only fired when _all_ targets were filtered — not when the kept targets later all turned out runtime-unavailable (circuit-open / cooldown / no credentials). So a combo could 503 while a compat-rejected-but-healthy provider sat unused. `handleRoundRobinCombo` now keeps the compat-rejected set and, when every compat-kept target was skipped without a single real attempt, probes those rejected targets as a **last-resort fallback tier** (via the new pure `open-sse/services/combo/comboCompatFallback.ts`) before crystallizing the 503. Regression guard: `tests/unit/combo-roundrobin-compat-fallback-6238.test.ts`. (thanks @ThongAccount)
+
+- **fix(startup):** best-effort self-heal for a corrupted Turbopack dev cache on Windows ([#6289](https://github.com/diegosouzapw/OmniRoute/issues/6289)). On Windows, `pnpm dev` can fail at startup when Turbopack `mmap`s a persistent-cache SST file and the OS refuses the mapping (`os error 1455` — "paging file too small"), which Turbopack surfaces as a misleading `Module not found: Can't resolve '@/shared/utils/machine'`. This is a **known upstream Turbopack cache-corruption bug — not our code**. The dev launcher (`scripts/dev/run-next.mjs`) now wraps `nextApp.prepare()` and, when it rejects with that signature (`isTurbopackCacheCorruption` in the new `scripts/dev/turbopackCacheHeal.mjs`), purges `.build/next/**/cache/turbopack` and retries **once** with a clear log. **Caveat — best-effort only:** the corruption often surfaces as a runtime overlay rather than a `prepare()` rejection, so this cannot always intercept it; the reliable remedy remains manually deleting the Turbopack cache dir. Regression guard: `tests/unit/turbopack-cache-heal-6289.test.ts`. (thanks @chirag127)
+
+- **fix(providers):** qodercli PAT auth no longer fails with `spawn qodercli ENOENT` on Windows ([#6263](https://github.com/diegosouzapw/OmniRoute/issues/6263)) — `spawnQoderCli` spawned the bare `qodercli` name with `shell:false` and an unenriched env, so the npm `.cmd` wrapper under `%APPDATA%\npm` (a user-PATH directory) was never resolved. It now resolves the absolute `.cmd`/`.exe` path through the existing `getCliRuntimeStatus("qoder")` resolver in `src/shared/services/cliRuntime.ts` (memoized), spawns with `shell` when the target is a `.cmd`/`.bat`, and uses the cliRuntime-enriched env (PATH + PATHEXT + APPDATA); the ENOENT error now lists the searched paths plus the `CLI_QODER_BIN` override. End-to-end spawn on a real Windows host is host-only (Hard Rule #18); the path-resolution logic is unit-tested. Regression guard: `tests/unit/qodercli-windows-resolve-6263.test.ts`. (thanks @chirag127)
+
+- **fix(sse):** the reasoning-token buffer no longer inflates **probe-sized `max_tokens`** ([#6274](https://github.com/diegosouzapw/OmniRoute/issues/6274)) — Claude Code's `/model` capability check sends `max_tokens: 1`, but for a thinking-capable model with a large output cap (e.g. `glm-5.2`) the #3587 headroom heuristic (`max(current + 1000, ceil(current * 1.5))`) rewrote it to `1001` and forwarded that upstream, wasting tokens on a request that was never a genuine reasoning budget. `resolveReasoningBufferedMaxTokens()` (`open-sse/services/reasoningTokenBuffer.ts`) now short-circuits and returns the caller's value verbatim when it is below the new `REASONING_BUFFER_MIN_TRIGGER` (256) threshold — a tiny explicit limit is a probe, not a reasoning request. Real budgets still receive the #3587 headroom unchanged, and the guard runs after the existing capability checks so unknown / non-reasoning models keep returning `null`. Regression guard: `tests/unit/reasoning-token-buffer-6274.test.ts`. (thanks @brightfiscalband)
+
+- **fix(cli):** `omniroute reset-password` now works as a real subcommand, and password resets over piped (non-TTY) stdin actually apply ([#6261](https://github.com/diegosouzapw/OmniRoute/issues/6261), [#6258](https://github.com/diegosouzapw/OmniRoute/issues/6258)). Two coupled defects: (1) **#6261** — `bin/omniroute.mjs` routed everything through Commander with only two pre-Commander bypasses (`--mcp`, `reset-encrypted-columns`), so `omniroute reset-password` was rejected as an unknown command; only the separate `omniroute-reset-password` bin worked, while the docs falsely advertised the subcommand (incl. a bogus "legacy alias still works"). A pre-Commander bypass mirroring `reset-encrypted-columns` now dynamically imports `bin/reset-password.mjs` (which self-executes) before Commander parses; the three doc lines were corrected. (2) **#6258** — `bin/reset-password.mjs` issued two sequential `rl.question` prompts; under piped stdin the second read never settled at EOF, so `main()` never reached `resetManagementPassword` and the reset was a silent no-op (both prompts printed, no success, password unchanged). The CLI now detects non-TTY stdin and reads it once (first line = password, second line = confirm if present, else reused), adds a `--password-stdin` flag (entire stdin is the password, no confirmation), and exits `0` explicitly so the success line always flushes; interactive TTY behavior is unchanged. Regression guard: `tests/unit/reset-password-cli-6261-6258.test.ts` (3). (thanks @chirag127)
+
+- **fix(db):** the mass-migration **safety abort** now tells the operator how to bypass it and stops flooding the log ([#6260](https://github.com/diegosouzapw/OmniRoute/issues/6260)) — after restoring a backup that wiped the migration tracking table, `runMigrations()` threw the abort on every downstream `ensureDbInitialized()`, re-logging the full banner 11+ times, and the message never mentioned the existing `OMNIROUTE_MAX_PENDING_MIGRATIONS` escape hatch. The abort text now appends a bypass hint (set `OMNIROUTE_MAX_PENDING_MIGRATIONS=0` in `server.env` / `DATA_DIR/.env`), and a new `MigrationSafetyAbortError` is memoized so repeated calls in the same process throw the same instance and emit a single concise line instead of the full cascade. Regression guard: `tests/unit/migration-safety-abort-6260.test.ts`. (thanks @chirag127)
+
+- **fix(auth):** importing a **distinct** Codex/ChatGPT OAuth `auth.json` is no longer falsely rejected as "already exists" when it belongs to a different user in the same workspace ([#6301](https://github.com/diegosouzapw/OmniRoute/issues/6301)). `findExistingCodexConnection` (in `src/lib/oauth/utils/codexAuthImport.ts`) deduped **only** on `providerSpecificData.workspaceId === accountId`, where `accountId` is the shared `chatgpt_account_id`/`tokens.account_id` — so two members of the same ChatGPT Team collapsed onto a single connection (409 `duplicate_account`). The id_token's `https://api.openai.com/auth` claim carries a per-user `chatgpt_user_id` alongside the workspace id (the device-flow path already persisted it as `chatgptUserId`, but the import path did not). Now `parseAndValidateCodexAuth` extracts `userId` (`chatgpt_user_id` → `user_id` → JWT `sub`) into `ParsedCodexAuth`, the create/update paths persist `chatgptUserId` in `providerSpecificData` (mirroring `codex.ts`), and dedup keys on `workspaceId` **AND** `chatgptUserId` — with a backward-compat fallback to legacy accountId-only matching when no stored connection for that workspace records a `chatgptUserId`, so genuinely-same accounts still dedup. Regression guard: `tests/unit/codex-auth-import-userid-dedup-6301.test.ts` (4). (thanks @anungma)
+
+- **fix(providers):** importing models for the **venice-web** provider no longer fails with a red "Provider venice-web does not support models listing" ([#6269](https://github.com/diegosouzapw/OmniRoute/issues/6269)). `venice-web` is a web-cookie provider with an executor but no upstream `/v1/models` endpoint and no registry `models`, so the models route fell through to the tail `400`. Mirroring the `jules`/`linkup-search`/`ollama-search` fix (#5569), it now ships a static local catalog entry in `src/lib/providers/staticModels.ts` — seeding the current Venice lineup (`venice-uncensored`, `llama-3.3-70b`, `qwen3-235b`, `qwen3-4b`, `deepseek-r1-671b`; Venice rotates its catalog, see docs.venice.ai/models/overview) — so the route returns `200` with `source:"local_catalog"`, `intentional:true`. Regression guard: `tests/unit/static-models-venice-web-6269.test.ts`. (thanks @chirag127)
+
+- **fix(api):** the specialty model catalogs (`/v1/embeddings`, `/v1/images`, `/v1/music`, `/v1/videos` model lists) are now derived from the **unified catalog filtered by a predicate** (`getSpecialtyModelsResponse`) instead of ad-hoc per-route logic, so they consistently respect active-credential visibility and stay in sync with the main catalog ([#6303](https://github.com/diegosouzapw/OmniRoute/pull/6303)). Regression guard: `tests/unit/specialty-model-catalog-routes.test.ts`. (thanks @makcimbx)
+- **fix(api):** the agent-bridge server route now resolves the MITM manager via a **dynamic `import("@/mitm/manager.runtime")`** so Turbopack does not statically pull the stub (or over-bundle the manager), and the agent-skills generator anchors its output base path with `path.join(process.cwd(), …)` so Turbopack's static analyzer stops tracing the whole project root ([#6329](https://github.com/diegosouzapw/OmniRoute/issues/6329), [#6366](https://github.com/diegosouzapw/OmniRoute/pull/6366)). Regression guard: `tests/unit/agent-bridge-server-route-dynamic-import.test.ts`. (thanks @Iammilansoni)
+- **fix(api):** internal probes (combo-test, cloud-sync verify) now pick a **management-scoped / allow-all API key** instead of naively grabbing `getApiKeys()[0]` — a restricted `self:usage` first row made the probe fail with "Model X is not allowed for this API key" even when the combo path was healthy (`pickApiKeyForInternalUse` in `src/lib/db/apiKeys.ts`). The API-manager model editor also falls back to `/api/models?all=true` when `/v1/models` is catalog-protected ([#6372](https://github.com/diegosouzapw/OmniRoute/pull/6372)). Regression guard: `tests/unit/pick-internal-api-key-6372.test.ts`. (thanks @jmengit)
+- **fix(live-ws):** the Live Dashboard WebSocket server now **rejects on bind failure** (e.g. `EADDRINUSE` when the API bridge already holds the port) instead of letting the error surface as an unhandled `error` event that crash-loops the process — the `error` listener is attached to `wss` (not `server`) and releases the EventBus subscription on a failed start ([#6324](https://github.com/diegosouzapw/OmniRoute/issues/6324)). Regression guard: `tests/unit/live-ws-eaddrinuse-6324.test.ts`. (thanks @vinayakkulkarni)
+- **fix(dashboard):** the Home provider-topology widget now trusts the live provider-metrics snapshot — it uses `topology.errorProvider` and live `activeRequests` directly instead of re-deriving state from a stale `lastErrorAt` or applying a frontend timeout filter, so the topology reflects real-time provider health ([#6322](https://github.com/diegosouzapw/OmniRoute/pull/6322)). Regression guard: `tests/unit/home-provider-topology-live-state.test.ts`. (thanks @xz-dev)
+- **fix(sse):** strip zero-width markers from streamed **tool-call arguments** — a follow-up to [#5857](https://github.com/diegosouzapw/OmniRoute/pull/5857). That PR removed injected zero-width joiners (U+200D) from streamed assistant text/reasoning but deliberately left tool-call argument JSON byte-exact. The request-side obfuscation (`open-sse/services/claudeCodeObfuscation.ts`) injects ZWJ into agent words — including the temp path inside the Bash tool description — and Claude models copy that verbatim into generated commands, which are delivered as tool-call arguments rather than assistant text. As a result the ZWJ survived and corrupted code blocks (e.g. a temp path rendered with an invisible joiner). Now `open-sse/handlers/responseSanitizer.ts` strips zero-width code points from tool-call argument strings at every emit site (OpenAI non-stream/stream chat `tool_calls` + legacy `function_call`, native Responses `function_call` items, the OpenAI→Responses conversion, and the native Responses streaming `response.function_call_arguments.delta/.done` events). Only zero-width code points are removed; JSON structure and all other bytes stay identical (no parse/restringify), so normal arguments remain byte-exact. Regression guard: 6 new cases in `tests/unit/response-sanitizer.test.ts` (suite 50/50).
+
+- **fix(nodejs):** the default app log path now resolves under `DATA_DIR` (`~/.omniroute/logs/application/app.log`) instead of `process.cwd()` ([#6197](https://github.com/diegosouzapw/OmniRoute/issues/6197)) — the globally-installed CLI runs from an arbitrary working directory, so anchoring the default to cwd made file logging silently write to (or no-op under) an unrelated directory, contradicting the documented `.env.example` default. `getAppLogFilePath()` now computes the default lazily via the pure `resolveDataDir()` resolver (honours a per-process `DATA_DIR`, no directory-creation side effect); an explicit `APP_LOG_FILE_PATH` still wins. Regression guard: `tests/unit/logenv-datadir-path-6197.test.ts` (3).
+
+- **fix(docker):** AgentBridge/`startMitm` no longer aborts in containers/headless when the Antigravity-default DNS step can't write `/etc/hosts` ([#6127](https://github.com/diegosouzapw/OmniRoute/issues/6127)), and the privileged command's stderr now reaches `app.log` instead of only a bare exit code hitting the toast ([#6198](https://github.com/diegosouzapw/OmniRoute/issues/6198)). The default DNS step (`addDNSEntry`) was called unguarded while cert install and the two sibling DNS steps were each best-effort — in the runtime Docker image (`USER node`, no `sudo`, read-only `/etc/hosts`) it threw `Command failed with code 1` out of `startMitmInternal` and killed the whole start, discarding the stderr. The three DNS steps are extracted into a best-effort `provisionDnsEntries()` where each failure is logged with the full `err` (stderr included, folded in by `systemCommands.ts`) and never aborts the start. Regression guard: `tests/unit/mitm-dns-graceful-degrade-6127.test.ts` (4).
+
+- **fix(providers):** copilot-m365-web now supports the M365 Education "Starter / OfficeWebIncludedCopilot" tier and no longer returns an empty `content:null` stream ([#6210](https://github.com/diegosouzapw/OmniRoute/issues/6210)). Two gaps: (1) `buildWsUrl()` hardcoded the individual-consumer scenario (`OfficeWebPaidConsumerCopilot`, `isEdu=false`) — the EDU tier is now opt-in via `providerSpecificData.tier="edu"`, emitting `scenario=OfficeWebIncludedCopilot`/`isEdu=true` (the individual path is unchanged); (2) the EDU/GPT-5.5 path streams deltas via `arguments[0].writeAtCursor` (incremental) instead of only `messages[].text` (accumulated snapshots), which the parser dropped — a new `accumulateBotContent()` folds both formats, with `type:2 item.result.message` as a last-resort fallback. Regression guard: `tests/unit/copilot-m365-edu-writeatcursor-6210.test.ts` (10). (thanks @qpeyba)
+
+- **fix(providers):** GitLab Duo executor now feeds tool results back into the prompt instead of looping ([#6220](https://github.com/diegosouzapw/OmniRoute/issues/6220)) — `buildPrompt()` branched only on `system`/`user` and took `userParts.at(-1)`, silently dropping the `assistant{tool_calls}` + `tool{result}` turns the client appended, so the reconstructed prompt was byte-identical to turn 1 and the model re-emitted the same `<tool>` call forever. When a tool exchange is present the full conversation is now serialized, folding each tool result back keyed by its `tool_call_id`; simple conversations keep the legacy shape. Complements the tool_call emission from [#6051](https://github.com/diegosouzapw/OmniRoute/issues/6051) (the `kilo-duplicate` label was a false positive — different, sequential defect). Regression guard: `tests/unit/gitlab-tool-result-feedback-6220.test.ts` (4).
+
+- **fix(providers):** opencode-go/opencode-zen can now synthesize the OpenCode CLI identity headers Cloudflare requires on VPS egress ([#5997](https://github.com/diegosouzapw/OmniRoute/issues/5997)) — on a datacenter VPS, `opencode.ai/zen/go/v1/chat/completions` 403s (HTML challenge) requests lacking CLI identity, while the reporter's control curl proved that `User-Agent: opencode-cli/1.0.0` + `x-opencode-client: cli` + `x-opencode-project: default` + fresh request/session UUIDs succeed. Opt-in via `OPENCODE_SYNTHESIZE_CLI_HEADERS=true` (values overridable via `OPENCODE_GO_USER_AGENT`/`OPENCODE_USER_AGENT`/`OPENCODE_CLIENT`/`OPENCODE_PROJECT`); it fills only headers the client did not already send. Kept **off by default** — the forward-only path is deliberate (fabricating a wrong value risks upstream rejection; a prior dedup regressed with `opencode/local`), so this replaces the fragile local header-injection shim without changing default behavior. Regression guard: `tests/unit/opencode-cli-headers-synthesis-5997.test.ts` (6). (thanks @aleksesipenko)
+- fix(resilience): sticky session affinity now evicts and fails over to another account when the pinned account is exhausted/unavailable (#6219)
+
+- fix(sse): Responses API passthrough now drops internal commentary-phase output before forwarding to clients (gated by RESPONSES_PASSTHROUGH_DROP_COMMENTARY, default on) (#6199)
+
+- **fix(sse):** tool-call function schemas with a root `type: null` are now coerced to `type: "object"` before dispatch ([#6359](https://github.com/diegosouzapw/OmniRoute/issues/6359)) — clients like the Codex app emit `parameters: { type: null, ... }` for some tools, which OpenAI-compatible upstreams reject with `400 Invalid schema for function '...': schema must be a JSON Schema of 'type: "object"', got 'type: null'`, failing the whole request. `toolSchemaSanitizer` already stripped the null; it now re-adds the mandatory root `"object"` type (and empty `properties`/open `additionalProperties` when absent). Combinator roots (`anyOf`/`oneOf`/`allOf`) and explicit root types are left untouched. Regression guard: 5 new cases in `tests/unit/tool-schema-sanitizer.test.mjs`.
+
+- **fix(docker):** AgentBridge no longer fails to start on npm/Electron/VPS installs with "MITM manager stub reached at runtime" ([#6344](https://github.com/diegosouzapw/OmniRoute/issues/6344)) — v3.8.45 flipped the production bundler default to Turbopack, but `next.config.mjs` aliased `@/mitm/manager` to its Docker-only degraded stub **unconditionally**. That was harmless while Docker (which sets the alias intentionally for #3390 graceful degradation) was the sole Turbopack consumer, but once every artifact built with Turbopack the stub shipped to all non-Docker users and `startMitm` threw on the first Agent-Bridge start. The alias is now opt-in via `OMNIROUTE_MITM_STUB=1` (set only by the Dockerfile) through the shared `scripts/build/mitm-stub-flag.mjs` helper; default builds bundle the real manager. Regression guard: `tests/unit/mitm-stub-alias-6344.test.mjs` (4).
+
+- **fix(proxy):** stop the v3.8.44 proxy regression that leaked the real IP and disabled healthy proxies ([#6246](https://github.com/diegosouzapw/OmniRoute/issues/6246)). Two coupled defects from the new health scheduler: (1) **IP leak** — when a proxy assigned to a connection was marked `inactive`, resolution fell through to a **direct** egress instead of blocking, exposing the operator's real IP; (2) **over-deactivation** — the sweep flipped a proxy to `inactive` on the **first** failed probe and counted our own 5s timeout / a probe-target `5xx` as the proxy's fault, so healthy paid proxies vanished from egress selection ("my proxies are not being used anymore"). Fix: the sweep decision is extracted into a pure, network-free `decideProxyHealthAction` (`src/lib/proxyHealth/decision.ts`) — by default the health check now **only counts/logs and never downgrades status** (a proxy is downgraded/removed only with `PROXY_AUTO_REMOVE=true`, after `PROXY_AUTO_REMOVE_AFTER` **consecutive** conclusive failures); probes are classified tri-state so an inconclusive result (our timeout, or a `5xx` from the probe target) never penalizes the proxy, and the probe timeout is raised 5s→15s. Separately, `safeResolveProxy` now **fails closed** via the existing policy: a connection whose assigned proxy is dead is blocked instead of leaking direct (`hasBlockingProxyAssignment`), honoring the explicit `proxy off` toggles and the `PROXY_FAIL_OPEN=true` opt-out. Existing proxies stuck `inactive` by the old behavior need a one-time manual re-activate (the operator owns proxy status). Regression guards: `tests/unit/proxy-health-decide-action-6246.test.ts`, `tests/unit/proxy-assigned-unavailable-6246.test.ts`.
+
+- **fix(proxy):** make "Test All" read-only and add bulk enable/disable ([#6246](https://github.com/diegosouzapw/OmniRoute/issues/6246)). Complements the core fail-closed / scheduler fix (#6296) with the two remaining reporter asks. (1) The **"Test All" button** (`POST /api/settings/proxies/auto-test`) used to flip a proxy to `inactive` on a failed reachability probe; since the egress selector excludes `inactive` proxies, a flaky probe (an unreachable `httpbin.org`, a proxy that blocks `HEAD`, or a slow paid proxy) silently disabled every proxy that failed — "Test All" is now **read-only by default** (only the operator sets a proxy active/inactive; opt back into the legacy test-and-set with `PROXY_HEALTH_AUTO_DEACTIVATE=true`). (2) Adds a **bulk enable/disable** proxies endpoint + toolbar action (`POST /api/settings/proxies/batch-activate`) so an operator can re-activate proxies in one click. Regression guard: `tests/unit/proxy-health-6246.test.ts`. (thanks @tenshiak)
+
+- **chatcore (tools): stop the default 128-tool cap from silently dropping opencode's `task`/MCP tools.** opencode (used as an MCP/agent host) sends a large tool list; when it exceeds the speculative `MAX_TOOLS_LIMIT` (128) default, `truncateToolList` did a blind `tools.slice(0, 128)`, dropping every tool past index 128 — including opencode's built-in `task` tool (subagent launch) and many MCP tools, so models routed through OmniRoute could no longer spawn subagents or reach part of their tools. The cap exists to avoid upstream `400`s for providers with real hard limits (e.g. grok-cli 200), so it is kept for those: detection of the opencode client (`isOpencodeClient` — any `x-opencode-*` header, or `opencode` in the user-agent) now only bypasses the **speculative 128 default**, never a known provider ceiling. Precedence is explicit — a proactive/detected provider limit always truncates (even for opencode); otherwise opencode forwards its full tool list; otherwise the unchanged 128 default applies to every other client. Refactors `getEffectiveToolLimit` into `getKnownToolLimit(provider) ?? DEFAULT_LIMIT` (byte-identical for existing callers) and fixes a cosmetic debug-log that reported the truncated count instead of the original. Regression guard: `tests/unit/tool-limit-detector.test.ts`.
+
+- **fix(mitm):** the macOS MITM-cert install check now matches the system keychain again. `security find-certificate -a -Z` prints the SHA-1 as a colon-less hex string, but the installed-check compared it against `getCertFingerprint()`'s colon-separated form, so the substring match never hit — the cert was reported as not-installed and re-prompted for the sudo install on every run. Fingerprints are now normalized (colons stripped, upper-cased) on both sides via the extracted `macCertOutputHasFingerprint` helper. Regression guard: `tests/unit/mitm-cert-mac-fingerprint.test.ts`. ([#6204](https://github.com/diegosouzapw/OmniRoute/pull/6204), closes [#6134](https://github.com/diegosouzapw/OmniRoute/issues/6134) — thanks @rianonehub)
+
+- **fix(api):** `/v1/messages/count_tokens` now counts `tool_use`, `tool_result` and `thinking` content blocks (and array-form `system` prompts) in the local-estimation path, instead of only `text`. Real agentic conversations keep ~95% of their tokens inside tool results; the previous estimate returned near-zero for them, which silently broke Claude Code's auto-compaction (context grew past the window with no compaction until the upstream API rejected the request). The real provider-side count path is unchanged. Regression guard: `tests/unit/messages-count-tokens-route.test.ts`. ([#6221](https://github.com/diegosouzapw/OmniRoute/pull/6221) — thanks @luweiCN)
+
+- **fix(antigravity):** strip a trailing assistant prefill turn for Vertex Claude models to avoid upstream 400s ([#6114](https://github.com/diegosouzapw/OmniRoute/pull/6114)). Regression guard: `tests/unit/antigravity-claude-prefill-strip.test.ts`. (thanks @anki1kr)
+
+- **fix(security):** the mutable cloud-agent routes (`/api/cloud/credentials/update`, `/api/cloud/models/alias`) now require management auth instead of being treated as public. They were classified as public API routes, so a request without management credentials could update stored cloud-agent credentials and model aliases. They are removed from the public-route set, classified as management routes in the authz pipeline, and gated by `requireManagementAuth`; cloud **read**/auth routes stay public. Regression guards: `tests/unit/cloud-write-auth.test.ts`, `tests/unit/authz/classify.test.ts`, `tests/unit/public-api-routes.test.ts`. ([#6233](https://github.com/diegosouzapw/OmniRoute/pull/6233) — thanks @vittoroliveira-dev)
+
+- **refactor(dashboard):** extract the onboarding-wizard "Open provider details" link target into a pure, unit-tested `buildProviderDetailsHref(connection)` helper. The wizard already routes by `connection.id` (the node UUID) rather than the provider category slug (#6144/#6145); this hardens that behavior behind a tested helper that guards a missing id/connection. Regression guard: `tests/unit/provider-onboarding-href.test.ts`. ([#6166](https://github.com/diegosouzapw/OmniRoute/pull/6166) — thanks @KooshaPari)
+
+- **fix(security):** the doubao synthetic device-id generator now derives its digits via an unbiased crypto-random draw (rejection sampling over `crypto.randomBytes()`) instead of a `% 10` reduction, closing a CodeQL `js/biased-cryptographic-random` finding.
+- **fix(agentSkills):** the GitHub-skills generator now resolves `outputDir` to an absolute path before writing, fixing a regression introduced by #6366 (relative-to-cwd base path) that could write generated skill files to the wrong directory.
+- **fix(security):** `/api/keys/{id}/devices` now checks the HTTP method before auth/validation, returning a `405` for non-GET/DELETE verbs instead of a misleading `401`/`500` (closes a `dast-smoke` QUERY-method finding).
+- **fix(quality):** clear the last 2 heavy quality-gate reds on the release tip (cycle pre-flight).
+- **fix(mitm):** the test suite and CI can never mutate the OS trust store — `OMNIROUTE_SKIP_SYSTEM_TRUST=1` is set globally for tests/CI so `installCert`/`uninstallCert`/`installTproxyCa` skip the privileged OS dispatch ([#6310](https://github.com/diegosouzapw/OmniRoute/pull/6310); full detail is under the [3.8.45] section below — this branch received it via the parallel-cycle sync-back).
+- **fix(api):** `POST /api/github-skills` now Zod-validates its request body; documented the new quality-gate env vars and pinned the merge-integrity GitHub Actions to a commit SHA.
+- **fix(skills):** generate the missing `omni-github-skills` registry entry and align the agent-skills catalog-count tests (follow-up to #6186).
+- **fix(quality):** clear the cycle's 11 net-new ESLint errors and make `validate-release-green` suppressions-aware.
+
+### 📝 Maintenance
+
+- **i18n(it):** add 118 missing Italian (`it`) translations (net-additive — no existing keys dropped, valid JSON), improving Italian UI coverage. ([#6212](https://github.com/diegosouzapw/OmniRoute/pull/6212) — thanks @serverless83)
+- **chore(providers):** remove deprecated **MiMo V2** model entries from the catalogs (xiaomi-mimo, opencode-go, zenmux-free, audio TTS) — the upstream V2 line is superseded by MiMo V2.5; drops `mimo-v2-tts`, `mimo-v2-pro`, `mimo-v2-omni`, `mimo-v2-flash`, `mimo-v2-flash-free` and realigns the provider-catalog tests. ([#6248](https://github.com/diegosouzapw/OmniRoute/pull/6248) — thanks @backryun)
+- **chore(release):** ~50 commits on this branch are v3.8.45 pre-flight/hardening fixes and CI-perf work that landed here via the parallel-cycle sync-back (`sync-next-cycle.mjs`, Hard Rule #21) **after** the `v3.8.45` git tag was cut, and are already fully documented under the **[3.8.45]** section below — listed here only so the per-cycle commit-coverage check (`npm run release:uncovered`) doesn't flag them as gaps. Provider/catalog/UX/backend fixes: #6041, #6078, #6108, #6135, #6148, #6149, #6154, #6158, #6161, #6162, #6163, #6164, #6165, #6170, #6177, #6178, #6181, #6186, #6187, #6191, #6193, #6194, #6195, #6200, #6205, #6208, #6209, #6211, #6213, #6223, #6224, #6225, #6226, #6227, #6228, #6229, #6230, #6235, #6291, #6292. CI/release-pipeline work: #6167, #6203, #6214, #6215, #6218, #6273, #6275, #6283, #6284, #6285, #6300, #6305.
+- **chore(release):** additional zero-ref release-cycle plumbing on this branch, kept out of `release:uncovered` on purpose (no `#N` in the commit subject to cite): opening the v3.8.46 cycle, opening/closing the v3.8.45 cycle, the finalized [3.8.45] CHANGELOG i18n sync-back to 42 mirrors, the v3.8.45 cognitive/cyclomatic and file-size drift rebaselines, ESLint stale-suppression pruning (4,273 → 4,233), and clearing test-masking/docs-all pre-flight reds for v3.8.45.
+
+### ⚡ Performance & Infrastructure
+
+- **perf(release-green):** the pre-flight validator (`scripts/quality/validate-release-green.mjs`) now runs its 4 slow suites (unit / vitest / integration / pack-artifact) **concurrently** via `Promise.all` — pre-flight wall time drops from ~the sum of the suites to ~the slowest one (~30min saved per round; Phase 0 was the nº1 bottleneck of the v3.8.45 release benchmark, 2h54 of 6h34 e2e). Guard: `tests/unit/validate-release-green.test.ts` ("runs the slow suites CONCURRENTLY"). ([#6319](https://github.com/diegosouzapw/OmniRoute/pull/6319))
+- **fix(ci):** `scripts/release/sync-next-cycle.mjs` — two defects found live in its first production run (v3.8.45 Phase 5): (1) the `git()` helper's default 1 MiB `maxBuffer` crashed with `ENOBUFS` on `git show origin/main:CHANGELOG.md` (the CHANGELOG alone is >1 MiB) — widened to 64 MiB; (2) the i18n resync only propagated the `[NEXT]` (TBD) section, leaving the just-shipped finalized section as "— TBD" in all 42 mirrors — it now also syncs `[prevVersion]` bounded by the heading below it (new exported pure helper `versionAfter`). Guards: +5 tests in `tests/unit/sync-next-cycle.test.ts` (8/8). ([#6327](https://github.com/diegosouzapw/OmniRoute/pull/6327))
+- **test(ci):** concurrency-sensitive flaky tests are **quarantined into a serial pass** (`tests/unit/serial/`, `--test-concurrency=1`, appended to every unit runner incl. sharded variants — the serial pass is sharded too so concurrent shard jobs never self-collide). Initial set: `glm-coding-plan-monthly-3580`, `quota-division-blocks`, `provider-health-autopilot`, `combo-health-autopilot` — the class behind the ~28min CI wedges/re-runs (two live 1h+ wedges cancelled during this PR's own validation). Discovery + TIA gates track the new glob; systemic root cause (async logger writing after teardown) tracked in [#6360](https://github.com/diegosouzapw/OmniRoute/issues/6360). Guard: `tests/unit/test-serial-quarantine.test.ts` (4). ([#6347](https://github.com/diegosouzapw/OmniRoute/pull/6347))
+
+### 🙌 Contributors
+
+Thanks to everyone whose work landed in v3.8.46:
+
+| Contributor                                                    | PRs / Issues           |
+| -------------------------------------------------------------- | ---------------------- |
+| [@2220258345](https://github.com/2220258345)                   | direct commit / report |
+| [@abdofallah](https://github.com/abdofallah)                   | direct commit / report |
+| [@adentdk](https://github.com/adentdk)                         | direct commit / report |
+| [@aleksesipenko](https://github.com/aleksesipenko)             | direct commit / report |
+| [@anki1kr](https://github.com/anki1kr)                         | direct commit / report |
+| [@anungma](https://github.com/anungma)                         | direct commit / report |
+| [@arpicato](https://github.com/arpicato)                       | direct commit / report |
+| [@backryun](https://github.com/backryun)                       | #6248                  |
+| [@binsarjr](https://github.com/binsarjr)                       | direct commit / report |
+| [@brightfiscalband](https://github.com/brightfiscalband)       | direct commit / report |
+| [@chirag127](https://github.com/chirag127)                     | #6501, #6506           |
+| [@developerjillur](https://github.com/developerjillur)         | direct commit / report |
+| [@dilneiss](https://github.com/dilneiss)                       | #6499                  |
+| [@dtybnrj](https://github.com/dtybnrj)                         | direct commit / report |
+| [@Forcerecon](https://github.com/Forcerecon)                   | direct commit / report |
+| [@hao3039032](https://github.com/hao3039032)                   | direct commit / report |
+| [@Iammilansoni](https://github.com/Iammilansoni)               | #6150, #6245           |
+| [@jmengit](https://github.com/jmengit)                         | direct commit / report |
+| [@jordansilly77-stack](https://github.com/jordansilly77-stack) | direct commit / report |
+| [@JxnLexn](https://github.com/JxnLexn)                         | direct commit / report |
+| [@KooshaPari](https://github.com/KooshaPari)                   | #6166                  |
+| [@loopyd](https://github.com/loopyd)                           | direct commit / report |
+| [@luweiCN](https://github.com/luweiCN)                         | #6221                  |
+| [@makcimbx](https://github.com/makcimbx)                       | direct commit / report |
+| [@muflifadla38](https://github.com/muflifadla38)               | direct commit / report |
+| [@newnol](https://github.com/newnol)                           | direct commit / report |
+| [@ofekbetzalel](https://github.com/ofekbetzalel)               | direct commit / report |
+| [@ohahe52-dot](https://github.com/ohahe52-dot)                 | direct commit / report |
+| [@phidinhmanh](https://github.com/phidinhmanh)                 | direct commit / report |
+| [@powellnorma](https://github.com/powellnorma)                 | direct commit / report |
+| [@qpeyba](https://github.com/qpeyba)                           | direct commit / report |
+| [@RaviTharuma](https://github.com/RaviTharuma)                 | direct commit / report |
+| [@RCrushMe](https://github.com/RCrushMe)                       | direct commit / report |
+| [@rianonehub](https://github.com/rianonehub)                   | #6134, #6204           |
+| [@serverless83](https://github.com/serverless83)               | #6212                  |
+| [@swingtempo](https://github.com/swingtempo)                   | direct commit / report |
+| [@tenshiak](https://github.com/tenshiak)                       | direct commit / report |
+| [@ThongAccount](https://github.com/ThongAccount)               | direct commit / report |
+| [@UnrealAryan](https://github.com/UnrealAryan)                 | direct commit / report |
+| [@vinayakkulkarni](https://github.com/vinayakkulkarni)         | direct commit / report |
+| [@vittoroliveira-dev](https://github.com/vittoroliveira-dev)   | #6233                  |
+| [@warelik](https://github.com/warelik)                         | direct commit / report |
+| [@xxy9468615](https://github.com/xxy9468615)                   | direct commit / report |
+| [@xz-dev](https://github.com/xz-dev)                           | direct commit / report |
+| [@yanpaing007](https://github.com/yanpaing007)                 | direct commit / report |
+| [@diegosouzapw](https://github.com/diegosouzapw)               | maintainer             |
+
+---
+
 ## [3.8.45] — 2026-07-06
 
 ### ✨ New Features

docs/i18n/ta/CHANGELOG.md

@@ -6,6 +6,213 @@
 
 ## [3.8.31] — 2026-06-20
 
+## [3.8.46] — TBD
+
+### ✨ New Features
+
+- **feat(sse):** **hide paid-only models from `auto/*` routing** when `hidePaidModels` is on ([#6512](https://github.com/diegosouzapw/OmniRoute/issues/6512)) — follow-up to #6328/#6495. PR #6495 hid paid-only models from the `GET /v1/models` listing, but `auto/*` combos (`auto/best-coding`, `auto/glm`, …) could still pick a paid-only backend into their candidate pool → a 402/403 at request time. `createVirtualAutoCombo` now filters the candidate pool through the new pure `open-sse/services/autoCombo/paidModelFilter.ts` (`filterPaidOnlyCandidates`), applying the same free-model predicate #6495 uses in `catalog.ts` (`providerHasFreeModels(provider) && isFreeModel(provider, {id})`) whenever `settings.hidePaidModels === true`. Applied before the category/tier/family narrowing, so it covers every `auto/*` combo; an all-paid pool degrades to the existing graceful empty-pool path. **Opt-in — default OFF leaves the pool unchanged** (identity). Regression guard: `tests/unit/autoCombo/paid-model-filter-6512.test.ts` (4, incl. the default-off identity guard).
+- **feat(sse):** **provider-family auto combos** — `auto/glm`, `auto/minimax`, `auto/mimo`, `auto/zai`, `auto/gemma`, `auto/llama`, `auto/gemini` ([#6453](https://github.com/diegosouzapw/OmniRoute/issues/6453)) — new routable ids that materialize an on-demand virtual combo spanning whatever installed backends currently expose that model family, degrading gracefully as backends rotate. A new pure `open-sse/services/autoCombo/modelFamily.ts` (`detectModelFamily`) classifies by model-id prefix for six families; `zai` is instead resolved by provider id (z.ai's hosted API serves the same `glm-*` model ids as every other GLM backend, so `auto/zai` means "route to my z.ai backend specifically" vs `auto/glm`'s "any connected GLM backend"). Reuses the existing `createVirtualAutoCombo` on-demand materialization path (no DB writes) and the `/v1/models` catalog advertising loop. Regression guard: `tests/unit/autoCombo/provider-family-combos.test.ts` (11).
+- **feat(proxy):** native **proxy-pool round-robin / egress IP rotation** ([#6365](https://github.com/diegosouzapw/OmniRoute/issues/6365)) — a scope (global / provider / account) can now hold **multiple** proxies as a pool with a rotation strategy, so outbound requests cycle their egress IP instead of pinning one proxy per scope. Migration `117_proxy_pool_rotation.sql` lifts the `UNIQUE(scope, scope_id)` constraint (rebuild via the canonical rename/copy/drop; existing single assignments become 1-element pools) and adds a `proxy_scope_rotation` companion table holding the per-scope strategy + a persisted monotonic round-robin cursor. Strategies: `round-robin` (default, monotonic cursor — never `Math.random`), `random`, and `sticky-per-N-min`. Resolution (`resolveProxyForScopeFromRegistry` / `resolveProxyForConnectionFromRegistry`) now fetches the alive, position-ordered candidate set (unchanged `PROXY_ALIVE_PREDICATE`) and applies the strategy; an empty / all-dead pool still returns `null` — the #6246 fail-closed guard is **untouched** (never falls through to direct egress). Backend + DB only; dashboard pool-builder UI is a follow-up. Regression guard: `tests/unit/proxy-pool-rotation-6365.test.ts` (8, incl. fail-closed + backward-compat).
+- **feat(providers):** end-to-end **tool/function calling on the native Gemini `/v1beta` endpoint** ([#6222](https://github.com/diegosouzapw/OmniRoute/issues/6222)) — both directions of the Gemini↔OpenAI conversion now preserve tool data (previously silently dropped). Request side: `convertGeminiToInternal` (extracted to its own testable module) maps `tools[].functionDeclarations` → OpenAI `tools`, prior `functionCall` parts → assistant `tool_calls`, and `functionResponse` parts → `tool`-role messages. Response side: `convertOpenAIResponseToGemini` emits `parts[].functionCall {name,args}` from `message.tool_calls`, and the streaming `openAIChunkToGeminiChunk` accumulates fragmented `tool_calls` deltas by index into complete `functionCall` parts. The non-Gemini client paths (Claude, OpenAI-Responses) already preserved tool calls — this closes the gap specific to the native Gemini surface. Regression guard: `tests/unit/v1beta-gemini-tool-calling-6222.test.ts` (6, incl. a streaming SSE round-trip).
+- **feat(providers):** copilot-m365-web **enterprise / work tier** support ([#6334](https://github.com/diegosouzapw/OmniRoute/issues/6334)) — mirrors the EDU-tier pattern (#6210): `M365ConnectionParams` gains an `agent` field, a new opt-in `M365_ENTERPRISE_OVERRIDES` preset (`agent=work`, `scenario=officeweb`, `licenseType=Premium`) applies via `providerSpecificData.tier="enterprise"` (alias `"work"`), and `agent` is also overridable directly via `providerSpecificData.agent`. `buildWsUrl` was hardcoding `agent="web"` (the one enterprise-distinguishing param with no override path), so a Premium work account handshook then returned an empty stream. The individual and EDU paths are untouched. Kilo's dup flag vs #6210 (EDU tier) was a false positive — different tier. Regression guard: `tests/unit/copilot-m365-enterprise-6334.test.ts` (7). End-to-end confirmation on a real Premium work account is a live-VPS validation follow-up (Hard Rule #18). (thanks @Forcerecon)
+- **feat(api):** standardized, provider-agnostic **`effort` + `thinking` request params** ([#6241](https://github.com/diegosouzapw/OmniRoute/issues/6241)) — a thin standardization layer over the existing mature per-provider reasoning plumbing (no provider mapper touched). `providerChatCompletionSchema` gains a canonical `effort` (reusing the shared `none/low/medium/high/xhigh` vocabulary — the UI tiers `extra`/`max` collapse onto `xhigh`) and a boolean `thinking`. A pure `normalizeReasoningRequest` (wired once in `src/sse/handlers/chat.ts`, before any reasoning field is read) folds them onto the fields the translators already consume (`reasoning_effort` / `reasoning.effort` / `thinking`), so they fan out to Anthropic / Gemini / xAI / Responses — an explicit client `reasoning_effort` / object-shaped `thinking` always wins (backward-compatible). `/models` additively exposes `supportsThinking` + `effort_tiers` so the frontend can render the toggles (UI component is a follow-up). Regression guard: `tests/unit/effort-thinking-standardization-6241.test.ts` (12). (thanks @Iammilansoni, @shabeer)
+- **feat(combo):** new **`pipeline` (sequential) combo strategy** ([#6297](https://github.com/diegosouzapw/OmniRoute/issues/6297)) — the 18th routing strategy runs targets **in order**, threading each step's output into the next step's input, with an optional per-step `prompt` (system instruction); only the final step's response is returned. Distinct from `fusion` (parallel fan-out + judge). Implemented as a self-contained `open-sse/services/pipeline.ts` (sibling to `fusion.ts`), dispatched from `combo.ts`; the step list reuses `combo.models` order and reads an optional `prompt` off each target (backward-compatible — ignored by every other strategy). Intermediate steps run non-streaming with tools stripped (complete prose to thread forward); the final step keeps the client's `stream` flag + tools. A failing/empty/unparseable intermediate step fails the whole pipeline explicitly via a sanitized error (never silently swallowed). Kilo's dup flag vs #563 was a false positive (that's model→chain selection; this is a sequential chain). Regression guard: `tests/unit/combo-pipeline-strategy.test.ts` (5). (thanks @ofekbetzalel)
+- **feat(ci):** `check:test-masking` now flags **inline-reimplemented prod conditions** ([#6348](https://github.com/diegosouzapw/OmniRoute/issues/6348)) — a new **report-only** subcheck (v2, 6A.10 family) catches the wrong-shape contract test: a test that recomputes the condition under test inline instead of importing/exercising the real function (the #6216 class, where `=== 500` → `>= 500` stayed green because the test re-implemented the branch). For each added/modified test file it warns when the file textually duplicates a ≥3-token conditional from a production file touched in the same PR **and** does not import the symbol/module owning it, via a pure, fixture-tested `findReimplementedConditions()` with an allowlist mirroring `assertReductionAllowlist`. Report-only for now (does not fail the gate) — to be promoted to blocking after a triage cycle. Regression guard: `tests/unit/check-test-masking.test.ts` (45).
+- **feat(sse):** per-connection routing override (native vs CLIProxyAPI) ([#6339](https://github.com/diegosouzapw/OmniRoute/issues/6339)) — the previously-dead `isCliproxyapiDeepModeEnabled` helper is now wired into `resolveExecutorWithProxy`: a single connection can opt itself into the CLIProxyAPI passthrough executor via `providerSpecificData.cliproxyapiMode="claude-native"`, with precedence **connection override > provider `upstream_proxy_config` mode > default**. `resolveExecutorWithProxy` now receives the resolved connection's `providerSpecificData` (threaded from `chatCore.ts`), so one connection can deep-route while the provider's other connections stay native — no DB schema change (the toggle rides in `providerSpecificData`). Also resolves the same-provider mixing ask in #6340. Regression guard: `tests/unit/chatcore-executor-proxy.test.ts` (9). (thanks @RaviTharuma)
+- **feat(dashboard):** "Add session cookie" modal now shows a prominent **"Open ‹host› →"** link to the provider's own site ([#6268](https://github.com/diegosouzapw/OmniRoute/issues/6268)) — every `-web` cookie-session provider (chatgpt-web, claude-web, gemini-web, kimi-web, lmarena, qwen-web, m365-copilot-web, …) renders a one-click external link (opening the provider's login/home page in a new tab) so operators no longer tab away to retype the URL mid-setup. The host resolves from a pure, unit-tested `resolveWebProviderHost()` (prefers `WEB_COOKIE_PROVIDERS[id].website`, falls back to the registry `baseUrl` origin); non-web providers render exactly as before. Kilo's dup flag vs #6265 (modal-too-small-on-1080p) was a false positive — distinct concern. Regression guard: `tests/unit/resolve-web-provider-host.test.ts` (5). (thanks @chirag127)
+- **feat(providers):** add **DigitalOcean** AI (serverless inference) as an OpenAI-compatible API-key provider ([#6373](https://github.com/diegosouzapw/OmniRoute/pull/6373)) — base `https://inference.do-ai.run/v1`, wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/digitalocean/`, `src/shared/constants/providers/apikey/inference-hosts.ts`). Regression guard: `tests/unit/digitalocean-provider.test.ts`. (thanks @newnol)
+- **feat(providers):** add **Huancheng Public API** (`hcnsec`) as an OpenAI-compatible regional provider ([#6410](https://github.com/diegosouzapw/OmniRoute/pull/6410)) — Xinjiang Huancheng Cybersecurity's public LLM platform (base `https://api.hcnsec.cn/v1`, free credits via daily check-ins), wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/hcnsec/`, `src/shared/constants/providers/apikey/regional.ts`). Regression guard: `tests/unit/hcnsec-provider.test.ts`. (thanks @UnrealAryan)
+- **feat(dashboard):** the web-session credential guide now shows an **"Open {host}" link** ([#6316](https://github.com/diegosouzapw/OmniRoute/pull/6316)) to the provider's sign-in site (derived from the provider `website` via `getProviderWebsiteHost`), so you can jump straight to the page where the cookie/session must be captured. Regression guard: `tests/unit/web-session-provider-link-6316.test.ts`. (thanks @jordansilly77-stack)
+- **feat(cerebras):** add the **Gemma 4 31B** model (`gemma-4-31b`) to the Cerebras registry + pricing table ([#6331](https://github.com/diegosouzapw/OmniRoute/pull/6331)). Regression guard extends `tests/unit/t28-model-catalog-updates.test.ts`. (thanks @backryun)
+- **feat(providers):** add **Yuanbao (web)** as a cookie-session provider ([#6196](https://github.com/diegosouzapw/OmniRoute/issues/6196)) — `yuanbao-web` (Tencent Yuanbao, `yuanbao.tencent.com`) with cookie-only auth (`hy_user`/`hy_token` + public agent id), SSE→OpenAI translation incl. `reasoning_content`, exposing DeepSeek V3/R1 + Hunyuan / Hunyuan-T1. Regression guard: `tests/unit/providers-yuanbao-web.test.ts`. `together-web` was **deferred** (no verifiable web-session endpoint — needs a captured request) and `huggingchat-web` **dropped** (the existing `huggingchat` already is a web-cookie provider). (thanks @chirag127)
+- **feat(providers):** route the built-in **agentrouter** through the dynamic Claude-Code wire image ([#6056](https://github.com/diegosouzapw/OmniRoute/issues/6056)) — a small static allow-set (`CC_WIRE_IMAGE_BUILTINS` in `open-sse/services/ccWireImageBuiltins.ts`), consulted by `isClaudeCodeCompatible` / `isClaudeCodeCompatibleProvider` / `applyFingerprint`, makes agentrouter adopt the CC wire-image headers + fingerprint **while guarding the CC baseUrl/auth branches** so it keeps its own registry `baseUrl` and `x-api-key` auth. Regression guard: `tests/unit/agentrouter-cc-wire-image.test.ts` (asserts the wire image is applied AND agentrouter's baseUrl/auth are preserved). Live WAF-acceptance against agentrouter.org is a VPS validation follow-up (Hard Rule #18).
+- **feat(providers):** **bulk-add API keys for Cloudflare Workers AI** ([#6174](https://github.com/diegosouzapw/OmniRoute/issues/6174)) — `cloudflare-ai` is removed from the bulk-add exclusion list and the bulk parser gains a 3-field `name|accountId|apiKey` mode; the bulk route now builds a **per-entry** `providerSpecificData` so each key carries its own `accountId` (fixing the previous shared-object reuse), and both the create + key-validation paths receive it. Regression guard: `tests/unit/bulk-api-key-parser-cloudflare.test.ts`. (thanks @muflifadla38)
+- **feat(dashboard):** routing/settings UX clarity ([#6147](https://github.com/diegosouzapw/OmniRoute/issues/6147)) — (1) weighted combos show the **effective routing share %** next to each weight when weights don't sum to 100 (`WeightTotalBar.tsx`); (2) the status widget's user-facing **"Cloud Sync" label is renamed** to "Remote Settings Sync" (`CloudSyncStatus.tsx`; internal ids/state untouched); (3) built-in providers gain an **opt-in advanced base-URL override** (`isBaseUrlOverrideEligibleProvider`, hidden behind an "Advanced" toggle, reusing the existing `providerSpecificData.baseUrl` persistence — not globally widened). Regression guard: `tests/unit/routing-settings-ux-6147.test.ts`.
+- **feat(combo):** add an option to **disable session stickiness**, per-combo or globally — round-robin / random combos can rotate to a different connection on every request instead of pinning a whole conversation to one connection by its first-message hash. Resolution precedence per-combo `config.disableSessionStickiness` → global `settings.disableSessionStickiness` → default `false` (preserves the #3825 prompt-cache/504 fix); gates **both** stickiness call sites in `open-sse/services/combo.ts`. Exposed as a global toggle (Combo Defaults) and a per-combo Inherit/on/off control. ([#6168](https://github.com/diegosouzapw/OmniRoute/issues/6168)) Regression guard: `tests/unit/combo-disable-session-stickiness.test.ts`. (thanks @RCrushMe)
+- **feat(docker):** add the `OMNIROUTE_NO_SUDO` env flag for root-less / user-namespaced deployments — the MITM cert-trust command path (`resolveSudoSpawn` in `src/mitm/systemCommands.ts`) now strips the leading `sudo` when the flag is truthy, in addition to the existing root / sudo-missing cases, so the Proxy Agent runs without `sudo` (the operator trusts the CA manually, e.g. via `NODE_EXTRA_CA_CERTS`). Argv-array `spawn` preserved — no shell interpolation (Hard Rule #13). ([#6122](https://github.com/diegosouzapw/OmniRoute/issues/6122)) Regression guard: `tests/unit/mitm-systemCommands-no-sudo.test.ts`. (thanks @powellnorma)
+- **feat(providers):** add **Requesty** as an OpenAI-compatible gateway provider (BYOK, base `https://router.requesty.ai/v1`, ~200 free requests/day) — wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/requesty/`, `src/shared/constants/providers/apikey/gateways.ts`). ([#6120](https://github.com/diegosouzapw/OmniRoute/issues/6120)) Regression guard: `tests/unit/requesty-provider.test.ts`. (thanks @chirag127)
+- **feat(dashboard):** add **configured-only / available-only filters** to the Free Provider Rankings page ([#6150](https://github.com/diegosouzapw/OmniRoute/issues/6150)) — hide providers you haven't configured, or whose connections are all rate-limited / out of quota, via server-side query params (`?configuredOnly` / `?availableOnly` on `GET /api/free-provider-rankings`) backed by a testable lib helper reusing the in-process connection state (no Redis). Both filters default off, so the default view is unchanged; this supersedes the earlier client-side "Configured Only" toggle (#6245) with an available-only dimension and unit-tested logic. Regression guard: `tests/unit/freeProviderRankings-filters.test.ts`.
+
+### 🔧 Bug Fixes
+
+- **fix(dashboard):** adding a second API key connection for the same provider no longer silently overwrites the first — the Add-API-key modal now derives a unique default connection name (`main`, then `main-2`, `main-3`, …) so the backend name-based upsert can't collide (#6499 — thanks @dilneiss).
+- **fix(compression):** the session-dedup engine now also deduplicates a large multi-line block repeated **within a single message** (intra-message dedup), not just across turns; the compression-preview API surfaces a `fallbackReason`, and the fusion panel reports how many models were rate-limited vs failed on total-panel failure (#6501 — thanks @chirag127).
+- **fix(compression):** a stacked-pipeline step naming an unregistered engine now surfaces a `validationErrors` entry instead of silently no-op'ing, so misconfigured pipelines are visible in the preview API (#6506 — thanks @chirag127).
+- **feat(usage):** add a **Codex reset-credit redemption** flow to the Provider Limits UI ([#6361](https://github.com/diegosouzapw/OmniRoute/pull/6361)) — a `useCodexResetCreditRedemption` hook + `/api/usage/codex-reset-credit` route + `codexResetCredits` lib let you redeem banked Codex reset credits from the quota card. Regression guard: `tests/unit/codex-reset-credits.test.ts`. (thanks @JxnLexn)
+- **feat(glm):** add **team-plan quota settings** for `glm-cn` connections ([#6351](https://github.com/diegosouzapw/OmniRoute/pull/6351)) — a dedicated `GlmTeamQuotaFields` form section (team quota id / limits) threaded through the Add/Edit connection modals, persisted via `providerSpecificData`, with the GLM usage service reading the team quota. Regression guards: `tests/unit/glm-team-quota.test.ts`, `provider-specific-data-schema.test.ts`. (thanks @hao3039032)
+- **feat(providers):** add **TinyFish** web-fetch/search support ([#6349](https://github.com/diegosouzapw/OmniRoute/pull/6349)) — a `tinyfish-fetch` executor + `/v1/web/fetch` route + MCP web-fetch tool, registered as a specialty-media provider with request-validation and a search-provider catalog entry. Regression guards: `tests/unit/executor-tinyfish-fetch.test.ts`, `web-fetch-handler.test.ts`, `mcp-web-fetch-tool.test.ts`, `provider-validation-tinyfish.test.ts`. (thanks @dtybnrj)
+- **fix(cli):** `omniroute launch-codex` now spawns `codex.cmd` through a shell on Windows (the npm `.cmd` shim is unresolvable by bare `spawn` → ENOENT), mirroring the qodercli Windows fix (#6263) ([#6312](https://github.com/diegosouzapw/OmniRoute/pull/6312)). Regression guard: `tests/unit/launch-codex-windows-spawn-6312.test.ts`. (thanks @swingtempo)
+- **fix(codex):** isolate the **Spark** quota from the shared Codex quota and stabilize the quota UI ordering / hydration so per-scope limits render consistently ([#6336](https://github.com/diegosouzapw/OmniRoute/pull/6336)). Regression guards: `tests/unit/codex-quota-selection-hydration.test.ts`, `provider-limits-ui.test.ts` + 3 more. (thanks @xz-dev)
+- **feat(api):** add a `hidePaidModels` setting that filters paid-only models out of the `/v1/models` catalog. Regression guard: `tests/unit/models-catalog-hide-paid.test.ts`. (thanks @chirag127)
+- **fix(api-manager):** the fallback model picker now preserves combos instead of dropping them when a primary model is unavailable ([#6443](https://github.com/diegosouzapw/OmniRoute/pull/6443)). Regression guard: `tests/unit/api-manager-page-static.test.ts`. (thanks @jmengit)
+- **fix(providers):** recoverable Antigravity / Cloud-Code (Gemini Code Assist) `403` responses ([#6452](https://github.com/diegosouzapw/OmniRoute/pull/6452)) are now classified as a retryable project-config error instead of a terminal account ban, so a fixable project/API-disabled 403 no longer forces a ~1-year cooldown / full OAuth reconnect. Regression guard: `tests/unit/errorclassifier-antigravity-403.test.ts`. (thanks @developerjillur)
+- **fix(mitm):** `sanitizeHeaders` now redacts `Set-Cookie` response headers so upstream session cookies never leak into logs / diagnostics ([#6451](https://github.com/diegosouzapw/OmniRoute/pull/6451)). Regression guard: `tests/unit/mitm-sanitize-headers.test.ts`. (thanks @developerjillur)
+- **fix(api):** `/api/compression/preview` now accepts `mode: "caveman"` and correctly handles stacked / zero-compression previews ([#6425](https://github.com/diegosouzapw/OmniRoute/issues/6425)). Regression guard: `tests/unit/api/compression-preview-caveman-and-stacked-6425.test.ts`. (thanks @chirag127)
+- **feat(providers):** add **Zed** hosted LLM aggregator as a native-app provider ([#6118](https://github.com/diegosouzapw/OmniRoute/pull/6118)) — OAuth sign-in via the Zed hosted flow, registered through the shared provider registry + executor. Regression guards: `tests/unit/zed-oauth-provider.test.ts`, `zed-import-utils.test.ts`, `zed-docker-detect.test.ts`, `mitm-handler-zed.test.ts`. VPS-validated via live operator login (Hard Rule #18).
+- **fix(oauth):** the Kiro SSO-cache auto-import now **preserves the IDC region** — cross-region Amazon Q / Kiro profiles imported from the SSO cache are no longer collapsed to the default region ([#6113](https://github.com/diegosouzapw/OmniRoute/pull/6113)). Regression guard: `tests/unit/kiro-auto-import-idc-2059.test.ts`. VPS-validated via live operator login (Hard Rule #18).
+- **fix(dashboard):** passthrough model aliases no longer collide when two namespaced model ids share a last segment (port from 9router#1850, [#6431](https://github.com/diegosouzapw/OmniRoute/pull/6431)). `enx/gpt-5.5` and `enx/codebuddy/gpt-5.5` both auto-generated the alias `gpt-5.5`, so the second model could never be added (the UI just alerted "alias already exists"). Aliases are now disambiguated deterministically — bare last segment when free, then parent-qualified (`codebuddy-gpt-5.5`), then a numeric suffix — while re-adding the exact same model id is still blocked. Regression guard: `tests/unit/passthrough-alias-1850.test.ts`. (thanks @arpicato)
+- **fix(translator):** preserve a Gemini `functionResponse` co-located with other parts (another `functionCall`, or trailing `text`) in the same content when translating **Gemini → OpenAI** ([#6376](https://github.com/diegosouzapw/OmniRoute/pull/6376)). `convertGeminiContent()` early-returned the tool message on the first `functionResponse` part, dropping any co-located parts; such contents are now pre-split (one tool message per `functionResponse`, emitted first, plus one message for the remaining parts). Regression guard: `tests/unit/gemini-to-openai-function-response.test.ts`. (thanks @warelik)
+- **fix(headroom):** detect a python interpreter managed by **mise / pyenv / asdf / conda** (port from 9router#2353, [#6382](https://github.com/diegosouzapw/OmniRoute/pull/6382)). Headroom's python probe (`src/lib/headroom/detect.ts`) searched a hardcoded `PATH`, but version managers expose their interpreters via shim dirs that only join `PATH` through interactive-shell activation — which the non-interactive server never runs, so a managed python (≥3.10) was invisible and Headroom reported it missing. The search path now prepends the well-known shim/bin dirs (`~/.local/share/mise/shims`, `~/.pyenv/shims`, `~/.asdf/shims`, `$CONDA_PREFIX/bin`, `~/.local/bin`, respecting `MISE_DATA_DIR`/`PYENV_ROOT`/`ASDF_DATA_DIR` when set), and a new `HEADROOM_PYTHON` env override lets operators point straight at their interpreter (mirroring `HEADROOM_URL`). Still shell-free (`execFileSync`). Regression guard: `tests/unit/headroom-detect.test.ts` (5). (thanks @loopyd)
+- **fix(executors):** strip the OpenAI-Codex/Claude-CLI `client_metadata` passthrough field for **NVIDIA** requests (port from 9router#1887, [#6411](https://github.com/diegosouzapw/OmniRoute/pull/6411)). NVIDIA's OpenAI-compatible wrapper rejects it with `400 Unsupported parameter`, the same class already handled for `cerebras`/`mistral`; `nvidia` (executor `default`) was missing from the strip allowlist so Codex/Claude-Code passthrough requests 400'd. Regression guard: `tests/unit/executor-default-strip-client-metadata.test.ts` (+nvidia case). (thanks @phidinhmanh)
+- **fix(translator):** strip the Claude-style `thinking` field for **NVIDIA `z-ai/glm-5.2`** (port from 9router#2023, [#6413](https://github.com/diegosouzapw/OmniRoute/pull/6413)). NVIDIA's OpenAI-compatible wrapper 400s on `thinking` (a Claude-format client routed here leaves a `thinking:{type:"adaptive"}`); the existing strip rule only dropped `reasoning`. Same class already handled for `minimax-m2.7`. Regression guard: `tests/unit/nvidia-minimax-thinking-strip.test.ts` (+glm-5.2 case). (thanks @phidinhmanh)
+- **fix(translator):** suppress the streamed `</think>` close marker for the **Antigravity IDE** client (port from 9router#1061, [#6415](https://github.com/diegosouzapw/OmniRoute/pull/6415)). On thinking-only turns Antigravity rendered a bare `</think>` as the sole visible content, tripping its loop-detection and wasting requests. Antigravity's UA (`vscode/<v> (Antigravity/<v>)`) is added to the marker-suppress allowlist (alongside OpenCode); Claude Code / Cursor still get the marker, and `x-omniroute-thinking-marker: on` force-restores it. Regression guard: `tests/unit/think-close-marker-suppress-5245.test.ts`. (thanks @abdofallah)
+- **fix(executors):** strip nested `reasoning_content` from messages for **Mistral** (port from 9router#1649, [#6417](https://github.com/diegosouzapw/OmniRoute/pull/6417)). Mistral's API returns `422 extra_forbidden` when an assistant message carries `reasoning_content` (replayed thinking from a prior turn, e.g. via the Codex `/responses` path); the generic top-level 400 field-downgrade retry never covered the nested per-message field. `DefaultExecutor` now strips it for provider `mistral` only, so DeepSeek (which requires replayed `reasoning_content`) is unaffected. Regression guard: `tests/unit/mistral-strip-reasoning-content-1649.test.ts`. (thanks @xxy9468615)
+- **fix(executors):** strip the `client_metadata` passthrough field on the **OpenCode** path (port from 9router#1442, [#6418](https://github.com/diegosouzapw/OmniRoute/pull/6418)). OpenCode upstreams (e.g. `kimi-k2.6` via opencode-go) reject it with `400 "Extra inputs are not permitted, field: 'client_metadata'"`; the DefaultExecutor strip only covered cerebras/mistral and `OpencodeExecutor` extends `BaseExecutor` directly, so nothing removed it there. Regression guard: `tests/unit/opencode-strip-client-metadata-1442.test.ts`. (thanks @yanpaing007)
+- **fix(executors):** inject the `reasoning_content` echo for the native **Moonshot Kimi** provider (port from 9router#1480, [#6419](https://github.com/diegosouzapw/OmniRoute/pull/6419)). Kimi (executor `default`) is a thinking-mode upstream that 400s with "reasoning_content must be passed back" when a prior assistant turn lacks it; the placeholder injection was only wired into the OpenCode meta-provider, so direct multi-turn Kimi conversations failed. Scoped to `kimi` (gateway-served models matching the thinking-model name pattern are unaffected). Regression guard: `tests/unit/kimi-native-reasoning-injected-1480.test.ts`. (thanks @2220258345)
+- **fix(executors):** recover from a strict gateway's `context_management: Extra inputs are not permitted` 400 (port from 9router#1468, [#6420](https://github.com/diegosouzapw/OmniRoute/pull/6420)). **Claude Code** always sends a top-level `context_management` field; strict anthropic-compatible gateways reject it. The dedicated context-editing 400-fallback only fired when OmniRoute's own `contextEditing` feature was enabled (default off), so a client-sent field passed through untouched and 400'd. `context_management` is now in the generic reactive field-strip list, so it's stripped-and-retried once regardless of the feature flag (with correct request re-signing for claude-compatible relays). Regression guard: `tests/unit/provider-field-strips.test.ts`. (thanks @ohahe52-dot)
+- **fix(network):** enable **RFC 8305 Happy Eyeballs** (`autoSelectFamily`) on the direct-egress undici dispatcher (port from 9router#1237, [#6423](https://github.com/diegosouzapw/OmniRoute/pull/6423)). When DNS returns both IPv6 (AAAA) and IPv4 (A) and the IPv6 route is broken (e.g. a NAT64 `64:ff9b::` prefix without routing), undici tried IPv6 first and hung until `ETIMEDOUT` (then a 502 + account lockout), even though `curl` reached the same host. The direct dispatcher now races both families and uses whichever connects first. Proxy paths pin family via `proxyTls` and are unaffected. Regression guard: `tests/unit/direct-dispatcher-pipelining-4580.test.ts`. (thanks @adentdk)
+- **fix(combo):** round-robin now advances the rotation pointer past the model that **actually served**, not the eagerly-scheduled one (port from 9router#948, [#6428](https://github.com/diegosouzapw/OmniRoute/pull/6428)). With `stickyLimit: 1` (true round-robin), when the scheduled model failed and a _different_ model served via fallback, the counter had already advanced +1 from the scheduled index — so the next request reused the fallback-served model, degrading round-robin into hot-spotting on whichever model was healthy. The pointer now advances to the served index + 1 (mirroring the sticky-limit>1 path). Session-stickiness (#3825) and distribution are preserved. Regression guard: `tests/unit/combo-rr-fallback-advance-948.test.ts`. (thanks @binsarjr)
+- **fix(sse):** a non-string `model` field is now rejected with a `400` before the resolver, instead of crashing downstream `.toLowerCase()`/`.split()` calls into an empty-body `500` that escapes the error sanitizer ([#6407](https://github.com/diegosouzapw/OmniRoute/issues/6407)). Regression guard: `tests/unit/chat-non-string-model-6407.test.ts`. (thanks @chirag127)
+- **fix(api):** unknown `/api/*` routes now return a JSON `404` (instead of the dashboard HTML shell) and scalar chat params (`model`/`temperature`/etc.) are validated **before** the provider lookup so malformed requests fail fast with a clear `400` ([#6424](https://github.com/diegosouzapw/OmniRoute/issues/6424), [#6412](https://github.com/diegosouzapw/OmniRoute/issues/6412)). Regression guards: `tests/unit/api/api-catchall-json-404.test.ts`, `tests/unit/chat-early-schema-validation-6412.test.ts`. (thanks @chirag127)
+- **fix(api):** `/v1/chat/completions` now rejects a non-JSON `Content-Type` with a `400` before parsing the body ([#6414](https://github.com/diegosouzapw/OmniRoute/issues/6414)). Regression guard: `tests/unit/v1-chat-completions-content-type-6414.test.ts`. (thanks @chirag127)
+- **fix(api):** the `X-OmniRoute-Compression` response header is now echoed on `/v1/chat/completions` and `/v1/completions` ([#6422](https://github.com/diegosouzapw/OmniRoute/issues/6422)). Regression guard: `tests/unit/compression-header-echo-6422.test.ts`. (thanks @chirag127)
+- **fix(api):** concurrent `GET /v1/models` requests are coalesced into a single catalog build ([#6408](https://github.com/diegosouzapw/OmniRoute/issues/6408)). Regression guard: `tests/unit/v1-models-concurrent-6408.test.ts`. (thanks @chirag127)
+- **fix(api):** `/v1/completions` now echoes the requested `body.model` in its JSON + streamed responses ([#6429](https://github.com/diegosouzapw/OmniRoute/pull/6429)). Regression guard: `tests/unit/completions-body-model-echo.test.ts`. (thanks @chirag127)
+- **fix(api):** env-var master keys now see the full `/v1/models` catalog ([#6406](https://github.com/diegosouzapw/OmniRoute/issues/6406)). Regression guard: `tests/unit/models-catalog-envkey-6406.test.ts`. (thanks @chirag127)
+- **fix(api):** non-streaming `/v1/completions` responses now echo `body.model` aligned with the `X-OmniRoute-Model` header ([#6426](https://github.com/diegosouzapw/OmniRoute/issues/6426)). Regression guard: `tests/unit/v1-completions-model-header-match-6426.test.ts`. (thanks @chirag127)
+- **fix(api):** unknown `/v1/*` routes now return a JSON `404 not_found` instead of the Next.js dashboard HTML shell ([#6405](https://github.com/diegosouzapw/OmniRoute/issues/6405)). Regression guard: `tests/unit/api/v1-catchall-json-404.test.ts`. (thanks @chirag127)
+- **fix(api):** the per-connection provider models route now degrades to the shipped catalog when a provider's `/models` endpoint answers with a redirect ([#6267](https://github.com/diegosouzapw/OmniRoute/issues/6267)) — a `qwen-web` import failed with a raw `Redirect blocked … (307)` 503. `safeOutboundFetch` throws `REDIRECT_BLOCKED` on the 307, `getSafeOutboundFetchErrorStatus` maps it to 503, and `buildDiscoveryErrorFallbackResponse` treated every 503 as a hard error — so the non-empty `getModelsByProviderId("qwen-web")` catalog was never surfaced. A models-endpoint redirect is not a fixable-config error (unlike `URL_GUARD_BLOCKED`/`INVALID_URL`, which stay hard errors), so it now falls back to the local/cached catalog before the 503 short-circuit. General fix — covers any config-driven provider that 307s. Regression guard: `tests/unit/provider-models-qwen-web-redirect-6267.test.ts`. (thanks @chirag127)
+- **fix(api):** the per-connection provider models route (MCP `list_models_catalog` + the dashboard import view) now merges USER-ADDED custom models into its response ([#6247](https://github.com/diegosouzapw/OmniRoute/issues/6247)) — custom models live in the `key_value` namespace `customModels`, which the live REST `/api/v1/models` already merges, but `src/app/api/providers/[id]/models/route.ts` never read `getCustomModels`, so custom models were dropped on both the discovery-success and local_catalog paths. They are now folded into the returned model list (deduped by id, stamped `owned_by: provider`), fixing MCP + the dashboard import view in one place. Regression guard: `tests/unit/provider-models-custom-merge-6247.test.ts`. (thanks @RCrushMe)
+
+- **fix(providers):** GitLab Duo tool-calling follow-up turns no longer fail upstream with `422 {"detail":"Validation error"}` (tokens 0/0, rejected pre-inference). The #6234 tool-result-feedback fix serialized the **entire** multi-turn conversation into GitLab's single-file `code_suggestions` (`small_file`) generation endpoint — folded history that turn-N sent as an oversized `current_file.content_above_cursor` **and** duplicated verbatim into `user_instruction`, tripping the AI-Gateway's `small_file` validation guard. The executor now **bounds** that prompt: it keeps system + latest user message + the most-recent tool round (dropping older turns), caps oversized tool results, and stops duplicating the full prompt into `user_instruction` (which now carries only the short latest user message) — while still feeding the most-recent tool result back so the agent continues (`open-sse/executors/gitlab.ts`, [#6220](https://github.com/diegosouzapw/OmniRoute/issues/6220)). The unit test covers the bounding logic; the upstream 422→200 clearing is VPS-only (Hard Rule #18). Regression guard: `tests/unit/gitlab-tool-exchange-bounded-6220.test.ts`.
+
+- **fix(i18n):** the provider-detail (`/dashboard/providers/[id]`) connection-status filter labels no longer render as `__MISSING__:All` / `__MISSING__:Active` / `__MISSING__:Error` / `__MISSING__:Banned` / `__MISSING__:CreditsExhausted` in non-English locales (notably pt-BR) ([#6290](https://github.com/diegosouzapw/OmniRoute/issues/6290)). Root cause was **not** the namespace mismatch the issue guessed — the `providers.filter*` keys resolve correctly in `en.json`; the debt lived in the locale mirrors (`src/i18n/messages/*.json`), where these five keys carried the `__MISSING__:` sync sentinel in ~15 locales and were absent entirely in ~26 others, so next-intl found the key and echoed the sentinel verbatim. All 40 non-English/-Chinese mirrors now ship real translations for the five `providers.filter*` labels. Regression guard: `tests/unit/i18n-provider-filter-keys-6290.test.ts`. (thanks @diegosouzapw)
+
+- **fix(providers):** the `copilot-m365-web` streaming executor now emits `debug`-level WebSocket diagnostics ([#6210](https://github.com/diegosouzapw/OmniRoute/issues/6210)) — the outbound WS URL (with the `access_token` **redacted** via `redactWsUrl()`), handshake success/failure, and each received SignalR frame's `type`/`target`. Previously the streaming path logged nothing, so an empty `content:null` response (the M365 Education / Starter tier symptom fixed in #6234) was undiagnosable even at `APP_LOG_LEVEL=debug`. The change is debug-level and side-effect-free — it does not alter streaming behavior or the frame parser, and the token never reaches the logs. Regression guard: `tests/unit/copilot-m365-web-logging-6210.test.ts` (thanks @qpeyba)
+
+- **fix(resilience):** a round-robin combo no longer returns `503 all upstream accounts are unavailable` when a compatibility-rejected target is actually healthy ([#6238](https://github.com/diegosouzapw/OmniRoute/issues/6238)). `filterTargetsByRequestCompatibility` drops request-incompatible targets (tool/vision/structured-output unsupported, or below the required context window) **before** any availability check runs, and its `compatible.length === 0` safety net only fired when _all_ targets were filtered — not when the kept targets later all turned out runtime-unavailable (circuit-open / cooldown / no credentials). So a combo could 503 while a compat-rejected-but-healthy provider sat unused. `handleRoundRobinCombo` now keeps the compat-rejected set and, when every compat-kept target was skipped without a single real attempt, probes those rejected targets as a **last-resort fallback tier** (via the new pure `open-sse/services/combo/comboCompatFallback.ts`) before crystallizing the 503. Regression guard: `tests/unit/combo-roundrobin-compat-fallback-6238.test.ts`. (thanks @ThongAccount)
+
+- **fix(startup):** best-effort self-heal for a corrupted Turbopack dev cache on Windows ([#6289](https://github.com/diegosouzapw/OmniRoute/issues/6289)). On Windows, `pnpm dev` can fail at startup when Turbopack `mmap`s a persistent-cache SST file and the OS refuses the mapping (`os error 1455` — "paging file too small"), which Turbopack surfaces as a misleading `Module not found: Can't resolve '@/shared/utils/machine'`. This is a **known upstream Turbopack cache-corruption bug — not our code**. The dev launcher (`scripts/dev/run-next.mjs`) now wraps `nextApp.prepare()` and, when it rejects with that signature (`isTurbopackCacheCorruption` in the new `scripts/dev/turbopackCacheHeal.mjs`), purges `.build/next/**/cache/turbopack` and retries **once** with a clear log. **Caveat — best-effort only:** the corruption often surfaces as a runtime overlay rather than a `prepare()` rejection, so this cannot always intercept it; the reliable remedy remains manually deleting the Turbopack cache dir. Regression guard: `tests/unit/turbopack-cache-heal-6289.test.ts`. (thanks @chirag127)
+
+- **fix(providers):** qodercli PAT auth no longer fails with `spawn qodercli ENOENT` on Windows ([#6263](https://github.com/diegosouzapw/OmniRoute/issues/6263)) — `spawnQoderCli` spawned the bare `qodercli` name with `shell:false` and an unenriched env, so the npm `.cmd` wrapper under `%APPDATA%\npm` (a user-PATH directory) was never resolved. It now resolves the absolute `.cmd`/`.exe` path through the existing `getCliRuntimeStatus("qoder")` resolver in `src/shared/services/cliRuntime.ts` (memoized), spawns with `shell` when the target is a `.cmd`/`.bat`, and uses the cliRuntime-enriched env (PATH + PATHEXT + APPDATA); the ENOENT error now lists the searched paths plus the `CLI_QODER_BIN` override. End-to-end spawn on a real Windows host is host-only (Hard Rule #18); the path-resolution logic is unit-tested. Regression guard: `tests/unit/qodercli-windows-resolve-6263.test.ts`. (thanks @chirag127)
+
+- **fix(sse):** the reasoning-token buffer no longer inflates **probe-sized `max_tokens`** ([#6274](https://github.com/diegosouzapw/OmniRoute/issues/6274)) — Claude Code's `/model` capability check sends `max_tokens: 1`, but for a thinking-capable model with a large output cap (e.g. `glm-5.2`) the #3587 headroom heuristic (`max(current + 1000, ceil(current * 1.5))`) rewrote it to `1001` and forwarded that upstream, wasting tokens on a request that was never a genuine reasoning budget. `resolveReasoningBufferedMaxTokens()` (`open-sse/services/reasoningTokenBuffer.ts`) now short-circuits and returns the caller's value verbatim when it is below the new `REASONING_BUFFER_MIN_TRIGGER` (256) threshold — a tiny explicit limit is a probe, not a reasoning request. Real budgets still receive the #3587 headroom unchanged, and the guard runs after the existing capability checks so unknown / non-reasoning models keep returning `null`. Regression guard: `tests/unit/reasoning-token-buffer-6274.test.ts`. (thanks @brightfiscalband)
+
+- **fix(cli):** `omniroute reset-password` now works as a real subcommand, and password resets over piped (non-TTY) stdin actually apply ([#6261](https://github.com/diegosouzapw/OmniRoute/issues/6261), [#6258](https://github.com/diegosouzapw/OmniRoute/issues/6258)). Two coupled defects: (1) **#6261** — `bin/omniroute.mjs` routed everything through Commander with only two pre-Commander bypasses (`--mcp`, `reset-encrypted-columns`), so `omniroute reset-password` was rejected as an unknown command; only the separate `omniroute-reset-password` bin worked, while the docs falsely advertised the subcommand (incl. a bogus "legacy alias still works"). A pre-Commander bypass mirroring `reset-encrypted-columns` now dynamically imports `bin/reset-password.mjs` (which self-executes) before Commander parses; the three doc lines were corrected. (2) **#6258** — `bin/reset-password.mjs` issued two sequential `rl.question` prompts; under piped stdin the second read never settled at EOF, so `main()` never reached `resetManagementPassword` and the reset was a silent no-op (both prompts printed, no success, password unchanged). The CLI now detects non-TTY stdin and reads it once (first line = password, second line = confirm if present, else reused), adds a `--password-stdin` flag (entire stdin is the password, no confirmation), and exits `0` explicitly so the success line always flushes; interactive TTY behavior is unchanged. Regression guard: `tests/unit/reset-password-cli-6261-6258.test.ts` (3). (thanks @chirag127)
+
+- **fix(db):** the mass-migration **safety abort** now tells the operator how to bypass it and stops flooding the log ([#6260](https://github.com/diegosouzapw/OmniRoute/issues/6260)) — after restoring a backup that wiped the migration tracking table, `runMigrations()` threw the abort on every downstream `ensureDbInitialized()`, re-logging the full banner 11+ times, and the message never mentioned the existing `OMNIROUTE_MAX_PENDING_MIGRATIONS` escape hatch. The abort text now appends a bypass hint (set `OMNIROUTE_MAX_PENDING_MIGRATIONS=0` in `server.env` / `DATA_DIR/.env`), and a new `MigrationSafetyAbortError` is memoized so repeated calls in the same process throw the same instance and emit a single concise line instead of the full cascade. Regression guard: `tests/unit/migration-safety-abort-6260.test.ts`. (thanks @chirag127)
+
+- **fix(auth):** importing a **distinct** Codex/ChatGPT OAuth `auth.json` is no longer falsely rejected as "already exists" when it belongs to a different user in the same workspace ([#6301](https://github.com/diegosouzapw/OmniRoute/issues/6301)). `findExistingCodexConnection` (in `src/lib/oauth/utils/codexAuthImport.ts`) deduped **only** on `providerSpecificData.workspaceId === accountId`, where `accountId` is the shared `chatgpt_account_id`/`tokens.account_id` — so two members of the same ChatGPT Team collapsed onto a single connection (409 `duplicate_account`). The id_token's `https://api.openai.com/auth` claim carries a per-user `chatgpt_user_id` alongside the workspace id (the device-flow path already persisted it as `chatgptUserId`, but the import path did not). Now `parseAndValidateCodexAuth` extracts `userId` (`chatgpt_user_id` → `user_id` → JWT `sub`) into `ParsedCodexAuth`, the create/update paths persist `chatgptUserId` in `providerSpecificData` (mirroring `codex.ts`), and dedup keys on `workspaceId` **AND** `chatgptUserId` — with a backward-compat fallback to legacy accountId-only matching when no stored connection for that workspace records a `chatgptUserId`, so genuinely-same accounts still dedup. Regression guard: `tests/unit/codex-auth-import-userid-dedup-6301.test.ts` (4). (thanks @anungma)
+
+- **fix(providers):** importing models for the **venice-web** provider no longer fails with a red "Provider venice-web does not support models listing" ([#6269](https://github.com/diegosouzapw/OmniRoute/issues/6269)). `venice-web` is a web-cookie provider with an executor but no upstream `/v1/models` endpoint and no registry `models`, so the models route fell through to the tail `400`. Mirroring the `jules`/`linkup-search`/`ollama-search` fix (#5569), it now ships a static local catalog entry in `src/lib/providers/staticModels.ts` — seeding the current Venice lineup (`venice-uncensored`, `llama-3.3-70b`, `qwen3-235b`, `qwen3-4b`, `deepseek-r1-671b`; Venice rotates its catalog, see docs.venice.ai/models/overview) — so the route returns `200` with `source:"local_catalog"`, `intentional:true`. Regression guard: `tests/unit/static-models-venice-web-6269.test.ts`. (thanks @chirag127)
+
+- **fix(api):** the specialty model catalogs (`/v1/embeddings`, `/v1/images`, `/v1/music`, `/v1/videos` model lists) are now derived from the **unified catalog filtered by a predicate** (`getSpecialtyModelsResponse`) instead of ad-hoc per-route logic, so they consistently respect active-credential visibility and stay in sync with the main catalog ([#6303](https://github.com/diegosouzapw/OmniRoute/pull/6303)). Regression guard: `tests/unit/specialty-model-catalog-routes.test.ts`. (thanks @makcimbx)
+- **fix(api):** the agent-bridge server route now resolves the MITM manager via a **dynamic `import("@/mitm/manager.runtime")`** so Turbopack does not statically pull the stub (or over-bundle the manager), and the agent-skills generator anchors its output base path with `path.join(process.cwd(), …)` so Turbopack's static analyzer stops tracing the whole project root ([#6329](https://github.com/diegosouzapw/OmniRoute/issues/6329), [#6366](https://github.com/diegosouzapw/OmniRoute/pull/6366)). Regression guard: `tests/unit/agent-bridge-server-route-dynamic-import.test.ts`. (thanks @Iammilansoni)
+- **fix(api):** internal probes (combo-test, cloud-sync verify) now pick a **management-scoped / allow-all API key** instead of naively grabbing `getApiKeys()[0]` — a restricted `self:usage` first row made the probe fail with "Model X is not allowed for this API key" even when the combo path was healthy (`pickApiKeyForInternalUse` in `src/lib/db/apiKeys.ts`). The API-manager model editor also falls back to `/api/models?all=true` when `/v1/models` is catalog-protected ([#6372](https://github.com/diegosouzapw/OmniRoute/pull/6372)). Regression guard: `tests/unit/pick-internal-api-key-6372.test.ts`. (thanks @jmengit)
+- **fix(live-ws):** the Live Dashboard WebSocket server now **rejects on bind failure** (e.g. `EADDRINUSE` when the API bridge already holds the port) instead of letting the error surface as an unhandled `error` event that crash-loops the process — the `error` listener is attached to `wss` (not `server`) and releases the EventBus subscription on a failed start ([#6324](https://github.com/diegosouzapw/OmniRoute/issues/6324)). Regression guard: `tests/unit/live-ws-eaddrinuse-6324.test.ts`. (thanks @vinayakkulkarni)
+- **fix(dashboard):** the Home provider-topology widget now trusts the live provider-metrics snapshot — it uses `topology.errorProvider` and live `activeRequests` directly instead of re-deriving state from a stale `lastErrorAt` or applying a frontend timeout filter, so the topology reflects real-time provider health ([#6322](https://github.com/diegosouzapw/OmniRoute/pull/6322)). Regression guard: `tests/unit/home-provider-topology-live-state.test.ts`. (thanks @xz-dev)
+- **fix(sse):** strip zero-width markers from streamed **tool-call arguments** — a follow-up to [#5857](https://github.com/diegosouzapw/OmniRoute/pull/5857). That PR removed injected zero-width joiners (U+200D) from streamed assistant text/reasoning but deliberately left tool-call argument JSON byte-exact. The request-side obfuscation (`open-sse/services/claudeCodeObfuscation.ts`) injects ZWJ into agent words — including the temp path inside the Bash tool description — and Claude models copy that verbatim into generated commands, which are delivered as tool-call arguments rather than assistant text. As a result the ZWJ survived and corrupted code blocks (e.g. a temp path rendered with an invisible joiner). Now `open-sse/handlers/responseSanitizer.ts` strips zero-width code points from tool-call argument strings at every emit site (OpenAI non-stream/stream chat `tool_calls` + legacy `function_call`, native Responses `function_call` items, the OpenAI→Responses conversion, and the native Responses streaming `response.function_call_arguments.delta/.done` events). Only zero-width code points are removed; JSON structure and all other bytes stay identical (no parse/restringify), so normal arguments remain byte-exact. Regression guard: 6 new cases in `tests/unit/response-sanitizer.test.ts` (suite 50/50).
+
+- **fix(nodejs):** the default app log path now resolves under `DATA_DIR` (`~/.omniroute/logs/application/app.log`) instead of `process.cwd()` ([#6197](https://github.com/diegosouzapw/OmniRoute/issues/6197)) — the globally-installed CLI runs from an arbitrary working directory, so anchoring the default to cwd made file logging silently write to (or no-op under) an unrelated directory, contradicting the documented `.env.example` default. `getAppLogFilePath()` now computes the default lazily via the pure `resolveDataDir()` resolver (honours a per-process `DATA_DIR`, no directory-creation side effect); an explicit `APP_LOG_FILE_PATH` still wins. Regression guard: `tests/unit/logenv-datadir-path-6197.test.ts` (3).
+
+- **fix(docker):** AgentBridge/`startMitm` no longer aborts in containers/headless when the Antigravity-default DNS step can't write `/etc/hosts` ([#6127](https://github.com/diegosouzapw/OmniRoute/issues/6127)), and the privileged command's stderr now reaches `app.log` instead of only a bare exit code hitting the toast ([#6198](https://github.com/diegosouzapw/OmniRoute/issues/6198)). The default DNS step (`addDNSEntry`) was called unguarded while cert install and the two sibling DNS steps were each best-effort — in the runtime Docker image (`USER node`, no `sudo`, read-only `/etc/hosts`) it threw `Command failed with code 1` out of `startMitmInternal` and killed the whole start, discarding the stderr. The three DNS steps are extracted into a best-effort `provisionDnsEntries()` where each failure is logged with the full `err` (stderr included, folded in by `systemCommands.ts`) and never aborts the start. Regression guard: `tests/unit/mitm-dns-graceful-degrade-6127.test.ts` (4).
+
+- **fix(providers):** copilot-m365-web now supports the M365 Education "Starter / OfficeWebIncludedCopilot" tier and no longer returns an empty `content:null` stream ([#6210](https://github.com/diegosouzapw/OmniRoute/issues/6210)). Two gaps: (1) `buildWsUrl()` hardcoded the individual-consumer scenario (`OfficeWebPaidConsumerCopilot`, `isEdu=false`) — the EDU tier is now opt-in via `providerSpecificData.tier="edu"`, emitting `scenario=OfficeWebIncludedCopilot`/`isEdu=true` (the individual path is unchanged); (2) the EDU/GPT-5.5 path streams deltas via `arguments[0].writeAtCursor` (incremental) instead of only `messages[].text` (accumulated snapshots), which the parser dropped — a new `accumulateBotContent()` folds both formats, with `type:2 item.result.message` as a last-resort fallback. Regression guard: `tests/unit/copilot-m365-edu-writeatcursor-6210.test.ts` (10). (thanks @qpeyba)
+
+- **fix(providers):** GitLab Duo executor now feeds tool results back into the prompt instead of looping ([#6220](https://github.com/diegosouzapw/OmniRoute/issues/6220)) — `buildPrompt()` branched only on `system`/`user` and took `userParts.at(-1)`, silently dropping the `assistant{tool_calls}` + `tool{result}` turns the client appended, so the reconstructed prompt was byte-identical to turn 1 and the model re-emitted the same `<tool>` call forever. When a tool exchange is present the full conversation is now serialized, folding each tool result back keyed by its `tool_call_id`; simple conversations keep the legacy shape. Complements the tool_call emission from [#6051](https://github.com/diegosouzapw/OmniRoute/issues/6051) (the `kilo-duplicate` label was a false positive — different, sequential defect). Regression guard: `tests/unit/gitlab-tool-result-feedback-6220.test.ts` (4).
+
+- **fix(providers):** opencode-go/opencode-zen can now synthesize the OpenCode CLI identity headers Cloudflare requires on VPS egress ([#5997](https://github.com/diegosouzapw/OmniRoute/issues/5997)) — on a datacenter VPS, `opencode.ai/zen/go/v1/chat/completions` 403s (HTML challenge) requests lacking CLI identity, while the reporter's control curl proved that `User-Agent: opencode-cli/1.0.0` + `x-opencode-client: cli` + `x-opencode-project: default` + fresh request/session UUIDs succeed. Opt-in via `OPENCODE_SYNTHESIZE_CLI_HEADERS=true` (values overridable via `OPENCODE_GO_USER_AGENT`/`OPENCODE_USER_AGENT`/`OPENCODE_CLIENT`/`OPENCODE_PROJECT`); it fills only headers the client did not already send. Kept **off by default** — the forward-only path is deliberate (fabricating a wrong value risks upstream rejection; a prior dedup regressed with `opencode/local`), so this replaces the fragile local header-injection shim without changing default behavior. Regression guard: `tests/unit/opencode-cli-headers-synthesis-5997.test.ts` (6). (thanks @aleksesipenko)
+- fix(resilience): sticky session affinity now evicts and fails over to another account when the pinned account is exhausted/unavailable (#6219)
+
+- fix(sse): Responses API passthrough now drops internal commentary-phase output before forwarding to clients (gated by RESPONSES_PASSTHROUGH_DROP_COMMENTARY, default on) (#6199)
+
+- **fix(sse):** tool-call function schemas with a root `type: null` are now coerced to `type: "object"` before dispatch ([#6359](https://github.com/diegosouzapw/OmniRoute/issues/6359)) — clients like the Codex app emit `parameters: { type: null, ... }` for some tools, which OpenAI-compatible upstreams reject with `400 Invalid schema for function '...': schema must be a JSON Schema of 'type: "object"', got 'type: null'`, failing the whole request. `toolSchemaSanitizer` already stripped the null; it now re-adds the mandatory root `"object"` type (and empty `properties`/open `additionalProperties` when absent). Combinator roots (`anyOf`/`oneOf`/`allOf`) and explicit root types are left untouched. Regression guard: 5 new cases in `tests/unit/tool-schema-sanitizer.test.mjs`.
+
+- **fix(docker):** AgentBridge no longer fails to start on npm/Electron/VPS installs with "MITM manager stub reached at runtime" ([#6344](https://github.com/diegosouzapw/OmniRoute/issues/6344)) — v3.8.45 flipped the production bundler default to Turbopack, but `next.config.mjs` aliased `@/mitm/manager` to its Docker-only degraded stub **unconditionally**. That was harmless while Docker (which sets the alias intentionally for #3390 graceful degradation) was the sole Turbopack consumer, but once every artifact built with Turbopack the stub shipped to all non-Docker users and `startMitm` threw on the first Agent-Bridge start. The alias is now opt-in via `OMNIROUTE_MITM_STUB=1` (set only by the Dockerfile) through the shared `scripts/build/mitm-stub-flag.mjs` helper; default builds bundle the real manager. Regression guard: `tests/unit/mitm-stub-alias-6344.test.mjs` (4).
+
+- **fix(proxy):** stop the v3.8.44 proxy regression that leaked the real IP and disabled healthy proxies ([#6246](https://github.com/diegosouzapw/OmniRoute/issues/6246)). Two coupled defects from the new health scheduler: (1) **IP leak** — when a proxy assigned to a connection was marked `inactive`, resolution fell through to a **direct** egress instead of blocking, exposing the operator's real IP; (2) **over-deactivation** — the sweep flipped a proxy to `inactive` on the **first** failed probe and counted our own 5s timeout / a probe-target `5xx` as the proxy's fault, so healthy paid proxies vanished from egress selection ("my proxies are not being used anymore"). Fix: the sweep decision is extracted into a pure, network-free `decideProxyHealthAction` (`src/lib/proxyHealth/decision.ts`) — by default the health check now **only counts/logs and never downgrades status** (a proxy is downgraded/removed only with `PROXY_AUTO_REMOVE=true`, after `PROXY_AUTO_REMOVE_AFTER` **consecutive** conclusive failures); probes are classified tri-state so an inconclusive result (our timeout, or a `5xx` from the probe target) never penalizes the proxy, and the probe timeout is raised 5s→15s. Separately, `safeResolveProxy` now **fails closed** via the existing policy: a connection whose assigned proxy is dead is blocked instead of leaking direct (`hasBlockingProxyAssignment`), honoring the explicit `proxy off` toggles and the `PROXY_FAIL_OPEN=true` opt-out. Existing proxies stuck `inactive` by the old behavior need a one-time manual re-activate (the operator owns proxy status). Regression guards: `tests/unit/proxy-health-decide-action-6246.test.ts`, `tests/unit/proxy-assigned-unavailable-6246.test.ts`.
+
+- **fix(proxy):** make "Test All" read-only and add bulk enable/disable ([#6246](https://github.com/diegosouzapw/OmniRoute/issues/6246)). Complements the core fail-closed / scheduler fix (#6296) with the two remaining reporter asks. (1) The **"Test All" button** (`POST /api/settings/proxies/auto-test`) used to flip a proxy to `inactive` on a failed reachability probe; since the egress selector excludes `inactive` proxies, a flaky probe (an unreachable `httpbin.org`, a proxy that blocks `HEAD`, or a slow paid proxy) silently disabled every proxy that failed — "Test All" is now **read-only by default** (only the operator sets a proxy active/inactive; opt back into the legacy test-and-set with `PROXY_HEALTH_AUTO_DEACTIVATE=true`). (2) Adds a **bulk enable/disable** proxies endpoint + toolbar action (`POST /api/settings/proxies/batch-activate`) so an operator can re-activate proxies in one click. Regression guard: `tests/unit/proxy-health-6246.test.ts`. (thanks @tenshiak)
+
+- **chatcore (tools): stop the default 128-tool cap from silently dropping opencode's `task`/MCP tools.** opencode (used as an MCP/agent host) sends a large tool list; when it exceeds the speculative `MAX_TOOLS_LIMIT` (128) default, `truncateToolList` did a blind `tools.slice(0, 128)`, dropping every tool past index 128 — including opencode's built-in `task` tool (subagent launch) and many MCP tools, so models routed through OmniRoute could no longer spawn subagents or reach part of their tools. The cap exists to avoid upstream `400`s for providers with real hard limits (e.g. grok-cli 200), so it is kept for those: detection of the opencode client (`isOpencodeClient` — any `x-opencode-*` header, or `opencode` in the user-agent) now only bypasses the **speculative 128 default**, never a known provider ceiling. Precedence is explicit — a proactive/detected provider limit always truncates (even for opencode); otherwise opencode forwards its full tool list; otherwise the unchanged 128 default applies to every other client. Refactors `getEffectiveToolLimit` into `getKnownToolLimit(provider) ?? DEFAULT_LIMIT` (byte-identical for existing callers) and fixes a cosmetic debug-log that reported the truncated count instead of the original. Regression guard: `tests/unit/tool-limit-detector.test.ts`.
+
+- **fix(mitm):** the macOS MITM-cert install check now matches the system keychain again. `security find-certificate -a -Z` prints the SHA-1 as a colon-less hex string, but the installed-check compared it against `getCertFingerprint()`'s colon-separated form, so the substring match never hit — the cert was reported as not-installed and re-prompted for the sudo install on every run. Fingerprints are now normalized (colons stripped, upper-cased) on both sides via the extracted `macCertOutputHasFingerprint` helper. Regression guard: `tests/unit/mitm-cert-mac-fingerprint.test.ts`. ([#6204](https://github.com/diegosouzapw/OmniRoute/pull/6204), closes [#6134](https://github.com/diegosouzapw/OmniRoute/issues/6134) — thanks @rianonehub)
+
+- **fix(api):** `/v1/messages/count_tokens` now counts `tool_use`, `tool_result` and `thinking` content blocks (and array-form `system` prompts) in the local-estimation path, instead of only `text`. Real agentic conversations keep ~95% of their tokens inside tool results; the previous estimate returned near-zero for them, which silently broke Claude Code's auto-compaction (context grew past the window with no compaction until the upstream API rejected the request). The real provider-side count path is unchanged. Regression guard: `tests/unit/messages-count-tokens-route.test.ts`. ([#6221](https://github.com/diegosouzapw/OmniRoute/pull/6221) — thanks @luweiCN)
+
+- **fix(antigravity):** strip a trailing assistant prefill turn for Vertex Claude models to avoid upstream 400s ([#6114](https://github.com/diegosouzapw/OmniRoute/pull/6114)). Regression guard: `tests/unit/antigravity-claude-prefill-strip.test.ts`. (thanks @anki1kr)
+
+- **fix(security):** the mutable cloud-agent routes (`/api/cloud/credentials/update`, `/api/cloud/models/alias`) now require management auth instead of being treated as public. They were classified as public API routes, so a request without management credentials could update stored cloud-agent credentials and model aliases. They are removed from the public-route set, classified as management routes in the authz pipeline, and gated by `requireManagementAuth`; cloud **read**/auth routes stay public. Regression guards: `tests/unit/cloud-write-auth.test.ts`, `tests/unit/authz/classify.test.ts`, `tests/unit/public-api-routes.test.ts`. ([#6233](https://github.com/diegosouzapw/OmniRoute/pull/6233) — thanks @vittoroliveira-dev)
+
+- **refactor(dashboard):** extract the onboarding-wizard "Open provider details" link target into a pure, unit-tested `buildProviderDetailsHref(connection)` helper. The wizard already routes by `connection.id` (the node UUID) rather than the provider category slug (#6144/#6145); this hardens that behavior behind a tested helper that guards a missing id/connection. Regression guard: `tests/unit/provider-onboarding-href.test.ts`. ([#6166](https://github.com/diegosouzapw/OmniRoute/pull/6166) — thanks @KooshaPari)
+
+- **fix(security):** the doubao synthetic device-id generator now derives its digits via an unbiased crypto-random draw (rejection sampling over `crypto.randomBytes()`) instead of a `% 10` reduction, closing a CodeQL `js/biased-cryptographic-random` finding.
+- **fix(agentSkills):** the GitHub-skills generator now resolves `outputDir` to an absolute path before writing, fixing a regression introduced by #6366 (relative-to-cwd base path) that could write generated skill files to the wrong directory.
+- **fix(security):** `/api/keys/{id}/devices` now checks the HTTP method before auth/validation, returning a `405` for non-GET/DELETE verbs instead of a misleading `401`/`500` (closes a `dast-smoke` QUERY-method finding).
+- **fix(quality):** clear the last 2 heavy quality-gate reds on the release tip (cycle pre-flight).
+- **fix(mitm):** the test suite and CI can never mutate the OS trust store — `OMNIROUTE_SKIP_SYSTEM_TRUST=1` is set globally for tests/CI so `installCert`/`uninstallCert`/`installTproxyCa` skip the privileged OS dispatch ([#6310](https://github.com/diegosouzapw/OmniRoute/pull/6310); full detail is under the [3.8.45] section below — this branch received it via the parallel-cycle sync-back).
+- **fix(api):** `POST /api/github-skills` now Zod-validates its request body; documented the new quality-gate env vars and pinned the merge-integrity GitHub Actions to a commit SHA.
+- **fix(skills):** generate the missing `omni-github-skills` registry entry and align the agent-skills catalog-count tests (follow-up to #6186).
+- **fix(quality):** clear the cycle's 11 net-new ESLint errors and make `validate-release-green` suppressions-aware.
+
+### 📝 Maintenance
+
+- **i18n(it):** add 118 missing Italian (`it`) translations (net-additive — no existing keys dropped, valid JSON), improving Italian UI coverage. ([#6212](https://github.com/diegosouzapw/OmniRoute/pull/6212) — thanks @serverless83)
+- **chore(providers):** remove deprecated **MiMo V2** model entries from the catalogs (xiaomi-mimo, opencode-go, zenmux-free, audio TTS) — the upstream V2 line is superseded by MiMo V2.5; drops `mimo-v2-tts`, `mimo-v2-pro`, `mimo-v2-omni`, `mimo-v2-flash`, `mimo-v2-flash-free` and realigns the provider-catalog tests. ([#6248](https://github.com/diegosouzapw/OmniRoute/pull/6248) — thanks @backryun)
+- **chore(release):** ~50 commits on this branch are v3.8.45 pre-flight/hardening fixes and CI-perf work that landed here via the parallel-cycle sync-back (`sync-next-cycle.mjs`, Hard Rule #21) **after** the `v3.8.45` git tag was cut, and are already fully documented under the **[3.8.45]** section below — listed here only so the per-cycle commit-coverage check (`npm run release:uncovered`) doesn't flag them as gaps. Provider/catalog/UX/backend fixes: #6041, #6078, #6108, #6135, #6148, #6149, #6154, #6158, #6161, #6162, #6163, #6164, #6165, #6170, #6177, #6178, #6181, #6186, #6187, #6191, #6193, #6194, #6195, #6200, #6205, #6208, #6209, #6211, #6213, #6223, #6224, #6225, #6226, #6227, #6228, #6229, #6230, #6235, #6291, #6292. CI/release-pipeline work: #6167, #6203, #6214, #6215, #6218, #6273, #6275, #6283, #6284, #6285, #6300, #6305.
+- **chore(release):** additional zero-ref release-cycle plumbing on this branch, kept out of `release:uncovered` on purpose (no `#N` in the commit subject to cite): opening the v3.8.46 cycle, opening/closing the v3.8.45 cycle, the finalized [3.8.45] CHANGELOG i18n sync-back to 42 mirrors, the v3.8.45 cognitive/cyclomatic and file-size drift rebaselines, ESLint stale-suppression pruning (4,273 → 4,233), and clearing test-masking/docs-all pre-flight reds for v3.8.45.
+
+### ⚡ Performance & Infrastructure
+
+- **perf(release-green):** the pre-flight validator (`scripts/quality/validate-release-green.mjs`) now runs its 4 slow suites (unit / vitest / integration / pack-artifact) **concurrently** via `Promise.all` — pre-flight wall time drops from ~the sum of the suites to ~the slowest one (~30min saved per round; Phase 0 was the nº1 bottleneck of the v3.8.45 release benchmark, 2h54 of 6h34 e2e). Guard: `tests/unit/validate-release-green.test.ts` ("runs the slow suites CONCURRENTLY"). ([#6319](https://github.com/diegosouzapw/OmniRoute/pull/6319))
+- **fix(ci):** `scripts/release/sync-next-cycle.mjs` — two defects found live in its first production run (v3.8.45 Phase 5): (1) the `git()` helper's default 1 MiB `maxBuffer` crashed with `ENOBUFS` on `git show origin/main:CHANGELOG.md` (the CHANGELOG alone is >1 MiB) — widened to 64 MiB; (2) the i18n resync only propagated the `[NEXT]` (TBD) section, leaving the just-shipped finalized section as "— TBD" in all 42 mirrors — it now also syncs `[prevVersion]` bounded by the heading below it (new exported pure helper `versionAfter`). Guards: +5 tests in `tests/unit/sync-next-cycle.test.ts` (8/8). ([#6327](https://github.com/diegosouzapw/OmniRoute/pull/6327))
+- **test(ci):** concurrency-sensitive flaky tests are **quarantined into a serial pass** (`tests/unit/serial/`, `--test-concurrency=1`, appended to every unit runner incl. sharded variants — the serial pass is sharded too so concurrent shard jobs never self-collide). Initial set: `glm-coding-plan-monthly-3580`, `quota-division-blocks`, `provider-health-autopilot`, `combo-health-autopilot` — the class behind the ~28min CI wedges/re-runs (two live 1h+ wedges cancelled during this PR's own validation). Discovery + TIA gates track the new glob; systemic root cause (async logger writing after teardown) tracked in [#6360](https://github.com/diegosouzapw/OmniRoute/issues/6360). Guard: `tests/unit/test-serial-quarantine.test.ts` (4). ([#6347](https://github.com/diegosouzapw/OmniRoute/pull/6347))
+
+### 🙌 Contributors
+
+Thanks to everyone whose work landed in v3.8.46:
+
+| Contributor                                                    | PRs / Issues           |
+| -------------------------------------------------------------- | ---------------------- |
+| [@2220258345](https://github.com/2220258345)                   | direct commit / report |
+| [@abdofallah](https://github.com/abdofallah)                   | direct commit / report |
+| [@adentdk](https://github.com/adentdk)                         | direct commit / report |
+| [@aleksesipenko](https://github.com/aleksesipenko)             | direct commit / report |
+| [@anki1kr](https://github.com/anki1kr)                         | direct commit / report |
+| [@anungma](https://github.com/anungma)                         | direct commit / report |
+| [@arpicato](https://github.com/arpicato)                       | direct commit / report |
+| [@backryun](https://github.com/backryun)                       | #6248                  |
+| [@binsarjr](https://github.com/binsarjr)                       | direct commit / report |
+| [@brightfiscalband](https://github.com/brightfiscalband)       | direct commit / report |
+| [@chirag127](https://github.com/chirag127)                     | #6501, #6506           |
+| [@developerjillur](https://github.com/developerjillur)         | direct commit / report |
+| [@dilneiss](https://github.com/dilneiss)                       | #6499                  |
+| [@dtybnrj](https://github.com/dtybnrj)                         | direct commit / report |
+| [@Forcerecon](https://github.com/Forcerecon)                   | direct commit / report |
+| [@hao3039032](https://github.com/hao3039032)                   | direct commit / report |
+| [@Iammilansoni](https://github.com/Iammilansoni)               | #6150, #6245           |
+| [@jmengit](https://github.com/jmengit)                         | direct commit / report |
+| [@jordansilly77-stack](https://github.com/jordansilly77-stack) | direct commit / report |
+| [@JxnLexn](https://github.com/JxnLexn)                         | direct commit / report |
+| [@KooshaPari](https://github.com/KooshaPari)                   | #6166                  |
+| [@loopyd](https://github.com/loopyd)                           | direct commit / report |
+| [@luweiCN](https://github.com/luweiCN)                         | #6221                  |
+| [@makcimbx](https://github.com/makcimbx)                       | direct commit / report |
+| [@muflifadla38](https://github.com/muflifadla38)               | direct commit / report |
+| [@newnol](https://github.com/newnol)                           | direct commit / report |
+| [@ofekbetzalel](https://github.com/ofekbetzalel)               | direct commit / report |
+| [@ohahe52-dot](https://github.com/ohahe52-dot)                 | direct commit / report |
+| [@phidinhmanh](https://github.com/phidinhmanh)                 | direct commit / report |
+| [@powellnorma](https://github.com/powellnorma)                 | direct commit / report |
+| [@qpeyba](https://github.com/qpeyba)                           | direct commit / report |
+| [@RaviTharuma](https://github.com/RaviTharuma)                 | direct commit / report |
+| [@RCrushMe](https://github.com/RCrushMe)                       | direct commit / report |
+| [@rianonehub](https://github.com/rianonehub)                   | #6134, #6204           |
+| [@serverless83](https://github.com/serverless83)               | #6212                  |
+| [@swingtempo](https://github.com/swingtempo)                   | direct commit / report |
+| [@tenshiak](https://github.com/tenshiak)                       | direct commit / report |
+| [@ThongAccount](https://github.com/ThongAccount)               | direct commit / report |
+| [@UnrealAryan](https://github.com/UnrealAryan)                 | direct commit / report |
+| [@vinayakkulkarni](https://github.com/vinayakkulkarni)         | direct commit / report |
+| [@vittoroliveira-dev](https://github.com/vittoroliveira-dev)   | #6233                  |
+| [@warelik](https://github.com/warelik)                         | direct commit / report |
+| [@xxy9468615](https://github.com/xxy9468615)                   | direct commit / report |
+| [@xz-dev](https://github.com/xz-dev)                           | direct commit / report |
+| [@yanpaing007](https://github.com/yanpaing007)                 | direct commit / report |
+| [@diegosouzapw](https://github.com/diegosouzapw)               | maintainer             |
+
+---
+
 ## [3.8.45] — 2026-07-06
 
 ### ✨ New Features

docs/i18n/te/CHANGELOG.md

@@ -6,6 +6,213 @@
 
 ## [3.8.31] — 2026-06-20
 
+## [3.8.46] — TBD
+
+### ✨ New Features
+
+- **feat(sse):** **hide paid-only models from `auto/*` routing** when `hidePaidModels` is on ([#6512](https://github.com/diegosouzapw/OmniRoute/issues/6512)) — follow-up to #6328/#6495. PR #6495 hid paid-only models from the `GET /v1/models` listing, but `auto/*` combos (`auto/best-coding`, `auto/glm`, …) could still pick a paid-only backend into their candidate pool → a 402/403 at request time. `createVirtualAutoCombo` now filters the candidate pool through the new pure `open-sse/services/autoCombo/paidModelFilter.ts` (`filterPaidOnlyCandidates`), applying the same free-model predicate #6495 uses in `catalog.ts` (`providerHasFreeModels(provider) && isFreeModel(provider, {id})`) whenever `settings.hidePaidModels === true`. Applied before the category/tier/family narrowing, so it covers every `auto/*` combo; an all-paid pool degrades to the existing graceful empty-pool path. **Opt-in — default OFF leaves the pool unchanged** (identity). Regression guard: `tests/unit/autoCombo/paid-model-filter-6512.test.ts` (4, incl. the default-off identity guard).
+- **feat(sse):** **provider-family auto combos** — `auto/glm`, `auto/minimax`, `auto/mimo`, `auto/zai`, `auto/gemma`, `auto/llama`, `auto/gemini` ([#6453](https://github.com/diegosouzapw/OmniRoute/issues/6453)) — new routable ids that materialize an on-demand virtual combo spanning whatever installed backends currently expose that model family, degrading gracefully as backends rotate. A new pure `open-sse/services/autoCombo/modelFamily.ts` (`detectModelFamily`) classifies by model-id prefix for six families; `zai` is instead resolved by provider id (z.ai's hosted API serves the same `glm-*` model ids as every other GLM backend, so `auto/zai` means "route to my z.ai backend specifically" vs `auto/glm`'s "any connected GLM backend"). Reuses the existing `createVirtualAutoCombo` on-demand materialization path (no DB writes) and the `/v1/models` catalog advertising loop. Regression guard: `tests/unit/autoCombo/provider-family-combos.test.ts` (11).
+- **feat(proxy):** native **proxy-pool round-robin / egress IP rotation** ([#6365](https://github.com/diegosouzapw/OmniRoute/issues/6365)) — a scope (global / provider / account) can now hold **multiple** proxies as a pool with a rotation strategy, so outbound requests cycle their egress IP instead of pinning one proxy per scope. Migration `117_proxy_pool_rotation.sql` lifts the `UNIQUE(scope, scope_id)` constraint (rebuild via the canonical rename/copy/drop; existing single assignments become 1-element pools) and adds a `proxy_scope_rotation` companion table holding the per-scope strategy + a persisted monotonic round-robin cursor. Strategies: `round-robin` (default, monotonic cursor — never `Math.random`), `random`, and `sticky-per-N-min`. Resolution (`resolveProxyForScopeFromRegistry` / `resolveProxyForConnectionFromRegistry`) now fetches the alive, position-ordered candidate set (unchanged `PROXY_ALIVE_PREDICATE`) and applies the strategy; an empty / all-dead pool still returns `null` — the #6246 fail-closed guard is **untouched** (never falls through to direct egress). Backend + DB only; dashboard pool-builder UI is a follow-up. Regression guard: `tests/unit/proxy-pool-rotation-6365.test.ts` (8, incl. fail-closed + backward-compat).
+- **feat(providers):** end-to-end **tool/function calling on the native Gemini `/v1beta` endpoint** ([#6222](https://github.com/diegosouzapw/OmniRoute/issues/6222)) — both directions of the Gemini↔OpenAI conversion now preserve tool data (previously silently dropped). Request side: `convertGeminiToInternal` (extracted to its own testable module) maps `tools[].functionDeclarations` → OpenAI `tools`, prior `functionCall` parts → assistant `tool_calls`, and `functionResponse` parts → `tool`-role messages. Response side: `convertOpenAIResponseToGemini` emits `parts[].functionCall {name,args}` from `message.tool_calls`, and the streaming `openAIChunkToGeminiChunk` accumulates fragmented `tool_calls` deltas by index into complete `functionCall` parts. The non-Gemini client paths (Claude, OpenAI-Responses) already preserved tool calls — this closes the gap specific to the native Gemini surface. Regression guard: `tests/unit/v1beta-gemini-tool-calling-6222.test.ts` (6, incl. a streaming SSE round-trip).
+- **feat(providers):** copilot-m365-web **enterprise / work tier** support ([#6334](https://github.com/diegosouzapw/OmniRoute/issues/6334)) — mirrors the EDU-tier pattern (#6210): `M365ConnectionParams` gains an `agent` field, a new opt-in `M365_ENTERPRISE_OVERRIDES` preset (`agent=work`, `scenario=officeweb`, `licenseType=Premium`) applies via `providerSpecificData.tier="enterprise"` (alias `"work"`), and `agent` is also overridable directly via `providerSpecificData.agent`. `buildWsUrl` was hardcoding `agent="web"` (the one enterprise-distinguishing param with no override path), so a Premium work account handshook then returned an empty stream. The individual and EDU paths are untouched. Kilo's dup flag vs #6210 (EDU tier) was a false positive — different tier. Regression guard: `tests/unit/copilot-m365-enterprise-6334.test.ts` (7). End-to-end confirmation on a real Premium work account is a live-VPS validation follow-up (Hard Rule #18). (thanks @Forcerecon)
+- **feat(api):** standardized, provider-agnostic **`effort` + `thinking` request params** ([#6241](https://github.com/diegosouzapw/OmniRoute/issues/6241)) — a thin standardization layer over the existing mature per-provider reasoning plumbing (no provider mapper touched). `providerChatCompletionSchema` gains a canonical `effort` (reusing the shared `none/low/medium/high/xhigh` vocabulary — the UI tiers `extra`/`max` collapse onto `xhigh`) and a boolean `thinking`. A pure `normalizeReasoningRequest` (wired once in `src/sse/handlers/chat.ts`, before any reasoning field is read) folds them onto the fields the translators already consume (`reasoning_effort` / `reasoning.effort` / `thinking`), so they fan out to Anthropic / Gemini / xAI / Responses — an explicit client `reasoning_effort` / object-shaped `thinking` always wins (backward-compatible). `/models` additively exposes `supportsThinking` + `effort_tiers` so the frontend can render the toggles (UI component is a follow-up). Regression guard: `tests/unit/effort-thinking-standardization-6241.test.ts` (12). (thanks @Iammilansoni, @shabeer)
+- **feat(combo):** new **`pipeline` (sequential) combo strategy** ([#6297](https://github.com/diegosouzapw/OmniRoute/issues/6297)) — the 18th routing strategy runs targets **in order**, threading each step's output into the next step's input, with an optional per-step `prompt` (system instruction); only the final step's response is returned. Distinct from `fusion` (parallel fan-out + judge). Implemented as a self-contained `open-sse/services/pipeline.ts` (sibling to `fusion.ts`), dispatched from `combo.ts`; the step list reuses `combo.models` order and reads an optional `prompt` off each target (backward-compatible — ignored by every other strategy). Intermediate steps run non-streaming with tools stripped (complete prose to thread forward); the final step keeps the client's `stream` flag + tools. A failing/empty/unparseable intermediate step fails the whole pipeline explicitly via a sanitized error (never silently swallowed). Kilo's dup flag vs #563 was a false positive (that's model→chain selection; this is a sequential chain). Regression guard: `tests/unit/combo-pipeline-strategy.test.ts` (5). (thanks @ofekbetzalel)
+- **feat(ci):** `check:test-masking` now flags **inline-reimplemented prod conditions** ([#6348](https://github.com/diegosouzapw/OmniRoute/issues/6348)) — a new **report-only** subcheck (v2, 6A.10 family) catches the wrong-shape contract test: a test that recomputes the condition under test inline instead of importing/exercising the real function (the #6216 class, where `=== 500` → `>= 500` stayed green because the test re-implemented the branch). For each added/modified test file it warns when the file textually duplicates a ≥3-token conditional from a production file touched in the same PR **and** does not import the symbol/module owning it, via a pure, fixture-tested `findReimplementedConditions()` with an allowlist mirroring `assertReductionAllowlist`. Report-only for now (does not fail the gate) — to be promoted to blocking after a triage cycle. Regression guard: `tests/unit/check-test-masking.test.ts` (45).
+- **feat(sse):** per-connection routing override (native vs CLIProxyAPI) ([#6339](https://github.com/diegosouzapw/OmniRoute/issues/6339)) — the previously-dead `isCliproxyapiDeepModeEnabled` helper is now wired into `resolveExecutorWithProxy`: a single connection can opt itself into the CLIProxyAPI passthrough executor via `providerSpecificData.cliproxyapiMode="claude-native"`, with precedence **connection override > provider `upstream_proxy_config` mode > default**. `resolveExecutorWithProxy` now receives the resolved connection's `providerSpecificData` (threaded from `chatCore.ts`), so one connection can deep-route while the provider's other connections stay native — no DB schema change (the toggle rides in `providerSpecificData`). Also resolves the same-provider mixing ask in #6340. Regression guard: `tests/unit/chatcore-executor-proxy.test.ts` (9). (thanks @RaviTharuma)
+- **feat(dashboard):** "Add session cookie" modal now shows a prominent **"Open ‹host› →"** link to the provider's own site ([#6268](https://github.com/diegosouzapw/OmniRoute/issues/6268)) — every `-web` cookie-session provider (chatgpt-web, claude-web, gemini-web, kimi-web, lmarena, qwen-web, m365-copilot-web, …) renders a one-click external link (opening the provider's login/home page in a new tab) so operators no longer tab away to retype the URL mid-setup. The host resolves from a pure, unit-tested `resolveWebProviderHost()` (prefers `WEB_COOKIE_PROVIDERS[id].website`, falls back to the registry `baseUrl` origin); non-web providers render exactly as before. Kilo's dup flag vs #6265 (modal-too-small-on-1080p) was a false positive — distinct concern. Regression guard: `tests/unit/resolve-web-provider-host.test.ts` (5). (thanks @chirag127)
+- **feat(providers):** add **DigitalOcean** AI (serverless inference) as an OpenAI-compatible API-key provider ([#6373](https://github.com/diegosouzapw/OmniRoute/pull/6373)) — base `https://inference.do-ai.run/v1`, wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/digitalocean/`, `src/shared/constants/providers/apikey/inference-hosts.ts`). Regression guard: `tests/unit/digitalocean-provider.test.ts`. (thanks @newnol)
+- **feat(providers):** add **Huancheng Public API** (`hcnsec`) as an OpenAI-compatible regional provider ([#6410](https://github.com/diegosouzapw/OmniRoute/pull/6410)) — Xinjiang Huancheng Cybersecurity's public LLM platform (base `https://api.hcnsec.cn/v1`, free credits via daily check-ins), wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/hcnsec/`, `src/shared/constants/providers/apikey/regional.ts`). Regression guard: `tests/unit/hcnsec-provider.test.ts`. (thanks @UnrealAryan)
+- **feat(dashboard):** the web-session credential guide now shows an **"Open {host}" link** ([#6316](https://github.com/diegosouzapw/OmniRoute/pull/6316)) to the provider's sign-in site (derived from the provider `website` via `getProviderWebsiteHost`), so you can jump straight to the page where the cookie/session must be captured. Regression guard: `tests/unit/web-session-provider-link-6316.test.ts`. (thanks @jordansilly77-stack)
+- **feat(cerebras):** add the **Gemma 4 31B** model (`gemma-4-31b`) to the Cerebras registry + pricing table ([#6331](https://github.com/diegosouzapw/OmniRoute/pull/6331)). Regression guard extends `tests/unit/t28-model-catalog-updates.test.ts`. (thanks @backryun)
+- **feat(providers):** add **Yuanbao (web)** as a cookie-session provider ([#6196](https://github.com/diegosouzapw/OmniRoute/issues/6196)) — `yuanbao-web` (Tencent Yuanbao, `yuanbao.tencent.com`) with cookie-only auth (`hy_user`/`hy_token` + public agent id), SSE→OpenAI translation incl. `reasoning_content`, exposing DeepSeek V3/R1 + Hunyuan / Hunyuan-T1. Regression guard: `tests/unit/providers-yuanbao-web.test.ts`. `together-web` was **deferred** (no verifiable web-session endpoint — needs a captured request) and `huggingchat-web` **dropped** (the existing `huggingchat` already is a web-cookie provider). (thanks @chirag127)
+- **feat(providers):** route the built-in **agentrouter** through the dynamic Claude-Code wire image ([#6056](https://github.com/diegosouzapw/OmniRoute/issues/6056)) — a small static allow-set (`CC_WIRE_IMAGE_BUILTINS` in `open-sse/services/ccWireImageBuiltins.ts`), consulted by `isClaudeCodeCompatible` / `isClaudeCodeCompatibleProvider` / `applyFingerprint`, makes agentrouter adopt the CC wire-image headers + fingerprint **while guarding the CC baseUrl/auth branches** so it keeps its own registry `baseUrl` and `x-api-key` auth. Regression guard: `tests/unit/agentrouter-cc-wire-image.test.ts` (asserts the wire image is applied AND agentrouter's baseUrl/auth are preserved). Live WAF-acceptance against agentrouter.org is a VPS validation follow-up (Hard Rule #18).
+- **feat(providers):** **bulk-add API keys for Cloudflare Workers AI** ([#6174](https://github.com/diegosouzapw/OmniRoute/issues/6174)) — `cloudflare-ai` is removed from the bulk-add exclusion list and the bulk parser gains a 3-field `name|accountId|apiKey` mode; the bulk route now builds a **per-entry** `providerSpecificData` so each key carries its own `accountId` (fixing the previous shared-object reuse), and both the create + key-validation paths receive it. Regression guard: `tests/unit/bulk-api-key-parser-cloudflare.test.ts`. (thanks @muflifadla38)
+- **feat(dashboard):** routing/settings UX clarity ([#6147](https://github.com/diegosouzapw/OmniRoute/issues/6147)) — (1) weighted combos show the **effective routing share %** next to each weight when weights don't sum to 100 (`WeightTotalBar.tsx`); (2) the status widget's user-facing **"Cloud Sync" label is renamed** to "Remote Settings Sync" (`CloudSyncStatus.tsx`; internal ids/state untouched); (3) built-in providers gain an **opt-in advanced base-URL override** (`isBaseUrlOverrideEligibleProvider`, hidden behind an "Advanced" toggle, reusing the existing `providerSpecificData.baseUrl` persistence — not globally widened). Regression guard: `tests/unit/routing-settings-ux-6147.test.ts`.
+- **feat(combo):** add an option to **disable session stickiness**, per-combo or globally — round-robin / random combos can rotate to a different connection on every request instead of pinning a whole conversation to one connection by its first-message hash. Resolution precedence per-combo `config.disableSessionStickiness` → global `settings.disableSessionStickiness` → default `false` (preserves the #3825 prompt-cache/504 fix); gates **both** stickiness call sites in `open-sse/services/combo.ts`. Exposed as a global toggle (Combo Defaults) and a per-combo Inherit/on/off control. ([#6168](https://github.com/diegosouzapw/OmniRoute/issues/6168)) Regression guard: `tests/unit/combo-disable-session-stickiness.test.ts`. (thanks @RCrushMe)
+- **feat(docker):** add the `OMNIROUTE_NO_SUDO` env flag for root-less / user-namespaced deployments — the MITM cert-trust command path (`resolveSudoSpawn` in `src/mitm/systemCommands.ts`) now strips the leading `sudo` when the flag is truthy, in addition to the existing root / sudo-missing cases, so the Proxy Agent runs without `sudo` (the operator trusts the CA manually, e.g. via `NODE_EXTRA_CA_CERTS`). Argv-array `spawn` preserved — no shell interpolation (Hard Rule #13). ([#6122](https://github.com/diegosouzapw/OmniRoute/issues/6122)) Regression guard: `tests/unit/mitm-systemCommands-no-sudo.test.ts`. (thanks @powellnorma)
+- **feat(providers):** add **Requesty** as an OpenAI-compatible gateway provider (BYOK, base `https://router.requesty.ai/v1`, ~200 free requests/day) — wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/requesty/`, `src/shared/constants/providers/apikey/gateways.ts`). ([#6120](https://github.com/diegosouzapw/OmniRoute/issues/6120)) Regression guard: `tests/unit/requesty-provider.test.ts`. (thanks @chirag127)
+- **feat(dashboard):** add **configured-only / available-only filters** to the Free Provider Rankings page ([#6150](https://github.com/diegosouzapw/OmniRoute/issues/6150)) — hide providers you haven't configured, or whose connections are all rate-limited / out of quota, via server-side query params (`?configuredOnly` / `?availableOnly` on `GET /api/free-provider-rankings`) backed by a testable lib helper reusing the in-process connection state (no Redis). Both filters default off, so the default view is unchanged; this supersedes the earlier client-side "Configured Only" toggle (#6245) with an available-only dimension and unit-tested logic. Regression guard: `tests/unit/freeProviderRankings-filters.test.ts`.
+
+### 🔧 Bug Fixes
+
+- **fix(dashboard):** adding a second API key connection for the same provider no longer silently overwrites the first — the Add-API-key modal now derives a unique default connection name (`main`, then `main-2`, `main-3`, …) so the backend name-based upsert can't collide (#6499 — thanks @dilneiss).
+- **fix(compression):** the session-dedup engine now also deduplicates a large multi-line block repeated **within a single message** (intra-message dedup), not just across turns; the compression-preview API surfaces a `fallbackReason`, and the fusion panel reports how many models were rate-limited vs failed on total-panel failure (#6501 — thanks @chirag127).
+- **fix(compression):** a stacked-pipeline step naming an unregistered engine now surfaces a `validationErrors` entry instead of silently no-op'ing, so misconfigured pipelines are visible in the preview API (#6506 — thanks @chirag127).
+- **feat(usage):** add a **Codex reset-credit redemption** flow to the Provider Limits UI ([#6361](https://github.com/diegosouzapw/OmniRoute/pull/6361)) — a `useCodexResetCreditRedemption` hook + `/api/usage/codex-reset-credit` route + `codexResetCredits` lib let you redeem banked Codex reset credits from the quota card. Regression guard: `tests/unit/codex-reset-credits.test.ts`. (thanks @JxnLexn)
+- **feat(glm):** add **team-plan quota settings** for `glm-cn` connections ([#6351](https://github.com/diegosouzapw/OmniRoute/pull/6351)) — a dedicated `GlmTeamQuotaFields` form section (team quota id / limits) threaded through the Add/Edit connection modals, persisted via `providerSpecificData`, with the GLM usage service reading the team quota. Regression guards: `tests/unit/glm-team-quota.test.ts`, `provider-specific-data-schema.test.ts`. (thanks @hao3039032)
+- **feat(providers):** add **TinyFish** web-fetch/search support ([#6349](https://github.com/diegosouzapw/OmniRoute/pull/6349)) — a `tinyfish-fetch` executor + `/v1/web/fetch` route + MCP web-fetch tool, registered as a specialty-media provider with request-validation and a search-provider catalog entry. Regression guards: `tests/unit/executor-tinyfish-fetch.test.ts`, `web-fetch-handler.test.ts`, `mcp-web-fetch-tool.test.ts`, `provider-validation-tinyfish.test.ts`. (thanks @dtybnrj)
+- **fix(cli):** `omniroute launch-codex` now spawns `codex.cmd` through a shell on Windows (the npm `.cmd` shim is unresolvable by bare `spawn` → ENOENT), mirroring the qodercli Windows fix (#6263) ([#6312](https://github.com/diegosouzapw/OmniRoute/pull/6312)). Regression guard: `tests/unit/launch-codex-windows-spawn-6312.test.ts`. (thanks @swingtempo)
+- **fix(codex):** isolate the **Spark** quota from the shared Codex quota and stabilize the quota UI ordering / hydration so per-scope limits render consistently ([#6336](https://github.com/diegosouzapw/OmniRoute/pull/6336)). Regression guards: `tests/unit/codex-quota-selection-hydration.test.ts`, `provider-limits-ui.test.ts` + 3 more. (thanks @xz-dev)
+- **feat(api):** add a `hidePaidModels` setting that filters paid-only models out of the `/v1/models` catalog. Regression guard: `tests/unit/models-catalog-hide-paid.test.ts`. (thanks @chirag127)
+- **fix(api-manager):** the fallback model picker now preserves combos instead of dropping them when a primary model is unavailable ([#6443](https://github.com/diegosouzapw/OmniRoute/pull/6443)). Regression guard: `tests/unit/api-manager-page-static.test.ts`. (thanks @jmengit)
+- **fix(providers):** recoverable Antigravity / Cloud-Code (Gemini Code Assist) `403` responses ([#6452](https://github.com/diegosouzapw/OmniRoute/pull/6452)) are now classified as a retryable project-config error instead of a terminal account ban, so a fixable project/API-disabled 403 no longer forces a ~1-year cooldown / full OAuth reconnect. Regression guard: `tests/unit/errorclassifier-antigravity-403.test.ts`. (thanks @developerjillur)
+- **fix(mitm):** `sanitizeHeaders` now redacts `Set-Cookie` response headers so upstream session cookies never leak into logs / diagnostics ([#6451](https://github.com/diegosouzapw/OmniRoute/pull/6451)). Regression guard: `tests/unit/mitm-sanitize-headers.test.ts`. (thanks @developerjillur)
+- **fix(api):** `/api/compression/preview` now accepts `mode: "caveman"` and correctly handles stacked / zero-compression previews ([#6425](https://github.com/diegosouzapw/OmniRoute/issues/6425)). Regression guard: `tests/unit/api/compression-preview-caveman-and-stacked-6425.test.ts`. (thanks @chirag127)
+- **feat(providers):** add **Zed** hosted LLM aggregator as a native-app provider ([#6118](https://github.com/diegosouzapw/OmniRoute/pull/6118)) — OAuth sign-in via the Zed hosted flow, registered through the shared provider registry + executor. Regression guards: `tests/unit/zed-oauth-provider.test.ts`, `zed-import-utils.test.ts`, `zed-docker-detect.test.ts`, `mitm-handler-zed.test.ts`. VPS-validated via live operator login (Hard Rule #18).
+- **fix(oauth):** the Kiro SSO-cache auto-import now **preserves the IDC region** — cross-region Amazon Q / Kiro profiles imported from the SSO cache are no longer collapsed to the default region ([#6113](https://github.com/diegosouzapw/OmniRoute/pull/6113)). Regression guard: `tests/unit/kiro-auto-import-idc-2059.test.ts`. VPS-validated via live operator login (Hard Rule #18).
+- **fix(dashboard):** passthrough model aliases no longer collide when two namespaced model ids share a last segment (port from 9router#1850, [#6431](https://github.com/diegosouzapw/OmniRoute/pull/6431)). `enx/gpt-5.5` and `enx/codebuddy/gpt-5.5` both auto-generated the alias `gpt-5.5`, so the second model could never be added (the UI just alerted "alias already exists"). Aliases are now disambiguated deterministically — bare last segment when free, then parent-qualified (`codebuddy-gpt-5.5`), then a numeric suffix — while re-adding the exact same model id is still blocked. Regression guard: `tests/unit/passthrough-alias-1850.test.ts`. (thanks @arpicato)
+- **fix(translator):** preserve a Gemini `functionResponse` co-located with other parts (another `functionCall`, or trailing `text`) in the same content when translating **Gemini → OpenAI** ([#6376](https://github.com/diegosouzapw/OmniRoute/pull/6376)). `convertGeminiContent()` early-returned the tool message on the first `functionResponse` part, dropping any co-located parts; such contents are now pre-split (one tool message per `functionResponse`, emitted first, plus one message for the remaining parts). Regression guard: `tests/unit/gemini-to-openai-function-response.test.ts`. (thanks @warelik)
+- **fix(headroom):** detect a python interpreter managed by **mise / pyenv / asdf / conda** (port from 9router#2353, [#6382](https://github.com/diegosouzapw/OmniRoute/pull/6382)). Headroom's python probe (`src/lib/headroom/detect.ts`) searched a hardcoded `PATH`, but version managers expose their interpreters via shim dirs that only join `PATH` through interactive-shell activation — which the non-interactive server never runs, so a managed python (≥3.10) was invisible and Headroom reported it missing. The search path now prepends the well-known shim/bin dirs (`~/.local/share/mise/shims`, `~/.pyenv/shims`, `~/.asdf/shims`, `$CONDA_PREFIX/bin`, `~/.local/bin`, respecting `MISE_DATA_DIR`/`PYENV_ROOT`/`ASDF_DATA_DIR` when set), and a new `HEADROOM_PYTHON` env override lets operators point straight at their interpreter (mirroring `HEADROOM_URL`). Still shell-free (`execFileSync`). Regression guard: `tests/unit/headroom-detect.test.ts` (5). (thanks @loopyd)
+- **fix(executors):** strip the OpenAI-Codex/Claude-CLI `client_metadata` passthrough field for **NVIDIA** requests (port from 9router#1887, [#6411](https://github.com/diegosouzapw/OmniRoute/pull/6411)). NVIDIA's OpenAI-compatible wrapper rejects it with `400 Unsupported parameter`, the same class already handled for `cerebras`/`mistral`; `nvidia` (executor `default`) was missing from the strip allowlist so Codex/Claude-Code passthrough requests 400'd. Regression guard: `tests/unit/executor-default-strip-client-metadata.test.ts` (+nvidia case). (thanks @phidinhmanh)
+- **fix(translator):** strip the Claude-style `thinking` field for **NVIDIA `z-ai/glm-5.2`** (port from 9router#2023, [#6413](https://github.com/diegosouzapw/OmniRoute/pull/6413)). NVIDIA's OpenAI-compatible wrapper 400s on `thinking` (a Claude-format client routed here leaves a `thinking:{type:"adaptive"}`); the existing strip rule only dropped `reasoning`. Same class already handled for `minimax-m2.7`. Regression guard: `tests/unit/nvidia-minimax-thinking-strip.test.ts` (+glm-5.2 case). (thanks @phidinhmanh)
+- **fix(translator):** suppress the streamed `</think>` close marker for the **Antigravity IDE** client (port from 9router#1061, [#6415](https://github.com/diegosouzapw/OmniRoute/pull/6415)). On thinking-only turns Antigravity rendered a bare `</think>` as the sole visible content, tripping its loop-detection and wasting requests. Antigravity's UA (`vscode/<v> (Antigravity/<v>)`) is added to the marker-suppress allowlist (alongside OpenCode); Claude Code / Cursor still get the marker, and `x-omniroute-thinking-marker: on` force-restores it. Regression guard: `tests/unit/think-close-marker-suppress-5245.test.ts`. (thanks @abdofallah)
+- **fix(executors):** strip nested `reasoning_content` from messages for **Mistral** (port from 9router#1649, [#6417](https://github.com/diegosouzapw/OmniRoute/pull/6417)). Mistral's API returns `422 extra_forbidden` when an assistant message carries `reasoning_content` (replayed thinking from a prior turn, e.g. via the Codex `/responses` path); the generic top-level 400 field-downgrade retry never covered the nested per-message field. `DefaultExecutor` now strips it for provider `mistral` only, so DeepSeek (which requires replayed `reasoning_content`) is unaffected. Regression guard: `tests/unit/mistral-strip-reasoning-content-1649.test.ts`. (thanks @xxy9468615)
+- **fix(executors):** strip the `client_metadata` passthrough field on the **OpenCode** path (port from 9router#1442, [#6418](https://github.com/diegosouzapw/OmniRoute/pull/6418)). OpenCode upstreams (e.g. `kimi-k2.6` via opencode-go) reject it with `400 "Extra inputs are not permitted, field: 'client_metadata'"`; the DefaultExecutor strip only covered cerebras/mistral and `OpencodeExecutor` extends `BaseExecutor` directly, so nothing removed it there. Regression guard: `tests/unit/opencode-strip-client-metadata-1442.test.ts`. (thanks @yanpaing007)
+- **fix(executors):** inject the `reasoning_content` echo for the native **Moonshot Kimi** provider (port from 9router#1480, [#6419](https://github.com/diegosouzapw/OmniRoute/pull/6419)). Kimi (executor `default`) is a thinking-mode upstream that 400s with "reasoning_content must be passed back" when a prior assistant turn lacks it; the placeholder injection was only wired into the OpenCode meta-provider, so direct multi-turn Kimi conversations failed. Scoped to `kimi` (gateway-served models matching the thinking-model name pattern are unaffected). Regression guard: `tests/unit/kimi-native-reasoning-injected-1480.test.ts`. (thanks @2220258345)
+- **fix(executors):** recover from a strict gateway's `context_management: Extra inputs are not permitted` 400 (port from 9router#1468, [#6420](https://github.com/diegosouzapw/OmniRoute/pull/6420)). **Claude Code** always sends a top-level `context_management` field; strict anthropic-compatible gateways reject it. The dedicated context-editing 400-fallback only fired when OmniRoute's own `contextEditing` feature was enabled (default off), so a client-sent field passed through untouched and 400'd. `context_management` is now in the generic reactive field-strip list, so it's stripped-and-retried once regardless of the feature flag (with correct request re-signing for claude-compatible relays). Regression guard: `tests/unit/provider-field-strips.test.ts`. (thanks @ohahe52-dot)
+- **fix(network):** enable **RFC 8305 Happy Eyeballs** (`autoSelectFamily`) on the direct-egress undici dispatcher (port from 9router#1237, [#6423](https://github.com/diegosouzapw/OmniRoute/pull/6423)). When DNS returns both IPv6 (AAAA) and IPv4 (A) and the IPv6 route is broken (e.g. a NAT64 `64:ff9b::` prefix without routing), undici tried IPv6 first and hung until `ETIMEDOUT` (then a 502 + account lockout), even though `curl` reached the same host. The direct dispatcher now races both families and uses whichever connects first. Proxy paths pin family via `proxyTls` and are unaffected. Regression guard: `tests/unit/direct-dispatcher-pipelining-4580.test.ts`. (thanks @adentdk)
+- **fix(combo):** round-robin now advances the rotation pointer past the model that **actually served**, not the eagerly-scheduled one (port from 9router#948, [#6428](https://github.com/diegosouzapw/OmniRoute/pull/6428)). With `stickyLimit: 1` (true round-robin), when the scheduled model failed and a _different_ model served via fallback, the counter had already advanced +1 from the scheduled index — so the next request reused the fallback-served model, degrading round-robin into hot-spotting on whichever model was healthy. The pointer now advances to the served index + 1 (mirroring the sticky-limit>1 path). Session-stickiness (#3825) and distribution are preserved. Regression guard: `tests/unit/combo-rr-fallback-advance-948.test.ts`. (thanks @binsarjr)
+- **fix(sse):** a non-string `model` field is now rejected with a `400` before the resolver, instead of crashing downstream `.toLowerCase()`/`.split()` calls into an empty-body `500` that escapes the error sanitizer ([#6407](https://github.com/diegosouzapw/OmniRoute/issues/6407)). Regression guard: `tests/unit/chat-non-string-model-6407.test.ts`. (thanks @chirag127)
+- **fix(api):** unknown `/api/*` routes now return a JSON `404` (instead of the dashboard HTML shell) and scalar chat params (`model`/`temperature`/etc.) are validated **before** the provider lookup so malformed requests fail fast with a clear `400` ([#6424](https://github.com/diegosouzapw/OmniRoute/issues/6424), [#6412](https://github.com/diegosouzapw/OmniRoute/issues/6412)). Regression guards: `tests/unit/api/api-catchall-json-404.test.ts`, `tests/unit/chat-early-schema-validation-6412.test.ts`. (thanks @chirag127)
+- **fix(api):** `/v1/chat/completions` now rejects a non-JSON `Content-Type` with a `400` before parsing the body ([#6414](https://github.com/diegosouzapw/OmniRoute/issues/6414)). Regression guard: `tests/unit/v1-chat-completions-content-type-6414.test.ts`. (thanks @chirag127)
+- **fix(api):** the `X-OmniRoute-Compression` response header is now echoed on `/v1/chat/completions` and `/v1/completions` ([#6422](https://github.com/diegosouzapw/OmniRoute/issues/6422)). Regression guard: `tests/unit/compression-header-echo-6422.test.ts`. (thanks @chirag127)
+- **fix(api):** concurrent `GET /v1/models` requests are coalesced into a single catalog build ([#6408](https://github.com/diegosouzapw/OmniRoute/issues/6408)). Regression guard: `tests/unit/v1-models-concurrent-6408.test.ts`. (thanks @chirag127)
+- **fix(api):** `/v1/completions` now echoes the requested `body.model` in its JSON + streamed responses ([#6429](https://github.com/diegosouzapw/OmniRoute/pull/6429)). Regression guard: `tests/unit/completions-body-model-echo.test.ts`. (thanks @chirag127)
+- **fix(api):** env-var master keys now see the full `/v1/models` catalog ([#6406](https://github.com/diegosouzapw/OmniRoute/issues/6406)). Regression guard: `tests/unit/models-catalog-envkey-6406.test.ts`. (thanks @chirag127)
+- **fix(api):** non-streaming `/v1/completions` responses now echo `body.model` aligned with the `X-OmniRoute-Model` header ([#6426](https://github.com/diegosouzapw/OmniRoute/issues/6426)). Regression guard: `tests/unit/v1-completions-model-header-match-6426.test.ts`. (thanks @chirag127)
+- **fix(api):** unknown `/v1/*` routes now return a JSON `404 not_found` instead of the Next.js dashboard HTML shell ([#6405](https://github.com/diegosouzapw/OmniRoute/issues/6405)). Regression guard: `tests/unit/api/v1-catchall-json-404.test.ts`. (thanks @chirag127)
+- **fix(api):** the per-connection provider models route now degrades to the shipped catalog when a provider's `/models` endpoint answers with a redirect ([#6267](https://github.com/diegosouzapw/OmniRoute/issues/6267)) — a `qwen-web` import failed with a raw `Redirect blocked … (307)` 503. `safeOutboundFetch` throws `REDIRECT_BLOCKED` on the 307, `getSafeOutboundFetchErrorStatus` maps it to 503, and `buildDiscoveryErrorFallbackResponse` treated every 503 as a hard error — so the non-empty `getModelsByProviderId("qwen-web")` catalog was never surfaced. A models-endpoint redirect is not a fixable-config error (unlike `URL_GUARD_BLOCKED`/`INVALID_URL`, which stay hard errors), so it now falls back to the local/cached catalog before the 503 short-circuit. General fix — covers any config-driven provider that 307s. Regression guard: `tests/unit/provider-models-qwen-web-redirect-6267.test.ts`. (thanks @chirag127)
+- **fix(api):** the per-connection provider models route (MCP `list_models_catalog` + the dashboard import view) now merges USER-ADDED custom models into its response ([#6247](https://github.com/diegosouzapw/OmniRoute/issues/6247)) — custom models live in the `key_value` namespace `customModels`, which the live REST `/api/v1/models` already merges, but `src/app/api/providers/[id]/models/route.ts` never read `getCustomModels`, so custom models were dropped on both the discovery-success and local_catalog paths. They are now folded into the returned model list (deduped by id, stamped `owned_by: provider`), fixing MCP + the dashboard import view in one place. Regression guard: `tests/unit/provider-models-custom-merge-6247.test.ts`. (thanks @RCrushMe)
+
+- **fix(providers):** GitLab Duo tool-calling follow-up turns no longer fail upstream with `422 {"detail":"Validation error"}` (tokens 0/0, rejected pre-inference). The #6234 tool-result-feedback fix serialized the **entire** multi-turn conversation into GitLab's single-file `code_suggestions` (`small_file`) generation endpoint — folded history that turn-N sent as an oversized `current_file.content_above_cursor` **and** duplicated verbatim into `user_instruction`, tripping the AI-Gateway's `small_file` validation guard. The executor now **bounds** that prompt: it keeps system + latest user message + the most-recent tool round (dropping older turns), caps oversized tool results, and stops duplicating the full prompt into `user_instruction` (which now carries only the short latest user message) — while still feeding the most-recent tool result back so the agent continues (`open-sse/executors/gitlab.ts`, [#6220](https://github.com/diegosouzapw/OmniRoute/issues/6220)). The unit test covers the bounding logic; the upstream 422→200 clearing is VPS-only (Hard Rule #18). Regression guard: `tests/unit/gitlab-tool-exchange-bounded-6220.test.ts`.
+
+- **fix(i18n):** the provider-detail (`/dashboard/providers/[id]`) connection-status filter labels no longer render as `__MISSING__:All` / `__MISSING__:Active` / `__MISSING__:Error` / `__MISSING__:Banned` / `__MISSING__:CreditsExhausted` in non-English locales (notably pt-BR) ([#6290](https://github.com/diegosouzapw/OmniRoute/issues/6290)). Root cause was **not** the namespace mismatch the issue guessed — the `providers.filter*` keys resolve correctly in `en.json`; the debt lived in the locale mirrors (`src/i18n/messages/*.json`), where these five keys carried the `__MISSING__:` sync sentinel in ~15 locales and were absent entirely in ~26 others, so next-intl found the key and echoed the sentinel verbatim. All 40 non-English/-Chinese mirrors now ship real translations for the five `providers.filter*` labels. Regression guard: `tests/unit/i18n-provider-filter-keys-6290.test.ts`. (thanks @diegosouzapw)
+
+- **fix(providers):** the `copilot-m365-web` streaming executor now emits `debug`-level WebSocket diagnostics ([#6210](https://github.com/diegosouzapw/OmniRoute/issues/6210)) — the outbound WS URL (with the `access_token` **redacted** via `redactWsUrl()`), handshake success/failure, and each received SignalR frame's `type`/`target`. Previously the streaming path logged nothing, so an empty `content:null` response (the M365 Education / Starter tier symptom fixed in #6234) was undiagnosable even at `APP_LOG_LEVEL=debug`. The change is debug-level and side-effect-free — it does not alter streaming behavior or the frame parser, and the token never reaches the logs. Regression guard: `tests/unit/copilot-m365-web-logging-6210.test.ts` (thanks @qpeyba)
+
+- **fix(resilience):** a round-robin combo no longer returns `503 all upstream accounts are unavailable` when a compatibility-rejected target is actually healthy ([#6238](https://github.com/diegosouzapw/OmniRoute/issues/6238)). `filterTargetsByRequestCompatibility` drops request-incompatible targets (tool/vision/structured-output unsupported, or below the required context window) **before** any availability check runs, and its `compatible.length === 0` safety net only fired when _all_ targets were filtered — not when the kept targets later all turned out runtime-unavailable (circuit-open / cooldown / no credentials). So a combo could 503 while a compat-rejected-but-healthy provider sat unused. `handleRoundRobinCombo` now keeps the compat-rejected set and, when every compat-kept target was skipped without a single real attempt, probes those rejected targets as a **last-resort fallback tier** (via the new pure `open-sse/services/combo/comboCompatFallback.ts`) before crystallizing the 503. Regression guard: `tests/unit/combo-roundrobin-compat-fallback-6238.test.ts`. (thanks @ThongAccount)
+
+- **fix(startup):** best-effort self-heal for a corrupted Turbopack dev cache on Windows ([#6289](https://github.com/diegosouzapw/OmniRoute/issues/6289)). On Windows, `pnpm dev` can fail at startup when Turbopack `mmap`s a persistent-cache SST file and the OS refuses the mapping (`os error 1455` — "paging file too small"), which Turbopack surfaces as a misleading `Module not found: Can't resolve '@/shared/utils/machine'`. This is a **known upstream Turbopack cache-corruption bug — not our code**. The dev launcher (`scripts/dev/run-next.mjs`) now wraps `nextApp.prepare()` and, when it rejects with that signature (`isTurbopackCacheCorruption` in the new `scripts/dev/turbopackCacheHeal.mjs`), purges `.build/next/**/cache/turbopack` and retries **once** with a clear log. **Caveat — best-effort only:** the corruption often surfaces as a runtime overlay rather than a `prepare()` rejection, so this cannot always intercept it; the reliable remedy remains manually deleting the Turbopack cache dir. Regression guard: `tests/unit/turbopack-cache-heal-6289.test.ts`. (thanks @chirag127)
+
+- **fix(providers):** qodercli PAT auth no longer fails with `spawn qodercli ENOENT` on Windows ([#6263](https://github.com/diegosouzapw/OmniRoute/issues/6263)) — `spawnQoderCli` spawned the bare `qodercli` name with `shell:false` and an unenriched env, so the npm `.cmd` wrapper under `%APPDATA%\npm` (a user-PATH directory) was never resolved. It now resolves the absolute `.cmd`/`.exe` path through the existing `getCliRuntimeStatus("qoder")` resolver in `src/shared/services/cliRuntime.ts` (memoized), spawns with `shell` when the target is a `.cmd`/`.bat`, and uses the cliRuntime-enriched env (PATH + PATHEXT + APPDATA); the ENOENT error now lists the searched paths plus the `CLI_QODER_BIN` override. End-to-end spawn on a real Windows host is host-only (Hard Rule #18); the path-resolution logic is unit-tested. Regression guard: `tests/unit/qodercli-windows-resolve-6263.test.ts`. (thanks @chirag127)
+
+- **fix(sse):** the reasoning-token buffer no longer inflates **probe-sized `max_tokens`** ([#6274](https://github.com/diegosouzapw/OmniRoute/issues/6274)) — Claude Code's `/model` capability check sends `max_tokens: 1`, but for a thinking-capable model with a large output cap (e.g. `glm-5.2`) the #3587 headroom heuristic (`max(current + 1000, ceil(current * 1.5))`) rewrote it to `1001` and forwarded that upstream, wasting tokens on a request that was never a genuine reasoning budget. `resolveReasoningBufferedMaxTokens()` (`open-sse/services/reasoningTokenBuffer.ts`) now short-circuits and returns the caller's value verbatim when it is below the new `REASONING_BUFFER_MIN_TRIGGER` (256) threshold — a tiny explicit limit is a probe, not a reasoning request. Real budgets still receive the #3587 headroom unchanged, and the guard runs after the existing capability checks so unknown / non-reasoning models keep returning `null`. Regression guard: `tests/unit/reasoning-token-buffer-6274.test.ts`. (thanks @brightfiscalband)
+
+- **fix(cli):** `omniroute reset-password` now works as a real subcommand, and password resets over piped (non-TTY) stdin actually apply ([#6261](https://github.com/diegosouzapw/OmniRoute/issues/6261), [#6258](https://github.com/diegosouzapw/OmniRoute/issues/6258)). Two coupled defects: (1) **#6261** — `bin/omniroute.mjs` routed everything through Commander with only two pre-Commander bypasses (`--mcp`, `reset-encrypted-columns`), so `omniroute reset-password` was rejected as an unknown command; only the separate `omniroute-reset-password` bin worked, while the docs falsely advertised the subcommand (incl. a bogus "legacy alias still works"). A pre-Commander bypass mirroring `reset-encrypted-columns` now dynamically imports `bin/reset-password.mjs` (which self-executes) before Commander parses; the three doc lines were corrected. (2) **#6258** — `bin/reset-password.mjs` issued two sequential `rl.question` prompts; under piped stdin the second read never settled at EOF, so `main()` never reached `resetManagementPassword` and the reset was a silent no-op (both prompts printed, no success, password unchanged). The CLI now detects non-TTY stdin and reads it once (first line = password, second line = confirm if present, else reused), adds a `--password-stdin` flag (entire stdin is the password, no confirmation), and exits `0` explicitly so the success line always flushes; interactive TTY behavior is unchanged. Regression guard: `tests/unit/reset-password-cli-6261-6258.test.ts` (3). (thanks @chirag127)
+
+- **fix(db):** the mass-migration **safety abort** now tells the operator how to bypass it and stops flooding the log ([#6260](https://github.com/diegosouzapw/OmniRoute/issues/6260)) — after restoring a backup that wiped the migration tracking table, `runMigrations()` threw the abort on every downstream `ensureDbInitialized()`, re-logging the full banner 11+ times, and the message never mentioned the existing `OMNIROUTE_MAX_PENDING_MIGRATIONS` escape hatch. The abort text now appends a bypass hint (set `OMNIROUTE_MAX_PENDING_MIGRATIONS=0` in `server.env` / `DATA_DIR/.env`), and a new `MigrationSafetyAbortError` is memoized so repeated calls in the same process throw the same instance and emit a single concise line instead of the full cascade. Regression guard: `tests/unit/migration-safety-abort-6260.test.ts`. (thanks @chirag127)
+
+- **fix(auth):** importing a **distinct** Codex/ChatGPT OAuth `auth.json` is no longer falsely rejected as "already exists" when it belongs to a different user in the same workspace ([#6301](https://github.com/diegosouzapw/OmniRoute/issues/6301)). `findExistingCodexConnection` (in `src/lib/oauth/utils/codexAuthImport.ts`) deduped **only** on `providerSpecificData.workspaceId === accountId`, where `accountId` is the shared `chatgpt_account_id`/`tokens.account_id` — so two members of the same ChatGPT Team collapsed onto a single connection (409 `duplicate_account`). The id_token's `https://api.openai.com/auth` claim carries a per-user `chatgpt_user_id` alongside the workspace id (the device-flow path already persisted it as `chatgptUserId`, but the import path did not). Now `parseAndValidateCodexAuth` extracts `userId` (`chatgpt_user_id` → `user_id` → JWT `sub`) into `ParsedCodexAuth`, the create/update paths persist `chatgptUserId` in `providerSpecificData` (mirroring `codex.ts`), and dedup keys on `workspaceId` **AND** `chatgptUserId` — with a backward-compat fallback to legacy accountId-only matching when no stored connection for that workspace records a `chatgptUserId`, so genuinely-same accounts still dedup. Regression guard: `tests/unit/codex-auth-import-userid-dedup-6301.test.ts` (4). (thanks @anungma)
+
+- **fix(providers):** importing models for the **venice-web** provider no longer fails with a red "Provider venice-web does not support models listing" ([#6269](https://github.com/diegosouzapw/OmniRoute/issues/6269)). `venice-web` is a web-cookie provider with an executor but no upstream `/v1/models` endpoint and no registry `models`, so the models route fell through to the tail `400`. Mirroring the `jules`/`linkup-search`/`ollama-search` fix (#5569), it now ships a static local catalog entry in `src/lib/providers/staticModels.ts` — seeding the current Venice lineup (`venice-uncensored`, `llama-3.3-70b`, `qwen3-235b`, `qwen3-4b`, `deepseek-r1-671b`; Venice rotates its catalog, see docs.venice.ai/models/overview) — so the route returns `200` with `source:"local_catalog"`, `intentional:true`. Regression guard: `tests/unit/static-models-venice-web-6269.test.ts`. (thanks @chirag127)
+
+- **fix(api):** the specialty model catalogs (`/v1/embeddings`, `/v1/images`, `/v1/music`, `/v1/videos` model lists) are now derived from the **unified catalog filtered by a predicate** (`getSpecialtyModelsResponse`) instead of ad-hoc per-route logic, so they consistently respect active-credential visibility and stay in sync with the main catalog ([#6303](https://github.com/diegosouzapw/OmniRoute/pull/6303)). Regression guard: `tests/unit/specialty-model-catalog-routes.test.ts`. (thanks @makcimbx)
+- **fix(api):** the agent-bridge server route now resolves the MITM manager via a **dynamic `import("@/mitm/manager.runtime")`** so Turbopack does not statically pull the stub (or over-bundle the manager), and the agent-skills generator anchors its output base path with `path.join(process.cwd(), …)` so Turbopack's static analyzer stops tracing the whole project root ([#6329](https://github.com/diegosouzapw/OmniRoute/issues/6329), [#6366](https://github.com/diegosouzapw/OmniRoute/pull/6366)). Regression guard: `tests/unit/agent-bridge-server-route-dynamic-import.test.ts`. (thanks @Iammilansoni)
+- **fix(api):** internal probes (combo-test, cloud-sync verify) now pick a **management-scoped / allow-all API key** instead of naively grabbing `getApiKeys()[0]` — a restricted `self:usage` first row made the probe fail with "Model X is not allowed for this API key" even when the combo path was healthy (`pickApiKeyForInternalUse` in `src/lib/db/apiKeys.ts`). The API-manager model editor also falls back to `/api/models?all=true` when `/v1/models` is catalog-protected ([#6372](https://github.com/diegosouzapw/OmniRoute/pull/6372)). Regression guard: `tests/unit/pick-internal-api-key-6372.test.ts`. (thanks @jmengit)
+- **fix(live-ws):** the Live Dashboard WebSocket server now **rejects on bind failure** (e.g. `EADDRINUSE` when the API bridge already holds the port) instead of letting the error surface as an unhandled `error` event that crash-loops the process — the `error` listener is attached to `wss` (not `server`) and releases the EventBus subscription on a failed start ([#6324](https://github.com/diegosouzapw/OmniRoute/issues/6324)). Regression guard: `tests/unit/live-ws-eaddrinuse-6324.test.ts`. (thanks @vinayakkulkarni)
+- **fix(dashboard):** the Home provider-topology widget now trusts the live provider-metrics snapshot — it uses `topology.errorProvider` and live `activeRequests` directly instead of re-deriving state from a stale `lastErrorAt` or applying a frontend timeout filter, so the topology reflects real-time provider health ([#6322](https://github.com/diegosouzapw/OmniRoute/pull/6322)). Regression guard: `tests/unit/home-provider-topology-live-state.test.ts`. (thanks @xz-dev)
+- **fix(sse):** strip zero-width markers from streamed **tool-call arguments** — a follow-up to [#5857](https://github.com/diegosouzapw/OmniRoute/pull/5857). That PR removed injected zero-width joiners (U+200D) from streamed assistant text/reasoning but deliberately left tool-call argument JSON byte-exact. The request-side obfuscation (`open-sse/services/claudeCodeObfuscation.ts`) injects ZWJ into agent words — including the temp path inside the Bash tool description — and Claude models copy that verbatim into generated commands, which are delivered as tool-call arguments rather than assistant text. As a result the ZWJ survived and corrupted code blocks (e.g. a temp path rendered with an invisible joiner). Now `open-sse/handlers/responseSanitizer.ts` strips zero-width code points from tool-call argument strings at every emit site (OpenAI non-stream/stream chat `tool_calls` + legacy `function_call`, native Responses `function_call` items, the OpenAI→Responses conversion, and the native Responses streaming `response.function_call_arguments.delta/.done` events). Only zero-width code points are removed; JSON structure and all other bytes stay identical (no parse/restringify), so normal arguments remain byte-exact. Regression guard: 6 new cases in `tests/unit/response-sanitizer.test.ts` (suite 50/50).
+
+- **fix(nodejs):** the default app log path now resolves under `DATA_DIR` (`~/.omniroute/logs/application/app.log`) instead of `process.cwd()` ([#6197](https://github.com/diegosouzapw/OmniRoute/issues/6197)) — the globally-installed CLI runs from an arbitrary working directory, so anchoring the default to cwd made file logging silently write to (or no-op under) an unrelated directory, contradicting the documented `.env.example` default. `getAppLogFilePath()` now computes the default lazily via the pure `resolveDataDir()` resolver (honours a per-process `DATA_DIR`, no directory-creation side effect); an explicit `APP_LOG_FILE_PATH` still wins. Regression guard: `tests/unit/logenv-datadir-path-6197.test.ts` (3).
+
+- **fix(docker):** AgentBridge/`startMitm` no longer aborts in containers/headless when the Antigravity-default DNS step can't write `/etc/hosts` ([#6127](https://github.com/diegosouzapw/OmniRoute/issues/6127)), and the privileged command's stderr now reaches `app.log` instead of only a bare exit code hitting the toast ([#6198](https://github.com/diegosouzapw/OmniRoute/issues/6198)). The default DNS step (`addDNSEntry`) was called unguarded while cert install and the two sibling DNS steps were each best-effort — in the runtime Docker image (`USER node`, no `sudo`, read-only `/etc/hosts`) it threw `Command failed with code 1` out of `startMitmInternal` and killed the whole start, discarding the stderr. The three DNS steps are extracted into a best-effort `provisionDnsEntries()` where each failure is logged with the full `err` (stderr included, folded in by `systemCommands.ts`) and never aborts the start. Regression guard: `tests/unit/mitm-dns-graceful-degrade-6127.test.ts` (4).
+
+- **fix(providers):** copilot-m365-web now supports the M365 Education "Starter / OfficeWebIncludedCopilot" tier and no longer returns an empty `content:null` stream ([#6210](https://github.com/diegosouzapw/OmniRoute/issues/6210)). Two gaps: (1) `buildWsUrl()` hardcoded the individual-consumer scenario (`OfficeWebPaidConsumerCopilot`, `isEdu=false`) — the EDU tier is now opt-in via `providerSpecificData.tier="edu"`, emitting `scenario=OfficeWebIncludedCopilot`/`isEdu=true` (the individual path is unchanged); (2) the EDU/GPT-5.5 path streams deltas via `arguments[0].writeAtCursor` (incremental) instead of only `messages[].text` (accumulated snapshots), which the parser dropped — a new `accumulateBotContent()` folds both formats, with `type:2 item.result.message` as a last-resort fallback. Regression guard: `tests/unit/copilot-m365-edu-writeatcursor-6210.test.ts` (10). (thanks @qpeyba)
+
+- **fix(providers):** GitLab Duo executor now feeds tool results back into the prompt instead of looping ([#6220](https://github.com/diegosouzapw/OmniRoute/issues/6220)) — `buildPrompt()` branched only on `system`/`user` and took `userParts.at(-1)`, silently dropping the `assistant{tool_calls}` + `tool{result}` turns the client appended, so the reconstructed prompt was byte-identical to turn 1 and the model re-emitted the same `<tool>` call forever. When a tool exchange is present the full conversation is now serialized, folding each tool result back keyed by its `tool_call_id`; simple conversations keep the legacy shape. Complements the tool_call emission from [#6051](https://github.com/diegosouzapw/OmniRoute/issues/6051) (the `kilo-duplicate` label was a false positive — different, sequential defect). Regression guard: `tests/unit/gitlab-tool-result-feedback-6220.test.ts` (4).
+
+- **fix(providers):** opencode-go/opencode-zen can now synthesize the OpenCode CLI identity headers Cloudflare requires on VPS egress ([#5997](https://github.com/diegosouzapw/OmniRoute/issues/5997)) — on a datacenter VPS, `opencode.ai/zen/go/v1/chat/completions` 403s (HTML challenge) requests lacking CLI identity, while the reporter's control curl proved that `User-Agent: opencode-cli/1.0.0` + `x-opencode-client: cli` + `x-opencode-project: default` + fresh request/session UUIDs succeed. Opt-in via `OPENCODE_SYNTHESIZE_CLI_HEADERS=true` (values overridable via `OPENCODE_GO_USER_AGENT`/`OPENCODE_USER_AGENT`/`OPENCODE_CLIENT`/`OPENCODE_PROJECT`); it fills only headers the client did not already send. Kept **off by default** — the forward-only path is deliberate (fabricating a wrong value risks upstream rejection; a prior dedup regressed with `opencode/local`), so this replaces the fragile local header-injection shim without changing default behavior. Regression guard: `tests/unit/opencode-cli-headers-synthesis-5997.test.ts` (6). (thanks @aleksesipenko)
+- fix(resilience): sticky session affinity now evicts and fails over to another account when the pinned account is exhausted/unavailable (#6219)
+
+- fix(sse): Responses API passthrough now drops internal commentary-phase output before forwarding to clients (gated by RESPONSES_PASSTHROUGH_DROP_COMMENTARY, default on) (#6199)
+
+- **fix(sse):** tool-call function schemas with a root `type: null` are now coerced to `type: "object"` before dispatch ([#6359](https://github.com/diegosouzapw/OmniRoute/issues/6359)) — clients like the Codex app emit `parameters: { type: null, ... }` for some tools, which OpenAI-compatible upstreams reject with `400 Invalid schema for function '...': schema must be a JSON Schema of 'type: "object"', got 'type: null'`, failing the whole request. `toolSchemaSanitizer` already stripped the null; it now re-adds the mandatory root `"object"` type (and empty `properties`/open `additionalProperties` when absent). Combinator roots (`anyOf`/`oneOf`/`allOf`) and explicit root types are left untouched. Regression guard: 5 new cases in `tests/unit/tool-schema-sanitizer.test.mjs`.
+
+- **fix(docker):** AgentBridge no longer fails to start on npm/Electron/VPS installs with "MITM manager stub reached at runtime" ([#6344](https://github.com/diegosouzapw/OmniRoute/issues/6344)) — v3.8.45 flipped the production bundler default to Turbopack, but `next.config.mjs` aliased `@/mitm/manager` to its Docker-only degraded stub **unconditionally**. That was harmless while Docker (which sets the alias intentionally for #3390 graceful degradation) was the sole Turbopack consumer, but once every artifact built with Turbopack the stub shipped to all non-Docker users and `startMitm` threw on the first Agent-Bridge start. The alias is now opt-in via `OMNIROUTE_MITM_STUB=1` (set only by the Dockerfile) through the shared `scripts/build/mitm-stub-flag.mjs` helper; default builds bundle the real manager. Regression guard: `tests/unit/mitm-stub-alias-6344.test.mjs` (4).
+
+- **fix(proxy):** stop the v3.8.44 proxy regression that leaked the real IP and disabled healthy proxies ([#6246](https://github.com/diegosouzapw/OmniRoute/issues/6246)). Two coupled defects from the new health scheduler: (1) **IP leak** — when a proxy assigned to a connection was marked `inactive`, resolution fell through to a **direct** egress instead of blocking, exposing the operator's real IP; (2) **over-deactivation** — the sweep flipped a proxy to `inactive` on the **first** failed probe and counted our own 5s timeout / a probe-target `5xx` as the proxy's fault, so healthy paid proxies vanished from egress selection ("my proxies are not being used anymore"). Fix: the sweep decision is extracted into a pure, network-free `decideProxyHealthAction` (`src/lib/proxyHealth/decision.ts`) — by default the health check now **only counts/logs and never downgrades status** (a proxy is downgraded/removed only with `PROXY_AUTO_REMOVE=true`, after `PROXY_AUTO_REMOVE_AFTER` **consecutive** conclusive failures); probes are classified tri-state so an inconclusive result (our timeout, or a `5xx` from the probe target) never penalizes the proxy, and the probe timeout is raised 5s→15s. Separately, `safeResolveProxy` now **fails closed** via the existing policy: a connection whose assigned proxy is dead is blocked instead of leaking direct (`hasBlockingProxyAssignment`), honoring the explicit `proxy off` toggles and the `PROXY_FAIL_OPEN=true` opt-out. Existing proxies stuck `inactive` by the old behavior need a one-time manual re-activate (the operator owns proxy status). Regression guards: `tests/unit/proxy-health-decide-action-6246.test.ts`, `tests/unit/proxy-assigned-unavailable-6246.test.ts`.
+
+- **fix(proxy):** make "Test All" read-only and add bulk enable/disable ([#6246](https://github.com/diegosouzapw/OmniRoute/issues/6246)). Complements the core fail-closed / scheduler fix (#6296) with the two remaining reporter asks. (1) The **"Test All" button** (`POST /api/settings/proxies/auto-test`) used to flip a proxy to `inactive` on a failed reachability probe; since the egress selector excludes `inactive` proxies, a flaky probe (an unreachable `httpbin.org`, a proxy that blocks `HEAD`, or a slow paid proxy) silently disabled every proxy that failed — "Test All" is now **read-only by default** (only the operator sets a proxy active/inactive; opt back into the legacy test-and-set with `PROXY_HEALTH_AUTO_DEACTIVATE=true`). (2) Adds a **bulk enable/disable** proxies endpoint + toolbar action (`POST /api/settings/proxies/batch-activate`) so an operator can re-activate proxies in one click. Regression guard: `tests/unit/proxy-health-6246.test.ts`. (thanks @tenshiak)
+
+- **chatcore (tools): stop the default 128-tool cap from silently dropping opencode's `task`/MCP tools.** opencode (used as an MCP/agent host) sends a large tool list; when it exceeds the speculative `MAX_TOOLS_LIMIT` (128) default, `truncateToolList` did a blind `tools.slice(0, 128)`, dropping every tool past index 128 — including opencode's built-in `task` tool (subagent launch) and many MCP tools, so models routed through OmniRoute could no longer spawn subagents or reach part of their tools. The cap exists to avoid upstream `400`s for providers with real hard limits (e.g. grok-cli 200), so it is kept for those: detection of the opencode client (`isOpencodeClient` — any `x-opencode-*` header, or `opencode` in the user-agent) now only bypasses the **speculative 128 default**, never a known provider ceiling. Precedence is explicit — a proactive/detected provider limit always truncates (even for opencode); otherwise opencode forwards its full tool list; otherwise the unchanged 128 default applies to every other client. Refactors `getEffectiveToolLimit` into `getKnownToolLimit(provider) ?? DEFAULT_LIMIT` (byte-identical for existing callers) and fixes a cosmetic debug-log that reported the truncated count instead of the original. Regression guard: `tests/unit/tool-limit-detector.test.ts`.
+
+- **fix(mitm):** the macOS MITM-cert install check now matches the system keychain again. `security find-certificate -a -Z` prints the SHA-1 as a colon-less hex string, but the installed-check compared it against `getCertFingerprint()`'s colon-separated form, so the substring match never hit — the cert was reported as not-installed and re-prompted for the sudo install on every run. Fingerprints are now normalized (colons stripped, upper-cased) on both sides via the extracted `macCertOutputHasFingerprint` helper. Regression guard: `tests/unit/mitm-cert-mac-fingerprint.test.ts`. ([#6204](https://github.com/diegosouzapw/OmniRoute/pull/6204), closes [#6134](https://github.com/diegosouzapw/OmniRoute/issues/6134) — thanks @rianonehub)
+
+- **fix(api):** `/v1/messages/count_tokens` now counts `tool_use`, `tool_result` and `thinking` content blocks (and array-form `system` prompts) in the local-estimation path, instead of only `text`. Real agentic conversations keep ~95% of their tokens inside tool results; the previous estimate returned near-zero for them, which silently broke Claude Code's auto-compaction (context grew past the window with no compaction until the upstream API rejected the request). The real provider-side count path is unchanged. Regression guard: `tests/unit/messages-count-tokens-route.test.ts`. ([#6221](https://github.com/diegosouzapw/OmniRoute/pull/6221) — thanks @luweiCN)
+
+- **fix(antigravity):** strip a trailing assistant prefill turn for Vertex Claude models to avoid upstream 400s ([#6114](https://github.com/diegosouzapw/OmniRoute/pull/6114)). Regression guard: `tests/unit/antigravity-claude-prefill-strip.test.ts`. (thanks @anki1kr)
+
+- **fix(security):** the mutable cloud-agent routes (`/api/cloud/credentials/update`, `/api/cloud/models/alias`) now require management auth instead of being treated as public. They were classified as public API routes, so a request without management credentials could update stored cloud-agent credentials and model aliases. They are removed from the public-route set, classified as management routes in the authz pipeline, and gated by `requireManagementAuth`; cloud **read**/auth routes stay public. Regression guards: `tests/unit/cloud-write-auth.test.ts`, `tests/unit/authz/classify.test.ts`, `tests/unit/public-api-routes.test.ts`. ([#6233](https://github.com/diegosouzapw/OmniRoute/pull/6233) — thanks @vittoroliveira-dev)
+
+- **refactor(dashboard):** extract the onboarding-wizard "Open provider details" link target into a pure, unit-tested `buildProviderDetailsHref(connection)` helper. The wizard already routes by `connection.id` (the node UUID) rather than the provider category slug (#6144/#6145); this hardens that behavior behind a tested helper that guards a missing id/connection. Regression guard: `tests/unit/provider-onboarding-href.test.ts`. ([#6166](https://github.com/diegosouzapw/OmniRoute/pull/6166) — thanks @KooshaPari)
+
+- **fix(security):** the doubao synthetic device-id generator now derives its digits via an unbiased crypto-random draw (rejection sampling over `crypto.randomBytes()`) instead of a `% 10` reduction, closing a CodeQL `js/biased-cryptographic-random` finding.
+- **fix(agentSkills):** the GitHub-skills generator now resolves `outputDir` to an absolute path before writing, fixing a regression introduced by #6366 (relative-to-cwd base path) that could write generated skill files to the wrong directory.
+- **fix(security):** `/api/keys/{id}/devices` now checks the HTTP method before auth/validation, returning a `405` for non-GET/DELETE verbs instead of a misleading `401`/`500` (closes a `dast-smoke` QUERY-method finding).
+- **fix(quality):** clear the last 2 heavy quality-gate reds on the release tip (cycle pre-flight).
+- **fix(mitm):** the test suite and CI can never mutate the OS trust store — `OMNIROUTE_SKIP_SYSTEM_TRUST=1` is set globally for tests/CI so `installCert`/`uninstallCert`/`installTproxyCa` skip the privileged OS dispatch ([#6310](https://github.com/diegosouzapw/OmniRoute/pull/6310); full detail is under the [3.8.45] section below — this branch received it via the parallel-cycle sync-back).
+- **fix(api):** `POST /api/github-skills` now Zod-validates its request body; documented the new quality-gate env vars and pinned the merge-integrity GitHub Actions to a commit SHA.
+- **fix(skills):** generate the missing `omni-github-skills` registry entry and align the agent-skills catalog-count tests (follow-up to #6186).
+- **fix(quality):** clear the cycle's 11 net-new ESLint errors and make `validate-release-green` suppressions-aware.
+
+### 📝 Maintenance
+
+- **i18n(it):** add 118 missing Italian (`it`) translations (net-additive — no existing keys dropped, valid JSON), improving Italian UI coverage. ([#6212](https://github.com/diegosouzapw/OmniRoute/pull/6212) — thanks @serverless83)
+- **chore(providers):** remove deprecated **MiMo V2** model entries from the catalogs (xiaomi-mimo, opencode-go, zenmux-free, audio TTS) — the upstream V2 line is superseded by MiMo V2.5; drops `mimo-v2-tts`, `mimo-v2-pro`, `mimo-v2-omni`, `mimo-v2-flash`, `mimo-v2-flash-free` and realigns the provider-catalog tests. ([#6248](https://github.com/diegosouzapw/OmniRoute/pull/6248) — thanks @backryun)
+- **chore(release):** ~50 commits on this branch are v3.8.45 pre-flight/hardening fixes and CI-perf work that landed here via the parallel-cycle sync-back (`sync-next-cycle.mjs`, Hard Rule #21) **after** the `v3.8.45` git tag was cut, and are already fully documented under the **[3.8.45]** section below — listed here only so the per-cycle commit-coverage check (`npm run release:uncovered`) doesn't flag them as gaps. Provider/catalog/UX/backend fixes: #6041, #6078, #6108, #6135, #6148, #6149, #6154, #6158, #6161, #6162, #6163, #6164, #6165, #6170, #6177, #6178, #6181, #6186, #6187, #6191, #6193, #6194, #6195, #6200, #6205, #6208, #6209, #6211, #6213, #6223, #6224, #6225, #6226, #6227, #6228, #6229, #6230, #6235, #6291, #6292. CI/release-pipeline work: #6167, #6203, #6214, #6215, #6218, #6273, #6275, #6283, #6284, #6285, #6300, #6305.
+- **chore(release):** additional zero-ref release-cycle plumbing on this branch, kept out of `release:uncovered` on purpose (no `#N` in the commit subject to cite): opening the v3.8.46 cycle, opening/closing the v3.8.45 cycle, the finalized [3.8.45] CHANGELOG i18n sync-back to 42 mirrors, the v3.8.45 cognitive/cyclomatic and file-size drift rebaselines, ESLint stale-suppression pruning (4,273 → 4,233), and clearing test-masking/docs-all pre-flight reds for v3.8.45.
+
+### ⚡ Performance & Infrastructure
+
+- **perf(release-green):** the pre-flight validator (`scripts/quality/validate-release-green.mjs`) now runs its 4 slow suites (unit / vitest / integration / pack-artifact) **concurrently** via `Promise.all` — pre-flight wall time drops from ~the sum of the suites to ~the slowest one (~30min saved per round; Phase 0 was the nº1 bottleneck of the v3.8.45 release benchmark, 2h54 of 6h34 e2e). Guard: `tests/unit/validate-release-green.test.ts` ("runs the slow suites CONCURRENTLY"). ([#6319](https://github.com/diegosouzapw/OmniRoute/pull/6319))
+- **fix(ci):** `scripts/release/sync-next-cycle.mjs` — two defects found live in its first production run (v3.8.45 Phase 5): (1) the `git()` helper's default 1 MiB `maxBuffer` crashed with `ENOBUFS` on `git show origin/main:CHANGELOG.md` (the CHANGELOG alone is >1 MiB) — widened to 64 MiB; (2) the i18n resync only propagated the `[NEXT]` (TBD) section, leaving the just-shipped finalized section as "— TBD" in all 42 mirrors — it now also syncs `[prevVersion]` bounded by the heading below it (new exported pure helper `versionAfter`). Guards: +5 tests in `tests/unit/sync-next-cycle.test.ts` (8/8). ([#6327](https://github.com/diegosouzapw/OmniRoute/pull/6327))
+- **test(ci):** concurrency-sensitive flaky tests are **quarantined into a serial pass** (`tests/unit/serial/`, `--test-concurrency=1`, appended to every unit runner incl. sharded variants — the serial pass is sharded too so concurrent shard jobs never self-collide). Initial set: `glm-coding-plan-monthly-3580`, `quota-division-blocks`, `provider-health-autopilot`, `combo-health-autopilot` — the class behind the ~28min CI wedges/re-runs (two live 1h+ wedges cancelled during this PR's own validation). Discovery + TIA gates track the new glob; systemic root cause (async logger writing after teardown) tracked in [#6360](https://github.com/diegosouzapw/OmniRoute/issues/6360). Guard: `tests/unit/test-serial-quarantine.test.ts` (4). ([#6347](https://github.com/diegosouzapw/OmniRoute/pull/6347))
+
+### 🙌 Contributors
+
+Thanks to everyone whose work landed in v3.8.46:
+
+| Contributor                                                    | PRs / Issues           |
+| -------------------------------------------------------------- | ---------------------- |
+| [@2220258345](https://github.com/2220258345)                   | direct commit / report |
+| [@abdofallah](https://github.com/abdofallah)                   | direct commit / report |
+| [@adentdk](https://github.com/adentdk)                         | direct commit / report |
+| [@aleksesipenko](https://github.com/aleksesipenko)             | direct commit / report |
+| [@anki1kr](https://github.com/anki1kr)                         | direct commit / report |
+| [@anungma](https://github.com/anungma)                         | direct commit / report |
+| [@arpicato](https://github.com/arpicato)                       | direct commit / report |
+| [@backryun](https://github.com/backryun)                       | #6248                  |
+| [@binsarjr](https://github.com/binsarjr)                       | direct commit / report |
+| [@brightfiscalband](https://github.com/brightfiscalband)       | direct commit / report |
+| [@chirag127](https://github.com/chirag127)                     | #6501, #6506           |
+| [@developerjillur](https://github.com/developerjillur)         | direct commit / report |
+| [@dilneiss](https://github.com/dilneiss)                       | #6499                  |
+| [@dtybnrj](https://github.com/dtybnrj)                         | direct commit / report |
+| [@Forcerecon](https://github.com/Forcerecon)                   | direct commit / report |
+| [@hao3039032](https://github.com/hao3039032)                   | direct commit / report |
+| [@Iammilansoni](https://github.com/Iammilansoni)               | #6150, #6245           |
+| [@jmengit](https://github.com/jmengit)                         | direct commit / report |
+| [@jordansilly77-stack](https://github.com/jordansilly77-stack) | direct commit / report |
+| [@JxnLexn](https://github.com/JxnLexn)                         | direct commit / report |
+| [@KooshaPari](https://github.com/KooshaPari)                   | #6166                  |
+| [@loopyd](https://github.com/loopyd)                           | direct commit / report |
+| [@luweiCN](https://github.com/luweiCN)                         | #6221                  |
+| [@makcimbx](https://github.com/makcimbx)                       | direct commit / report |
+| [@muflifadla38](https://github.com/muflifadla38)               | direct commit / report |
+| [@newnol](https://github.com/newnol)                           | direct commit / report |
+| [@ofekbetzalel](https://github.com/ofekbetzalel)               | direct commit / report |
+| [@ohahe52-dot](https://github.com/ohahe52-dot)                 | direct commit / report |
+| [@phidinhmanh](https://github.com/phidinhmanh)                 | direct commit / report |
+| [@powellnorma](https://github.com/powellnorma)                 | direct commit / report |
+| [@qpeyba](https://github.com/qpeyba)                           | direct commit / report |
+| [@RaviTharuma](https://github.com/RaviTharuma)                 | direct commit / report |
+| [@RCrushMe](https://github.com/RCrushMe)                       | direct commit / report |
+| [@rianonehub](https://github.com/rianonehub)                   | #6134, #6204           |
+| [@serverless83](https://github.com/serverless83)               | #6212                  |
+| [@swingtempo](https://github.com/swingtempo)                   | direct commit / report |
+| [@tenshiak](https://github.com/tenshiak)                       | direct commit / report |
+| [@ThongAccount](https://github.com/ThongAccount)               | direct commit / report |
+| [@UnrealAryan](https://github.com/UnrealAryan)                 | direct commit / report |
+| [@vinayakkulkarni](https://github.com/vinayakkulkarni)         | direct commit / report |
+| [@vittoroliveira-dev](https://github.com/vittoroliveira-dev)   | #6233                  |
+| [@warelik](https://github.com/warelik)                         | direct commit / report |
+| [@xxy9468615](https://github.com/xxy9468615)                   | direct commit / report |
+| [@xz-dev](https://github.com/xz-dev)                           | direct commit / report |
+| [@yanpaing007](https://github.com/yanpaing007)                 | direct commit / report |
+| [@diegosouzapw](https://github.com/diegosouzapw)               | maintainer             |
+
+---
+
 ## [3.8.45] — 2026-07-06
 
 ### ✨ New Features

docs/i18n/th/CHANGELOG.md

@@ -6,6 +6,213 @@
 
 ## [3.8.31] — 2026-06-20
 
+## [3.8.46] — TBD
+
+### ✨ New Features
+
+- **feat(sse):** **hide paid-only models from `auto/*` routing** when `hidePaidModels` is on ([#6512](https://github.com/diegosouzapw/OmniRoute/issues/6512)) — follow-up to #6328/#6495. PR #6495 hid paid-only models from the `GET /v1/models` listing, but `auto/*` combos (`auto/best-coding`, `auto/glm`, …) could still pick a paid-only backend into their candidate pool → a 402/403 at request time. `createVirtualAutoCombo` now filters the candidate pool through the new pure `open-sse/services/autoCombo/paidModelFilter.ts` (`filterPaidOnlyCandidates`), applying the same free-model predicate #6495 uses in `catalog.ts` (`providerHasFreeModels(provider) && isFreeModel(provider, {id})`) whenever `settings.hidePaidModels === true`. Applied before the category/tier/family narrowing, so it covers every `auto/*` combo; an all-paid pool degrades to the existing graceful empty-pool path. **Opt-in — default OFF leaves the pool unchanged** (identity). Regression guard: `tests/unit/autoCombo/paid-model-filter-6512.test.ts` (4, incl. the default-off identity guard).
+- **feat(sse):** **provider-family auto combos** — `auto/glm`, `auto/minimax`, `auto/mimo`, `auto/zai`, `auto/gemma`, `auto/llama`, `auto/gemini` ([#6453](https://github.com/diegosouzapw/OmniRoute/issues/6453)) — new routable ids that materialize an on-demand virtual combo spanning whatever installed backends currently expose that model family, degrading gracefully as backends rotate. A new pure `open-sse/services/autoCombo/modelFamily.ts` (`detectModelFamily`) classifies by model-id prefix for six families; `zai` is instead resolved by provider id (z.ai's hosted API serves the same `glm-*` model ids as every other GLM backend, so `auto/zai` means "route to my z.ai backend specifically" vs `auto/glm`'s "any connected GLM backend"). Reuses the existing `createVirtualAutoCombo` on-demand materialization path (no DB writes) and the `/v1/models` catalog advertising loop. Regression guard: `tests/unit/autoCombo/provider-family-combos.test.ts` (11).
+- **feat(proxy):** native **proxy-pool round-robin / egress IP rotation** ([#6365](https://github.com/diegosouzapw/OmniRoute/issues/6365)) — a scope (global / provider / account) can now hold **multiple** proxies as a pool with a rotation strategy, so outbound requests cycle their egress IP instead of pinning one proxy per scope. Migration `117_proxy_pool_rotation.sql` lifts the `UNIQUE(scope, scope_id)` constraint (rebuild via the canonical rename/copy/drop; existing single assignments become 1-element pools) and adds a `proxy_scope_rotation` companion table holding the per-scope strategy + a persisted monotonic round-robin cursor. Strategies: `round-robin` (default, monotonic cursor — never `Math.random`), `random`, and `sticky-per-N-min`. Resolution (`resolveProxyForScopeFromRegistry` / `resolveProxyForConnectionFromRegistry`) now fetches the alive, position-ordered candidate set (unchanged `PROXY_ALIVE_PREDICATE`) and applies the strategy; an empty / all-dead pool still returns `null` — the #6246 fail-closed guard is **untouched** (never falls through to direct egress). Backend + DB only; dashboard pool-builder UI is a follow-up. Regression guard: `tests/unit/proxy-pool-rotation-6365.test.ts` (8, incl. fail-closed + backward-compat).
+- **feat(providers):** end-to-end **tool/function calling on the native Gemini `/v1beta` endpoint** ([#6222](https://github.com/diegosouzapw/OmniRoute/issues/6222)) — both directions of the Gemini↔OpenAI conversion now preserve tool data (previously silently dropped). Request side: `convertGeminiToInternal` (extracted to its own testable module) maps `tools[].functionDeclarations` → OpenAI `tools`, prior `functionCall` parts → assistant `tool_calls`, and `functionResponse` parts → `tool`-role messages. Response side: `convertOpenAIResponseToGemini` emits `parts[].functionCall {name,args}` from `message.tool_calls`, and the streaming `openAIChunkToGeminiChunk` accumulates fragmented `tool_calls` deltas by index into complete `functionCall` parts. The non-Gemini client paths (Claude, OpenAI-Responses) already preserved tool calls — this closes the gap specific to the native Gemini surface. Regression guard: `tests/unit/v1beta-gemini-tool-calling-6222.test.ts` (6, incl. a streaming SSE round-trip).
+- **feat(providers):** copilot-m365-web **enterprise / work tier** support ([#6334](https://github.com/diegosouzapw/OmniRoute/issues/6334)) — mirrors the EDU-tier pattern (#6210): `M365ConnectionParams` gains an `agent` field, a new opt-in `M365_ENTERPRISE_OVERRIDES` preset (`agent=work`, `scenario=officeweb`, `licenseType=Premium`) applies via `providerSpecificData.tier="enterprise"` (alias `"work"`), and `agent` is also overridable directly via `providerSpecificData.agent`. `buildWsUrl` was hardcoding `agent="web"` (the one enterprise-distinguishing param with no override path), so a Premium work account handshook then returned an empty stream. The individual and EDU paths are untouched. Kilo's dup flag vs #6210 (EDU tier) was a false positive — different tier. Regression guard: `tests/unit/copilot-m365-enterprise-6334.test.ts` (7). End-to-end confirmation on a real Premium work account is a live-VPS validation follow-up (Hard Rule #18). (thanks @Forcerecon)
+- **feat(api):** standardized, provider-agnostic **`effort` + `thinking` request params** ([#6241](https://github.com/diegosouzapw/OmniRoute/issues/6241)) — a thin standardization layer over the existing mature per-provider reasoning plumbing (no provider mapper touched). `providerChatCompletionSchema` gains a canonical `effort` (reusing the shared `none/low/medium/high/xhigh` vocabulary — the UI tiers `extra`/`max` collapse onto `xhigh`) and a boolean `thinking`. A pure `normalizeReasoningRequest` (wired once in `src/sse/handlers/chat.ts`, before any reasoning field is read) folds them onto the fields the translators already consume (`reasoning_effort` / `reasoning.effort` / `thinking`), so they fan out to Anthropic / Gemini / xAI / Responses — an explicit client `reasoning_effort` / object-shaped `thinking` always wins (backward-compatible). `/models` additively exposes `supportsThinking` + `effort_tiers` so the frontend can render the toggles (UI component is a follow-up). Regression guard: `tests/unit/effort-thinking-standardization-6241.test.ts` (12). (thanks @Iammilansoni, @shabeer)
+- **feat(combo):** new **`pipeline` (sequential) combo strategy** ([#6297](https://github.com/diegosouzapw/OmniRoute/issues/6297)) — the 18th routing strategy runs targets **in order**, threading each step's output into the next step's input, with an optional per-step `prompt` (system instruction); only the final step's response is returned. Distinct from `fusion` (parallel fan-out + judge). Implemented as a self-contained `open-sse/services/pipeline.ts` (sibling to `fusion.ts`), dispatched from `combo.ts`; the step list reuses `combo.models` order and reads an optional `prompt` off each target (backward-compatible — ignored by every other strategy). Intermediate steps run non-streaming with tools stripped (complete prose to thread forward); the final step keeps the client's `stream` flag + tools. A failing/empty/unparseable intermediate step fails the whole pipeline explicitly via a sanitized error (never silently swallowed). Kilo's dup flag vs #563 was a false positive (that's model→chain selection; this is a sequential chain). Regression guard: `tests/unit/combo-pipeline-strategy.test.ts` (5). (thanks @ofekbetzalel)
+- **feat(ci):** `check:test-masking` now flags **inline-reimplemented prod conditions** ([#6348](https://github.com/diegosouzapw/OmniRoute/issues/6348)) — a new **report-only** subcheck (v2, 6A.10 family) catches the wrong-shape contract test: a test that recomputes the condition under test inline instead of importing/exercising the real function (the #6216 class, where `=== 500` → `>= 500` stayed green because the test re-implemented the branch). For each added/modified test file it warns when the file textually duplicates a ≥3-token conditional from a production file touched in the same PR **and** does not import the symbol/module owning it, via a pure, fixture-tested `findReimplementedConditions()` with an allowlist mirroring `assertReductionAllowlist`. Report-only for now (does not fail the gate) — to be promoted to blocking after a triage cycle. Regression guard: `tests/unit/check-test-masking.test.ts` (45).
+- **feat(sse):** per-connection routing override (native vs CLIProxyAPI) ([#6339](https://github.com/diegosouzapw/OmniRoute/issues/6339)) — the previously-dead `isCliproxyapiDeepModeEnabled` helper is now wired into `resolveExecutorWithProxy`: a single connection can opt itself into the CLIProxyAPI passthrough executor via `providerSpecificData.cliproxyapiMode="claude-native"`, with precedence **connection override > provider `upstream_proxy_config` mode > default**. `resolveExecutorWithProxy` now receives the resolved connection's `providerSpecificData` (threaded from `chatCore.ts`), so one connection can deep-route while the provider's other connections stay native — no DB schema change (the toggle rides in `providerSpecificData`). Also resolves the same-provider mixing ask in #6340. Regression guard: `tests/unit/chatcore-executor-proxy.test.ts` (9). (thanks @RaviTharuma)
+- **feat(dashboard):** "Add session cookie" modal now shows a prominent **"Open ‹host› →"** link to the provider's own site ([#6268](https://github.com/diegosouzapw/OmniRoute/issues/6268)) — every `-web` cookie-session provider (chatgpt-web, claude-web, gemini-web, kimi-web, lmarena, qwen-web, m365-copilot-web, …) renders a one-click external link (opening the provider's login/home page in a new tab) so operators no longer tab away to retype the URL mid-setup. The host resolves from a pure, unit-tested `resolveWebProviderHost()` (prefers `WEB_COOKIE_PROVIDERS[id].website`, falls back to the registry `baseUrl` origin); non-web providers render exactly as before. Kilo's dup flag vs #6265 (modal-too-small-on-1080p) was a false positive — distinct concern. Regression guard: `tests/unit/resolve-web-provider-host.test.ts` (5). (thanks @chirag127)
+- **feat(providers):** add **DigitalOcean** AI (serverless inference) as an OpenAI-compatible API-key provider ([#6373](https://github.com/diegosouzapw/OmniRoute/pull/6373)) — base `https://inference.do-ai.run/v1`, wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/digitalocean/`, `src/shared/constants/providers/apikey/inference-hosts.ts`). Regression guard: `tests/unit/digitalocean-provider.test.ts`. (thanks @newnol)
+- **feat(providers):** add **Huancheng Public API** (`hcnsec`) as an OpenAI-compatible regional provider ([#6410](https://github.com/diegosouzapw/OmniRoute/pull/6410)) — Xinjiang Huancheng Cybersecurity's public LLM platform (base `https://api.hcnsec.cn/v1`, free credits via daily check-ins), wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/hcnsec/`, `src/shared/constants/providers/apikey/regional.ts`). Regression guard: `tests/unit/hcnsec-provider.test.ts`. (thanks @UnrealAryan)
+- **feat(dashboard):** the web-session credential guide now shows an **"Open {host}" link** ([#6316](https://github.com/diegosouzapw/OmniRoute/pull/6316)) to the provider's sign-in site (derived from the provider `website` via `getProviderWebsiteHost`), so you can jump straight to the page where the cookie/session must be captured. Regression guard: `tests/unit/web-session-provider-link-6316.test.ts`. (thanks @jordansilly77-stack)
+- **feat(cerebras):** add the **Gemma 4 31B** model (`gemma-4-31b`) to the Cerebras registry + pricing table ([#6331](https://github.com/diegosouzapw/OmniRoute/pull/6331)). Regression guard extends `tests/unit/t28-model-catalog-updates.test.ts`. (thanks @backryun)
+- **feat(providers):** add **Yuanbao (web)** as a cookie-session provider ([#6196](https://github.com/diegosouzapw/OmniRoute/issues/6196)) — `yuanbao-web` (Tencent Yuanbao, `yuanbao.tencent.com`) with cookie-only auth (`hy_user`/`hy_token` + public agent id), SSE→OpenAI translation incl. `reasoning_content`, exposing DeepSeek V3/R1 + Hunyuan / Hunyuan-T1. Regression guard: `tests/unit/providers-yuanbao-web.test.ts`. `together-web` was **deferred** (no verifiable web-session endpoint — needs a captured request) and `huggingchat-web` **dropped** (the existing `huggingchat` already is a web-cookie provider). (thanks @chirag127)
+- **feat(providers):** route the built-in **agentrouter** through the dynamic Claude-Code wire image ([#6056](https://github.com/diegosouzapw/OmniRoute/issues/6056)) — a small static allow-set (`CC_WIRE_IMAGE_BUILTINS` in `open-sse/services/ccWireImageBuiltins.ts`), consulted by `isClaudeCodeCompatible` / `isClaudeCodeCompatibleProvider` / `applyFingerprint`, makes agentrouter adopt the CC wire-image headers + fingerprint **while guarding the CC baseUrl/auth branches** so it keeps its own registry `baseUrl` and `x-api-key` auth. Regression guard: `tests/unit/agentrouter-cc-wire-image.test.ts` (asserts the wire image is applied AND agentrouter's baseUrl/auth are preserved). Live WAF-acceptance against agentrouter.org is a VPS validation follow-up (Hard Rule #18).
+- **feat(providers):** **bulk-add API keys for Cloudflare Workers AI** ([#6174](https://github.com/diegosouzapw/OmniRoute/issues/6174)) — `cloudflare-ai` is removed from the bulk-add exclusion list and the bulk parser gains a 3-field `name|accountId|apiKey` mode; the bulk route now builds a **per-entry** `providerSpecificData` so each key carries its own `accountId` (fixing the previous shared-object reuse), and both the create + key-validation paths receive it. Regression guard: `tests/unit/bulk-api-key-parser-cloudflare.test.ts`. (thanks @muflifadla38)
+- **feat(dashboard):** routing/settings UX clarity ([#6147](https://github.com/diegosouzapw/OmniRoute/issues/6147)) — (1) weighted combos show the **effective routing share %** next to each weight when weights don't sum to 100 (`WeightTotalBar.tsx`); (2) the status widget's user-facing **"Cloud Sync" label is renamed** to "Remote Settings Sync" (`CloudSyncStatus.tsx`; internal ids/state untouched); (3) built-in providers gain an **opt-in advanced base-URL override** (`isBaseUrlOverrideEligibleProvider`, hidden behind an "Advanced" toggle, reusing the existing `providerSpecificData.baseUrl` persistence — not globally widened). Regression guard: `tests/unit/routing-settings-ux-6147.test.ts`.
+- **feat(combo):** add an option to **disable session stickiness**, per-combo or globally — round-robin / random combos can rotate to a different connection on every request instead of pinning a whole conversation to one connection by its first-message hash. Resolution precedence per-combo `config.disableSessionStickiness` → global `settings.disableSessionStickiness` → default `false` (preserves the #3825 prompt-cache/504 fix); gates **both** stickiness call sites in `open-sse/services/combo.ts`. Exposed as a global toggle (Combo Defaults) and a per-combo Inherit/on/off control. ([#6168](https://github.com/diegosouzapw/OmniRoute/issues/6168)) Regression guard: `tests/unit/combo-disable-session-stickiness.test.ts`. (thanks @RCrushMe)
+- **feat(docker):** add the `OMNIROUTE_NO_SUDO` env flag for root-less / user-namespaced deployments — the MITM cert-trust command path (`resolveSudoSpawn` in `src/mitm/systemCommands.ts`) now strips the leading `sudo` when the flag is truthy, in addition to the existing root / sudo-missing cases, so the Proxy Agent runs without `sudo` (the operator trusts the CA manually, e.g. via `NODE_EXTRA_CA_CERTS`). Argv-array `spawn` preserved — no shell interpolation (Hard Rule #13). ([#6122](https://github.com/diegosouzapw/OmniRoute/issues/6122)) Regression guard: `tests/unit/mitm-systemCommands-no-sudo.test.ts`. (thanks @powellnorma)
+- **feat(providers):** add **Requesty** as an OpenAI-compatible gateway provider (BYOK, base `https://router.requesty.ai/v1`, ~200 free requests/day) — wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/requesty/`, `src/shared/constants/providers/apikey/gateways.ts`). ([#6120](https://github.com/diegosouzapw/OmniRoute/issues/6120)) Regression guard: `tests/unit/requesty-provider.test.ts`. (thanks @chirag127)
+- **feat(dashboard):** add **configured-only / available-only filters** to the Free Provider Rankings page ([#6150](https://github.com/diegosouzapw/OmniRoute/issues/6150)) — hide providers you haven't configured, or whose connections are all rate-limited / out of quota, via server-side query params (`?configuredOnly` / `?availableOnly` on `GET /api/free-provider-rankings`) backed by a testable lib helper reusing the in-process connection state (no Redis). Both filters default off, so the default view is unchanged; this supersedes the earlier client-side "Configured Only" toggle (#6245) with an available-only dimension and unit-tested logic. Regression guard: `tests/unit/freeProviderRankings-filters.test.ts`.
+
+### 🔧 Bug Fixes
+
+- **fix(dashboard):** adding a second API key connection for the same provider no longer silently overwrites the first — the Add-API-key modal now derives a unique default connection name (`main`, then `main-2`, `main-3`, …) so the backend name-based upsert can't collide (#6499 — thanks @dilneiss).
+- **fix(compression):** the session-dedup engine now also deduplicates a large multi-line block repeated **within a single message** (intra-message dedup), not just across turns; the compression-preview API surfaces a `fallbackReason`, and the fusion panel reports how many models were rate-limited vs failed on total-panel failure (#6501 — thanks @chirag127).
+- **fix(compression):** a stacked-pipeline step naming an unregistered engine now surfaces a `validationErrors` entry instead of silently no-op'ing, so misconfigured pipelines are visible in the preview API (#6506 — thanks @chirag127).
+- **feat(usage):** add a **Codex reset-credit redemption** flow to the Provider Limits UI ([#6361](https://github.com/diegosouzapw/OmniRoute/pull/6361)) — a `useCodexResetCreditRedemption` hook + `/api/usage/codex-reset-credit` route + `codexResetCredits` lib let you redeem banked Codex reset credits from the quota card. Regression guard: `tests/unit/codex-reset-credits.test.ts`. (thanks @JxnLexn)
+- **feat(glm):** add **team-plan quota settings** for `glm-cn` connections ([#6351](https://github.com/diegosouzapw/OmniRoute/pull/6351)) — a dedicated `GlmTeamQuotaFields` form section (team quota id / limits) threaded through the Add/Edit connection modals, persisted via `providerSpecificData`, with the GLM usage service reading the team quota. Regression guards: `tests/unit/glm-team-quota.test.ts`, `provider-specific-data-schema.test.ts`. (thanks @hao3039032)
+- **feat(providers):** add **TinyFish** web-fetch/search support ([#6349](https://github.com/diegosouzapw/OmniRoute/pull/6349)) — a `tinyfish-fetch` executor + `/v1/web/fetch` route + MCP web-fetch tool, registered as a specialty-media provider with request-validation and a search-provider catalog entry. Regression guards: `tests/unit/executor-tinyfish-fetch.test.ts`, `web-fetch-handler.test.ts`, `mcp-web-fetch-tool.test.ts`, `provider-validation-tinyfish.test.ts`. (thanks @dtybnrj)
+- **fix(cli):** `omniroute launch-codex` now spawns `codex.cmd` through a shell on Windows (the npm `.cmd` shim is unresolvable by bare `spawn` → ENOENT), mirroring the qodercli Windows fix (#6263) ([#6312](https://github.com/diegosouzapw/OmniRoute/pull/6312)). Regression guard: `tests/unit/launch-codex-windows-spawn-6312.test.ts`. (thanks @swingtempo)
+- **fix(codex):** isolate the **Spark** quota from the shared Codex quota and stabilize the quota UI ordering / hydration so per-scope limits render consistently ([#6336](https://github.com/diegosouzapw/OmniRoute/pull/6336)). Regression guards: `tests/unit/codex-quota-selection-hydration.test.ts`, `provider-limits-ui.test.ts` + 3 more. (thanks @xz-dev)
+- **feat(api):** add a `hidePaidModels` setting that filters paid-only models out of the `/v1/models` catalog. Regression guard: `tests/unit/models-catalog-hide-paid.test.ts`. (thanks @chirag127)
+- **fix(api-manager):** the fallback model picker now preserves combos instead of dropping them when a primary model is unavailable ([#6443](https://github.com/diegosouzapw/OmniRoute/pull/6443)). Regression guard: `tests/unit/api-manager-page-static.test.ts`. (thanks @jmengit)
+- **fix(providers):** recoverable Antigravity / Cloud-Code (Gemini Code Assist) `403` responses ([#6452](https://github.com/diegosouzapw/OmniRoute/pull/6452)) are now classified as a retryable project-config error instead of a terminal account ban, so a fixable project/API-disabled 403 no longer forces a ~1-year cooldown / full OAuth reconnect. Regression guard: `tests/unit/errorclassifier-antigravity-403.test.ts`. (thanks @developerjillur)
+- **fix(mitm):** `sanitizeHeaders` now redacts `Set-Cookie` response headers so upstream session cookies never leak into logs / diagnostics ([#6451](https://github.com/diegosouzapw/OmniRoute/pull/6451)). Regression guard: `tests/unit/mitm-sanitize-headers.test.ts`. (thanks @developerjillur)
+- **fix(api):** `/api/compression/preview` now accepts `mode: "caveman"` and correctly handles stacked / zero-compression previews ([#6425](https://github.com/diegosouzapw/OmniRoute/issues/6425)). Regression guard: `tests/unit/api/compression-preview-caveman-and-stacked-6425.test.ts`. (thanks @chirag127)
+- **feat(providers):** add **Zed** hosted LLM aggregator as a native-app provider ([#6118](https://github.com/diegosouzapw/OmniRoute/pull/6118)) — OAuth sign-in via the Zed hosted flow, registered through the shared provider registry + executor. Regression guards: `tests/unit/zed-oauth-provider.test.ts`, `zed-import-utils.test.ts`, `zed-docker-detect.test.ts`, `mitm-handler-zed.test.ts`. VPS-validated via live operator login (Hard Rule #18).
+- **fix(oauth):** the Kiro SSO-cache auto-import now **preserves the IDC region** — cross-region Amazon Q / Kiro profiles imported from the SSO cache are no longer collapsed to the default region ([#6113](https://github.com/diegosouzapw/OmniRoute/pull/6113)). Regression guard: `tests/unit/kiro-auto-import-idc-2059.test.ts`. VPS-validated via live operator login (Hard Rule #18).
+- **fix(dashboard):** passthrough model aliases no longer collide when two namespaced model ids share a last segment (port from 9router#1850, [#6431](https://github.com/diegosouzapw/OmniRoute/pull/6431)). `enx/gpt-5.5` and `enx/codebuddy/gpt-5.5` both auto-generated the alias `gpt-5.5`, so the second model could never be added (the UI just alerted "alias already exists"). Aliases are now disambiguated deterministically — bare last segment when free, then parent-qualified (`codebuddy-gpt-5.5`), then a numeric suffix — while re-adding the exact same model id is still blocked. Regression guard: `tests/unit/passthrough-alias-1850.test.ts`. (thanks @arpicato)
+- **fix(translator):** preserve a Gemini `functionResponse` co-located with other parts (another `functionCall`, or trailing `text`) in the same content when translating **Gemini → OpenAI** ([#6376](https://github.com/diegosouzapw/OmniRoute/pull/6376)). `convertGeminiContent()` early-returned the tool message on the first `functionResponse` part, dropping any co-located parts; such contents are now pre-split (one tool message per `functionResponse`, emitted first, plus one message for the remaining parts). Regression guard: `tests/unit/gemini-to-openai-function-response.test.ts`. (thanks @warelik)
+- **fix(headroom):** detect a python interpreter managed by **mise / pyenv / asdf / conda** (port from 9router#2353, [#6382](https://github.com/diegosouzapw/OmniRoute/pull/6382)). Headroom's python probe (`src/lib/headroom/detect.ts`) searched a hardcoded `PATH`, but version managers expose their interpreters via shim dirs that only join `PATH` through interactive-shell activation — which the non-interactive server never runs, so a managed python (≥3.10) was invisible and Headroom reported it missing. The search path now prepends the well-known shim/bin dirs (`~/.local/share/mise/shims`, `~/.pyenv/shims`, `~/.asdf/shims`, `$CONDA_PREFIX/bin`, `~/.local/bin`, respecting `MISE_DATA_DIR`/`PYENV_ROOT`/`ASDF_DATA_DIR` when set), and a new `HEADROOM_PYTHON` env override lets operators point straight at their interpreter (mirroring `HEADROOM_URL`). Still shell-free (`execFileSync`). Regression guard: `tests/unit/headroom-detect.test.ts` (5). (thanks @loopyd)
+- **fix(executors):** strip the OpenAI-Codex/Claude-CLI `client_metadata` passthrough field for **NVIDIA** requests (port from 9router#1887, [#6411](https://github.com/diegosouzapw/OmniRoute/pull/6411)). NVIDIA's OpenAI-compatible wrapper rejects it with `400 Unsupported parameter`, the same class already handled for `cerebras`/`mistral`; `nvidia` (executor `default`) was missing from the strip allowlist so Codex/Claude-Code passthrough requests 400'd. Regression guard: `tests/unit/executor-default-strip-client-metadata.test.ts` (+nvidia case). (thanks @phidinhmanh)
+- **fix(translator):** strip the Claude-style `thinking` field for **NVIDIA `z-ai/glm-5.2`** (port from 9router#2023, [#6413](https://github.com/diegosouzapw/OmniRoute/pull/6413)). NVIDIA's OpenAI-compatible wrapper 400s on `thinking` (a Claude-format client routed here leaves a `thinking:{type:"adaptive"}`); the existing strip rule only dropped `reasoning`. Same class already handled for `minimax-m2.7`. Regression guard: `tests/unit/nvidia-minimax-thinking-strip.test.ts` (+glm-5.2 case). (thanks @phidinhmanh)
+- **fix(translator):** suppress the streamed `</think>` close marker for the **Antigravity IDE** client (port from 9router#1061, [#6415](https://github.com/diegosouzapw/OmniRoute/pull/6415)). On thinking-only turns Antigravity rendered a bare `</think>` as the sole visible content, tripping its loop-detection and wasting requests. Antigravity's UA (`vscode/<v> (Antigravity/<v>)`) is added to the marker-suppress allowlist (alongside OpenCode); Claude Code / Cursor still get the marker, and `x-omniroute-thinking-marker: on` force-restores it. Regression guard: `tests/unit/think-close-marker-suppress-5245.test.ts`. (thanks @abdofallah)
+- **fix(executors):** strip nested `reasoning_content` from messages for **Mistral** (port from 9router#1649, [#6417](https://github.com/diegosouzapw/OmniRoute/pull/6417)). Mistral's API returns `422 extra_forbidden` when an assistant message carries `reasoning_content` (replayed thinking from a prior turn, e.g. via the Codex `/responses` path); the generic top-level 400 field-downgrade retry never covered the nested per-message field. `DefaultExecutor` now strips it for provider `mistral` only, so DeepSeek (which requires replayed `reasoning_content`) is unaffected. Regression guard: `tests/unit/mistral-strip-reasoning-content-1649.test.ts`. (thanks @xxy9468615)
+- **fix(executors):** strip the `client_metadata` passthrough field on the **OpenCode** path (port from 9router#1442, [#6418](https://github.com/diegosouzapw/OmniRoute/pull/6418)). OpenCode upstreams (e.g. `kimi-k2.6` via opencode-go) reject it with `400 "Extra inputs are not permitted, field: 'client_metadata'"`; the DefaultExecutor strip only covered cerebras/mistral and `OpencodeExecutor` extends `BaseExecutor` directly, so nothing removed it there. Regression guard: `tests/unit/opencode-strip-client-metadata-1442.test.ts`. (thanks @yanpaing007)
+- **fix(executors):** inject the `reasoning_content` echo for the native **Moonshot Kimi** provider (port from 9router#1480, [#6419](https://github.com/diegosouzapw/OmniRoute/pull/6419)). Kimi (executor `default`) is a thinking-mode upstream that 400s with "reasoning_content must be passed back" when a prior assistant turn lacks it; the placeholder injection was only wired into the OpenCode meta-provider, so direct multi-turn Kimi conversations failed. Scoped to `kimi` (gateway-served models matching the thinking-model name pattern are unaffected). Regression guard: `tests/unit/kimi-native-reasoning-injected-1480.test.ts`. (thanks @2220258345)
+- **fix(executors):** recover from a strict gateway's `context_management: Extra inputs are not permitted` 400 (port from 9router#1468, [#6420](https://github.com/diegosouzapw/OmniRoute/pull/6420)). **Claude Code** always sends a top-level `context_management` field; strict anthropic-compatible gateways reject it. The dedicated context-editing 400-fallback only fired when OmniRoute's own `contextEditing` feature was enabled (default off), so a client-sent field passed through untouched and 400'd. `context_management` is now in the generic reactive field-strip list, so it's stripped-and-retried once regardless of the feature flag (with correct request re-signing for claude-compatible relays). Regression guard: `tests/unit/provider-field-strips.test.ts`. (thanks @ohahe52-dot)
+- **fix(network):** enable **RFC 8305 Happy Eyeballs** (`autoSelectFamily`) on the direct-egress undici dispatcher (port from 9router#1237, [#6423](https://github.com/diegosouzapw/OmniRoute/pull/6423)). When DNS returns both IPv6 (AAAA) and IPv4 (A) and the IPv6 route is broken (e.g. a NAT64 `64:ff9b::` prefix without routing), undici tried IPv6 first and hung until `ETIMEDOUT` (then a 502 + account lockout), even though `curl` reached the same host. The direct dispatcher now races both families and uses whichever connects first. Proxy paths pin family via `proxyTls` and are unaffected. Regression guard: `tests/unit/direct-dispatcher-pipelining-4580.test.ts`. (thanks @adentdk)
+- **fix(combo):** round-robin now advances the rotation pointer past the model that **actually served**, not the eagerly-scheduled one (port from 9router#948, [#6428](https://github.com/diegosouzapw/OmniRoute/pull/6428)). With `stickyLimit: 1` (true round-robin), when the scheduled model failed and a _different_ model served via fallback, the counter had already advanced +1 from the scheduled index — so the next request reused the fallback-served model, degrading round-robin into hot-spotting on whichever model was healthy. The pointer now advances to the served index + 1 (mirroring the sticky-limit>1 path). Session-stickiness (#3825) and distribution are preserved. Regression guard: `tests/unit/combo-rr-fallback-advance-948.test.ts`. (thanks @binsarjr)
+- **fix(sse):** a non-string `model` field is now rejected with a `400` before the resolver, instead of crashing downstream `.toLowerCase()`/`.split()` calls into an empty-body `500` that escapes the error sanitizer ([#6407](https://github.com/diegosouzapw/OmniRoute/issues/6407)). Regression guard: `tests/unit/chat-non-string-model-6407.test.ts`. (thanks @chirag127)
+- **fix(api):** unknown `/api/*` routes now return a JSON `404` (instead of the dashboard HTML shell) and scalar chat params (`model`/`temperature`/etc.) are validated **before** the provider lookup so malformed requests fail fast with a clear `400` ([#6424](https://github.com/diegosouzapw/OmniRoute/issues/6424), [#6412](https://github.com/diegosouzapw/OmniRoute/issues/6412)). Regression guards: `tests/unit/api/api-catchall-json-404.test.ts`, `tests/unit/chat-early-schema-validation-6412.test.ts`. (thanks @chirag127)
+- **fix(api):** `/v1/chat/completions` now rejects a non-JSON `Content-Type` with a `400` before parsing the body ([#6414](https://github.com/diegosouzapw/OmniRoute/issues/6414)). Regression guard: `tests/unit/v1-chat-completions-content-type-6414.test.ts`. (thanks @chirag127)
+- **fix(api):** the `X-OmniRoute-Compression` response header is now echoed on `/v1/chat/completions` and `/v1/completions` ([#6422](https://github.com/diegosouzapw/OmniRoute/issues/6422)). Regression guard: `tests/unit/compression-header-echo-6422.test.ts`. (thanks @chirag127)
+- **fix(api):** concurrent `GET /v1/models` requests are coalesced into a single catalog build ([#6408](https://github.com/diegosouzapw/OmniRoute/issues/6408)). Regression guard: `tests/unit/v1-models-concurrent-6408.test.ts`. (thanks @chirag127)
+- **fix(api):** `/v1/completions` now echoes the requested `body.model` in its JSON + streamed responses ([#6429](https://github.com/diegosouzapw/OmniRoute/pull/6429)). Regression guard: `tests/unit/completions-body-model-echo.test.ts`. (thanks @chirag127)
+- **fix(api):** env-var master keys now see the full `/v1/models` catalog ([#6406](https://github.com/diegosouzapw/OmniRoute/issues/6406)). Regression guard: `tests/unit/models-catalog-envkey-6406.test.ts`. (thanks @chirag127)
+- **fix(api):** non-streaming `/v1/completions` responses now echo `body.model` aligned with the `X-OmniRoute-Model` header ([#6426](https://github.com/diegosouzapw/OmniRoute/issues/6426)). Regression guard: `tests/unit/v1-completions-model-header-match-6426.test.ts`. (thanks @chirag127)
+- **fix(api):** unknown `/v1/*` routes now return a JSON `404 not_found` instead of the Next.js dashboard HTML shell ([#6405](https://github.com/diegosouzapw/OmniRoute/issues/6405)). Regression guard: `tests/unit/api/v1-catchall-json-404.test.ts`. (thanks @chirag127)
+- **fix(api):** the per-connection provider models route now degrades to the shipped catalog when a provider's `/models` endpoint answers with a redirect ([#6267](https://github.com/diegosouzapw/OmniRoute/issues/6267)) — a `qwen-web` import failed with a raw `Redirect blocked … (307)` 503. `safeOutboundFetch` throws `REDIRECT_BLOCKED` on the 307, `getSafeOutboundFetchErrorStatus` maps it to 503, and `buildDiscoveryErrorFallbackResponse` treated every 503 as a hard error — so the non-empty `getModelsByProviderId("qwen-web")` catalog was never surfaced. A models-endpoint redirect is not a fixable-config error (unlike `URL_GUARD_BLOCKED`/`INVALID_URL`, which stay hard errors), so it now falls back to the local/cached catalog before the 503 short-circuit. General fix — covers any config-driven provider that 307s. Regression guard: `tests/unit/provider-models-qwen-web-redirect-6267.test.ts`. (thanks @chirag127)
+- **fix(api):** the per-connection provider models route (MCP `list_models_catalog` + the dashboard import view) now merges USER-ADDED custom models into its response ([#6247](https://github.com/diegosouzapw/OmniRoute/issues/6247)) — custom models live in the `key_value` namespace `customModels`, which the live REST `/api/v1/models` already merges, but `src/app/api/providers/[id]/models/route.ts` never read `getCustomModels`, so custom models were dropped on both the discovery-success and local_catalog paths. They are now folded into the returned model list (deduped by id, stamped `owned_by: provider`), fixing MCP + the dashboard import view in one place. Regression guard: `tests/unit/provider-models-custom-merge-6247.test.ts`. (thanks @RCrushMe)
+
+- **fix(providers):** GitLab Duo tool-calling follow-up turns no longer fail upstream with `422 {"detail":"Validation error"}` (tokens 0/0, rejected pre-inference). The #6234 tool-result-feedback fix serialized the **entire** multi-turn conversation into GitLab's single-file `code_suggestions` (`small_file`) generation endpoint — folded history that turn-N sent as an oversized `current_file.content_above_cursor` **and** duplicated verbatim into `user_instruction`, tripping the AI-Gateway's `small_file` validation guard. The executor now **bounds** that prompt: it keeps system + latest user message + the most-recent tool round (dropping older turns), caps oversized tool results, and stops duplicating the full prompt into `user_instruction` (which now carries only the short latest user message) — while still feeding the most-recent tool result back so the agent continues (`open-sse/executors/gitlab.ts`, [#6220](https://github.com/diegosouzapw/OmniRoute/issues/6220)). The unit test covers the bounding logic; the upstream 422→200 clearing is VPS-only (Hard Rule #18). Regression guard: `tests/unit/gitlab-tool-exchange-bounded-6220.test.ts`.
+
+- **fix(i18n):** the provider-detail (`/dashboard/providers/[id]`) connection-status filter labels no longer render as `__MISSING__:All` / `__MISSING__:Active` / `__MISSING__:Error` / `__MISSING__:Banned` / `__MISSING__:CreditsExhausted` in non-English locales (notably pt-BR) ([#6290](https://github.com/diegosouzapw/OmniRoute/issues/6290)). Root cause was **not** the namespace mismatch the issue guessed — the `providers.filter*` keys resolve correctly in `en.json`; the debt lived in the locale mirrors (`src/i18n/messages/*.json`), where these five keys carried the `__MISSING__:` sync sentinel in ~15 locales and were absent entirely in ~26 others, so next-intl found the key and echoed the sentinel verbatim. All 40 non-English/-Chinese mirrors now ship real translations for the five `providers.filter*` labels. Regression guard: `tests/unit/i18n-provider-filter-keys-6290.test.ts`. (thanks @diegosouzapw)
+
+- **fix(providers):** the `copilot-m365-web` streaming executor now emits `debug`-level WebSocket diagnostics ([#6210](https://github.com/diegosouzapw/OmniRoute/issues/6210)) — the outbound WS URL (with the `access_token` **redacted** via `redactWsUrl()`), handshake success/failure, and each received SignalR frame's `type`/`target`. Previously the streaming path logged nothing, so an empty `content:null` response (the M365 Education / Starter tier symptom fixed in #6234) was undiagnosable even at `APP_LOG_LEVEL=debug`. The change is debug-level and side-effect-free — it does not alter streaming behavior or the frame parser, and the token never reaches the logs. Regression guard: `tests/unit/copilot-m365-web-logging-6210.test.ts` (thanks @qpeyba)
+
+- **fix(resilience):** a round-robin combo no longer returns `503 all upstream accounts are unavailable` when a compatibility-rejected target is actually healthy ([#6238](https://github.com/diegosouzapw/OmniRoute/issues/6238)). `filterTargetsByRequestCompatibility` drops request-incompatible targets (tool/vision/structured-output unsupported, or below the required context window) **before** any availability check runs, and its `compatible.length === 0` safety net only fired when _all_ targets were filtered — not when the kept targets later all turned out runtime-unavailable (circuit-open / cooldown / no credentials). So a combo could 503 while a compat-rejected-but-healthy provider sat unused. `handleRoundRobinCombo` now keeps the compat-rejected set and, when every compat-kept target was skipped without a single real attempt, probes those rejected targets as a **last-resort fallback tier** (via the new pure `open-sse/services/combo/comboCompatFallback.ts`) before crystallizing the 503. Regression guard: `tests/unit/combo-roundrobin-compat-fallback-6238.test.ts`. (thanks @ThongAccount)
+
+- **fix(startup):** best-effort self-heal for a corrupted Turbopack dev cache on Windows ([#6289](https://github.com/diegosouzapw/OmniRoute/issues/6289)). On Windows, `pnpm dev` can fail at startup when Turbopack `mmap`s a persistent-cache SST file and the OS refuses the mapping (`os error 1455` — "paging file too small"), which Turbopack surfaces as a misleading `Module not found: Can't resolve '@/shared/utils/machine'`. This is a **known upstream Turbopack cache-corruption bug — not our code**. The dev launcher (`scripts/dev/run-next.mjs`) now wraps `nextApp.prepare()` and, when it rejects with that signature (`isTurbopackCacheCorruption` in the new `scripts/dev/turbopackCacheHeal.mjs`), purges `.build/next/**/cache/turbopack` and retries **once** with a clear log. **Caveat — best-effort only:** the corruption often surfaces as a runtime overlay rather than a `prepare()` rejection, so this cannot always intercept it; the reliable remedy remains manually deleting the Turbopack cache dir. Regression guard: `tests/unit/turbopack-cache-heal-6289.test.ts`. (thanks @chirag127)
+
+- **fix(providers):** qodercli PAT auth no longer fails with `spawn qodercli ENOENT` on Windows ([#6263](https://github.com/diegosouzapw/OmniRoute/issues/6263)) — `spawnQoderCli` spawned the bare `qodercli` name with `shell:false` and an unenriched env, so the npm `.cmd` wrapper under `%APPDATA%\npm` (a user-PATH directory) was never resolved. It now resolves the absolute `.cmd`/`.exe` path through the existing `getCliRuntimeStatus("qoder")` resolver in `src/shared/services/cliRuntime.ts` (memoized), spawns with `shell` when the target is a `.cmd`/`.bat`, and uses the cliRuntime-enriched env (PATH + PATHEXT + APPDATA); the ENOENT error now lists the searched paths plus the `CLI_QODER_BIN` override. End-to-end spawn on a real Windows host is host-only (Hard Rule #18); the path-resolution logic is unit-tested. Regression guard: `tests/unit/qodercli-windows-resolve-6263.test.ts`. (thanks @chirag127)
+
+- **fix(sse):** the reasoning-token buffer no longer inflates **probe-sized `max_tokens`** ([#6274](https://github.com/diegosouzapw/OmniRoute/issues/6274)) — Claude Code's `/model` capability check sends `max_tokens: 1`, but for a thinking-capable model with a large output cap (e.g. `glm-5.2`) the #3587 headroom heuristic (`max(current + 1000, ceil(current * 1.5))`) rewrote it to `1001` and forwarded that upstream, wasting tokens on a request that was never a genuine reasoning budget. `resolveReasoningBufferedMaxTokens()` (`open-sse/services/reasoningTokenBuffer.ts`) now short-circuits and returns the caller's value verbatim when it is below the new `REASONING_BUFFER_MIN_TRIGGER` (256) threshold — a tiny explicit limit is a probe, not a reasoning request. Real budgets still receive the #3587 headroom unchanged, and the guard runs after the existing capability checks so unknown / non-reasoning models keep returning `null`. Regression guard: `tests/unit/reasoning-token-buffer-6274.test.ts`. (thanks @brightfiscalband)
+
+- **fix(cli):** `omniroute reset-password` now works as a real subcommand, and password resets over piped (non-TTY) stdin actually apply ([#6261](https://github.com/diegosouzapw/OmniRoute/issues/6261), [#6258](https://github.com/diegosouzapw/OmniRoute/issues/6258)). Two coupled defects: (1) **#6261** — `bin/omniroute.mjs` routed everything through Commander with only two pre-Commander bypasses (`--mcp`, `reset-encrypted-columns`), so `omniroute reset-password` was rejected as an unknown command; only the separate `omniroute-reset-password` bin worked, while the docs falsely advertised the subcommand (incl. a bogus "legacy alias still works"). A pre-Commander bypass mirroring `reset-encrypted-columns` now dynamically imports `bin/reset-password.mjs` (which self-executes) before Commander parses; the three doc lines were corrected. (2) **#6258** — `bin/reset-password.mjs` issued two sequential `rl.question` prompts; under piped stdin the second read never settled at EOF, so `main()` never reached `resetManagementPassword` and the reset was a silent no-op (both prompts printed, no success, password unchanged). The CLI now detects non-TTY stdin and reads it once (first line = password, second line = confirm if present, else reused), adds a `--password-stdin` flag (entire stdin is the password, no confirmation), and exits `0` explicitly so the success line always flushes; interactive TTY behavior is unchanged. Regression guard: `tests/unit/reset-password-cli-6261-6258.test.ts` (3). (thanks @chirag127)
+
+- **fix(db):** the mass-migration **safety abort** now tells the operator how to bypass it and stops flooding the log ([#6260](https://github.com/diegosouzapw/OmniRoute/issues/6260)) — after restoring a backup that wiped the migration tracking table, `runMigrations()` threw the abort on every downstream `ensureDbInitialized()`, re-logging the full banner 11+ times, and the message never mentioned the existing `OMNIROUTE_MAX_PENDING_MIGRATIONS` escape hatch. The abort text now appends a bypass hint (set `OMNIROUTE_MAX_PENDING_MIGRATIONS=0` in `server.env` / `DATA_DIR/.env`), and a new `MigrationSafetyAbortError` is memoized so repeated calls in the same process throw the same instance and emit a single concise line instead of the full cascade. Regression guard: `tests/unit/migration-safety-abort-6260.test.ts`. (thanks @chirag127)
+
+- **fix(auth):** importing a **distinct** Codex/ChatGPT OAuth `auth.json` is no longer falsely rejected as "already exists" when it belongs to a different user in the same workspace ([#6301](https://github.com/diegosouzapw/OmniRoute/issues/6301)). `findExistingCodexConnection` (in `src/lib/oauth/utils/codexAuthImport.ts`) deduped **only** on `providerSpecificData.workspaceId === accountId`, where `accountId` is the shared `chatgpt_account_id`/`tokens.account_id` — so two members of the same ChatGPT Team collapsed onto a single connection (409 `duplicate_account`). The id_token's `https://api.openai.com/auth` claim carries a per-user `chatgpt_user_id` alongside the workspace id (the device-flow path already persisted it as `chatgptUserId`, but the import path did not). Now `parseAndValidateCodexAuth` extracts `userId` (`chatgpt_user_id` → `user_id` → JWT `sub`) into `ParsedCodexAuth`, the create/update paths persist `chatgptUserId` in `providerSpecificData` (mirroring `codex.ts`), and dedup keys on `workspaceId` **AND** `chatgptUserId` — with a backward-compat fallback to legacy accountId-only matching when no stored connection for that workspace records a `chatgptUserId`, so genuinely-same accounts still dedup. Regression guard: `tests/unit/codex-auth-import-userid-dedup-6301.test.ts` (4). (thanks @anungma)
+
+- **fix(providers):** importing models for the **venice-web** provider no longer fails with a red "Provider venice-web does not support models listing" ([#6269](https://github.com/diegosouzapw/OmniRoute/issues/6269)). `venice-web` is a web-cookie provider with an executor but no upstream `/v1/models` endpoint and no registry `models`, so the models route fell through to the tail `400`. Mirroring the `jules`/`linkup-search`/`ollama-search` fix (#5569), it now ships a static local catalog entry in `src/lib/providers/staticModels.ts` — seeding the current Venice lineup (`venice-uncensored`, `llama-3.3-70b`, `qwen3-235b`, `qwen3-4b`, `deepseek-r1-671b`; Venice rotates its catalog, see docs.venice.ai/models/overview) — so the route returns `200` with `source:"local_catalog"`, `intentional:true`. Regression guard: `tests/unit/static-models-venice-web-6269.test.ts`. (thanks @chirag127)
+
+- **fix(api):** the specialty model catalogs (`/v1/embeddings`, `/v1/images`, `/v1/music`, `/v1/videos` model lists) are now derived from the **unified catalog filtered by a predicate** (`getSpecialtyModelsResponse`) instead of ad-hoc per-route logic, so they consistently respect active-credential visibility and stay in sync with the main catalog ([#6303](https://github.com/diegosouzapw/OmniRoute/pull/6303)). Regression guard: `tests/unit/specialty-model-catalog-routes.test.ts`. (thanks @makcimbx)
+- **fix(api):** the agent-bridge server route now resolves the MITM manager via a **dynamic `import("@/mitm/manager.runtime")`** so Turbopack does not statically pull the stub (or over-bundle the manager), and the agent-skills generator anchors its output base path with `path.join(process.cwd(), …)` so Turbopack's static analyzer stops tracing the whole project root ([#6329](https://github.com/diegosouzapw/OmniRoute/issues/6329), [#6366](https://github.com/diegosouzapw/OmniRoute/pull/6366)). Regression guard: `tests/unit/agent-bridge-server-route-dynamic-import.test.ts`. (thanks @Iammilansoni)
+- **fix(api):** internal probes (combo-test, cloud-sync verify) now pick a **management-scoped / allow-all API key** instead of naively grabbing `getApiKeys()[0]` — a restricted `self:usage` first row made the probe fail with "Model X is not allowed for this API key" even when the combo path was healthy (`pickApiKeyForInternalUse` in `src/lib/db/apiKeys.ts`). The API-manager model editor also falls back to `/api/models?all=true` when `/v1/models` is catalog-protected ([#6372](https://github.com/diegosouzapw/OmniRoute/pull/6372)). Regression guard: `tests/unit/pick-internal-api-key-6372.test.ts`. (thanks @jmengit)
+- **fix(live-ws):** the Live Dashboard WebSocket server now **rejects on bind failure** (e.g. `EADDRINUSE` when the API bridge already holds the port) instead of letting the error surface as an unhandled `error` event that crash-loops the process — the `error` listener is attached to `wss` (not `server`) and releases the EventBus subscription on a failed start ([#6324](https://github.com/diegosouzapw/OmniRoute/issues/6324)). Regression guard: `tests/unit/live-ws-eaddrinuse-6324.test.ts`. (thanks @vinayakkulkarni)
+- **fix(dashboard):** the Home provider-topology widget now trusts the live provider-metrics snapshot — it uses `topology.errorProvider` and live `activeRequests` directly instead of re-deriving state from a stale `lastErrorAt` or applying a frontend timeout filter, so the topology reflects real-time provider health ([#6322](https://github.com/diegosouzapw/OmniRoute/pull/6322)). Regression guard: `tests/unit/home-provider-topology-live-state.test.ts`. (thanks @xz-dev)
+- **fix(sse):** strip zero-width markers from streamed **tool-call arguments** — a follow-up to [#5857](https://github.com/diegosouzapw/OmniRoute/pull/5857). That PR removed injected zero-width joiners (U+200D) from streamed assistant text/reasoning but deliberately left tool-call argument JSON byte-exact. The request-side obfuscation (`open-sse/services/claudeCodeObfuscation.ts`) injects ZWJ into agent words — including the temp path inside the Bash tool description — and Claude models copy that verbatim into generated commands, which are delivered as tool-call arguments rather than assistant text. As a result the ZWJ survived and corrupted code blocks (e.g. a temp path rendered with an invisible joiner). Now `open-sse/handlers/responseSanitizer.ts` strips zero-width code points from tool-call argument strings at every emit site (OpenAI non-stream/stream chat `tool_calls` + legacy `function_call`, native Responses `function_call` items, the OpenAI→Responses conversion, and the native Responses streaming `response.function_call_arguments.delta/.done` events). Only zero-width code points are removed; JSON structure and all other bytes stay identical (no parse/restringify), so normal arguments remain byte-exact. Regression guard: 6 new cases in `tests/unit/response-sanitizer.test.ts` (suite 50/50).
+
+- **fix(nodejs):** the default app log path now resolves under `DATA_DIR` (`~/.omniroute/logs/application/app.log`) instead of `process.cwd()` ([#6197](https://github.com/diegosouzapw/OmniRoute/issues/6197)) — the globally-installed CLI runs from an arbitrary working directory, so anchoring the default to cwd made file logging silently write to (or no-op under) an unrelated directory, contradicting the documented `.env.example` default. `getAppLogFilePath()` now computes the default lazily via the pure `resolveDataDir()` resolver (honours a per-process `DATA_DIR`, no directory-creation side effect); an explicit `APP_LOG_FILE_PATH` still wins. Regression guard: `tests/unit/logenv-datadir-path-6197.test.ts` (3).
+
+- **fix(docker):** AgentBridge/`startMitm` no longer aborts in containers/headless when the Antigravity-default DNS step can't write `/etc/hosts` ([#6127](https://github.com/diegosouzapw/OmniRoute/issues/6127)), and the privileged command's stderr now reaches `app.log` instead of only a bare exit code hitting the toast ([#6198](https://github.com/diegosouzapw/OmniRoute/issues/6198)). The default DNS step (`addDNSEntry`) was called unguarded while cert install and the two sibling DNS steps were each best-effort — in the runtime Docker image (`USER node`, no `sudo`, read-only `/etc/hosts`) it threw `Command failed with code 1` out of `startMitmInternal` and killed the whole start, discarding the stderr. The three DNS steps are extracted into a best-effort `provisionDnsEntries()` where each failure is logged with the full `err` (stderr included, folded in by `systemCommands.ts`) and never aborts the start. Regression guard: `tests/unit/mitm-dns-graceful-degrade-6127.test.ts` (4).
+
+- **fix(providers):** copilot-m365-web now supports the M365 Education "Starter / OfficeWebIncludedCopilot" tier and no longer returns an empty `content:null` stream ([#6210](https://github.com/diegosouzapw/OmniRoute/issues/6210)). Two gaps: (1) `buildWsUrl()` hardcoded the individual-consumer scenario (`OfficeWebPaidConsumerCopilot`, `isEdu=false`) — the EDU tier is now opt-in via `providerSpecificData.tier="edu"`, emitting `scenario=OfficeWebIncludedCopilot`/`isEdu=true` (the individual path is unchanged); (2) the EDU/GPT-5.5 path streams deltas via `arguments[0].writeAtCursor` (incremental) instead of only `messages[].text` (accumulated snapshots), which the parser dropped — a new `accumulateBotContent()` folds both formats, with `type:2 item.result.message` as a last-resort fallback. Regression guard: `tests/unit/copilot-m365-edu-writeatcursor-6210.test.ts` (10). (thanks @qpeyba)
+
+- **fix(providers):** GitLab Duo executor now feeds tool results back into the prompt instead of looping ([#6220](https://github.com/diegosouzapw/OmniRoute/issues/6220)) — `buildPrompt()` branched only on `system`/`user` and took `userParts.at(-1)`, silently dropping the `assistant{tool_calls}` + `tool{result}` turns the client appended, so the reconstructed prompt was byte-identical to turn 1 and the model re-emitted the same `<tool>` call forever. When a tool exchange is present the full conversation is now serialized, folding each tool result back keyed by its `tool_call_id`; simple conversations keep the legacy shape. Complements the tool_call emission from [#6051](https://github.com/diegosouzapw/OmniRoute/issues/6051) (the `kilo-duplicate` label was a false positive — different, sequential defect). Regression guard: `tests/unit/gitlab-tool-result-feedback-6220.test.ts` (4).
+
+- **fix(providers):** opencode-go/opencode-zen can now synthesize the OpenCode CLI identity headers Cloudflare requires on VPS egress ([#5997](https://github.com/diegosouzapw/OmniRoute/issues/5997)) — on a datacenter VPS, `opencode.ai/zen/go/v1/chat/completions` 403s (HTML challenge) requests lacking CLI identity, while the reporter's control curl proved that `User-Agent: opencode-cli/1.0.0` + `x-opencode-client: cli` + `x-opencode-project: default` + fresh request/session UUIDs succeed. Opt-in via `OPENCODE_SYNTHESIZE_CLI_HEADERS=true` (values overridable via `OPENCODE_GO_USER_AGENT`/`OPENCODE_USER_AGENT`/`OPENCODE_CLIENT`/`OPENCODE_PROJECT`); it fills only headers the client did not already send. Kept **off by default** — the forward-only path is deliberate (fabricating a wrong value risks upstream rejection; a prior dedup regressed with `opencode/local`), so this replaces the fragile local header-injection shim without changing default behavior. Regression guard: `tests/unit/opencode-cli-headers-synthesis-5997.test.ts` (6). (thanks @aleksesipenko)
+- fix(resilience): sticky session affinity now evicts and fails over to another account when the pinned account is exhausted/unavailable (#6219)
+
+- fix(sse): Responses API passthrough now drops internal commentary-phase output before forwarding to clients (gated by RESPONSES_PASSTHROUGH_DROP_COMMENTARY, default on) (#6199)
+
+- **fix(sse):** tool-call function schemas with a root `type: null` are now coerced to `type: "object"` before dispatch ([#6359](https://github.com/diegosouzapw/OmniRoute/issues/6359)) — clients like the Codex app emit `parameters: { type: null, ... }` for some tools, which OpenAI-compatible upstreams reject with `400 Invalid schema for function '...': schema must be a JSON Schema of 'type: "object"', got 'type: null'`, failing the whole request. `toolSchemaSanitizer` already stripped the null; it now re-adds the mandatory root `"object"` type (and empty `properties`/open `additionalProperties` when absent). Combinator roots (`anyOf`/`oneOf`/`allOf`) and explicit root types are left untouched. Regression guard: 5 new cases in `tests/unit/tool-schema-sanitizer.test.mjs`.
+
+- **fix(docker):** AgentBridge no longer fails to start on npm/Electron/VPS installs with "MITM manager stub reached at runtime" ([#6344](https://github.com/diegosouzapw/OmniRoute/issues/6344)) — v3.8.45 flipped the production bundler default to Turbopack, but `next.config.mjs` aliased `@/mitm/manager` to its Docker-only degraded stub **unconditionally**. That was harmless while Docker (which sets the alias intentionally for #3390 graceful degradation) was the sole Turbopack consumer, but once every artifact built with Turbopack the stub shipped to all non-Docker users and `startMitm` threw on the first Agent-Bridge start. The alias is now opt-in via `OMNIROUTE_MITM_STUB=1` (set only by the Dockerfile) through the shared `scripts/build/mitm-stub-flag.mjs` helper; default builds bundle the real manager. Regression guard: `tests/unit/mitm-stub-alias-6344.test.mjs` (4).
+
+- **fix(proxy):** stop the v3.8.44 proxy regression that leaked the real IP and disabled healthy proxies ([#6246](https://github.com/diegosouzapw/OmniRoute/issues/6246)). Two coupled defects from the new health scheduler: (1) **IP leak** — when a proxy assigned to a connection was marked `inactive`, resolution fell through to a **direct** egress instead of blocking, exposing the operator's real IP; (2) **over-deactivation** — the sweep flipped a proxy to `inactive` on the **first** failed probe and counted our own 5s timeout / a probe-target `5xx` as the proxy's fault, so healthy paid proxies vanished from egress selection ("my proxies are not being used anymore"). Fix: the sweep decision is extracted into a pure, network-free `decideProxyHealthAction` (`src/lib/proxyHealth/decision.ts`) — by default the health check now **only counts/logs and never downgrades status** (a proxy is downgraded/removed only with `PROXY_AUTO_REMOVE=true`, after `PROXY_AUTO_REMOVE_AFTER` **consecutive** conclusive failures); probes are classified tri-state so an inconclusive result (our timeout, or a `5xx` from the probe target) never penalizes the proxy, and the probe timeout is raised 5s→15s. Separately, `safeResolveProxy` now **fails closed** via the existing policy: a connection whose assigned proxy is dead is blocked instead of leaking direct (`hasBlockingProxyAssignment`), honoring the explicit `proxy off` toggles and the `PROXY_FAIL_OPEN=true` opt-out. Existing proxies stuck `inactive` by the old behavior need a one-time manual re-activate (the operator owns proxy status). Regression guards: `tests/unit/proxy-health-decide-action-6246.test.ts`, `tests/unit/proxy-assigned-unavailable-6246.test.ts`.
+
+- **fix(proxy):** make "Test All" read-only and add bulk enable/disable ([#6246](https://github.com/diegosouzapw/OmniRoute/issues/6246)). Complements the core fail-closed / scheduler fix (#6296) with the two remaining reporter asks. (1) The **"Test All" button** (`POST /api/settings/proxies/auto-test`) used to flip a proxy to `inactive` on a failed reachability probe; since the egress selector excludes `inactive` proxies, a flaky probe (an unreachable `httpbin.org`, a proxy that blocks `HEAD`, or a slow paid proxy) silently disabled every proxy that failed — "Test All" is now **read-only by default** (only the operator sets a proxy active/inactive; opt back into the legacy test-and-set with `PROXY_HEALTH_AUTO_DEACTIVATE=true`). (2) Adds a **bulk enable/disable** proxies endpoint + toolbar action (`POST /api/settings/proxies/batch-activate`) so an operator can re-activate proxies in one click. Regression guard: `tests/unit/proxy-health-6246.test.ts`. (thanks @tenshiak)
+
+- **chatcore (tools): stop the default 128-tool cap from silently dropping opencode's `task`/MCP tools.** opencode (used as an MCP/agent host) sends a large tool list; when it exceeds the speculative `MAX_TOOLS_LIMIT` (128) default, `truncateToolList` did a blind `tools.slice(0, 128)`, dropping every tool past index 128 — including opencode's built-in `task` tool (subagent launch) and many MCP tools, so models routed through OmniRoute could no longer spawn subagents or reach part of their tools. The cap exists to avoid upstream `400`s for providers with real hard limits (e.g. grok-cli 200), so it is kept for those: detection of the opencode client (`isOpencodeClient` — any `x-opencode-*` header, or `opencode` in the user-agent) now only bypasses the **speculative 128 default**, never a known provider ceiling. Precedence is explicit — a proactive/detected provider limit always truncates (even for opencode); otherwise opencode forwards its full tool list; otherwise the unchanged 128 default applies to every other client. Refactors `getEffectiveToolLimit` into `getKnownToolLimit(provider) ?? DEFAULT_LIMIT` (byte-identical for existing callers) and fixes a cosmetic debug-log that reported the truncated count instead of the original. Regression guard: `tests/unit/tool-limit-detector.test.ts`.
+
+- **fix(mitm):** the macOS MITM-cert install check now matches the system keychain again. `security find-certificate -a -Z` prints the SHA-1 as a colon-less hex string, but the installed-check compared it against `getCertFingerprint()`'s colon-separated form, so the substring match never hit — the cert was reported as not-installed and re-prompted for the sudo install on every run. Fingerprints are now normalized (colons stripped, upper-cased) on both sides via the extracted `macCertOutputHasFingerprint` helper. Regression guard: `tests/unit/mitm-cert-mac-fingerprint.test.ts`. ([#6204](https://github.com/diegosouzapw/OmniRoute/pull/6204), closes [#6134](https://github.com/diegosouzapw/OmniRoute/issues/6134) — thanks @rianonehub)
+
+- **fix(api):** `/v1/messages/count_tokens` now counts `tool_use`, `tool_result` and `thinking` content blocks (and array-form `system` prompts) in the local-estimation path, instead of only `text`. Real agentic conversations keep ~95% of their tokens inside tool results; the previous estimate returned near-zero for them, which silently broke Claude Code's auto-compaction (context grew past the window with no compaction until the upstream API rejected the request). The real provider-side count path is unchanged. Regression guard: `tests/unit/messages-count-tokens-route.test.ts`. ([#6221](https://github.com/diegosouzapw/OmniRoute/pull/6221) — thanks @luweiCN)
+
+- **fix(antigravity):** strip a trailing assistant prefill turn for Vertex Claude models to avoid upstream 400s ([#6114](https://github.com/diegosouzapw/OmniRoute/pull/6114)). Regression guard: `tests/unit/antigravity-claude-prefill-strip.test.ts`. (thanks @anki1kr)
+
+- **fix(security):** the mutable cloud-agent routes (`/api/cloud/credentials/update`, `/api/cloud/models/alias`) now require management auth instead of being treated as public. They were classified as public API routes, so a request without management credentials could update stored cloud-agent credentials and model aliases. They are removed from the public-route set, classified as management routes in the authz pipeline, and gated by `requireManagementAuth`; cloud **read**/auth routes stay public. Regression guards: `tests/unit/cloud-write-auth.test.ts`, `tests/unit/authz/classify.test.ts`, `tests/unit/public-api-routes.test.ts`. ([#6233](https://github.com/diegosouzapw/OmniRoute/pull/6233) — thanks @vittoroliveira-dev)
+
+- **refactor(dashboard):** extract the onboarding-wizard "Open provider details" link target into a pure, unit-tested `buildProviderDetailsHref(connection)` helper. The wizard already routes by `connection.id` (the node UUID) rather than the provider category slug (#6144/#6145); this hardens that behavior behind a tested helper that guards a missing id/connection. Regression guard: `tests/unit/provider-onboarding-href.test.ts`. ([#6166](https://github.com/diegosouzapw/OmniRoute/pull/6166) — thanks @KooshaPari)
+
+- **fix(security):** the doubao synthetic device-id generator now derives its digits via an unbiased crypto-random draw (rejection sampling over `crypto.randomBytes()`) instead of a `% 10` reduction, closing a CodeQL `js/biased-cryptographic-random` finding.
+- **fix(agentSkills):** the GitHub-skills generator now resolves `outputDir` to an absolute path before writing, fixing a regression introduced by #6366 (relative-to-cwd base path) that could write generated skill files to the wrong directory.
+- **fix(security):** `/api/keys/{id}/devices` now checks the HTTP method before auth/validation, returning a `405` for non-GET/DELETE verbs instead of a misleading `401`/`500` (closes a `dast-smoke` QUERY-method finding).
+- **fix(quality):** clear the last 2 heavy quality-gate reds on the release tip (cycle pre-flight).
+- **fix(mitm):** the test suite and CI can never mutate the OS trust store — `OMNIROUTE_SKIP_SYSTEM_TRUST=1` is set globally for tests/CI so `installCert`/`uninstallCert`/`installTproxyCa` skip the privileged OS dispatch ([#6310](https://github.com/diegosouzapw/OmniRoute/pull/6310); full detail is under the [3.8.45] section below — this branch received it via the parallel-cycle sync-back).
+- **fix(api):** `POST /api/github-skills` now Zod-validates its request body; documented the new quality-gate env vars and pinned the merge-integrity GitHub Actions to a commit SHA.
+- **fix(skills):** generate the missing `omni-github-skills` registry entry and align the agent-skills catalog-count tests (follow-up to #6186).
+- **fix(quality):** clear the cycle's 11 net-new ESLint errors and make `validate-release-green` suppressions-aware.
+
+### 📝 Maintenance
+
+- **i18n(it):** add 118 missing Italian (`it`) translations (net-additive — no existing keys dropped, valid JSON), improving Italian UI coverage. ([#6212](https://github.com/diegosouzapw/OmniRoute/pull/6212) — thanks @serverless83)
+- **chore(providers):** remove deprecated **MiMo V2** model entries from the catalogs (xiaomi-mimo, opencode-go, zenmux-free, audio TTS) — the upstream V2 line is superseded by MiMo V2.5; drops `mimo-v2-tts`, `mimo-v2-pro`, `mimo-v2-omni`, `mimo-v2-flash`, `mimo-v2-flash-free` and realigns the provider-catalog tests. ([#6248](https://github.com/diegosouzapw/OmniRoute/pull/6248) — thanks @backryun)
+- **chore(release):** ~50 commits on this branch are v3.8.45 pre-flight/hardening fixes and CI-perf work that landed here via the parallel-cycle sync-back (`sync-next-cycle.mjs`, Hard Rule #21) **after** the `v3.8.45` git tag was cut, and are already fully documented under the **[3.8.45]** section below — listed here only so the per-cycle commit-coverage check (`npm run release:uncovered`) doesn't flag them as gaps. Provider/catalog/UX/backend fixes: #6041, #6078, #6108, #6135, #6148, #6149, #6154, #6158, #6161, #6162, #6163, #6164, #6165, #6170, #6177, #6178, #6181, #6186, #6187, #6191, #6193, #6194, #6195, #6200, #6205, #6208, #6209, #6211, #6213, #6223, #6224, #6225, #6226, #6227, #6228, #6229, #6230, #6235, #6291, #6292. CI/release-pipeline work: #6167, #6203, #6214, #6215, #6218, #6273, #6275, #6283, #6284, #6285, #6300, #6305.
+- **chore(release):** additional zero-ref release-cycle plumbing on this branch, kept out of `release:uncovered` on purpose (no `#N` in the commit subject to cite): opening the v3.8.46 cycle, opening/closing the v3.8.45 cycle, the finalized [3.8.45] CHANGELOG i18n sync-back to 42 mirrors, the v3.8.45 cognitive/cyclomatic and file-size drift rebaselines, ESLint stale-suppression pruning (4,273 → 4,233), and clearing test-masking/docs-all pre-flight reds for v3.8.45.
+
+### ⚡ Performance & Infrastructure
+
+- **perf(release-green):** the pre-flight validator (`scripts/quality/validate-release-green.mjs`) now runs its 4 slow suites (unit / vitest / integration / pack-artifact) **concurrently** via `Promise.all` — pre-flight wall time drops from ~the sum of the suites to ~the slowest one (~30min saved per round; Phase 0 was the nº1 bottleneck of the v3.8.45 release benchmark, 2h54 of 6h34 e2e). Guard: `tests/unit/validate-release-green.test.ts` ("runs the slow suites CONCURRENTLY"). ([#6319](https://github.com/diegosouzapw/OmniRoute/pull/6319))
+- **fix(ci):** `scripts/release/sync-next-cycle.mjs` — two defects found live in its first production run (v3.8.45 Phase 5): (1) the `git()` helper's default 1 MiB `maxBuffer` crashed with `ENOBUFS` on `git show origin/main:CHANGELOG.md` (the CHANGELOG alone is >1 MiB) — widened to 64 MiB; (2) the i18n resync only propagated the `[NEXT]` (TBD) section, leaving the just-shipped finalized section as "— TBD" in all 42 mirrors — it now also syncs `[prevVersion]` bounded by the heading below it (new exported pure helper `versionAfter`). Guards: +5 tests in `tests/unit/sync-next-cycle.test.ts` (8/8). ([#6327](https://github.com/diegosouzapw/OmniRoute/pull/6327))
+- **test(ci):** concurrency-sensitive flaky tests are **quarantined into a serial pass** (`tests/unit/serial/`, `--test-concurrency=1`, appended to every unit runner incl. sharded variants — the serial pass is sharded too so concurrent shard jobs never self-collide). Initial set: `glm-coding-plan-monthly-3580`, `quota-division-blocks`, `provider-health-autopilot`, `combo-health-autopilot` — the class behind the ~28min CI wedges/re-runs (two live 1h+ wedges cancelled during this PR's own validation). Discovery + TIA gates track the new glob; systemic root cause (async logger writing after teardown) tracked in [#6360](https://github.com/diegosouzapw/OmniRoute/issues/6360). Guard: `tests/unit/test-serial-quarantine.test.ts` (4). ([#6347](https://github.com/diegosouzapw/OmniRoute/pull/6347))
+
+### 🙌 Contributors
+
+Thanks to everyone whose work landed in v3.8.46:
+
+| Contributor                                                    | PRs / Issues           |
+| -------------------------------------------------------------- | ---------------------- |
+| [@2220258345](https://github.com/2220258345)                   | direct commit / report |
+| [@abdofallah](https://github.com/abdofallah)                   | direct commit / report |
+| [@adentdk](https://github.com/adentdk)                         | direct commit / report |
+| [@aleksesipenko](https://github.com/aleksesipenko)             | direct commit / report |
+| [@anki1kr](https://github.com/anki1kr)                         | direct commit / report |
+| [@anungma](https://github.com/anungma)                         | direct commit / report |
+| [@arpicato](https://github.com/arpicato)                       | direct commit / report |
+| [@backryun](https://github.com/backryun)                       | #6248                  |
+| [@binsarjr](https://github.com/binsarjr)                       | direct commit / report |
+| [@brightfiscalband](https://github.com/brightfiscalband)       | direct commit / report |
+| [@chirag127](https://github.com/chirag127)                     | #6501, #6506           |
+| [@developerjillur](https://github.com/developerjillur)         | direct commit / report |
+| [@dilneiss](https://github.com/dilneiss)                       | #6499                  |
+| [@dtybnrj](https://github.com/dtybnrj)                         | direct commit / report |
+| [@Forcerecon](https://github.com/Forcerecon)                   | direct commit / report |
+| [@hao3039032](https://github.com/hao3039032)                   | direct commit / report |
+| [@Iammilansoni](https://github.com/Iammilansoni)               | #6150, #6245           |
+| [@jmengit](https://github.com/jmengit)                         | direct commit / report |
+| [@jordansilly77-stack](https://github.com/jordansilly77-stack) | direct commit / report |
+| [@JxnLexn](https://github.com/JxnLexn)                         | direct commit / report |
+| [@KooshaPari](https://github.com/KooshaPari)                   | #6166                  |
+| [@loopyd](https://github.com/loopyd)                           | direct commit / report |
+| [@luweiCN](https://github.com/luweiCN)                         | #6221                  |
+| [@makcimbx](https://github.com/makcimbx)                       | direct commit / report |
+| [@muflifadla38](https://github.com/muflifadla38)               | direct commit / report |
+| [@newnol](https://github.com/newnol)                           | direct commit / report |
+| [@ofekbetzalel](https://github.com/ofekbetzalel)               | direct commit / report |
+| [@ohahe52-dot](https://github.com/ohahe52-dot)                 | direct commit / report |
+| [@phidinhmanh](https://github.com/phidinhmanh)                 | direct commit / report |
+| [@powellnorma](https://github.com/powellnorma)                 | direct commit / report |
+| [@qpeyba](https://github.com/qpeyba)                           | direct commit / report |
+| [@RaviTharuma](https://github.com/RaviTharuma)                 | direct commit / report |
+| [@RCrushMe](https://github.com/RCrushMe)                       | direct commit / report |
+| [@rianonehub](https://github.com/rianonehub)                   | #6134, #6204           |
+| [@serverless83](https://github.com/serverless83)               | #6212                  |
+| [@swingtempo](https://github.com/swingtempo)                   | direct commit / report |
+| [@tenshiak](https://github.com/tenshiak)                       | direct commit / report |
+| [@ThongAccount](https://github.com/ThongAccount)               | direct commit / report |
+| [@UnrealAryan](https://github.com/UnrealAryan)                 | direct commit / report |
+| [@vinayakkulkarni](https://github.com/vinayakkulkarni)         | direct commit / report |
+| [@vittoroliveira-dev](https://github.com/vittoroliveira-dev)   | #6233                  |
+| [@warelik](https://github.com/warelik)                         | direct commit / report |
+| [@xxy9468615](https://github.com/xxy9468615)                   | direct commit / report |
+| [@xz-dev](https://github.com/xz-dev)                           | direct commit / report |
+| [@yanpaing007](https://github.com/yanpaing007)                 | direct commit / report |
+| [@diegosouzapw](https://github.com/diegosouzapw)               | maintainer             |
+
+---
+
 ## [3.8.45] — 2026-07-06
 
 ### ✨ New Features

docs/i18n/tr/CHANGELOG.md

@@ -6,6 +6,213 @@
 
 ## [3.8.31] — 2026-06-20
 
+## [3.8.46] — TBD
+
+### ✨ New Features
+
+- **feat(sse):** **hide paid-only models from `auto/*` routing** when `hidePaidModels` is on ([#6512](https://github.com/diegosouzapw/OmniRoute/issues/6512)) — follow-up to #6328/#6495. PR #6495 hid paid-only models from the `GET /v1/models` listing, but `auto/*` combos (`auto/best-coding`, `auto/glm`, …) could still pick a paid-only backend into their candidate pool → a 402/403 at request time. `createVirtualAutoCombo` now filters the candidate pool through the new pure `open-sse/services/autoCombo/paidModelFilter.ts` (`filterPaidOnlyCandidates`), applying the same free-model predicate #6495 uses in `catalog.ts` (`providerHasFreeModels(provider) && isFreeModel(provider, {id})`) whenever `settings.hidePaidModels === true`. Applied before the category/tier/family narrowing, so it covers every `auto/*` combo; an all-paid pool degrades to the existing graceful empty-pool path. **Opt-in — default OFF leaves the pool unchanged** (identity). Regression guard: `tests/unit/autoCombo/paid-model-filter-6512.test.ts` (4, incl. the default-off identity guard).
+- **feat(sse):** **provider-family auto combos** — `auto/glm`, `auto/minimax`, `auto/mimo`, `auto/zai`, `auto/gemma`, `auto/llama`, `auto/gemini` ([#6453](https://github.com/diegosouzapw/OmniRoute/issues/6453)) — new routable ids that materialize an on-demand virtual combo spanning whatever installed backends currently expose that model family, degrading gracefully as backends rotate. A new pure `open-sse/services/autoCombo/modelFamily.ts` (`detectModelFamily`) classifies by model-id prefix for six families; `zai` is instead resolved by provider id (z.ai's hosted API serves the same `glm-*` model ids as every other GLM backend, so `auto/zai` means "route to my z.ai backend specifically" vs `auto/glm`'s "any connected GLM backend"). Reuses the existing `createVirtualAutoCombo` on-demand materialization path (no DB writes) and the `/v1/models` catalog advertising loop. Regression guard: `tests/unit/autoCombo/provider-family-combos.test.ts` (11).
+- **feat(proxy):** native **proxy-pool round-robin / egress IP rotation** ([#6365](https://github.com/diegosouzapw/OmniRoute/issues/6365)) — a scope (global / provider / account) can now hold **multiple** proxies as a pool with a rotation strategy, so outbound requests cycle their egress IP instead of pinning one proxy per scope. Migration `117_proxy_pool_rotation.sql` lifts the `UNIQUE(scope, scope_id)` constraint (rebuild via the canonical rename/copy/drop; existing single assignments become 1-element pools) and adds a `proxy_scope_rotation` companion table holding the per-scope strategy + a persisted monotonic round-robin cursor. Strategies: `round-robin` (default, monotonic cursor — never `Math.random`), `random`, and `sticky-per-N-min`. Resolution (`resolveProxyForScopeFromRegistry` / `resolveProxyForConnectionFromRegistry`) now fetches the alive, position-ordered candidate set (unchanged `PROXY_ALIVE_PREDICATE`) and applies the strategy; an empty / all-dead pool still returns `null` — the #6246 fail-closed guard is **untouched** (never falls through to direct egress). Backend + DB only; dashboard pool-builder UI is a follow-up. Regression guard: `tests/unit/proxy-pool-rotation-6365.test.ts` (8, incl. fail-closed + backward-compat).
+- **feat(providers):** end-to-end **tool/function calling on the native Gemini `/v1beta` endpoint** ([#6222](https://github.com/diegosouzapw/OmniRoute/issues/6222)) — both directions of the Gemini↔OpenAI conversion now preserve tool data (previously silently dropped). Request side: `convertGeminiToInternal` (extracted to its own testable module) maps `tools[].functionDeclarations` → OpenAI `tools`, prior `functionCall` parts → assistant `tool_calls`, and `functionResponse` parts → `tool`-role messages. Response side: `convertOpenAIResponseToGemini` emits `parts[].functionCall {name,args}` from `message.tool_calls`, and the streaming `openAIChunkToGeminiChunk` accumulates fragmented `tool_calls` deltas by index into complete `functionCall` parts. The non-Gemini client paths (Claude, OpenAI-Responses) already preserved tool calls — this closes the gap specific to the native Gemini surface. Regression guard: `tests/unit/v1beta-gemini-tool-calling-6222.test.ts` (6, incl. a streaming SSE round-trip).
+- **feat(providers):** copilot-m365-web **enterprise / work tier** support ([#6334](https://github.com/diegosouzapw/OmniRoute/issues/6334)) — mirrors the EDU-tier pattern (#6210): `M365ConnectionParams` gains an `agent` field, a new opt-in `M365_ENTERPRISE_OVERRIDES` preset (`agent=work`, `scenario=officeweb`, `licenseType=Premium`) applies via `providerSpecificData.tier="enterprise"` (alias `"work"`), and `agent` is also overridable directly via `providerSpecificData.agent`. `buildWsUrl` was hardcoding `agent="web"` (the one enterprise-distinguishing param with no override path), so a Premium work account handshook then returned an empty stream. The individual and EDU paths are untouched. Kilo's dup flag vs #6210 (EDU tier) was a false positive — different tier. Regression guard: `tests/unit/copilot-m365-enterprise-6334.test.ts` (7). End-to-end confirmation on a real Premium work account is a live-VPS validation follow-up (Hard Rule #18). (thanks @Forcerecon)
+- **feat(api):** standardized, provider-agnostic **`effort` + `thinking` request params** ([#6241](https://github.com/diegosouzapw/OmniRoute/issues/6241)) — a thin standardization layer over the existing mature per-provider reasoning plumbing (no provider mapper touched). `providerChatCompletionSchema` gains a canonical `effort` (reusing the shared `none/low/medium/high/xhigh` vocabulary — the UI tiers `extra`/`max` collapse onto `xhigh`) and a boolean `thinking`. A pure `normalizeReasoningRequest` (wired once in `src/sse/handlers/chat.ts`, before any reasoning field is read) folds them onto the fields the translators already consume (`reasoning_effort` / `reasoning.effort` / `thinking`), so they fan out to Anthropic / Gemini / xAI / Responses — an explicit client `reasoning_effort` / object-shaped `thinking` always wins (backward-compatible). `/models` additively exposes `supportsThinking` + `effort_tiers` so the frontend can render the toggles (UI component is a follow-up). Regression guard: `tests/unit/effort-thinking-standardization-6241.test.ts` (12). (thanks @Iammilansoni, @shabeer)
+- **feat(combo):** new **`pipeline` (sequential) combo strategy** ([#6297](https://github.com/diegosouzapw/OmniRoute/issues/6297)) — the 18th routing strategy runs targets **in order**, threading each step's output into the next step's input, with an optional per-step `prompt` (system instruction); only the final step's response is returned. Distinct from `fusion` (parallel fan-out + judge). Implemented as a self-contained `open-sse/services/pipeline.ts` (sibling to `fusion.ts`), dispatched from `combo.ts`; the step list reuses `combo.models` order and reads an optional `prompt` off each target (backward-compatible — ignored by every other strategy). Intermediate steps run non-streaming with tools stripped (complete prose to thread forward); the final step keeps the client's `stream` flag + tools. A failing/empty/unparseable intermediate step fails the whole pipeline explicitly via a sanitized error (never silently swallowed). Kilo's dup flag vs #563 was a false positive (that's model→chain selection; this is a sequential chain). Regression guard: `tests/unit/combo-pipeline-strategy.test.ts` (5). (thanks @ofekbetzalel)
+- **feat(ci):** `check:test-masking` now flags **inline-reimplemented prod conditions** ([#6348](https://github.com/diegosouzapw/OmniRoute/issues/6348)) — a new **report-only** subcheck (v2, 6A.10 family) catches the wrong-shape contract test: a test that recomputes the condition under test inline instead of importing/exercising the real function (the #6216 class, where `=== 500` → `>= 500` stayed green because the test re-implemented the branch). For each added/modified test file it warns when the file textually duplicates a ≥3-token conditional from a production file touched in the same PR **and** does not import the symbol/module owning it, via a pure, fixture-tested `findReimplementedConditions()` with an allowlist mirroring `assertReductionAllowlist`. Report-only for now (does not fail the gate) — to be promoted to blocking after a triage cycle. Regression guard: `tests/unit/check-test-masking.test.ts` (45).
+- **feat(sse):** per-connection routing override (native vs CLIProxyAPI) ([#6339](https://github.com/diegosouzapw/OmniRoute/issues/6339)) — the previously-dead `isCliproxyapiDeepModeEnabled` helper is now wired into `resolveExecutorWithProxy`: a single connection can opt itself into the CLIProxyAPI passthrough executor via `providerSpecificData.cliproxyapiMode="claude-native"`, with precedence **connection override > provider `upstream_proxy_config` mode > default**. `resolveExecutorWithProxy` now receives the resolved connection's `providerSpecificData` (threaded from `chatCore.ts`), so one connection can deep-route while the provider's other connections stay native — no DB schema change (the toggle rides in `providerSpecificData`). Also resolves the same-provider mixing ask in #6340. Regression guard: `tests/unit/chatcore-executor-proxy.test.ts` (9). (thanks @RaviTharuma)
+- **feat(dashboard):** "Add session cookie" modal now shows a prominent **"Open ‹host› →"** link to the provider's own site ([#6268](https://github.com/diegosouzapw/OmniRoute/issues/6268)) — every `-web` cookie-session provider (chatgpt-web, claude-web, gemini-web, kimi-web, lmarena, qwen-web, m365-copilot-web, …) renders a one-click external link (opening the provider's login/home page in a new tab) so operators no longer tab away to retype the URL mid-setup. The host resolves from a pure, unit-tested `resolveWebProviderHost()` (prefers `WEB_COOKIE_PROVIDERS[id].website`, falls back to the registry `baseUrl` origin); non-web providers render exactly as before. Kilo's dup flag vs #6265 (modal-too-small-on-1080p) was a false positive — distinct concern. Regression guard: `tests/unit/resolve-web-provider-host.test.ts` (5). (thanks @chirag127)
+- **feat(providers):** add **DigitalOcean** AI (serverless inference) as an OpenAI-compatible API-key provider ([#6373](https://github.com/diegosouzapw/OmniRoute/pull/6373)) — base `https://inference.do-ai.run/v1`, wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/digitalocean/`, `src/shared/constants/providers/apikey/inference-hosts.ts`). Regression guard: `tests/unit/digitalocean-provider.test.ts`. (thanks @newnol)
+- **feat(providers):** add **Huancheng Public API** (`hcnsec`) as an OpenAI-compatible regional provider ([#6410](https://github.com/diegosouzapw/OmniRoute/pull/6410)) — Xinjiang Huancheng Cybersecurity's public LLM platform (base `https://api.hcnsec.cn/v1`, free credits via daily check-ins), wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/hcnsec/`, `src/shared/constants/providers/apikey/regional.ts`). Regression guard: `tests/unit/hcnsec-provider.test.ts`. (thanks @UnrealAryan)
+- **feat(dashboard):** the web-session credential guide now shows an **"Open {host}" link** ([#6316](https://github.com/diegosouzapw/OmniRoute/pull/6316)) to the provider's sign-in site (derived from the provider `website` via `getProviderWebsiteHost`), so you can jump straight to the page where the cookie/session must be captured. Regression guard: `tests/unit/web-session-provider-link-6316.test.ts`. (thanks @jordansilly77-stack)
+- **feat(cerebras):** add the **Gemma 4 31B** model (`gemma-4-31b`) to the Cerebras registry + pricing table ([#6331](https://github.com/diegosouzapw/OmniRoute/pull/6331)). Regression guard extends `tests/unit/t28-model-catalog-updates.test.ts`. (thanks @backryun)
+- **feat(providers):** add **Yuanbao (web)** as a cookie-session provider ([#6196](https://github.com/diegosouzapw/OmniRoute/issues/6196)) — `yuanbao-web` (Tencent Yuanbao, `yuanbao.tencent.com`) with cookie-only auth (`hy_user`/`hy_token` + public agent id), SSE→OpenAI translation incl. `reasoning_content`, exposing DeepSeek V3/R1 + Hunyuan / Hunyuan-T1. Regression guard: `tests/unit/providers-yuanbao-web.test.ts`. `together-web` was **deferred** (no verifiable web-session endpoint — needs a captured request) and `huggingchat-web` **dropped** (the existing `huggingchat` already is a web-cookie provider). (thanks @chirag127)
+- **feat(providers):** route the built-in **agentrouter** through the dynamic Claude-Code wire image ([#6056](https://github.com/diegosouzapw/OmniRoute/issues/6056)) — a small static allow-set (`CC_WIRE_IMAGE_BUILTINS` in `open-sse/services/ccWireImageBuiltins.ts`), consulted by `isClaudeCodeCompatible` / `isClaudeCodeCompatibleProvider` / `applyFingerprint`, makes agentrouter adopt the CC wire-image headers + fingerprint **while guarding the CC baseUrl/auth branches** so it keeps its own registry `baseUrl` and `x-api-key` auth. Regression guard: `tests/unit/agentrouter-cc-wire-image.test.ts` (asserts the wire image is applied AND agentrouter's baseUrl/auth are preserved). Live WAF-acceptance against agentrouter.org is a VPS validation follow-up (Hard Rule #18).
+- **feat(providers):** **bulk-add API keys for Cloudflare Workers AI** ([#6174](https://github.com/diegosouzapw/OmniRoute/issues/6174)) — `cloudflare-ai` is removed from the bulk-add exclusion list and the bulk parser gains a 3-field `name|accountId|apiKey` mode; the bulk route now builds a **per-entry** `providerSpecificData` so each key carries its own `accountId` (fixing the previous shared-object reuse), and both the create + key-validation paths receive it. Regression guard: `tests/unit/bulk-api-key-parser-cloudflare.test.ts`. (thanks @muflifadla38)
+- **feat(dashboard):** routing/settings UX clarity ([#6147](https://github.com/diegosouzapw/OmniRoute/issues/6147)) — (1) weighted combos show the **effective routing share %** next to each weight when weights don't sum to 100 (`WeightTotalBar.tsx`); (2) the status widget's user-facing **"Cloud Sync" label is renamed** to "Remote Settings Sync" (`CloudSyncStatus.tsx`; internal ids/state untouched); (3) built-in providers gain an **opt-in advanced base-URL override** (`isBaseUrlOverrideEligibleProvider`, hidden behind an "Advanced" toggle, reusing the existing `providerSpecificData.baseUrl` persistence — not globally widened). Regression guard: `tests/unit/routing-settings-ux-6147.test.ts`.
+- **feat(combo):** add an option to **disable session stickiness**, per-combo or globally — round-robin / random combos can rotate to a different connection on every request instead of pinning a whole conversation to one connection by its first-message hash. Resolution precedence per-combo `config.disableSessionStickiness` → global `settings.disableSessionStickiness` → default `false` (preserves the #3825 prompt-cache/504 fix); gates **both** stickiness call sites in `open-sse/services/combo.ts`. Exposed as a global toggle (Combo Defaults) and a per-combo Inherit/on/off control. ([#6168](https://github.com/diegosouzapw/OmniRoute/issues/6168)) Regression guard: `tests/unit/combo-disable-session-stickiness.test.ts`. (thanks @RCrushMe)
+- **feat(docker):** add the `OMNIROUTE_NO_SUDO` env flag for root-less / user-namespaced deployments — the MITM cert-trust command path (`resolveSudoSpawn` in `src/mitm/systemCommands.ts`) now strips the leading `sudo` when the flag is truthy, in addition to the existing root / sudo-missing cases, so the Proxy Agent runs without `sudo` (the operator trusts the CA manually, e.g. via `NODE_EXTRA_CA_CERTS`). Argv-array `spawn` preserved — no shell interpolation (Hard Rule #13). ([#6122](https://github.com/diegosouzapw/OmniRoute/issues/6122)) Regression guard: `tests/unit/mitm-systemCommands-no-sudo.test.ts`. (thanks @powellnorma)
+- **feat(providers):** add **Requesty** as an OpenAI-compatible gateway provider (BYOK, base `https://router.requesty.ai/v1`, ~200 free requests/day) — wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/requesty/`, `src/shared/constants/providers/apikey/gateways.ts`). ([#6120](https://github.com/diegosouzapw/OmniRoute/issues/6120)) Regression guard: `tests/unit/requesty-provider.test.ts`. (thanks @chirag127)
+- **feat(dashboard):** add **configured-only / available-only filters** to the Free Provider Rankings page ([#6150](https://github.com/diegosouzapw/OmniRoute/issues/6150)) — hide providers you haven't configured, or whose connections are all rate-limited / out of quota, via server-side query params (`?configuredOnly` / `?availableOnly` on `GET /api/free-provider-rankings`) backed by a testable lib helper reusing the in-process connection state (no Redis). Both filters default off, so the default view is unchanged; this supersedes the earlier client-side "Configured Only" toggle (#6245) with an available-only dimension and unit-tested logic. Regression guard: `tests/unit/freeProviderRankings-filters.test.ts`.
+
+### 🔧 Bug Fixes
+
+- **fix(dashboard):** adding a second API key connection for the same provider no longer silently overwrites the first — the Add-API-key modal now derives a unique default connection name (`main`, then `main-2`, `main-3`, …) so the backend name-based upsert can't collide (#6499 — thanks @dilneiss).
+- **fix(compression):** the session-dedup engine now also deduplicates a large multi-line block repeated **within a single message** (intra-message dedup), not just across turns; the compression-preview API surfaces a `fallbackReason`, and the fusion panel reports how many models were rate-limited vs failed on total-panel failure (#6501 — thanks @chirag127).
+- **fix(compression):** a stacked-pipeline step naming an unregistered engine now surfaces a `validationErrors` entry instead of silently no-op'ing, so misconfigured pipelines are visible in the preview API (#6506 — thanks @chirag127).
+- **feat(usage):** add a **Codex reset-credit redemption** flow to the Provider Limits UI ([#6361](https://github.com/diegosouzapw/OmniRoute/pull/6361)) — a `useCodexResetCreditRedemption` hook + `/api/usage/codex-reset-credit` route + `codexResetCredits` lib let you redeem banked Codex reset credits from the quota card. Regression guard: `tests/unit/codex-reset-credits.test.ts`. (thanks @JxnLexn)
+- **feat(glm):** add **team-plan quota settings** for `glm-cn` connections ([#6351](https://github.com/diegosouzapw/OmniRoute/pull/6351)) — a dedicated `GlmTeamQuotaFields` form section (team quota id / limits) threaded through the Add/Edit connection modals, persisted via `providerSpecificData`, with the GLM usage service reading the team quota. Regression guards: `tests/unit/glm-team-quota.test.ts`, `provider-specific-data-schema.test.ts`. (thanks @hao3039032)
+- **feat(providers):** add **TinyFish** web-fetch/search support ([#6349](https://github.com/diegosouzapw/OmniRoute/pull/6349)) — a `tinyfish-fetch` executor + `/v1/web/fetch` route + MCP web-fetch tool, registered as a specialty-media provider with request-validation and a search-provider catalog entry. Regression guards: `tests/unit/executor-tinyfish-fetch.test.ts`, `web-fetch-handler.test.ts`, `mcp-web-fetch-tool.test.ts`, `provider-validation-tinyfish.test.ts`. (thanks @dtybnrj)
+- **fix(cli):** `omniroute launch-codex` now spawns `codex.cmd` through a shell on Windows (the npm `.cmd` shim is unresolvable by bare `spawn` → ENOENT), mirroring the qodercli Windows fix (#6263) ([#6312](https://github.com/diegosouzapw/OmniRoute/pull/6312)). Regression guard: `tests/unit/launch-codex-windows-spawn-6312.test.ts`. (thanks @swingtempo)
+- **fix(codex):** isolate the **Spark** quota from the shared Codex quota and stabilize the quota UI ordering / hydration so per-scope limits render consistently ([#6336](https://github.com/diegosouzapw/OmniRoute/pull/6336)). Regression guards: `tests/unit/codex-quota-selection-hydration.test.ts`, `provider-limits-ui.test.ts` + 3 more. (thanks @xz-dev)
+- **feat(api):** add a `hidePaidModels` setting that filters paid-only models out of the `/v1/models` catalog. Regression guard: `tests/unit/models-catalog-hide-paid.test.ts`. (thanks @chirag127)
+- **fix(api-manager):** the fallback model picker now preserves combos instead of dropping them when a primary model is unavailable ([#6443](https://github.com/diegosouzapw/OmniRoute/pull/6443)). Regression guard: `tests/unit/api-manager-page-static.test.ts`. (thanks @jmengit)
+- **fix(providers):** recoverable Antigravity / Cloud-Code (Gemini Code Assist) `403` responses ([#6452](https://github.com/diegosouzapw/OmniRoute/pull/6452)) are now classified as a retryable project-config error instead of a terminal account ban, so a fixable project/API-disabled 403 no longer forces a ~1-year cooldown / full OAuth reconnect. Regression guard: `tests/unit/errorclassifier-antigravity-403.test.ts`. (thanks @developerjillur)
+- **fix(mitm):** `sanitizeHeaders` now redacts `Set-Cookie` response headers so upstream session cookies never leak into logs / diagnostics ([#6451](https://github.com/diegosouzapw/OmniRoute/pull/6451)). Regression guard: `tests/unit/mitm-sanitize-headers.test.ts`. (thanks @developerjillur)
+- **fix(api):** `/api/compression/preview` now accepts `mode: "caveman"` and correctly handles stacked / zero-compression previews ([#6425](https://github.com/diegosouzapw/OmniRoute/issues/6425)). Regression guard: `tests/unit/api/compression-preview-caveman-and-stacked-6425.test.ts`. (thanks @chirag127)
+- **feat(providers):** add **Zed** hosted LLM aggregator as a native-app provider ([#6118](https://github.com/diegosouzapw/OmniRoute/pull/6118)) — OAuth sign-in via the Zed hosted flow, registered through the shared provider registry + executor. Regression guards: `tests/unit/zed-oauth-provider.test.ts`, `zed-import-utils.test.ts`, `zed-docker-detect.test.ts`, `mitm-handler-zed.test.ts`. VPS-validated via live operator login (Hard Rule #18).
+- **fix(oauth):** the Kiro SSO-cache auto-import now **preserves the IDC region** — cross-region Amazon Q / Kiro profiles imported from the SSO cache are no longer collapsed to the default region ([#6113](https://github.com/diegosouzapw/OmniRoute/pull/6113)). Regression guard: `tests/unit/kiro-auto-import-idc-2059.test.ts`. VPS-validated via live operator login (Hard Rule #18).
+- **fix(dashboard):** passthrough model aliases no longer collide when two namespaced model ids share a last segment (port from 9router#1850, [#6431](https://github.com/diegosouzapw/OmniRoute/pull/6431)). `enx/gpt-5.5` and `enx/codebuddy/gpt-5.5` both auto-generated the alias `gpt-5.5`, so the second model could never be added (the UI just alerted "alias already exists"). Aliases are now disambiguated deterministically — bare last segment when free, then parent-qualified (`codebuddy-gpt-5.5`), then a numeric suffix — while re-adding the exact same model id is still blocked. Regression guard: `tests/unit/passthrough-alias-1850.test.ts`. (thanks @arpicato)
+- **fix(translator):** preserve a Gemini `functionResponse` co-located with other parts (another `functionCall`, or trailing `text`) in the same content when translating **Gemini → OpenAI** ([#6376](https://github.com/diegosouzapw/OmniRoute/pull/6376)). `convertGeminiContent()` early-returned the tool message on the first `functionResponse` part, dropping any co-located parts; such contents are now pre-split (one tool message per `functionResponse`, emitted first, plus one message for the remaining parts). Regression guard: `tests/unit/gemini-to-openai-function-response.test.ts`. (thanks @warelik)
+- **fix(headroom):** detect a python interpreter managed by **mise / pyenv / asdf / conda** (port from 9router#2353, [#6382](https://github.com/diegosouzapw/OmniRoute/pull/6382)). Headroom's python probe (`src/lib/headroom/detect.ts`) searched a hardcoded `PATH`, but version managers expose their interpreters via shim dirs that only join `PATH` through interactive-shell activation — which the non-interactive server never runs, so a managed python (≥3.10) was invisible and Headroom reported it missing. The search path now prepends the well-known shim/bin dirs (`~/.local/share/mise/shims`, `~/.pyenv/shims`, `~/.asdf/shims`, `$CONDA_PREFIX/bin`, `~/.local/bin`, respecting `MISE_DATA_DIR`/`PYENV_ROOT`/`ASDF_DATA_DIR` when set), and a new `HEADROOM_PYTHON` env override lets operators point straight at their interpreter (mirroring `HEADROOM_URL`). Still shell-free (`execFileSync`). Regression guard: `tests/unit/headroom-detect.test.ts` (5). (thanks @loopyd)
+- **fix(executors):** strip the OpenAI-Codex/Claude-CLI `client_metadata` passthrough field for **NVIDIA** requests (port from 9router#1887, [#6411](https://github.com/diegosouzapw/OmniRoute/pull/6411)). NVIDIA's OpenAI-compatible wrapper rejects it with `400 Unsupported parameter`, the same class already handled for `cerebras`/`mistral`; `nvidia` (executor `default`) was missing from the strip allowlist so Codex/Claude-Code passthrough requests 400'd. Regression guard: `tests/unit/executor-default-strip-client-metadata.test.ts` (+nvidia case). (thanks @phidinhmanh)
+- **fix(translator):** strip the Claude-style `thinking` field for **NVIDIA `z-ai/glm-5.2`** (port from 9router#2023, [#6413](https://github.com/diegosouzapw/OmniRoute/pull/6413)). NVIDIA's OpenAI-compatible wrapper 400s on `thinking` (a Claude-format client routed here leaves a `thinking:{type:"adaptive"}`); the existing strip rule only dropped `reasoning`. Same class already handled for `minimax-m2.7`. Regression guard: `tests/unit/nvidia-minimax-thinking-strip.test.ts` (+glm-5.2 case). (thanks @phidinhmanh)
+- **fix(translator):** suppress the streamed `</think>` close marker for the **Antigravity IDE** client (port from 9router#1061, [#6415](https://github.com/diegosouzapw/OmniRoute/pull/6415)). On thinking-only turns Antigravity rendered a bare `</think>` as the sole visible content, tripping its loop-detection and wasting requests. Antigravity's UA (`vscode/<v> (Antigravity/<v>)`) is added to the marker-suppress allowlist (alongside OpenCode); Claude Code / Cursor still get the marker, and `x-omniroute-thinking-marker: on` force-restores it. Regression guard: `tests/unit/think-close-marker-suppress-5245.test.ts`. (thanks @abdofallah)
+- **fix(executors):** strip nested `reasoning_content` from messages for **Mistral** (port from 9router#1649, [#6417](https://github.com/diegosouzapw/OmniRoute/pull/6417)). Mistral's API returns `422 extra_forbidden` when an assistant message carries `reasoning_content` (replayed thinking from a prior turn, e.g. via the Codex `/responses` path); the generic top-level 400 field-downgrade retry never covered the nested per-message field. `DefaultExecutor` now strips it for provider `mistral` only, so DeepSeek (which requires replayed `reasoning_content`) is unaffected. Regression guard: `tests/unit/mistral-strip-reasoning-content-1649.test.ts`. (thanks @xxy9468615)
+- **fix(executors):** strip the `client_metadata` passthrough field on the **OpenCode** path (port from 9router#1442, [#6418](https://github.com/diegosouzapw/OmniRoute/pull/6418)). OpenCode upstreams (e.g. `kimi-k2.6` via opencode-go) reject it with `400 "Extra inputs are not permitted, field: 'client_metadata'"`; the DefaultExecutor strip only covered cerebras/mistral and `OpencodeExecutor` extends `BaseExecutor` directly, so nothing removed it there. Regression guard: `tests/unit/opencode-strip-client-metadata-1442.test.ts`. (thanks @yanpaing007)
+- **fix(executors):** inject the `reasoning_content` echo for the native **Moonshot Kimi** provider (port from 9router#1480, [#6419](https://github.com/diegosouzapw/OmniRoute/pull/6419)). Kimi (executor `default`) is a thinking-mode upstream that 400s with "reasoning_content must be passed back" when a prior assistant turn lacks it; the placeholder injection was only wired into the OpenCode meta-provider, so direct multi-turn Kimi conversations failed. Scoped to `kimi` (gateway-served models matching the thinking-model name pattern are unaffected). Regression guard: `tests/unit/kimi-native-reasoning-injected-1480.test.ts`. (thanks @2220258345)
+- **fix(executors):** recover from a strict gateway's `context_management: Extra inputs are not permitted` 400 (port from 9router#1468, [#6420](https://github.com/diegosouzapw/OmniRoute/pull/6420)). **Claude Code** always sends a top-level `context_management` field; strict anthropic-compatible gateways reject it. The dedicated context-editing 400-fallback only fired when OmniRoute's own `contextEditing` feature was enabled (default off), so a client-sent field passed through untouched and 400'd. `context_management` is now in the generic reactive field-strip list, so it's stripped-and-retried once regardless of the feature flag (with correct request re-signing for claude-compatible relays). Regression guard: `tests/unit/provider-field-strips.test.ts`. (thanks @ohahe52-dot)
+- **fix(network):** enable **RFC 8305 Happy Eyeballs** (`autoSelectFamily`) on the direct-egress undici dispatcher (port from 9router#1237, [#6423](https://github.com/diegosouzapw/OmniRoute/pull/6423)). When DNS returns both IPv6 (AAAA) and IPv4 (A) and the IPv6 route is broken (e.g. a NAT64 `64:ff9b::` prefix without routing), undici tried IPv6 first and hung until `ETIMEDOUT` (then a 502 + account lockout), even though `curl` reached the same host. The direct dispatcher now races both families and uses whichever connects first. Proxy paths pin family via `proxyTls` and are unaffected. Regression guard: `tests/unit/direct-dispatcher-pipelining-4580.test.ts`. (thanks @adentdk)
+- **fix(combo):** round-robin now advances the rotation pointer past the model that **actually served**, not the eagerly-scheduled one (port from 9router#948, [#6428](https://github.com/diegosouzapw/OmniRoute/pull/6428)). With `stickyLimit: 1` (true round-robin), when the scheduled model failed and a _different_ model served via fallback, the counter had already advanced +1 from the scheduled index — so the next request reused the fallback-served model, degrading round-robin into hot-spotting on whichever model was healthy. The pointer now advances to the served index + 1 (mirroring the sticky-limit>1 path). Session-stickiness (#3825) and distribution are preserved. Regression guard: `tests/unit/combo-rr-fallback-advance-948.test.ts`. (thanks @binsarjr)
+- **fix(sse):** a non-string `model` field is now rejected with a `400` before the resolver, instead of crashing downstream `.toLowerCase()`/`.split()` calls into an empty-body `500` that escapes the error sanitizer ([#6407](https://github.com/diegosouzapw/OmniRoute/issues/6407)). Regression guard: `tests/unit/chat-non-string-model-6407.test.ts`. (thanks @chirag127)
+- **fix(api):** unknown `/api/*` routes now return a JSON `404` (instead of the dashboard HTML shell) and scalar chat params (`model`/`temperature`/etc.) are validated **before** the provider lookup so malformed requests fail fast with a clear `400` ([#6424](https://github.com/diegosouzapw/OmniRoute/issues/6424), [#6412](https://github.com/diegosouzapw/OmniRoute/issues/6412)). Regression guards: `tests/unit/api/api-catchall-json-404.test.ts`, `tests/unit/chat-early-schema-validation-6412.test.ts`. (thanks @chirag127)
+- **fix(api):** `/v1/chat/completions` now rejects a non-JSON `Content-Type` with a `400` before parsing the body ([#6414](https://github.com/diegosouzapw/OmniRoute/issues/6414)). Regression guard: `tests/unit/v1-chat-completions-content-type-6414.test.ts`. (thanks @chirag127)
+- **fix(api):** the `X-OmniRoute-Compression` response header is now echoed on `/v1/chat/completions` and `/v1/completions` ([#6422](https://github.com/diegosouzapw/OmniRoute/issues/6422)). Regression guard: `tests/unit/compression-header-echo-6422.test.ts`. (thanks @chirag127)
+- **fix(api):** concurrent `GET /v1/models` requests are coalesced into a single catalog build ([#6408](https://github.com/diegosouzapw/OmniRoute/issues/6408)). Regression guard: `tests/unit/v1-models-concurrent-6408.test.ts`. (thanks @chirag127)
+- **fix(api):** `/v1/completions` now echoes the requested `body.model` in its JSON + streamed responses ([#6429](https://github.com/diegosouzapw/OmniRoute/pull/6429)). Regression guard: `tests/unit/completions-body-model-echo.test.ts`. (thanks @chirag127)
+- **fix(api):** env-var master keys now see the full `/v1/models` catalog ([#6406](https://github.com/diegosouzapw/OmniRoute/issues/6406)). Regression guard: `tests/unit/models-catalog-envkey-6406.test.ts`. (thanks @chirag127)
+- **fix(api):** non-streaming `/v1/completions` responses now echo `body.model` aligned with the `X-OmniRoute-Model` header ([#6426](https://github.com/diegosouzapw/OmniRoute/issues/6426)). Regression guard: `tests/unit/v1-completions-model-header-match-6426.test.ts`. (thanks @chirag127)
+- **fix(api):** unknown `/v1/*` routes now return a JSON `404 not_found` instead of the Next.js dashboard HTML shell ([#6405](https://github.com/diegosouzapw/OmniRoute/issues/6405)). Regression guard: `tests/unit/api/v1-catchall-json-404.test.ts`. (thanks @chirag127)
+- **fix(api):** the per-connection provider models route now degrades to the shipped catalog when a provider's `/models` endpoint answers with a redirect ([#6267](https://github.com/diegosouzapw/OmniRoute/issues/6267)) — a `qwen-web` import failed with a raw `Redirect blocked … (307)` 503. `safeOutboundFetch` throws `REDIRECT_BLOCKED` on the 307, `getSafeOutboundFetchErrorStatus` maps it to 503, and `buildDiscoveryErrorFallbackResponse` treated every 503 as a hard error — so the non-empty `getModelsByProviderId("qwen-web")` catalog was never surfaced. A models-endpoint redirect is not a fixable-config error (unlike `URL_GUARD_BLOCKED`/`INVALID_URL`, which stay hard errors), so it now falls back to the local/cached catalog before the 503 short-circuit. General fix — covers any config-driven provider that 307s. Regression guard: `tests/unit/provider-models-qwen-web-redirect-6267.test.ts`. (thanks @chirag127)
+- **fix(api):** the per-connection provider models route (MCP `list_models_catalog` + the dashboard import view) now merges USER-ADDED custom models into its response ([#6247](https://github.com/diegosouzapw/OmniRoute/issues/6247)) — custom models live in the `key_value` namespace `customModels`, which the live REST `/api/v1/models` already merges, but `src/app/api/providers/[id]/models/route.ts` never read `getCustomModels`, so custom models were dropped on both the discovery-success and local_catalog paths. They are now folded into the returned model list (deduped by id, stamped `owned_by: provider`), fixing MCP + the dashboard import view in one place. Regression guard: `tests/unit/provider-models-custom-merge-6247.test.ts`. (thanks @RCrushMe)
+
+- **fix(providers):** GitLab Duo tool-calling follow-up turns no longer fail upstream with `422 {"detail":"Validation error"}` (tokens 0/0, rejected pre-inference). The #6234 tool-result-feedback fix serialized the **entire** multi-turn conversation into GitLab's single-file `code_suggestions` (`small_file`) generation endpoint — folded history that turn-N sent as an oversized `current_file.content_above_cursor` **and** duplicated verbatim into `user_instruction`, tripping the AI-Gateway's `small_file` validation guard. The executor now **bounds** that prompt: it keeps system + latest user message + the most-recent tool round (dropping older turns), caps oversized tool results, and stops duplicating the full prompt into `user_instruction` (which now carries only the short latest user message) — while still feeding the most-recent tool result back so the agent continues (`open-sse/executors/gitlab.ts`, [#6220](https://github.com/diegosouzapw/OmniRoute/issues/6220)). The unit test covers the bounding logic; the upstream 422→200 clearing is VPS-only (Hard Rule #18). Regression guard: `tests/unit/gitlab-tool-exchange-bounded-6220.test.ts`.
+
+- **fix(i18n):** the provider-detail (`/dashboard/providers/[id]`) connection-status filter labels no longer render as `__MISSING__:All` / `__MISSING__:Active` / `__MISSING__:Error` / `__MISSING__:Banned` / `__MISSING__:CreditsExhausted` in non-English locales (notably pt-BR) ([#6290](https://github.com/diegosouzapw/OmniRoute/issues/6290)). Root cause was **not** the namespace mismatch the issue guessed — the `providers.filter*` keys resolve correctly in `en.json`; the debt lived in the locale mirrors (`src/i18n/messages/*.json`), where these five keys carried the `__MISSING__:` sync sentinel in ~15 locales and were absent entirely in ~26 others, so next-intl found the key and echoed the sentinel verbatim. All 40 non-English/-Chinese mirrors now ship real translations for the five `providers.filter*` labels. Regression guard: `tests/unit/i18n-provider-filter-keys-6290.test.ts`. (thanks @diegosouzapw)
+
+- **fix(providers):** the `copilot-m365-web` streaming executor now emits `debug`-level WebSocket diagnostics ([#6210](https://github.com/diegosouzapw/OmniRoute/issues/6210)) — the outbound WS URL (with the `access_token` **redacted** via `redactWsUrl()`), handshake success/failure, and each received SignalR frame's `type`/`target`. Previously the streaming path logged nothing, so an empty `content:null` response (the M365 Education / Starter tier symptom fixed in #6234) was undiagnosable even at `APP_LOG_LEVEL=debug`. The change is debug-level and side-effect-free — it does not alter streaming behavior or the frame parser, and the token never reaches the logs. Regression guard: `tests/unit/copilot-m365-web-logging-6210.test.ts` (thanks @qpeyba)
+
+- **fix(resilience):** a round-robin combo no longer returns `503 all upstream accounts are unavailable` when a compatibility-rejected target is actually healthy ([#6238](https://github.com/diegosouzapw/OmniRoute/issues/6238)). `filterTargetsByRequestCompatibility` drops request-incompatible targets (tool/vision/structured-output unsupported, or below the required context window) **before** any availability check runs, and its `compatible.length === 0` safety net only fired when _all_ targets were filtered — not when the kept targets later all turned out runtime-unavailable (circuit-open / cooldown / no credentials). So a combo could 503 while a compat-rejected-but-healthy provider sat unused. `handleRoundRobinCombo` now keeps the compat-rejected set and, when every compat-kept target was skipped without a single real attempt, probes those rejected targets as a **last-resort fallback tier** (via the new pure `open-sse/services/combo/comboCompatFallback.ts`) before crystallizing the 503. Regression guard: `tests/unit/combo-roundrobin-compat-fallback-6238.test.ts`. (thanks @ThongAccount)
+
+- **fix(startup):** best-effort self-heal for a corrupted Turbopack dev cache on Windows ([#6289](https://github.com/diegosouzapw/OmniRoute/issues/6289)). On Windows, `pnpm dev` can fail at startup when Turbopack `mmap`s a persistent-cache SST file and the OS refuses the mapping (`os error 1455` — "paging file too small"), which Turbopack surfaces as a misleading `Module not found: Can't resolve '@/shared/utils/machine'`. This is a **known upstream Turbopack cache-corruption bug — not our code**. The dev launcher (`scripts/dev/run-next.mjs`) now wraps `nextApp.prepare()` and, when it rejects with that signature (`isTurbopackCacheCorruption` in the new `scripts/dev/turbopackCacheHeal.mjs`), purges `.build/next/**/cache/turbopack` and retries **once** with a clear log. **Caveat — best-effort only:** the corruption often surfaces as a runtime overlay rather than a `prepare()` rejection, so this cannot always intercept it; the reliable remedy remains manually deleting the Turbopack cache dir. Regression guard: `tests/unit/turbopack-cache-heal-6289.test.ts`. (thanks @chirag127)
+
+- **fix(providers):** qodercli PAT auth no longer fails with `spawn qodercli ENOENT` on Windows ([#6263](https://github.com/diegosouzapw/OmniRoute/issues/6263)) — `spawnQoderCli` spawned the bare `qodercli` name with `shell:false` and an unenriched env, so the npm `.cmd` wrapper under `%APPDATA%\npm` (a user-PATH directory) was never resolved. It now resolves the absolute `.cmd`/`.exe` path through the existing `getCliRuntimeStatus("qoder")` resolver in `src/shared/services/cliRuntime.ts` (memoized), spawns with `shell` when the target is a `.cmd`/`.bat`, and uses the cliRuntime-enriched env (PATH + PATHEXT + APPDATA); the ENOENT error now lists the searched paths plus the `CLI_QODER_BIN` override. End-to-end spawn on a real Windows host is host-only (Hard Rule #18); the path-resolution logic is unit-tested. Regression guard: `tests/unit/qodercli-windows-resolve-6263.test.ts`. (thanks @chirag127)
+
+- **fix(sse):** the reasoning-token buffer no longer inflates **probe-sized `max_tokens`** ([#6274](https://github.com/diegosouzapw/OmniRoute/issues/6274)) — Claude Code's `/model` capability check sends `max_tokens: 1`, but for a thinking-capable model with a large output cap (e.g. `glm-5.2`) the #3587 headroom heuristic (`max(current + 1000, ceil(current * 1.5))`) rewrote it to `1001` and forwarded that upstream, wasting tokens on a request that was never a genuine reasoning budget. `resolveReasoningBufferedMaxTokens()` (`open-sse/services/reasoningTokenBuffer.ts`) now short-circuits and returns the caller's value verbatim when it is below the new `REASONING_BUFFER_MIN_TRIGGER` (256) threshold — a tiny explicit limit is a probe, not a reasoning request. Real budgets still receive the #3587 headroom unchanged, and the guard runs after the existing capability checks so unknown / non-reasoning models keep returning `null`. Regression guard: `tests/unit/reasoning-token-buffer-6274.test.ts`. (thanks @brightfiscalband)
+
+- **fix(cli):** `omniroute reset-password` now works as a real subcommand, and password resets over piped (non-TTY) stdin actually apply ([#6261](https://github.com/diegosouzapw/OmniRoute/issues/6261), [#6258](https://github.com/diegosouzapw/OmniRoute/issues/6258)). Two coupled defects: (1) **#6261** — `bin/omniroute.mjs` routed everything through Commander with only two pre-Commander bypasses (`--mcp`, `reset-encrypted-columns`), so `omniroute reset-password` was rejected as an unknown command; only the separate `omniroute-reset-password` bin worked, while the docs falsely advertised the subcommand (incl. a bogus "legacy alias still works"). A pre-Commander bypass mirroring `reset-encrypted-columns` now dynamically imports `bin/reset-password.mjs` (which self-executes) before Commander parses; the three doc lines were corrected. (2) **#6258** — `bin/reset-password.mjs` issued two sequential `rl.question` prompts; under piped stdin the second read never settled at EOF, so `main()` never reached `resetManagementPassword` and the reset was a silent no-op (both prompts printed, no success, password unchanged). The CLI now detects non-TTY stdin and reads it once (first line = password, second line = confirm if present, else reused), adds a `--password-stdin` flag (entire stdin is the password, no confirmation), and exits `0` explicitly so the success line always flushes; interactive TTY behavior is unchanged. Regression guard: `tests/unit/reset-password-cli-6261-6258.test.ts` (3). (thanks @chirag127)
+
+- **fix(db):** the mass-migration **safety abort** now tells the operator how to bypass it and stops flooding the log ([#6260](https://github.com/diegosouzapw/OmniRoute/issues/6260)) — after restoring a backup that wiped the migration tracking table, `runMigrations()` threw the abort on every downstream `ensureDbInitialized()`, re-logging the full banner 11+ times, and the message never mentioned the existing `OMNIROUTE_MAX_PENDING_MIGRATIONS` escape hatch. The abort text now appends a bypass hint (set `OMNIROUTE_MAX_PENDING_MIGRATIONS=0` in `server.env` / `DATA_DIR/.env`), and a new `MigrationSafetyAbortError` is memoized so repeated calls in the same process throw the same instance and emit a single concise line instead of the full cascade. Regression guard: `tests/unit/migration-safety-abort-6260.test.ts`. (thanks @chirag127)
+
+- **fix(auth):** importing a **distinct** Codex/ChatGPT OAuth `auth.json` is no longer falsely rejected as "already exists" when it belongs to a different user in the same workspace ([#6301](https://github.com/diegosouzapw/OmniRoute/issues/6301)). `findExistingCodexConnection` (in `src/lib/oauth/utils/codexAuthImport.ts`) deduped **only** on `providerSpecificData.workspaceId === accountId`, where `accountId` is the shared `chatgpt_account_id`/`tokens.account_id` — so two members of the same ChatGPT Team collapsed onto a single connection (409 `duplicate_account`). The id_token's `https://api.openai.com/auth` claim carries a per-user `chatgpt_user_id` alongside the workspace id (the device-flow path already persisted it as `chatgptUserId`, but the import path did not). Now `parseAndValidateCodexAuth` extracts `userId` (`chatgpt_user_id` → `user_id` → JWT `sub`) into `ParsedCodexAuth`, the create/update paths persist `chatgptUserId` in `providerSpecificData` (mirroring `codex.ts`), and dedup keys on `workspaceId` **AND** `chatgptUserId` — with a backward-compat fallback to legacy accountId-only matching when no stored connection for that workspace records a `chatgptUserId`, so genuinely-same accounts still dedup. Regression guard: `tests/unit/codex-auth-import-userid-dedup-6301.test.ts` (4). (thanks @anungma)
+
+- **fix(providers):** importing models for the **venice-web** provider no longer fails with a red "Provider venice-web does not support models listing" ([#6269](https://github.com/diegosouzapw/OmniRoute/issues/6269)). `venice-web` is a web-cookie provider with an executor but no upstream `/v1/models` endpoint and no registry `models`, so the models route fell through to the tail `400`. Mirroring the `jules`/`linkup-search`/`ollama-search` fix (#5569), it now ships a static local catalog entry in `src/lib/providers/staticModels.ts` — seeding the current Venice lineup (`venice-uncensored`, `llama-3.3-70b`, `qwen3-235b`, `qwen3-4b`, `deepseek-r1-671b`; Venice rotates its catalog, see docs.venice.ai/models/overview) — so the route returns `200` with `source:"local_catalog"`, `intentional:true`. Regression guard: `tests/unit/static-models-venice-web-6269.test.ts`. (thanks @chirag127)
+
+- **fix(api):** the specialty model catalogs (`/v1/embeddings`, `/v1/images`, `/v1/music`, `/v1/videos` model lists) are now derived from the **unified catalog filtered by a predicate** (`getSpecialtyModelsResponse`) instead of ad-hoc per-route logic, so they consistently respect active-credential visibility and stay in sync with the main catalog ([#6303](https://github.com/diegosouzapw/OmniRoute/pull/6303)). Regression guard: `tests/unit/specialty-model-catalog-routes.test.ts`. (thanks @makcimbx)
+- **fix(api):** the agent-bridge server route now resolves the MITM manager via a **dynamic `import("@/mitm/manager.runtime")`** so Turbopack does not statically pull the stub (or over-bundle the manager), and the agent-skills generator anchors its output base path with `path.join(process.cwd(), …)` so Turbopack's static analyzer stops tracing the whole project root ([#6329](https://github.com/diegosouzapw/OmniRoute/issues/6329), [#6366](https://github.com/diegosouzapw/OmniRoute/pull/6366)). Regression guard: `tests/unit/agent-bridge-server-route-dynamic-import.test.ts`. (thanks @Iammilansoni)
+- **fix(api):** internal probes (combo-test, cloud-sync verify) now pick a **management-scoped / allow-all API key** instead of naively grabbing `getApiKeys()[0]` — a restricted `self:usage` first row made the probe fail with "Model X is not allowed for this API key" even when the combo path was healthy (`pickApiKeyForInternalUse` in `src/lib/db/apiKeys.ts`). The API-manager model editor also falls back to `/api/models?all=true` when `/v1/models` is catalog-protected ([#6372](https://github.com/diegosouzapw/OmniRoute/pull/6372)). Regression guard: `tests/unit/pick-internal-api-key-6372.test.ts`. (thanks @jmengit)
+- **fix(live-ws):** the Live Dashboard WebSocket server now **rejects on bind failure** (e.g. `EADDRINUSE` when the API bridge already holds the port) instead of letting the error surface as an unhandled `error` event that crash-loops the process — the `error` listener is attached to `wss` (not `server`) and releases the EventBus subscription on a failed start ([#6324](https://github.com/diegosouzapw/OmniRoute/issues/6324)). Regression guard: `tests/unit/live-ws-eaddrinuse-6324.test.ts`. (thanks @vinayakkulkarni)
+- **fix(dashboard):** the Home provider-topology widget now trusts the live provider-metrics snapshot — it uses `topology.errorProvider` and live `activeRequests` directly instead of re-deriving state from a stale `lastErrorAt` or applying a frontend timeout filter, so the topology reflects real-time provider health ([#6322](https://github.com/diegosouzapw/OmniRoute/pull/6322)). Regression guard: `tests/unit/home-provider-topology-live-state.test.ts`. (thanks @xz-dev)
+- **fix(sse):** strip zero-width markers from streamed **tool-call arguments** — a follow-up to [#5857](https://github.com/diegosouzapw/OmniRoute/pull/5857). That PR removed injected zero-width joiners (U+200D) from streamed assistant text/reasoning but deliberately left tool-call argument JSON byte-exact. The request-side obfuscation (`open-sse/services/claudeCodeObfuscation.ts`) injects ZWJ into agent words — including the temp path inside the Bash tool description — and Claude models copy that verbatim into generated commands, which are delivered as tool-call arguments rather than assistant text. As a result the ZWJ survived and corrupted code blocks (e.g. a temp path rendered with an invisible joiner). Now `open-sse/handlers/responseSanitizer.ts` strips zero-width code points from tool-call argument strings at every emit site (OpenAI non-stream/stream chat `tool_calls` + legacy `function_call`, native Responses `function_call` items, the OpenAI→Responses conversion, and the native Responses streaming `response.function_call_arguments.delta/.done` events). Only zero-width code points are removed; JSON structure and all other bytes stay identical (no parse/restringify), so normal arguments remain byte-exact. Regression guard: 6 new cases in `tests/unit/response-sanitizer.test.ts` (suite 50/50).
+
+- **fix(nodejs):** the default app log path now resolves under `DATA_DIR` (`~/.omniroute/logs/application/app.log`) instead of `process.cwd()` ([#6197](https://github.com/diegosouzapw/OmniRoute/issues/6197)) — the globally-installed CLI runs from an arbitrary working directory, so anchoring the default to cwd made file logging silently write to (or no-op under) an unrelated directory, contradicting the documented `.env.example` default. `getAppLogFilePath()` now computes the default lazily via the pure `resolveDataDir()` resolver (honours a per-process `DATA_DIR`, no directory-creation side effect); an explicit `APP_LOG_FILE_PATH` still wins. Regression guard: `tests/unit/logenv-datadir-path-6197.test.ts` (3).
+
+- **fix(docker):** AgentBridge/`startMitm` no longer aborts in containers/headless when the Antigravity-default DNS step can't write `/etc/hosts` ([#6127](https://github.com/diegosouzapw/OmniRoute/issues/6127)), and the privileged command's stderr now reaches `app.log` instead of only a bare exit code hitting the toast ([#6198](https://github.com/diegosouzapw/OmniRoute/issues/6198)). The default DNS step (`addDNSEntry`) was called unguarded while cert install and the two sibling DNS steps were each best-effort — in the runtime Docker image (`USER node`, no `sudo`, read-only `/etc/hosts`) it threw `Command failed with code 1` out of `startMitmInternal` and killed the whole start, discarding the stderr. The three DNS steps are extracted into a best-effort `provisionDnsEntries()` where each failure is logged with the full `err` (stderr included, folded in by `systemCommands.ts`) and never aborts the start. Regression guard: `tests/unit/mitm-dns-graceful-degrade-6127.test.ts` (4).
+
+- **fix(providers):** copilot-m365-web now supports the M365 Education "Starter / OfficeWebIncludedCopilot" tier and no longer returns an empty `content:null` stream ([#6210](https://github.com/diegosouzapw/OmniRoute/issues/6210)). Two gaps: (1) `buildWsUrl()` hardcoded the individual-consumer scenario (`OfficeWebPaidConsumerCopilot`, `isEdu=false`) — the EDU tier is now opt-in via `providerSpecificData.tier="edu"`, emitting `scenario=OfficeWebIncludedCopilot`/`isEdu=true` (the individual path is unchanged); (2) the EDU/GPT-5.5 path streams deltas via `arguments[0].writeAtCursor` (incremental) instead of only `messages[].text` (accumulated snapshots), which the parser dropped — a new `accumulateBotContent()` folds both formats, with `type:2 item.result.message` as a last-resort fallback. Regression guard: `tests/unit/copilot-m365-edu-writeatcursor-6210.test.ts` (10). (thanks @qpeyba)
+
+- **fix(providers):** GitLab Duo executor now feeds tool results back into the prompt instead of looping ([#6220](https://github.com/diegosouzapw/OmniRoute/issues/6220)) — `buildPrompt()` branched only on `system`/`user` and took `userParts.at(-1)`, silently dropping the `assistant{tool_calls}` + `tool{result}` turns the client appended, so the reconstructed prompt was byte-identical to turn 1 and the model re-emitted the same `<tool>` call forever. When a tool exchange is present the full conversation is now serialized, folding each tool result back keyed by its `tool_call_id`; simple conversations keep the legacy shape. Complements the tool_call emission from [#6051](https://github.com/diegosouzapw/OmniRoute/issues/6051) (the `kilo-duplicate` label was a false positive — different, sequential defect). Regression guard: `tests/unit/gitlab-tool-result-feedback-6220.test.ts` (4).
+
+- **fix(providers):** opencode-go/opencode-zen can now synthesize the OpenCode CLI identity headers Cloudflare requires on VPS egress ([#5997](https://github.com/diegosouzapw/OmniRoute/issues/5997)) — on a datacenter VPS, `opencode.ai/zen/go/v1/chat/completions` 403s (HTML challenge) requests lacking CLI identity, while the reporter's control curl proved that `User-Agent: opencode-cli/1.0.0` + `x-opencode-client: cli` + `x-opencode-project: default` + fresh request/session UUIDs succeed. Opt-in via `OPENCODE_SYNTHESIZE_CLI_HEADERS=true` (values overridable via `OPENCODE_GO_USER_AGENT`/`OPENCODE_USER_AGENT`/`OPENCODE_CLIENT`/`OPENCODE_PROJECT`); it fills only headers the client did not already send. Kept **off by default** — the forward-only path is deliberate (fabricating a wrong value risks upstream rejection; a prior dedup regressed with `opencode/local`), so this replaces the fragile local header-injection shim without changing default behavior. Regression guard: `tests/unit/opencode-cli-headers-synthesis-5997.test.ts` (6). (thanks @aleksesipenko)
+- fix(resilience): sticky session affinity now evicts and fails over to another account when the pinned account is exhausted/unavailable (#6219)
+
+- fix(sse): Responses API passthrough now drops internal commentary-phase output before forwarding to clients (gated by RESPONSES_PASSTHROUGH_DROP_COMMENTARY, default on) (#6199)
+
+- **fix(sse):** tool-call function schemas with a root `type: null` are now coerced to `type: "object"` before dispatch ([#6359](https://github.com/diegosouzapw/OmniRoute/issues/6359)) — clients like the Codex app emit `parameters: { type: null, ... }` for some tools, which OpenAI-compatible upstreams reject with `400 Invalid schema for function '...': schema must be a JSON Schema of 'type: "object"', got 'type: null'`, failing the whole request. `toolSchemaSanitizer` already stripped the null; it now re-adds the mandatory root `"object"` type (and empty `properties`/open `additionalProperties` when absent). Combinator roots (`anyOf`/`oneOf`/`allOf`) and explicit root types are left untouched. Regression guard: 5 new cases in `tests/unit/tool-schema-sanitizer.test.mjs`.
+
+- **fix(docker):** AgentBridge no longer fails to start on npm/Electron/VPS installs with "MITM manager stub reached at runtime" ([#6344](https://github.com/diegosouzapw/OmniRoute/issues/6344)) — v3.8.45 flipped the production bundler default to Turbopack, but `next.config.mjs` aliased `@/mitm/manager` to its Docker-only degraded stub **unconditionally**. That was harmless while Docker (which sets the alias intentionally for #3390 graceful degradation) was the sole Turbopack consumer, but once every artifact built with Turbopack the stub shipped to all non-Docker users and `startMitm` threw on the first Agent-Bridge start. The alias is now opt-in via `OMNIROUTE_MITM_STUB=1` (set only by the Dockerfile) through the shared `scripts/build/mitm-stub-flag.mjs` helper; default builds bundle the real manager. Regression guard: `tests/unit/mitm-stub-alias-6344.test.mjs` (4).
+
+- **fix(proxy):** stop the v3.8.44 proxy regression that leaked the real IP and disabled healthy proxies ([#6246](https://github.com/diegosouzapw/OmniRoute/issues/6246)). Two coupled defects from the new health scheduler: (1) **IP leak** — when a proxy assigned to a connection was marked `inactive`, resolution fell through to a **direct** egress instead of blocking, exposing the operator's real IP; (2) **over-deactivation** — the sweep flipped a proxy to `inactive` on the **first** failed probe and counted our own 5s timeout / a probe-target `5xx` as the proxy's fault, so healthy paid proxies vanished from egress selection ("my proxies are not being used anymore"). Fix: the sweep decision is extracted into a pure, network-free `decideProxyHealthAction` (`src/lib/proxyHealth/decision.ts`) — by default the health check now **only counts/logs and never downgrades status** (a proxy is downgraded/removed only with `PROXY_AUTO_REMOVE=true`, after `PROXY_AUTO_REMOVE_AFTER` **consecutive** conclusive failures); probes are classified tri-state so an inconclusive result (our timeout, or a `5xx` from the probe target) never penalizes the proxy, and the probe timeout is raised 5s→15s. Separately, `safeResolveProxy` now **fails closed** via the existing policy: a connection whose assigned proxy is dead is blocked instead of leaking direct (`hasBlockingProxyAssignment`), honoring the explicit `proxy off` toggles and the `PROXY_FAIL_OPEN=true` opt-out. Existing proxies stuck `inactive` by the old behavior need a one-time manual re-activate (the operator owns proxy status). Regression guards: `tests/unit/proxy-health-decide-action-6246.test.ts`, `tests/unit/proxy-assigned-unavailable-6246.test.ts`.
+
+- **fix(proxy):** make "Test All" read-only and add bulk enable/disable ([#6246](https://github.com/diegosouzapw/OmniRoute/issues/6246)). Complements the core fail-closed / scheduler fix (#6296) with the two remaining reporter asks. (1) The **"Test All" button** (`POST /api/settings/proxies/auto-test`) used to flip a proxy to `inactive` on a failed reachability probe; since the egress selector excludes `inactive` proxies, a flaky probe (an unreachable `httpbin.org`, a proxy that blocks `HEAD`, or a slow paid proxy) silently disabled every proxy that failed — "Test All" is now **read-only by default** (only the operator sets a proxy active/inactive; opt back into the legacy test-and-set with `PROXY_HEALTH_AUTO_DEACTIVATE=true`). (2) Adds a **bulk enable/disable** proxies endpoint + toolbar action (`POST /api/settings/proxies/batch-activate`) so an operator can re-activate proxies in one click. Regression guard: `tests/unit/proxy-health-6246.test.ts`. (thanks @tenshiak)
+
+- **chatcore (tools): stop the default 128-tool cap from silently dropping opencode's `task`/MCP tools.** opencode (used as an MCP/agent host) sends a large tool list; when it exceeds the speculative `MAX_TOOLS_LIMIT` (128) default, `truncateToolList` did a blind `tools.slice(0, 128)`, dropping every tool past index 128 — including opencode's built-in `task` tool (subagent launch) and many MCP tools, so models routed through OmniRoute could no longer spawn subagents or reach part of their tools. The cap exists to avoid upstream `400`s for providers with real hard limits (e.g. grok-cli 200), so it is kept for those: detection of the opencode client (`isOpencodeClient` — any `x-opencode-*` header, or `opencode` in the user-agent) now only bypasses the **speculative 128 default**, never a known provider ceiling. Precedence is explicit — a proactive/detected provider limit always truncates (even for opencode); otherwise opencode forwards its full tool list; otherwise the unchanged 128 default applies to every other client. Refactors `getEffectiveToolLimit` into `getKnownToolLimit(provider) ?? DEFAULT_LIMIT` (byte-identical for existing callers) and fixes a cosmetic debug-log that reported the truncated count instead of the original. Regression guard: `tests/unit/tool-limit-detector.test.ts`.
+
+- **fix(mitm):** the macOS MITM-cert install check now matches the system keychain again. `security find-certificate -a -Z` prints the SHA-1 as a colon-less hex string, but the installed-check compared it against `getCertFingerprint()`'s colon-separated form, so the substring match never hit — the cert was reported as not-installed and re-prompted for the sudo install on every run. Fingerprints are now normalized (colons stripped, upper-cased) on both sides via the extracted `macCertOutputHasFingerprint` helper. Regression guard: `tests/unit/mitm-cert-mac-fingerprint.test.ts`. ([#6204](https://github.com/diegosouzapw/OmniRoute/pull/6204), closes [#6134](https://github.com/diegosouzapw/OmniRoute/issues/6134) — thanks @rianonehub)
+
+- **fix(api):** `/v1/messages/count_tokens` now counts `tool_use`, `tool_result` and `thinking` content blocks (and array-form `system` prompts) in the local-estimation path, instead of only `text`. Real agentic conversations keep ~95% of their tokens inside tool results; the previous estimate returned near-zero for them, which silently broke Claude Code's auto-compaction (context grew past the window with no compaction until the upstream API rejected the request). The real provider-side count path is unchanged. Regression guard: `tests/unit/messages-count-tokens-route.test.ts`. ([#6221](https://github.com/diegosouzapw/OmniRoute/pull/6221) — thanks @luweiCN)
+
+- **fix(antigravity):** strip a trailing assistant prefill turn for Vertex Claude models to avoid upstream 400s ([#6114](https://github.com/diegosouzapw/OmniRoute/pull/6114)). Regression guard: `tests/unit/antigravity-claude-prefill-strip.test.ts`. (thanks @anki1kr)
+
+- **fix(security):** the mutable cloud-agent routes (`/api/cloud/credentials/update`, `/api/cloud/models/alias`) now require management auth instead of being treated as public. They were classified as public API routes, so a request without management credentials could update stored cloud-agent credentials and model aliases. They are removed from the public-route set, classified as management routes in the authz pipeline, and gated by `requireManagementAuth`; cloud **read**/auth routes stay public. Regression guards: `tests/unit/cloud-write-auth.test.ts`, `tests/unit/authz/classify.test.ts`, `tests/unit/public-api-routes.test.ts`. ([#6233](https://github.com/diegosouzapw/OmniRoute/pull/6233) — thanks @vittoroliveira-dev)
+
+- **refactor(dashboard):** extract the onboarding-wizard "Open provider details" link target into a pure, unit-tested `buildProviderDetailsHref(connection)` helper. The wizard already routes by `connection.id` (the node UUID) rather than the provider category slug (#6144/#6145); this hardens that behavior behind a tested helper that guards a missing id/connection. Regression guard: `tests/unit/provider-onboarding-href.test.ts`. ([#6166](https://github.com/diegosouzapw/OmniRoute/pull/6166) — thanks @KooshaPari)
+
+- **fix(security):** the doubao synthetic device-id generator now derives its digits via an unbiased crypto-random draw (rejection sampling over `crypto.randomBytes()`) instead of a `% 10` reduction, closing a CodeQL `js/biased-cryptographic-random` finding.
+- **fix(agentSkills):** the GitHub-skills generator now resolves `outputDir` to an absolute path before writing, fixing a regression introduced by #6366 (relative-to-cwd base path) that could write generated skill files to the wrong directory.
+- **fix(security):** `/api/keys/{id}/devices` now checks the HTTP method before auth/validation, returning a `405` for non-GET/DELETE verbs instead of a misleading `401`/`500` (closes a `dast-smoke` QUERY-method finding).
+- **fix(quality):** clear the last 2 heavy quality-gate reds on the release tip (cycle pre-flight).
+- **fix(mitm):** the test suite and CI can never mutate the OS trust store — `OMNIROUTE_SKIP_SYSTEM_TRUST=1` is set globally for tests/CI so `installCert`/`uninstallCert`/`installTproxyCa` skip the privileged OS dispatch ([#6310](https://github.com/diegosouzapw/OmniRoute/pull/6310); full detail is under the [3.8.45] section below — this branch received it via the parallel-cycle sync-back).
+- **fix(api):** `POST /api/github-skills` now Zod-validates its request body; documented the new quality-gate env vars and pinned the merge-integrity GitHub Actions to a commit SHA.
+- **fix(skills):** generate the missing `omni-github-skills` registry entry and align the agent-skills catalog-count tests (follow-up to #6186).
+- **fix(quality):** clear the cycle's 11 net-new ESLint errors and make `validate-release-green` suppressions-aware.
+
+### 📝 Maintenance
+
+- **i18n(it):** add 118 missing Italian (`it`) translations (net-additive — no existing keys dropped, valid JSON), improving Italian UI coverage. ([#6212](https://github.com/diegosouzapw/OmniRoute/pull/6212) — thanks @serverless83)
+- **chore(providers):** remove deprecated **MiMo V2** model entries from the catalogs (xiaomi-mimo, opencode-go, zenmux-free, audio TTS) — the upstream V2 line is superseded by MiMo V2.5; drops `mimo-v2-tts`, `mimo-v2-pro`, `mimo-v2-omni`, `mimo-v2-flash`, `mimo-v2-flash-free` and realigns the provider-catalog tests. ([#6248](https://github.com/diegosouzapw/OmniRoute/pull/6248) — thanks @backryun)
+- **chore(release):** ~50 commits on this branch are v3.8.45 pre-flight/hardening fixes and CI-perf work that landed here via the parallel-cycle sync-back (`sync-next-cycle.mjs`, Hard Rule #21) **after** the `v3.8.45` git tag was cut, and are already fully documented under the **[3.8.45]** section below — listed here only so the per-cycle commit-coverage check (`npm run release:uncovered`) doesn't flag them as gaps. Provider/catalog/UX/backend fixes: #6041, #6078, #6108, #6135, #6148, #6149, #6154, #6158, #6161, #6162, #6163, #6164, #6165, #6170, #6177, #6178, #6181, #6186, #6187, #6191, #6193, #6194, #6195, #6200, #6205, #6208, #6209, #6211, #6213, #6223, #6224, #6225, #6226, #6227, #6228, #6229, #6230, #6235, #6291, #6292. CI/release-pipeline work: #6167, #6203, #6214, #6215, #6218, #6273, #6275, #6283, #6284, #6285, #6300, #6305.
+- **chore(release):** additional zero-ref release-cycle plumbing on this branch, kept out of `release:uncovered` on purpose (no `#N` in the commit subject to cite): opening the v3.8.46 cycle, opening/closing the v3.8.45 cycle, the finalized [3.8.45] CHANGELOG i18n sync-back to 42 mirrors, the v3.8.45 cognitive/cyclomatic and file-size drift rebaselines, ESLint stale-suppression pruning (4,273 → 4,233), and clearing test-masking/docs-all pre-flight reds for v3.8.45.
+
+### ⚡ Performance & Infrastructure
+
+- **perf(release-green):** the pre-flight validator (`scripts/quality/validate-release-green.mjs`) now runs its 4 slow suites (unit / vitest / integration / pack-artifact) **concurrently** via `Promise.all` — pre-flight wall time drops from ~the sum of the suites to ~the slowest one (~30min saved per round; Phase 0 was the nº1 bottleneck of the v3.8.45 release benchmark, 2h54 of 6h34 e2e). Guard: `tests/unit/validate-release-green.test.ts` ("runs the slow suites CONCURRENTLY"). ([#6319](https://github.com/diegosouzapw/OmniRoute/pull/6319))
+- **fix(ci):** `scripts/release/sync-next-cycle.mjs` — two defects found live in its first production run (v3.8.45 Phase 5): (1) the `git()` helper's default 1 MiB `maxBuffer` crashed with `ENOBUFS` on `git show origin/main:CHANGELOG.md` (the CHANGELOG alone is >1 MiB) — widened to 64 MiB; (2) the i18n resync only propagated the `[NEXT]` (TBD) section, leaving the just-shipped finalized section as "— TBD" in all 42 mirrors — it now also syncs `[prevVersion]` bounded by the heading below it (new exported pure helper `versionAfter`). Guards: +5 tests in `tests/unit/sync-next-cycle.test.ts` (8/8). ([#6327](https://github.com/diegosouzapw/OmniRoute/pull/6327))
+- **test(ci):** concurrency-sensitive flaky tests are **quarantined into a serial pass** (`tests/unit/serial/`, `--test-concurrency=1`, appended to every unit runner incl. sharded variants — the serial pass is sharded too so concurrent shard jobs never self-collide). Initial set: `glm-coding-plan-monthly-3580`, `quota-division-blocks`, `provider-health-autopilot`, `combo-health-autopilot` — the class behind the ~28min CI wedges/re-runs (two live 1h+ wedges cancelled during this PR's own validation). Discovery + TIA gates track the new glob; systemic root cause (async logger writing after teardown) tracked in [#6360](https://github.com/diegosouzapw/OmniRoute/issues/6360). Guard: `tests/unit/test-serial-quarantine.test.ts` (4). ([#6347](https://github.com/diegosouzapw/OmniRoute/pull/6347))
+
+### 🙌 Contributors
+
+Thanks to everyone whose work landed in v3.8.46:
+
+| Contributor                                                    | PRs / Issues           |
+| -------------------------------------------------------------- | ---------------------- |
+| [@2220258345](https://github.com/2220258345)                   | direct commit / report |
+| [@abdofallah](https://github.com/abdofallah)                   | direct commit / report |
+| [@adentdk](https://github.com/adentdk)                         | direct commit / report |
+| [@aleksesipenko](https://github.com/aleksesipenko)             | direct commit / report |
+| [@anki1kr](https://github.com/anki1kr)                         | direct commit / report |
+| [@anungma](https://github.com/anungma)                         | direct commit / report |
+| [@arpicato](https://github.com/arpicato)                       | direct commit / report |
+| [@backryun](https://github.com/backryun)                       | #6248                  |
+| [@binsarjr](https://github.com/binsarjr)                       | direct commit / report |
+| [@brightfiscalband](https://github.com/brightfiscalband)       | direct commit / report |
+| [@chirag127](https://github.com/chirag127)                     | #6501, #6506           |
+| [@developerjillur](https://github.com/developerjillur)         | direct commit / report |
+| [@dilneiss](https://github.com/dilneiss)                       | #6499                  |
+| [@dtybnrj](https://github.com/dtybnrj)                         | direct commit / report |
+| [@Forcerecon](https://github.com/Forcerecon)                   | direct commit / report |
+| [@hao3039032](https://github.com/hao3039032)                   | direct commit / report |
+| [@Iammilansoni](https://github.com/Iammilansoni)               | #6150, #6245           |
+| [@jmengit](https://github.com/jmengit)                         | direct commit / report |
+| [@jordansilly77-stack](https://github.com/jordansilly77-stack) | direct commit / report |
+| [@JxnLexn](https://github.com/JxnLexn)                         | direct commit / report |
+| [@KooshaPari](https://github.com/KooshaPari)                   | #6166                  |
+| [@loopyd](https://github.com/loopyd)                           | direct commit / report |
+| [@luweiCN](https://github.com/luweiCN)                         | #6221                  |
+| [@makcimbx](https://github.com/makcimbx)                       | direct commit / report |
+| [@muflifadla38](https://github.com/muflifadla38)               | direct commit / report |
+| [@newnol](https://github.com/newnol)                           | direct commit / report |
+| [@ofekbetzalel](https://github.com/ofekbetzalel)               | direct commit / report |
+| [@ohahe52-dot](https://github.com/ohahe52-dot)                 | direct commit / report |
+| [@phidinhmanh](https://github.com/phidinhmanh)                 | direct commit / report |
+| [@powellnorma](https://github.com/powellnorma)                 | direct commit / report |
+| [@qpeyba](https://github.com/qpeyba)                           | direct commit / report |
+| [@RaviTharuma](https://github.com/RaviTharuma)                 | direct commit / report |
+| [@RCrushMe](https://github.com/RCrushMe)                       | direct commit / report |
+| [@rianonehub](https://github.com/rianonehub)                   | #6134, #6204           |
+| [@serverless83](https://github.com/serverless83)               | #6212                  |
+| [@swingtempo](https://github.com/swingtempo)                   | direct commit / report |
+| [@tenshiak](https://github.com/tenshiak)                       | direct commit / report |
+| [@ThongAccount](https://github.com/ThongAccount)               | direct commit / report |
+| [@UnrealAryan](https://github.com/UnrealAryan)                 | direct commit / report |
+| [@vinayakkulkarni](https://github.com/vinayakkulkarni)         | direct commit / report |
+| [@vittoroliveira-dev](https://github.com/vittoroliveira-dev)   | #6233                  |
+| [@warelik](https://github.com/warelik)                         | direct commit / report |
+| [@xxy9468615](https://github.com/xxy9468615)                   | direct commit / report |
+| [@xz-dev](https://github.com/xz-dev)                           | direct commit / report |
+| [@yanpaing007](https://github.com/yanpaing007)                 | direct commit / report |
+| [@diegosouzapw](https://github.com/diegosouzapw)               | maintainer             |
+
+---
+
 ## [3.8.45] — 2026-07-06
 
 ### ✨ New Features

docs/i18n/uk-UA/CHANGELOG.md

@@ -6,6 +6,213 @@
 
 ## [3.8.31] — 2026-06-20
 
+## [3.8.46] — TBD
+
+### ✨ New Features
+
+- **feat(sse):** **hide paid-only models from `auto/*` routing** when `hidePaidModels` is on ([#6512](https://github.com/diegosouzapw/OmniRoute/issues/6512)) — follow-up to #6328/#6495. PR #6495 hid paid-only models from the `GET /v1/models` listing, but `auto/*` combos (`auto/best-coding`, `auto/glm`, …) could still pick a paid-only backend into their candidate pool → a 402/403 at request time. `createVirtualAutoCombo` now filters the candidate pool through the new pure `open-sse/services/autoCombo/paidModelFilter.ts` (`filterPaidOnlyCandidates`), applying the same free-model predicate #6495 uses in `catalog.ts` (`providerHasFreeModels(provider) && isFreeModel(provider, {id})`) whenever `settings.hidePaidModels === true`. Applied before the category/tier/family narrowing, so it covers every `auto/*` combo; an all-paid pool degrades to the existing graceful empty-pool path. **Opt-in — default OFF leaves the pool unchanged** (identity). Regression guard: `tests/unit/autoCombo/paid-model-filter-6512.test.ts` (4, incl. the default-off identity guard).
+- **feat(sse):** **provider-family auto combos** — `auto/glm`, `auto/minimax`, `auto/mimo`, `auto/zai`, `auto/gemma`, `auto/llama`, `auto/gemini` ([#6453](https://github.com/diegosouzapw/OmniRoute/issues/6453)) — new routable ids that materialize an on-demand virtual combo spanning whatever installed backends currently expose that model family, degrading gracefully as backends rotate. A new pure `open-sse/services/autoCombo/modelFamily.ts` (`detectModelFamily`) classifies by model-id prefix for six families; `zai` is instead resolved by provider id (z.ai's hosted API serves the same `glm-*` model ids as every other GLM backend, so `auto/zai` means "route to my z.ai backend specifically" vs `auto/glm`'s "any connected GLM backend"). Reuses the existing `createVirtualAutoCombo` on-demand materialization path (no DB writes) and the `/v1/models` catalog advertising loop. Regression guard: `tests/unit/autoCombo/provider-family-combos.test.ts` (11).
+- **feat(proxy):** native **proxy-pool round-robin / egress IP rotation** ([#6365](https://github.com/diegosouzapw/OmniRoute/issues/6365)) — a scope (global / provider / account) can now hold **multiple** proxies as a pool with a rotation strategy, so outbound requests cycle their egress IP instead of pinning one proxy per scope. Migration `117_proxy_pool_rotation.sql` lifts the `UNIQUE(scope, scope_id)` constraint (rebuild via the canonical rename/copy/drop; existing single assignments become 1-element pools) and adds a `proxy_scope_rotation` companion table holding the per-scope strategy + a persisted monotonic round-robin cursor. Strategies: `round-robin` (default, monotonic cursor — never `Math.random`), `random`, and `sticky-per-N-min`. Resolution (`resolveProxyForScopeFromRegistry` / `resolveProxyForConnectionFromRegistry`) now fetches the alive, position-ordered candidate set (unchanged `PROXY_ALIVE_PREDICATE`) and applies the strategy; an empty / all-dead pool still returns `null` — the #6246 fail-closed guard is **untouched** (never falls through to direct egress). Backend + DB only; dashboard pool-builder UI is a follow-up. Regression guard: `tests/unit/proxy-pool-rotation-6365.test.ts` (8, incl. fail-closed + backward-compat).
+- **feat(providers):** end-to-end **tool/function calling on the native Gemini `/v1beta` endpoint** ([#6222](https://github.com/diegosouzapw/OmniRoute/issues/6222)) — both directions of the Gemini↔OpenAI conversion now preserve tool data (previously silently dropped). Request side: `convertGeminiToInternal` (extracted to its own testable module) maps `tools[].functionDeclarations` → OpenAI `tools`, prior `functionCall` parts → assistant `tool_calls`, and `functionResponse` parts → `tool`-role messages. Response side: `convertOpenAIResponseToGemini` emits `parts[].functionCall {name,args}` from `message.tool_calls`, and the streaming `openAIChunkToGeminiChunk` accumulates fragmented `tool_calls` deltas by index into complete `functionCall` parts. The non-Gemini client paths (Claude, OpenAI-Responses) already preserved tool calls — this closes the gap specific to the native Gemini surface. Regression guard: `tests/unit/v1beta-gemini-tool-calling-6222.test.ts` (6, incl. a streaming SSE round-trip).
+- **feat(providers):** copilot-m365-web **enterprise / work tier** support ([#6334](https://github.com/diegosouzapw/OmniRoute/issues/6334)) — mirrors the EDU-tier pattern (#6210): `M365ConnectionParams` gains an `agent` field, a new opt-in `M365_ENTERPRISE_OVERRIDES` preset (`agent=work`, `scenario=officeweb`, `licenseType=Premium`) applies via `providerSpecificData.tier="enterprise"` (alias `"work"`), and `agent` is also overridable directly via `providerSpecificData.agent`. `buildWsUrl` was hardcoding `agent="web"` (the one enterprise-distinguishing param with no override path), so a Premium work account handshook then returned an empty stream. The individual and EDU paths are untouched. Kilo's dup flag vs #6210 (EDU tier) was a false positive — different tier. Regression guard: `tests/unit/copilot-m365-enterprise-6334.test.ts` (7). End-to-end confirmation on a real Premium work account is a live-VPS validation follow-up (Hard Rule #18). (thanks @Forcerecon)
+- **feat(api):** standardized, provider-agnostic **`effort` + `thinking` request params** ([#6241](https://github.com/diegosouzapw/OmniRoute/issues/6241)) — a thin standardization layer over the existing mature per-provider reasoning plumbing (no provider mapper touched). `providerChatCompletionSchema` gains a canonical `effort` (reusing the shared `none/low/medium/high/xhigh` vocabulary — the UI tiers `extra`/`max` collapse onto `xhigh`) and a boolean `thinking`. A pure `normalizeReasoningRequest` (wired once in `src/sse/handlers/chat.ts`, before any reasoning field is read) folds them onto the fields the translators already consume (`reasoning_effort` / `reasoning.effort` / `thinking`), so they fan out to Anthropic / Gemini / xAI / Responses — an explicit client `reasoning_effort` / object-shaped `thinking` always wins (backward-compatible). `/models` additively exposes `supportsThinking` + `effort_tiers` so the frontend can render the toggles (UI component is a follow-up). Regression guard: `tests/unit/effort-thinking-standardization-6241.test.ts` (12). (thanks @Iammilansoni, @shabeer)
+- **feat(combo):** new **`pipeline` (sequential) combo strategy** ([#6297](https://github.com/diegosouzapw/OmniRoute/issues/6297)) — the 18th routing strategy runs targets **in order**, threading each step's output into the next step's input, with an optional per-step `prompt` (system instruction); only the final step's response is returned. Distinct from `fusion` (parallel fan-out + judge). Implemented as a self-contained `open-sse/services/pipeline.ts` (sibling to `fusion.ts`), dispatched from `combo.ts`; the step list reuses `combo.models` order and reads an optional `prompt` off each target (backward-compatible — ignored by every other strategy). Intermediate steps run non-streaming with tools stripped (complete prose to thread forward); the final step keeps the client's `stream` flag + tools. A failing/empty/unparseable intermediate step fails the whole pipeline explicitly via a sanitized error (never silently swallowed). Kilo's dup flag vs #563 was a false positive (that's model→chain selection; this is a sequential chain). Regression guard: `tests/unit/combo-pipeline-strategy.test.ts` (5). (thanks @ofekbetzalel)
+- **feat(ci):** `check:test-masking` now flags **inline-reimplemented prod conditions** ([#6348](https://github.com/diegosouzapw/OmniRoute/issues/6348)) — a new **report-only** subcheck (v2, 6A.10 family) catches the wrong-shape contract test: a test that recomputes the condition under test inline instead of importing/exercising the real function (the #6216 class, where `=== 500` → `>= 500` stayed green because the test re-implemented the branch). For each added/modified test file it warns when the file textually duplicates a ≥3-token conditional from a production file touched in the same PR **and** does not import the symbol/module owning it, via a pure, fixture-tested `findReimplementedConditions()` with an allowlist mirroring `assertReductionAllowlist`. Report-only for now (does not fail the gate) — to be promoted to blocking after a triage cycle. Regression guard: `tests/unit/check-test-masking.test.ts` (45).
+- **feat(sse):** per-connection routing override (native vs CLIProxyAPI) ([#6339](https://github.com/diegosouzapw/OmniRoute/issues/6339)) — the previously-dead `isCliproxyapiDeepModeEnabled` helper is now wired into `resolveExecutorWithProxy`: a single connection can opt itself into the CLIProxyAPI passthrough executor via `providerSpecificData.cliproxyapiMode="claude-native"`, with precedence **connection override > provider `upstream_proxy_config` mode > default**. `resolveExecutorWithProxy` now receives the resolved connection's `providerSpecificData` (threaded from `chatCore.ts`), so one connection can deep-route while the provider's other connections stay native — no DB schema change (the toggle rides in `providerSpecificData`). Also resolves the same-provider mixing ask in #6340. Regression guard: `tests/unit/chatcore-executor-proxy.test.ts` (9). (thanks @RaviTharuma)
+- **feat(dashboard):** "Add session cookie" modal now shows a prominent **"Open ‹host› →"** link to the provider's own site ([#6268](https://github.com/diegosouzapw/OmniRoute/issues/6268)) — every `-web` cookie-session provider (chatgpt-web, claude-web, gemini-web, kimi-web, lmarena, qwen-web, m365-copilot-web, …) renders a one-click external link (opening the provider's login/home page in a new tab) so operators no longer tab away to retype the URL mid-setup. The host resolves from a pure, unit-tested `resolveWebProviderHost()` (prefers `WEB_COOKIE_PROVIDERS[id].website`, falls back to the registry `baseUrl` origin); non-web providers render exactly as before. Kilo's dup flag vs #6265 (modal-too-small-on-1080p) was a false positive — distinct concern. Regression guard: `tests/unit/resolve-web-provider-host.test.ts` (5). (thanks @chirag127)
+- **feat(providers):** add **DigitalOcean** AI (serverless inference) as an OpenAI-compatible API-key provider ([#6373](https://github.com/diegosouzapw/OmniRoute/pull/6373)) — base `https://inference.do-ai.run/v1`, wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/digitalocean/`, `src/shared/constants/providers/apikey/inference-hosts.ts`). Regression guard: `tests/unit/digitalocean-provider.test.ts`. (thanks @newnol)
+- **feat(providers):** add **Huancheng Public API** (`hcnsec`) as an OpenAI-compatible regional provider ([#6410](https://github.com/diegosouzapw/OmniRoute/pull/6410)) — Xinjiang Huancheng Cybersecurity's public LLM platform (base `https://api.hcnsec.cn/v1`, free credits via daily check-ins), wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/hcnsec/`, `src/shared/constants/providers/apikey/regional.ts`). Regression guard: `tests/unit/hcnsec-provider.test.ts`. (thanks @UnrealAryan)
+- **feat(dashboard):** the web-session credential guide now shows an **"Open {host}" link** ([#6316](https://github.com/diegosouzapw/OmniRoute/pull/6316)) to the provider's sign-in site (derived from the provider `website` via `getProviderWebsiteHost`), so you can jump straight to the page where the cookie/session must be captured. Regression guard: `tests/unit/web-session-provider-link-6316.test.ts`. (thanks @jordansilly77-stack)
+- **feat(cerebras):** add the **Gemma 4 31B** model (`gemma-4-31b`) to the Cerebras registry + pricing table ([#6331](https://github.com/diegosouzapw/OmniRoute/pull/6331)). Regression guard extends `tests/unit/t28-model-catalog-updates.test.ts`. (thanks @backryun)
+- **feat(providers):** add **Yuanbao (web)** as a cookie-session provider ([#6196](https://github.com/diegosouzapw/OmniRoute/issues/6196)) — `yuanbao-web` (Tencent Yuanbao, `yuanbao.tencent.com`) with cookie-only auth (`hy_user`/`hy_token` + public agent id), SSE→OpenAI translation incl. `reasoning_content`, exposing DeepSeek V3/R1 + Hunyuan / Hunyuan-T1. Regression guard: `tests/unit/providers-yuanbao-web.test.ts`. `together-web` was **deferred** (no verifiable web-session endpoint — needs a captured request) and `huggingchat-web` **dropped** (the existing `huggingchat` already is a web-cookie provider). (thanks @chirag127)
+- **feat(providers):** route the built-in **agentrouter** through the dynamic Claude-Code wire image ([#6056](https://github.com/diegosouzapw/OmniRoute/issues/6056)) — a small static allow-set (`CC_WIRE_IMAGE_BUILTINS` in `open-sse/services/ccWireImageBuiltins.ts`), consulted by `isClaudeCodeCompatible` / `isClaudeCodeCompatibleProvider` / `applyFingerprint`, makes agentrouter adopt the CC wire-image headers + fingerprint **while guarding the CC baseUrl/auth branches** so it keeps its own registry `baseUrl` and `x-api-key` auth. Regression guard: `tests/unit/agentrouter-cc-wire-image.test.ts` (asserts the wire image is applied AND agentrouter's baseUrl/auth are preserved). Live WAF-acceptance against agentrouter.org is a VPS validation follow-up (Hard Rule #18).
+- **feat(providers):** **bulk-add API keys for Cloudflare Workers AI** ([#6174](https://github.com/diegosouzapw/OmniRoute/issues/6174)) — `cloudflare-ai` is removed from the bulk-add exclusion list and the bulk parser gains a 3-field `name|accountId|apiKey` mode; the bulk route now builds a **per-entry** `providerSpecificData` so each key carries its own `accountId` (fixing the previous shared-object reuse), and both the create + key-validation paths receive it. Regression guard: `tests/unit/bulk-api-key-parser-cloudflare.test.ts`. (thanks @muflifadla38)
+- **feat(dashboard):** routing/settings UX clarity ([#6147](https://github.com/diegosouzapw/OmniRoute/issues/6147)) — (1) weighted combos show the **effective routing share %** next to each weight when weights don't sum to 100 (`WeightTotalBar.tsx`); (2) the status widget's user-facing **"Cloud Sync" label is renamed** to "Remote Settings Sync" (`CloudSyncStatus.tsx`; internal ids/state untouched); (3) built-in providers gain an **opt-in advanced base-URL override** (`isBaseUrlOverrideEligibleProvider`, hidden behind an "Advanced" toggle, reusing the existing `providerSpecificData.baseUrl` persistence — not globally widened). Regression guard: `tests/unit/routing-settings-ux-6147.test.ts`.
+- **feat(combo):** add an option to **disable session stickiness**, per-combo or globally — round-robin / random combos can rotate to a different connection on every request instead of pinning a whole conversation to one connection by its first-message hash. Resolution precedence per-combo `config.disableSessionStickiness` → global `settings.disableSessionStickiness` → default `false` (preserves the #3825 prompt-cache/504 fix); gates **both** stickiness call sites in `open-sse/services/combo.ts`. Exposed as a global toggle (Combo Defaults) and a per-combo Inherit/on/off control. ([#6168](https://github.com/diegosouzapw/OmniRoute/issues/6168)) Regression guard: `tests/unit/combo-disable-session-stickiness.test.ts`. (thanks @RCrushMe)
+- **feat(docker):** add the `OMNIROUTE_NO_SUDO` env flag for root-less / user-namespaced deployments — the MITM cert-trust command path (`resolveSudoSpawn` in `src/mitm/systemCommands.ts`) now strips the leading `sudo` when the flag is truthy, in addition to the existing root / sudo-missing cases, so the Proxy Agent runs without `sudo` (the operator trusts the CA manually, e.g. via `NODE_EXTRA_CA_CERTS`). Argv-array `spawn` preserved — no shell interpolation (Hard Rule #13). ([#6122](https://github.com/diegosouzapw/OmniRoute/issues/6122)) Regression guard: `tests/unit/mitm-systemCommands-no-sudo.test.ts`. (thanks @powellnorma)
+- **feat(providers):** add **Requesty** as an OpenAI-compatible gateway provider (BYOK, base `https://router.requesty.ai/v1`, ~200 free requests/day) — wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/requesty/`, `src/shared/constants/providers/apikey/gateways.ts`). ([#6120](https://github.com/diegosouzapw/OmniRoute/issues/6120)) Regression guard: `tests/unit/requesty-provider.test.ts`. (thanks @chirag127)
+- **feat(dashboard):** add **configured-only / available-only filters** to the Free Provider Rankings page ([#6150](https://github.com/diegosouzapw/OmniRoute/issues/6150)) — hide providers you haven't configured, or whose connections are all rate-limited / out of quota, via server-side query params (`?configuredOnly` / `?availableOnly` on `GET /api/free-provider-rankings`) backed by a testable lib helper reusing the in-process connection state (no Redis). Both filters default off, so the default view is unchanged; this supersedes the earlier client-side "Configured Only" toggle (#6245) with an available-only dimension and unit-tested logic. Regression guard: `tests/unit/freeProviderRankings-filters.test.ts`.
+
+### 🔧 Bug Fixes
+
+- **fix(dashboard):** adding a second API key connection for the same provider no longer silently overwrites the first — the Add-API-key modal now derives a unique default connection name (`main`, then `main-2`, `main-3`, …) so the backend name-based upsert can't collide (#6499 — thanks @dilneiss).
+- **fix(compression):** the session-dedup engine now also deduplicates a large multi-line block repeated **within a single message** (intra-message dedup), not just across turns; the compression-preview API surfaces a `fallbackReason`, and the fusion panel reports how many models were rate-limited vs failed on total-panel failure (#6501 — thanks @chirag127).
+- **fix(compression):** a stacked-pipeline step naming an unregistered engine now surfaces a `validationErrors` entry instead of silently no-op'ing, so misconfigured pipelines are visible in the preview API (#6506 — thanks @chirag127).
+- **feat(usage):** add a **Codex reset-credit redemption** flow to the Provider Limits UI ([#6361](https://github.com/diegosouzapw/OmniRoute/pull/6361)) — a `useCodexResetCreditRedemption` hook + `/api/usage/codex-reset-credit` route + `codexResetCredits` lib let you redeem banked Codex reset credits from the quota card. Regression guard: `tests/unit/codex-reset-credits.test.ts`. (thanks @JxnLexn)
+- **feat(glm):** add **team-plan quota settings** for `glm-cn` connections ([#6351](https://github.com/diegosouzapw/OmniRoute/pull/6351)) — a dedicated `GlmTeamQuotaFields` form section (team quota id / limits) threaded through the Add/Edit connection modals, persisted via `providerSpecificData`, with the GLM usage service reading the team quota. Regression guards: `tests/unit/glm-team-quota.test.ts`, `provider-specific-data-schema.test.ts`. (thanks @hao3039032)
+- **feat(providers):** add **TinyFish** web-fetch/search support ([#6349](https://github.com/diegosouzapw/OmniRoute/pull/6349)) — a `tinyfish-fetch` executor + `/v1/web/fetch` route + MCP web-fetch tool, registered as a specialty-media provider with request-validation and a search-provider catalog entry. Regression guards: `tests/unit/executor-tinyfish-fetch.test.ts`, `web-fetch-handler.test.ts`, `mcp-web-fetch-tool.test.ts`, `provider-validation-tinyfish.test.ts`. (thanks @dtybnrj)
+- **fix(cli):** `omniroute launch-codex` now spawns `codex.cmd` through a shell on Windows (the npm `.cmd` shim is unresolvable by bare `spawn` → ENOENT), mirroring the qodercli Windows fix (#6263) ([#6312](https://github.com/diegosouzapw/OmniRoute/pull/6312)). Regression guard: `tests/unit/launch-codex-windows-spawn-6312.test.ts`. (thanks @swingtempo)
+- **fix(codex):** isolate the **Spark** quota from the shared Codex quota and stabilize the quota UI ordering / hydration so per-scope limits render consistently ([#6336](https://github.com/diegosouzapw/OmniRoute/pull/6336)). Regression guards: `tests/unit/codex-quota-selection-hydration.test.ts`, `provider-limits-ui.test.ts` + 3 more. (thanks @xz-dev)
+- **feat(api):** add a `hidePaidModels` setting that filters paid-only models out of the `/v1/models` catalog. Regression guard: `tests/unit/models-catalog-hide-paid.test.ts`. (thanks @chirag127)
+- **fix(api-manager):** the fallback model picker now preserves combos instead of dropping them when a primary model is unavailable ([#6443](https://github.com/diegosouzapw/OmniRoute/pull/6443)). Regression guard: `tests/unit/api-manager-page-static.test.ts`. (thanks @jmengit)
+- **fix(providers):** recoverable Antigravity / Cloud-Code (Gemini Code Assist) `403` responses ([#6452](https://github.com/diegosouzapw/OmniRoute/pull/6452)) are now classified as a retryable project-config error instead of a terminal account ban, so a fixable project/API-disabled 403 no longer forces a ~1-year cooldown / full OAuth reconnect. Regression guard: `tests/unit/errorclassifier-antigravity-403.test.ts`. (thanks @developerjillur)
+- **fix(mitm):** `sanitizeHeaders` now redacts `Set-Cookie` response headers so upstream session cookies never leak into logs / diagnostics ([#6451](https://github.com/diegosouzapw/OmniRoute/pull/6451)). Regression guard: `tests/unit/mitm-sanitize-headers.test.ts`. (thanks @developerjillur)
+- **fix(api):** `/api/compression/preview` now accepts `mode: "caveman"` and correctly handles stacked / zero-compression previews ([#6425](https://github.com/diegosouzapw/OmniRoute/issues/6425)). Regression guard: `tests/unit/api/compression-preview-caveman-and-stacked-6425.test.ts`. (thanks @chirag127)
+- **feat(providers):** add **Zed** hosted LLM aggregator as a native-app provider ([#6118](https://github.com/diegosouzapw/OmniRoute/pull/6118)) — OAuth sign-in via the Zed hosted flow, registered through the shared provider registry + executor. Regression guards: `tests/unit/zed-oauth-provider.test.ts`, `zed-import-utils.test.ts`, `zed-docker-detect.test.ts`, `mitm-handler-zed.test.ts`. VPS-validated via live operator login (Hard Rule #18).
+- **fix(oauth):** the Kiro SSO-cache auto-import now **preserves the IDC region** — cross-region Amazon Q / Kiro profiles imported from the SSO cache are no longer collapsed to the default region ([#6113](https://github.com/diegosouzapw/OmniRoute/pull/6113)). Regression guard: `tests/unit/kiro-auto-import-idc-2059.test.ts`. VPS-validated via live operator login (Hard Rule #18).
+- **fix(dashboard):** passthrough model aliases no longer collide when two namespaced model ids share a last segment (port from 9router#1850, [#6431](https://github.com/diegosouzapw/OmniRoute/pull/6431)). `enx/gpt-5.5` and `enx/codebuddy/gpt-5.5` both auto-generated the alias `gpt-5.5`, so the second model could never be added (the UI just alerted "alias already exists"). Aliases are now disambiguated deterministically — bare last segment when free, then parent-qualified (`codebuddy-gpt-5.5`), then a numeric suffix — while re-adding the exact same model id is still blocked. Regression guard: `tests/unit/passthrough-alias-1850.test.ts`. (thanks @arpicato)
+- **fix(translator):** preserve a Gemini `functionResponse` co-located with other parts (another `functionCall`, or trailing `text`) in the same content when translating **Gemini → OpenAI** ([#6376](https://github.com/diegosouzapw/OmniRoute/pull/6376)). `convertGeminiContent()` early-returned the tool message on the first `functionResponse` part, dropping any co-located parts; such contents are now pre-split (one tool message per `functionResponse`, emitted first, plus one message for the remaining parts). Regression guard: `tests/unit/gemini-to-openai-function-response.test.ts`. (thanks @warelik)
+- **fix(headroom):** detect a python interpreter managed by **mise / pyenv / asdf / conda** (port from 9router#2353, [#6382](https://github.com/diegosouzapw/OmniRoute/pull/6382)). Headroom's python probe (`src/lib/headroom/detect.ts`) searched a hardcoded `PATH`, but version managers expose their interpreters via shim dirs that only join `PATH` through interactive-shell activation — which the non-interactive server never runs, so a managed python (≥3.10) was invisible and Headroom reported it missing. The search path now prepends the well-known shim/bin dirs (`~/.local/share/mise/shims`, `~/.pyenv/shims`, `~/.asdf/shims`, `$CONDA_PREFIX/bin`, `~/.local/bin`, respecting `MISE_DATA_DIR`/`PYENV_ROOT`/`ASDF_DATA_DIR` when set), and a new `HEADROOM_PYTHON` env override lets operators point straight at their interpreter (mirroring `HEADROOM_URL`). Still shell-free (`execFileSync`). Regression guard: `tests/unit/headroom-detect.test.ts` (5). (thanks @loopyd)
+- **fix(executors):** strip the OpenAI-Codex/Claude-CLI `client_metadata` passthrough field for **NVIDIA** requests (port from 9router#1887, [#6411](https://github.com/diegosouzapw/OmniRoute/pull/6411)). NVIDIA's OpenAI-compatible wrapper rejects it with `400 Unsupported parameter`, the same class already handled for `cerebras`/`mistral`; `nvidia` (executor `default`) was missing from the strip allowlist so Codex/Claude-Code passthrough requests 400'd. Regression guard: `tests/unit/executor-default-strip-client-metadata.test.ts` (+nvidia case). (thanks @phidinhmanh)
+- **fix(translator):** strip the Claude-style `thinking` field for **NVIDIA `z-ai/glm-5.2`** (port from 9router#2023, [#6413](https://github.com/diegosouzapw/OmniRoute/pull/6413)). NVIDIA's OpenAI-compatible wrapper 400s on `thinking` (a Claude-format client routed here leaves a `thinking:{type:"adaptive"}`); the existing strip rule only dropped `reasoning`. Same class already handled for `minimax-m2.7`. Regression guard: `tests/unit/nvidia-minimax-thinking-strip.test.ts` (+glm-5.2 case). (thanks @phidinhmanh)
+- **fix(translator):** suppress the streamed `</think>` close marker for the **Antigravity IDE** client (port from 9router#1061, [#6415](https://github.com/diegosouzapw/OmniRoute/pull/6415)). On thinking-only turns Antigravity rendered a bare `</think>` as the sole visible content, tripping its loop-detection and wasting requests. Antigravity's UA (`vscode/<v> (Antigravity/<v>)`) is added to the marker-suppress allowlist (alongside OpenCode); Claude Code / Cursor still get the marker, and `x-omniroute-thinking-marker: on` force-restores it. Regression guard: `tests/unit/think-close-marker-suppress-5245.test.ts`. (thanks @abdofallah)
+- **fix(executors):** strip nested `reasoning_content` from messages for **Mistral** (port from 9router#1649, [#6417](https://github.com/diegosouzapw/OmniRoute/pull/6417)). Mistral's API returns `422 extra_forbidden` when an assistant message carries `reasoning_content` (replayed thinking from a prior turn, e.g. via the Codex `/responses` path); the generic top-level 400 field-downgrade retry never covered the nested per-message field. `DefaultExecutor` now strips it for provider `mistral` only, so DeepSeek (which requires replayed `reasoning_content`) is unaffected. Regression guard: `tests/unit/mistral-strip-reasoning-content-1649.test.ts`. (thanks @xxy9468615)
+- **fix(executors):** strip the `client_metadata` passthrough field on the **OpenCode** path (port from 9router#1442, [#6418](https://github.com/diegosouzapw/OmniRoute/pull/6418)). OpenCode upstreams (e.g. `kimi-k2.6` via opencode-go) reject it with `400 "Extra inputs are not permitted, field: 'client_metadata'"`; the DefaultExecutor strip only covered cerebras/mistral and `OpencodeExecutor` extends `BaseExecutor` directly, so nothing removed it there. Regression guard: `tests/unit/opencode-strip-client-metadata-1442.test.ts`. (thanks @yanpaing007)
+- **fix(executors):** inject the `reasoning_content` echo for the native **Moonshot Kimi** provider (port from 9router#1480, [#6419](https://github.com/diegosouzapw/OmniRoute/pull/6419)). Kimi (executor `default`) is a thinking-mode upstream that 400s with "reasoning_content must be passed back" when a prior assistant turn lacks it; the placeholder injection was only wired into the OpenCode meta-provider, so direct multi-turn Kimi conversations failed. Scoped to `kimi` (gateway-served models matching the thinking-model name pattern are unaffected). Regression guard: `tests/unit/kimi-native-reasoning-injected-1480.test.ts`. (thanks @2220258345)
+- **fix(executors):** recover from a strict gateway's `context_management: Extra inputs are not permitted` 400 (port from 9router#1468, [#6420](https://github.com/diegosouzapw/OmniRoute/pull/6420)). **Claude Code** always sends a top-level `context_management` field; strict anthropic-compatible gateways reject it. The dedicated context-editing 400-fallback only fired when OmniRoute's own `contextEditing` feature was enabled (default off), so a client-sent field passed through untouched and 400'd. `context_management` is now in the generic reactive field-strip list, so it's stripped-and-retried once regardless of the feature flag (with correct request re-signing for claude-compatible relays). Regression guard: `tests/unit/provider-field-strips.test.ts`. (thanks @ohahe52-dot)
+- **fix(network):** enable **RFC 8305 Happy Eyeballs** (`autoSelectFamily`) on the direct-egress undici dispatcher (port from 9router#1237, [#6423](https://github.com/diegosouzapw/OmniRoute/pull/6423)). When DNS returns both IPv6 (AAAA) and IPv4 (A) and the IPv6 route is broken (e.g. a NAT64 `64:ff9b::` prefix without routing), undici tried IPv6 first and hung until `ETIMEDOUT` (then a 502 + account lockout), even though `curl` reached the same host. The direct dispatcher now races both families and uses whichever connects first. Proxy paths pin family via `proxyTls` and are unaffected. Regression guard: `tests/unit/direct-dispatcher-pipelining-4580.test.ts`. (thanks @adentdk)
+- **fix(combo):** round-robin now advances the rotation pointer past the model that **actually served**, not the eagerly-scheduled one (port from 9router#948, [#6428](https://github.com/diegosouzapw/OmniRoute/pull/6428)). With `stickyLimit: 1` (true round-robin), when the scheduled model failed and a _different_ model served via fallback, the counter had already advanced +1 from the scheduled index — so the next request reused the fallback-served model, degrading round-robin into hot-spotting on whichever model was healthy. The pointer now advances to the served index + 1 (mirroring the sticky-limit>1 path). Session-stickiness (#3825) and distribution are preserved. Regression guard: `tests/unit/combo-rr-fallback-advance-948.test.ts`. (thanks @binsarjr)
+- **fix(sse):** a non-string `model` field is now rejected with a `400` before the resolver, instead of crashing downstream `.toLowerCase()`/`.split()` calls into an empty-body `500` that escapes the error sanitizer ([#6407](https://github.com/diegosouzapw/OmniRoute/issues/6407)). Regression guard: `tests/unit/chat-non-string-model-6407.test.ts`. (thanks @chirag127)
+- **fix(api):** unknown `/api/*` routes now return a JSON `404` (instead of the dashboard HTML shell) and scalar chat params (`model`/`temperature`/etc.) are validated **before** the provider lookup so malformed requests fail fast with a clear `400` ([#6424](https://github.com/diegosouzapw/OmniRoute/issues/6424), [#6412](https://github.com/diegosouzapw/OmniRoute/issues/6412)). Regression guards: `tests/unit/api/api-catchall-json-404.test.ts`, `tests/unit/chat-early-schema-validation-6412.test.ts`. (thanks @chirag127)
+- **fix(api):** `/v1/chat/completions` now rejects a non-JSON `Content-Type` with a `400` before parsing the body ([#6414](https://github.com/diegosouzapw/OmniRoute/issues/6414)). Regression guard: `tests/unit/v1-chat-completions-content-type-6414.test.ts`. (thanks @chirag127)
+- **fix(api):** the `X-OmniRoute-Compression` response header is now echoed on `/v1/chat/completions` and `/v1/completions` ([#6422](https://github.com/diegosouzapw/OmniRoute/issues/6422)). Regression guard: `tests/unit/compression-header-echo-6422.test.ts`. (thanks @chirag127)
+- **fix(api):** concurrent `GET /v1/models` requests are coalesced into a single catalog build ([#6408](https://github.com/diegosouzapw/OmniRoute/issues/6408)). Regression guard: `tests/unit/v1-models-concurrent-6408.test.ts`. (thanks @chirag127)
+- **fix(api):** `/v1/completions` now echoes the requested `body.model` in its JSON + streamed responses ([#6429](https://github.com/diegosouzapw/OmniRoute/pull/6429)). Regression guard: `tests/unit/completions-body-model-echo.test.ts`. (thanks @chirag127)
+- **fix(api):** env-var master keys now see the full `/v1/models` catalog ([#6406](https://github.com/diegosouzapw/OmniRoute/issues/6406)). Regression guard: `tests/unit/models-catalog-envkey-6406.test.ts`. (thanks @chirag127)
+- **fix(api):** non-streaming `/v1/completions` responses now echo `body.model` aligned with the `X-OmniRoute-Model` header ([#6426](https://github.com/diegosouzapw/OmniRoute/issues/6426)). Regression guard: `tests/unit/v1-completions-model-header-match-6426.test.ts`. (thanks @chirag127)
+- **fix(api):** unknown `/v1/*` routes now return a JSON `404 not_found` instead of the Next.js dashboard HTML shell ([#6405](https://github.com/diegosouzapw/OmniRoute/issues/6405)). Regression guard: `tests/unit/api/v1-catchall-json-404.test.ts`. (thanks @chirag127)
+- **fix(api):** the per-connection provider models route now degrades to the shipped catalog when a provider's `/models` endpoint answers with a redirect ([#6267](https://github.com/diegosouzapw/OmniRoute/issues/6267)) — a `qwen-web` import failed with a raw `Redirect blocked … (307)` 503. `safeOutboundFetch` throws `REDIRECT_BLOCKED` on the 307, `getSafeOutboundFetchErrorStatus` maps it to 503, and `buildDiscoveryErrorFallbackResponse` treated every 503 as a hard error — so the non-empty `getModelsByProviderId("qwen-web")` catalog was never surfaced. A models-endpoint redirect is not a fixable-config error (unlike `URL_GUARD_BLOCKED`/`INVALID_URL`, which stay hard errors), so it now falls back to the local/cached catalog before the 503 short-circuit. General fix — covers any config-driven provider that 307s. Regression guard: `tests/unit/provider-models-qwen-web-redirect-6267.test.ts`. (thanks @chirag127)
+- **fix(api):** the per-connection provider models route (MCP `list_models_catalog` + the dashboard import view) now merges USER-ADDED custom models into its response ([#6247](https://github.com/diegosouzapw/OmniRoute/issues/6247)) — custom models live in the `key_value` namespace `customModels`, which the live REST `/api/v1/models` already merges, but `src/app/api/providers/[id]/models/route.ts` never read `getCustomModels`, so custom models were dropped on both the discovery-success and local_catalog paths. They are now folded into the returned model list (deduped by id, stamped `owned_by: provider`), fixing MCP + the dashboard import view in one place. Regression guard: `tests/unit/provider-models-custom-merge-6247.test.ts`. (thanks @RCrushMe)
+
+- **fix(providers):** GitLab Duo tool-calling follow-up turns no longer fail upstream with `422 {"detail":"Validation error"}` (tokens 0/0, rejected pre-inference). The #6234 tool-result-feedback fix serialized the **entire** multi-turn conversation into GitLab's single-file `code_suggestions` (`small_file`) generation endpoint — folded history that turn-N sent as an oversized `current_file.content_above_cursor` **and** duplicated verbatim into `user_instruction`, tripping the AI-Gateway's `small_file` validation guard. The executor now **bounds** that prompt: it keeps system + latest user message + the most-recent tool round (dropping older turns), caps oversized tool results, and stops duplicating the full prompt into `user_instruction` (which now carries only the short latest user message) — while still feeding the most-recent tool result back so the agent continues (`open-sse/executors/gitlab.ts`, [#6220](https://github.com/diegosouzapw/OmniRoute/issues/6220)). The unit test covers the bounding logic; the upstream 422→200 clearing is VPS-only (Hard Rule #18). Regression guard: `tests/unit/gitlab-tool-exchange-bounded-6220.test.ts`.
+
+- **fix(i18n):** the provider-detail (`/dashboard/providers/[id]`) connection-status filter labels no longer render as `__MISSING__:All` / `__MISSING__:Active` / `__MISSING__:Error` / `__MISSING__:Banned` / `__MISSING__:CreditsExhausted` in non-English locales (notably pt-BR) ([#6290](https://github.com/diegosouzapw/OmniRoute/issues/6290)). Root cause was **not** the namespace mismatch the issue guessed — the `providers.filter*` keys resolve correctly in `en.json`; the debt lived in the locale mirrors (`src/i18n/messages/*.json`), where these five keys carried the `__MISSING__:` sync sentinel in ~15 locales and were absent entirely in ~26 others, so next-intl found the key and echoed the sentinel verbatim. All 40 non-English/-Chinese mirrors now ship real translations for the five `providers.filter*` labels. Regression guard: `tests/unit/i18n-provider-filter-keys-6290.test.ts`. (thanks @diegosouzapw)
+
+- **fix(providers):** the `copilot-m365-web` streaming executor now emits `debug`-level WebSocket diagnostics ([#6210](https://github.com/diegosouzapw/OmniRoute/issues/6210)) — the outbound WS URL (with the `access_token` **redacted** via `redactWsUrl()`), handshake success/failure, and each received SignalR frame's `type`/`target`. Previously the streaming path logged nothing, so an empty `content:null` response (the M365 Education / Starter tier symptom fixed in #6234) was undiagnosable even at `APP_LOG_LEVEL=debug`. The change is debug-level and side-effect-free — it does not alter streaming behavior or the frame parser, and the token never reaches the logs. Regression guard: `tests/unit/copilot-m365-web-logging-6210.test.ts` (thanks @qpeyba)
+
+- **fix(resilience):** a round-robin combo no longer returns `503 all upstream accounts are unavailable` when a compatibility-rejected target is actually healthy ([#6238](https://github.com/diegosouzapw/OmniRoute/issues/6238)). `filterTargetsByRequestCompatibility` drops request-incompatible targets (tool/vision/structured-output unsupported, or below the required context window) **before** any availability check runs, and its `compatible.length === 0` safety net only fired when _all_ targets were filtered — not when the kept targets later all turned out runtime-unavailable (circuit-open / cooldown / no credentials). So a combo could 503 while a compat-rejected-but-healthy provider sat unused. `handleRoundRobinCombo` now keeps the compat-rejected set and, when every compat-kept target was skipped without a single real attempt, probes those rejected targets as a **last-resort fallback tier** (via the new pure `open-sse/services/combo/comboCompatFallback.ts`) before crystallizing the 503. Regression guard: `tests/unit/combo-roundrobin-compat-fallback-6238.test.ts`. (thanks @ThongAccount)
+
+- **fix(startup):** best-effort self-heal for a corrupted Turbopack dev cache on Windows ([#6289](https://github.com/diegosouzapw/OmniRoute/issues/6289)). On Windows, `pnpm dev` can fail at startup when Turbopack `mmap`s a persistent-cache SST file and the OS refuses the mapping (`os error 1455` — "paging file too small"), which Turbopack surfaces as a misleading `Module not found: Can't resolve '@/shared/utils/machine'`. This is a **known upstream Turbopack cache-corruption bug — not our code**. The dev launcher (`scripts/dev/run-next.mjs`) now wraps `nextApp.prepare()` and, when it rejects with that signature (`isTurbopackCacheCorruption` in the new `scripts/dev/turbopackCacheHeal.mjs`), purges `.build/next/**/cache/turbopack` and retries **once** with a clear log. **Caveat — best-effort only:** the corruption often surfaces as a runtime overlay rather than a `prepare()` rejection, so this cannot always intercept it; the reliable remedy remains manually deleting the Turbopack cache dir. Regression guard: `tests/unit/turbopack-cache-heal-6289.test.ts`. (thanks @chirag127)
+
+- **fix(providers):** qodercli PAT auth no longer fails with `spawn qodercli ENOENT` on Windows ([#6263](https://github.com/diegosouzapw/OmniRoute/issues/6263)) — `spawnQoderCli` spawned the bare `qodercli` name with `shell:false` and an unenriched env, so the npm `.cmd` wrapper under `%APPDATA%\npm` (a user-PATH directory) was never resolved. It now resolves the absolute `.cmd`/`.exe` path through the existing `getCliRuntimeStatus("qoder")` resolver in `src/shared/services/cliRuntime.ts` (memoized), spawns with `shell` when the target is a `.cmd`/`.bat`, and uses the cliRuntime-enriched env (PATH + PATHEXT + APPDATA); the ENOENT error now lists the searched paths plus the `CLI_QODER_BIN` override. End-to-end spawn on a real Windows host is host-only (Hard Rule #18); the path-resolution logic is unit-tested. Regression guard: `tests/unit/qodercli-windows-resolve-6263.test.ts`. (thanks @chirag127)
+
+- **fix(sse):** the reasoning-token buffer no longer inflates **probe-sized `max_tokens`** ([#6274](https://github.com/diegosouzapw/OmniRoute/issues/6274)) — Claude Code's `/model` capability check sends `max_tokens: 1`, but for a thinking-capable model with a large output cap (e.g. `glm-5.2`) the #3587 headroom heuristic (`max(current + 1000, ceil(current * 1.5))`) rewrote it to `1001` and forwarded that upstream, wasting tokens on a request that was never a genuine reasoning budget. `resolveReasoningBufferedMaxTokens()` (`open-sse/services/reasoningTokenBuffer.ts`) now short-circuits and returns the caller's value verbatim when it is below the new `REASONING_BUFFER_MIN_TRIGGER` (256) threshold — a tiny explicit limit is a probe, not a reasoning request. Real budgets still receive the #3587 headroom unchanged, and the guard runs after the existing capability checks so unknown / non-reasoning models keep returning `null`. Regression guard: `tests/unit/reasoning-token-buffer-6274.test.ts`. (thanks @brightfiscalband)
+
+- **fix(cli):** `omniroute reset-password` now works as a real subcommand, and password resets over piped (non-TTY) stdin actually apply ([#6261](https://github.com/diegosouzapw/OmniRoute/issues/6261), [#6258](https://github.com/diegosouzapw/OmniRoute/issues/6258)). Two coupled defects: (1) **#6261** — `bin/omniroute.mjs` routed everything through Commander with only two pre-Commander bypasses (`--mcp`, `reset-encrypted-columns`), so `omniroute reset-password` was rejected as an unknown command; only the separate `omniroute-reset-password` bin worked, while the docs falsely advertised the subcommand (incl. a bogus "legacy alias still works"). A pre-Commander bypass mirroring `reset-encrypted-columns` now dynamically imports `bin/reset-password.mjs` (which self-executes) before Commander parses; the three doc lines were corrected. (2) **#6258** — `bin/reset-password.mjs` issued two sequential `rl.question` prompts; under piped stdin the second read never settled at EOF, so `main()` never reached `resetManagementPassword` and the reset was a silent no-op (both prompts printed, no success, password unchanged). The CLI now detects non-TTY stdin and reads it once (first line = password, second line = confirm if present, else reused), adds a `--password-stdin` flag (entire stdin is the password, no confirmation), and exits `0` explicitly so the success line always flushes; interactive TTY behavior is unchanged. Regression guard: `tests/unit/reset-password-cli-6261-6258.test.ts` (3). (thanks @chirag127)
+
+- **fix(db):** the mass-migration **safety abort** now tells the operator how to bypass it and stops flooding the log ([#6260](https://github.com/diegosouzapw/OmniRoute/issues/6260)) — after restoring a backup that wiped the migration tracking table, `runMigrations()` threw the abort on every downstream `ensureDbInitialized()`, re-logging the full banner 11+ times, and the message never mentioned the existing `OMNIROUTE_MAX_PENDING_MIGRATIONS` escape hatch. The abort text now appends a bypass hint (set `OMNIROUTE_MAX_PENDING_MIGRATIONS=0` in `server.env` / `DATA_DIR/.env`), and a new `MigrationSafetyAbortError` is memoized so repeated calls in the same process throw the same instance and emit a single concise line instead of the full cascade. Regression guard: `tests/unit/migration-safety-abort-6260.test.ts`. (thanks @chirag127)
+
+- **fix(auth):** importing a **distinct** Codex/ChatGPT OAuth `auth.json` is no longer falsely rejected as "already exists" when it belongs to a different user in the same workspace ([#6301](https://github.com/diegosouzapw/OmniRoute/issues/6301)). `findExistingCodexConnection` (in `src/lib/oauth/utils/codexAuthImport.ts`) deduped **only** on `providerSpecificData.workspaceId === accountId`, where `accountId` is the shared `chatgpt_account_id`/`tokens.account_id` — so two members of the same ChatGPT Team collapsed onto a single connection (409 `duplicate_account`). The id_token's `https://api.openai.com/auth` claim carries a per-user `chatgpt_user_id` alongside the workspace id (the device-flow path already persisted it as `chatgptUserId`, but the import path did not). Now `parseAndValidateCodexAuth` extracts `userId` (`chatgpt_user_id` → `user_id` → JWT `sub`) into `ParsedCodexAuth`, the create/update paths persist `chatgptUserId` in `providerSpecificData` (mirroring `codex.ts`), and dedup keys on `workspaceId` **AND** `chatgptUserId` — with a backward-compat fallback to legacy accountId-only matching when no stored connection for that workspace records a `chatgptUserId`, so genuinely-same accounts still dedup. Regression guard: `tests/unit/codex-auth-import-userid-dedup-6301.test.ts` (4). (thanks @anungma)
+
+- **fix(providers):** importing models for the **venice-web** provider no longer fails with a red "Provider venice-web does not support models listing" ([#6269](https://github.com/diegosouzapw/OmniRoute/issues/6269)). `venice-web` is a web-cookie provider with an executor but no upstream `/v1/models` endpoint and no registry `models`, so the models route fell through to the tail `400`. Mirroring the `jules`/`linkup-search`/`ollama-search` fix (#5569), it now ships a static local catalog entry in `src/lib/providers/staticModels.ts` — seeding the current Venice lineup (`venice-uncensored`, `llama-3.3-70b`, `qwen3-235b`, `qwen3-4b`, `deepseek-r1-671b`; Venice rotates its catalog, see docs.venice.ai/models/overview) — so the route returns `200` with `source:"local_catalog"`, `intentional:true`. Regression guard: `tests/unit/static-models-venice-web-6269.test.ts`. (thanks @chirag127)
+
+- **fix(api):** the specialty model catalogs (`/v1/embeddings`, `/v1/images`, `/v1/music`, `/v1/videos` model lists) are now derived from the **unified catalog filtered by a predicate** (`getSpecialtyModelsResponse`) instead of ad-hoc per-route logic, so they consistently respect active-credential visibility and stay in sync with the main catalog ([#6303](https://github.com/diegosouzapw/OmniRoute/pull/6303)). Regression guard: `tests/unit/specialty-model-catalog-routes.test.ts`. (thanks @makcimbx)
+- **fix(api):** the agent-bridge server route now resolves the MITM manager via a **dynamic `import("@/mitm/manager.runtime")`** so Turbopack does not statically pull the stub (or over-bundle the manager), and the agent-skills generator anchors its output base path with `path.join(process.cwd(), …)` so Turbopack's static analyzer stops tracing the whole project root ([#6329](https://github.com/diegosouzapw/OmniRoute/issues/6329), [#6366](https://github.com/diegosouzapw/OmniRoute/pull/6366)). Regression guard: `tests/unit/agent-bridge-server-route-dynamic-import.test.ts`. (thanks @Iammilansoni)
+- **fix(api):** internal probes (combo-test, cloud-sync verify) now pick a **management-scoped / allow-all API key** instead of naively grabbing `getApiKeys()[0]` — a restricted `self:usage` first row made the probe fail with "Model X is not allowed for this API key" even when the combo path was healthy (`pickApiKeyForInternalUse` in `src/lib/db/apiKeys.ts`). The API-manager model editor also falls back to `/api/models?all=true` when `/v1/models` is catalog-protected ([#6372](https://github.com/diegosouzapw/OmniRoute/pull/6372)). Regression guard: `tests/unit/pick-internal-api-key-6372.test.ts`. (thanks @jmengit)
+- **fix(live-ws):** the Live Dashboard WebSocket server now **rejects on bind failure** (e.g. `EADDRINUSE` when the API bridge already holds the port) instead of letting the error surface as an unhandled `error` event that crash-loops the process — the `error` listener is attached to `wss` (not `server`) and releases the EventBus subscription on a failed start ([#6324](https://github.com/diegosouzapw/OmniRoute/issues/6324)). Regression guard: `tests/unit/live-ws-eaddrinuse-6324.test.ts`. (thanks @vinayakkulkarni)
+- **fix(dashboard):** the Home provider-topology widget now trusts the live provider-metrics snapshot — it uses `topology.errorProvider` and live `activeRequests` directly instead of re-deriving state from a stale `lastErrorAt` or applying a frontend timeout filter, so the topology reflects real-time provider health ([#6322](https://github.com/diegosouzapw/OmniRoute/pull/6322)). Regression guard: `tests/unit/home-provider-topology-live-state.test.ts`. (thanks @xz-dev)
+- **fix(sse):** strip zero-width markers from streamed **tool-call arguments** — a follow-up to [#5857](https://github.com/diegosouzapw/OmniRoute/pull/5857). That PR removed injected zero-width joiners (U+200D) from streamed assistant text/reasoning but deliberately left tool-call argument JSON byte-exact. The request-side obfuscation (`open-sse/services/claudeCodeObfuscation.ts`) injects ZWJ into agent words — including the temp path inside the Bash tool description — and Claude models copy that verbatim into generated commands, which are delivered as tool-call arguments rather than assistant text. As a result the ZWJ survived and corrupted code blocks (e.g. a temp path rendered with an invisible joiner). Now `open-sse/handlers/responseSanitizer.ts` strips zero-width code points from tool-call argument strings at every emit site (OpenAI non-stream/stream chat `tool_calls` + legacy `function_call`, native Responses `function_call` items, the OpenAI→Responses conversion, and the native Responses streaming `response.function_call_arguments.delta/.done` events). Only zero-width code points are removed; JSON structure and all other bytes stay identical (no parse/restringify), so normal arguments remain byte-exact. Regression guard: 6 new cases in `tests/unit/response-sanitizer.test.ts` (suite 50/50).
+
+- **fix(nodejs):** the default app log path now resolves under `DATA_DIR` (`~/.omniroute/logs/application/app.log`) instead of `process.cwd()` ([#6197](https://github.com/diegosouzapw/OmniRoute/issues/6197)) — the globally-installed CLI runs from an arbitrary working directory, so anchoring the default to cwd made file logging silently write to (or no-op under) an unrelated directory, contradicting the documented `.env.example` default. `getAppLogFilePath()` now computes the default lazily via the pure `resolveDataDir()` resolver (honours a per-process `DATA_DIR`, no directory-creation side effect); an explicit `APP_LOG_FILE_PATH` still wins. Regression guard: `tests/unit/logenv-datadir-path-6197.test.ts` (3).
+
+- **fix(docker):** AgentBridge/`startMitm` no longer aborts in containers/headless when the Antigravity-default DNS step can't write `/etc/hosts` ([#6127](https://github.com/diegosouzapw/OmniRoute/issues/6127)), and the privileged command's stderr now reaches `app.log` instead of only a bare exit code hitting the toast ([#6198](https://github.com/diegosouzapw/OmniRoute/issues/6198)). The default DNS step (`addDNSEntry`) was called unguarded while cert install and the two sibling DNS steps were each best-effort — in the runtime Docker image (`USER node`, no `sudo`, read-only `/etc/hosts`) it threw `Command failed with code 1` out of `startMitmInternal` and killed the whole start, discarding the stderr. The three DNS steps are extracted into a best-effort `provisionDnsEntries()` where each failure is logged with the full `err` (stderr included, folded in by `systemCommands.ts`) and never aborts the start. Regression guard: `tests/unit/mitm-dns-graceful-degrade-6127.test.ts` (4).
+
+- **fix(providers):** copilot-m365-web now supports the M365 Education "Starter / OfficeWebIncludedCopilot" tier and no longer returns an empty `content:null` stream ([#6210](https://github.com/diegosouzapw/OmniRoute/issues/6210)). Two gaps: (1) `buildWsUrl()` hardcoded the individual-consumer scenario (`OfficeWebPaidConsumerCopilot`, `isEdu=false`) — the EDU tier is now opt-in via `providerSpecificData.tier="edu"`, emitting `scenario=OfficeWebIncludedCopilot`/`isEdu=true` (the individual path is unchanged); (2) the EDU/GPT-5.5 path streams deltas via `arguments[0].writeAtCursor` (incremental) instead of only `messages[].text` (accumulated snapshots), which the parser dropped — a new `accumulateBotContent()` folds both formats, with `type:2 item.result.message` as a last-resort fallback. Regression guard: `tests/unit/copilot-m365-edu-writeatcursor-6210.test.ts` (10). (thanks @qpeyba)
+
+- **fix(providers):** GitLab Duo executor now feeds tool results back into the prompt instead of looping ([#6220](https://github.com/diegosouzapw/OmniRoute/issues/6220)) — `buildPrompt()` branched only on `system`/`user` and took `userParts.at(-1)`, silently dropping the `assistant{tool_calls}` + `tool{result}` turns the client appended, so the reconstructed prompt was byte-identical to turn 1 and the model re-emitted the same `<tool>` call forever. When a tool exchange is present the full conversation is now serialized, folding each tool result back keyed by its `tool_call_id`; simple conversations keep the legacy shape. Complements the tool_call emission from [#6051](https://github.com/diegosouzapw/OmniRoute/issues/6051) (the `kilo-duplicate` label was a false positive — different, sequential defect). Regression guard: `tests/unit/gitlab-tool-result-feedback-6220.test.ts` (4).
+
+- **fix(providers):** opencode-go/opencode-zen can now synthesize the OpenCode CLI identity headers Cloudflare requires on VPS egress ([#5997](https://github.com/diegosouzapw/OmniRoute/issues/5997)) — on a datacenter VPS, `opencode.ai/zen/go/v1/chat/completions` 403s (HTML challenge) requests lacking CLI identity, while the reporter's control curl proved that `User-Agent: opencode-cli/1.0.0` + `x-opencode-client: cli` + `x-opencode-project: default` + fresh request/session UUIDs succeed. Opt-in via `OPENCODE_SYNTHESIZE_CLI_HEADERS=true` (values overridable via `OPENCODE_GO_USER_AGENT`/`OPENCODE_USER_AGENT`/`OPENCODE_CLIENT`/`OPENCODE_PROJECT`); it fills only headers the client did not already send. Kept **off by default** — the forward-only path is deliberate (fabricating a wrong value risks upstream rejection; a prior dedup regressed with `opencode/local`), so this replaces the fragile local header-injection shim without changing default behavior. Regression guard: `tests/unit/opencode-cli-headers-synthesis-5997.test.ts` (6). (thanks @aleksesipenko)
+- fix(resilience): sticky session affinity now evicts and fails over to another account when the pinned account is exhausted/unavailable (#6219)
+
+- fix(sse): Responses API passthrough now drops internal commentary-phase output before forwarding to clients (gated by RESPONSES_PASSTHROUGH_DROP_COMMENTARY, default on) (#6199)
+
+- **fix(sse):** tool-call function schemas with a root `type: null` are now coerced to `type: "object"` before dispatch ([#6359](https://github.com/diegosouzapw/OmniRoute/issues/6359)) — clients like the Codex app emit `parameters: { type: null, ... }` for some tools, which OpenAI-compatible upstreams reject with `400 Invalid schema for function '...': schema must be a JSON Schema of 'type: "object"', got 'type: null'`, failing the whole request. `toolSchemaSanitizer` already stripped the null; it now re-adds the mandatory root `"object"` type (and empty `properties`/open `additionalProperties` when absent). Combinator roots (`anyOf`/`oneOf`/`allOf`) and explicit root types are left untouched. Regression guard: 5 new cases in `tests/unit/tool-schema-sanitizer.test.mjs`.
+
+- **fix(docker):** AgentBridge no longer fails to start on npm/Electron/VPS installs with "MITM manager stub reached at runtime" ([#6344](https://github.com/diegosouzapw/OmniRoute/issues/6344)) — v3.8.45 flipped the production bundler default to Turbopack, but `next.config.mjs` aliased `@/mitm/manager` to its Docker-only degraded stub **unconditionally**. That was harmless while Docker (which sets the alias intentionally for #3390 graceful degradation) was the sole Turbopack consumer, but once every artifact built with Turbopack the stub shipped to all non-Docker users and `startMitm` threw on the first Agent-Bridge start. The alias is now opt-in via `OMNIROUTE_MITM_STUB=1` (set only by the Dockerfile) through the shared `scripts/build/mitm-stub-flag.mjs` helper; default builds bundle the real manager. Regression guard: `tests/unit/mitm-stub-alias-6344.test.mjs` (4).
+
+- **fix(proxy):** stop the v3.8.44 proxy regression that leaked the real IP and disabled healthy proxies ([#6246](https://github.com/diegosouzapw/OmniRoute/issues/6246)). Two coupled defects from the new health scheduler: (1) **IP leak** — when a proxy assigned to a connection was marked `inactive`, resolution fell through to a **direct** egress instead of blocking, exposing the operator's real IP; (2) **over-deactivation** — the sweep flipped a proxy to `inactive` on the **first** failed probe and counted our own 5s timeout / a probe-target `5xx` as the proxy's fault, so healthy paid proxies vanished from egress selection ("my proxies are not being used anymore"). Fix: the sweep decision is extracted into a pure, network-free `decideProxyHealthAction` (`src/lib/proxyHealth/decision.ts`) — by default the health check now **only counts/logs and never downgrades status** (a proxy is downgraded/removed only with `PROXY_AUTO_REMOVE=true`, after `PROXY_AUTO_REMOVE_AFTER` **consecutive** conclusive failures); probes are classified tri-state so an inconclusive result (our timeout, or a `5xx` from the probe target) never penalizes the proxy, and the probe timeout is raised 5s→15s. Separately, `safeResolveProxy` now **fails closed** via the existing policy: a connection whose assigned proxy is dead is blocked instead of leaking direct (`hasBlockingProxyAssignment`), honoring the explicit `proxy off` toggles and the `PROXY_FAIL_OPEN=true` opt-out. Existing proxies stuck `inactive` by the old behavior need a one-time manual re-activate (the operator owns proxy status). Regression guards: `tests/unit/proxy-health-decide-action-6246.test.ts`, `tests/unit/proxy-assigned-unavailable-6246.test.ts`.
+
+- **fix(proxy):** make "Test All" read-only and add bulk enable/disable ([#6246](https://github.com/diegosouzapw/OmniRoute/issues/6246)). Complements the core fail-closed / scheduler fix (#6296) with the two remaining reporter asks. (1) The **"Test All" button** (`POST /api/settings/proxies/auto-test`) used to flip a proxy to `inactive` on a failed reachability probe; since the egress selector excludes `inactive` proxies, a flaky probe (an unreachable `httpbin.org`, a proxy that blocks `HEAD`, or a slow paid proxy) silently disabled every proxy that failed — "Test All" is now **read-only by default** (only the operator sets a proxy active/inactive; opt back into the legacy test-and-set with `PROXY_HEALTH_AUTO_DEACTIVATE=true`). (2) Adds a **bulk enable/disable** proxies endpoint + toolbar action (`POST /api/settings/proxies/batch-activate`) so an operator can re-activate proxies in one click. Regression guard: `tests/unit/proxy-health-6246.test.ts`. (thanks @tenshiak)
+
+- **chatcore (tools): stop the default 128-tool cap from silently dropping opencode's `task`/MCP tools.** opencode (used as an MCP/agent host) sends a large tool list; when it exceeds the speculative `MAX_TOOLS_LIMIT` (128) default, `truncateToolList` did a blind `tools.slice(0, 128)`, dropping every tool past index 128 — including opencode's built-in `task` tool (subagent launch) and many MCP tools, so models routed through OmniRoute could no longer spawn subagents or reach part of their tools. The cap exists to avoid upstream `400`s for providers with real hard limits (e.g. grok-cli 200), so it is kept for those: detection of the opencode client (`isOpencodeClient` — any `x-opencode-*` header, or `opencode` in the user-agent) now only bypasses the **speculative 128 default**, never a known provider ceiling. Precedence is explicit — a proactive/detected provider limit always truncates (even for opencode); otherwise opencode forwards its full tool list; otherwise the unchanged 128 default applies to every other client. Refactors `getEffectiveToolLimit` into `getKnownToolLimit(provider) ?? DEFAULT_LIMIT` (byte-identical for existing callers) and fixes a cosmetic debug-log that reported the truncated count instead of the original. Regression guard: `tests/unit/tool-limit-detector.test.ts`.
+
+- **fix(mitm):** the macOS MITM-cert install check now matches the system keychain again. `security find-certificate -a -Z` prints the SHA-1 as a colon-less hex string, but the installed-check compared it against `getCertFingerprint()`'s colon-separated form, so the substring match never hit — the cert was reported as not-installed and re-prompted for the sudo install on every run. Fingerprints are now normalized (colons stripped, upper-cased) on both sides via the extracted `macCertOutputHasFingerprint` helper. Regression guard: `tests/unit/mitm-cert-mac-fingerprint.test.ts`. ([#6204](https://github.com/diegosouzapw/OmniRoute/pull/6204), closes [#6134](https://github.com/diegosouzapw/OmniRoute/issues/6134) — thanks @rianonehub)
+
+- **fix(api):** `/v1/messages/count_tokens` now counts `tool_use`, `tool_result` and `thinking` content blocks (and array-form `system` prompts) in the local-estimation path, instead of only `text`. Real agentic conversations keep ~95% of their tokens inside tool results; the previous estimate returned near-zero for them, which silently broke Claude Code's auto-compaction (context grew past the window with no compaction until the upstream API rejected the request). The real provider-side count path is unchanged. Regression guard: `tests/unit/messages-count-tokens-route.test.ts`. ([#6221](https://github.com/diegosouzapw/OmniRoute/pull/6221) — thanks @luweiCN)
+
+- **fix(antigravity):** strip a trailing assistant prefill turn for Vertex Claude models to avoid upstream 400s ([#6114](https://github.com/diegosouzapw/OmniRoute/pull/6114)). Regression guard: `tests/unit/antigravity-claude-prefill-strip.test.ts`. (thanks @anki1kr)
+
+- **fix(security):** the mutable cloud-agent routes (`/api/cloud/credentials/update`, `/api/cloud/models/alias`) now require management auth instead of being treated as public. They were classified as public API routes, so a request without management credentials could update stored cloud-agent credentials and model aliases. They are removed from the public-route set, classified as management routes in the authz pipeline, and gated by `requireManagementAuth`; cloud **read**/auth routes stay public. Regression guards: `tests/unit/cloud-write-auth.test.ts`, `tests/unit/authz/classify.test.ts`, `tests/unit/public-api-routes.test.ts`. ([#6233](https://github.com/diegosouzapw/OmniRoute/pull/6233) — thanks @vittoroliveira-dev)
+
+- **refactor(dashboard):** extract the onboarding-wizard "Open provider details" link target into a pure, unit-tested `buildProviderDetailsHref(connection)` helper. The wizard already routes by `connection.id` (the node UUID) rather than the provider category slug (#6144/#6145); this hardens that behavior behind a tested helper that guards a missing id/connection. Regression guard: `tests/unit/provider-onboarding-href.test.ts`. ([#6166](https://github.com/diegosouzapw/OmniRoute/pull/6166) — thanks @KooshaPari)
+
+- **fix(security):** the doubao synthetic device-id generator now derives its digits via an unbiased crypto-random draw (rejection sampling over `crypto.randomBytes()`) instead of a `% 10` reduction, closing a CodeQL `js/biased-cryptographic-random` finding.
+- **fix(agentSkills):** the GitHub-skills generator now resolves `outputDir` to an absolute path before writing, fixing a regression introduced by #6366 (relative-to-cwd base path) that could write generated skill files to the wrong directory.
+- **fix(security):** `/api/keys/{id}/devices` now checks the HTTP method before auth/validation, returning a `405` for non-GET/DELETE verbs instead of a misleading `401`/`500` (closes a `dast-smoke` QUERY-method finding).
+- **fix(quality):** clear the last 2 heavy quality-gate reds on the release tip (cycle pre-flight).
+- **fix(mitm):** the test suite and CI can never mutate the OS trust store — `OMNIROUTE_SKIP_SYSTEM_TRUST=1` is set globally for tests/CI so `installCert`/`uninstallCert`/`installTproxyCa` skip the privileged OS dispatch ([#6310](https://github.com/diegosouzapw/OmniRoute/pull/6310); full detail is under the [3.8.45] section below — this branch received it via the parallel-cycle sync-back).
+- **fix(api):** `POST /api/github-skills` now Zod-validates its request body; documented the new quality-gate env vars and pinned the merge-integrity GitHub Actions to a commit SHA.
+- **fix(skills):** generate the missing `omni-github-skills` registry entry and align the agent-skills catalog-count tests (follow-up to #6186).
+- **fix(quality):** clear the cycle's 11 net-new ESLint errors and make `validate-release-green` suppressions-aware.
+
+### 📝 Maintenance
+
+- **i18n(it):** add 118 missing Italian (`it`) translations (net-additive — no existing keys dropped, valid JSON), improving Italian UI coverage. ([#6212](https://github.com/diegosouzapw/OmniRoute/pull/6212) — thanks @serverless83)
+- **chore(providers):** remove deprecated **MiMo V2** model entries from the catalogs (xiaomi-mimo, opencode-go, zenmux-free, audio TTS) — the upstream V2 line is superseded by MiMo V2.5; drops `mimo-v2-tts`, `mimo-v2-pro`, `mimo-v2-omni`, `mimo-v2-flash`, `mimo-v2-flash-free` and realigns the provider-catalog tests. ([#6248](https://github.com/diegosouzapw/OmniRoute/pull/6248) — thanks @backryun)
+- **chore(release):** ~50 commits on this branch are v3.8.45 pre-flight/hardening fixes and CI-perf work that landed here via the parallel-cycle sync-back (`sync-next-cycle.mjs`, Hard Rule #21) **after** the `v3.8.45` git tag was cut, and are already fully documented under the **[3.8.45]** section below — listed here only so the per-cycle commit-coverage check (`npm run release:uncovered`) doesn't flag them as gaps. Provider/catalog/UX/backend fixes: #6041, #6078, #6108, #6135, #6148, #6149, #6154, #6158, #6161, #6162, #6163, #6164, #6165, #6170, #6177, #6178, #6181, #6186, #6187, #6191, #6193, #6194, #6195, #6200, #6205, #6208, #6209, #6211, #6213, #6223, #6224, #6225, #6226, #6227, #6228, #6229, #6230, #6235, #6291, #6292. CI/release-pipeline work: #6167, #6203, #6214, #6215, #6218, #6273, #6275, #6283, #6284, #6285, #6300, #6305.
+- **chore(release):** additional zero-ref release-cycle plumbing on this branch, kept out of `release:uncovered` on purpose (no `#N` in the commit subject to cite): opening the v3.8.46 cycle, opening/closing the v3.8.45 cycle, the finalized [3.8.45] CHANGELOG i18n sync-back to 42 mirrors, the v3.8.45 cognitive/cyclomatic and file-size drift rebaselines, ESLint stale-suppression pruning (4,273 → 4,233), and clearing test-masking/docs-all pre-flight reds for v3.8.45.
+
+### ⚡ Performance & Infrastructure
+
+- **perf(release-green):** the pre-flight validator (`scripts/quality/validate-release-green.mjs`) now runs its 4 slow suites (unit / vitest / integration / pack-artifact) **concurrently** via `Promise.all` — pre-flight wall time drops from ~the sum of the suites to ~the slowest one (~30min saved per round; Phase 0 was the nº1 bottleneck of the v3.8.45 release benchmark, 2h54 of 6h34 e2e). Guard: `tests/unit/validate-release-green.test.ts` ("runs the slow suites CONCURRENTLY"). ([#6319](https://github.com/diegosouzapw/OmniRoute/pull/6319))
+- **fix(ci):** `scripts/release/sync-next-cycle.mjs` — two defects found live in its first production run (v3.8.45 Phase 5): (1) the `git()` helper's default 1 MiB `maxBuffer` crashed with `ENOBUFS` on `git show origin/main:CHANGELOG.md` (the CHANGELOG alone is >1 MiB) — widened to 64 MiB; (2) the i18n resync only propagated the `[NEXT]` (TBD) section, leaving the just-shipped finalized section as "— TBD" in all 42 mirrors — it now also syncs `[prevVersion]` bounded by the heading below it (new exported pure helper `versionAfter`). Guards: +5 tests in `tests/unit/sync-next-cycle.test.ts` (8/8). ([#6327](https://github.com/diegosouzapw/OmniRoute/pull/6327))
+- **test(ci):** concurrency-sensitive flaky tests are **quarantined into a serial pass** (`tests/unit/serial/`, `--test-concurrency=1`, appended to every unit runner incl. sharded variants — the serial pass is sharded too so concurrent shard jobs never self-collide). Initial set: `glm-coding-plan-monthly-3580`, `quota-division-blocks`, `provider-health-autopilot`, `combo-health-autopilot` — the class behind the ~28min CI wedges/re-runs (two live 1h+ wedges cancelled during this PR's own validation). Discovery + TIA gates track the new glob; systemic root cause (async logger writing after teardown) tracked in [#6360](https://github.com/diegosouzapw/OmniRoute/issues/6360). Guard: `tests/unit/test-serial-quarantine.test.ts` (4). ([#6347](https://github.com/diegosouzapw/OmniRoute/pull/6347))
+
+### 🙌 Contributors
+
+Thanks to everyone whose work landed in v3.8.46:
+
+| Contributor                                                    | PRs / Issues           |
+| -------------------------------------------------------------- | ---------------------- |
+| [@2220258345](https://github.com/2220258345)                   | direct commit / report |
+| [@abdofallah](https://github.com/abdofallah)                   | direct commit / report |
+| [@adentdk](https://github.com/adentdk)                         | direct commit / report |
+| [@aleksesipenko](https://github.com/aleksesipenko)             | direct commit / report |
+| [@anki1kr](https://github.com/anki1kr)                         | direct commit / report |
+| [@anungma](https://github.com/anungma)                         | direct commit / report |
+| [@arpicato](https://github.com/arpicato)                       | direct commit / report |
+| [@backryun](https://github.com/backryun)                       | #6248                  |
+| [@binsarjr](https://github.com/binsarjr)                       | direct commit / report |
+| [@brightfiscalband](https://github.com/brightfiscalband)       | direct commit / report |
+| [@chirag127](https://github.com/chirag127)                     | #6501, #6506           |
+| [@developerjillur](https://github.com/developerjillur)         | direct commit / report |
+| [@dilneiss](https://github.com/dilneiss)                       | #6499                  |
+| [@dtybnrj](https://github.com/dtybnrj)                         | direct commit / report |
+| [@Forcerecon](https://github.com/Forcerecon)                   | direct commit / report |
+| [@hao3039032](https://github.com/hao3039032)                   | direct commit / report |
+| [@Iammilansoni](https://github.com/Iammilansoni)               | #6150, #6245           |
+| [@jmengit](https://github.com/jmengit)                         | direct commit / report |
+| [@jordansilly77-stack](https://github.com/jordansilly77-stack) | direct commit / report |
+| [@JxnLexn](https://github.com/JxnLexn)                         | direct commit / report |
+| [@KooshaPari](https://github.com/KooshaPari)                   | #6166                  |
+| [@loopyd](https://github.com/loopyd)                           | direct commit / report |
+| [@luweiCN](https://github.com/luweiCN)                         | #6221                  |
+| [@makcimbx](https://github.com/makcimbx)                       | direct commit / report |
+| [@muflifadla38](https://github.com/muflifadla38)               | direct commit / report |
+| [@newnol](https://github.com/newnol)                           | direct commit / report |
+| [@ofekbetzalel](https://github.com/ofekbetzalel)               | direct commit / report |
+| [@ohahe52-dot](https://github.com/ohahe52-dot)                 | direct commit / report |
+| [@phidinhmanh](https://github.com/phidinhmanh)                 | direct commit / report |
+| [@powellnorma](https://github.com/powellnorma)                 | direct commit / report |
+| [@qpeyba](https://github.com/qpeyba)                           | direct commit / report |
+| [@RaviTharuma](https://github.com/RaviTharuma)                 | direct commit / report |
+| [@RCrushMe](https://github.com/RCrushMe)                       | direct commit / report |
+| [@rianonehub](https://github.com/rianonehub)                   | #6134, #6204           |
+| [@serverless83](https://github.com/serverless83)               | #6212                  |
+| [@swingtempo](https://github.com/swingtempo)                   | direct commit / report |
+| [@tenshiak](https://github.com/tenshiak)                       | direct commit / report |
+| [@ThongAccount](https://github.com/ThongAccount)               | direct commit / report |
+| [@UnrealAryan](https://github.com/UnrealAryan)                 | direct commit / report |
+| [@vinayakkulkarni](https://github.com/vinayakkulkarni)         | direct commit / report |
+| [@vittoroliveira-dev](https://github.com/vittoroliveira-dev)   | #6233                  |
+| [@warelik](https://github.com/warelik)                         | direct commit / report |
+| [@xxy9468615](https://github.com/xxy9468615)                   | direct commit / report |
+| [@xz-dev](https://github.com/xz-dev)                           | direct commit / report |
+| [@yanpaing007](https://github.com/yanpaing007)                 | direct commit / report |
+| [@diegosouzapw](https://github.com/diegosouzapw)               | maintainer             |
+
+---
+
 ## [3.8.45] — 2026-07-06
 
 ### ✨ New Features

docs/i18n/ur/CHANGELOG.md

@@ -6,6 +6,213 @@
 
 ## [3.8.31] — 2026-06-20
 
+## [3.8.46] — TBD
+
+### ✨ New Features
+
+- **feat(sse):** **hide paid-only models from `auto/*` routing** when `hidePaidModels` is on ([#6512](https://github.com/diegosouzapw/OmniRoute/issues/6512)) — follow-up to #6328/#6495. PR #6495 hid paid-only models from the `GET /v1/models` listing, but `auto/*` combos (`auto/best-coding`, `auto/glm`, …) could still pick a paid-only backend into their candidate pool → a 402/403 at request time. `createVirtualAutoCombo` now filters the candidate pool through the new pure `open-sse/services/autoCombo/paidModelFilter.ts` (`filterPaidOnlyCandidates`), applying the same free-model predicate #6495 uses in `catalog.ts` (`providerHasFreeModels(provider) && isFreeModel(provider, {id})`) whenever `settings.hidePaidModels === true`. Applied before the category/tier/family narrowing, so it covers every `auto/*` combo; an all-paid pool degrades to the existing graceful empty-pool path. **Opt-in — default OFF leaves the pool unchanged** (identity). Regression guard: `tests/unit/autoCombo/paid-model-filter-6512.test.ts` (4, incl. the default-off identity guard).
+- **feat(sse):** **provider-family auto combos** — `auto/glm`, `auto/minimax`, `auto/mimo`, `auto/zai`, `auto/gemma`, `auto/llama`, `auto/gemini` ([#6453](https://github.com/diegosouzapw/OmniRoute/issues/6453)) — new routable ids that materialize an on-demand virtual combo spanning whatever installed backends currently expose that model family, degrading gracefully as backends rotate. A new pure `open-sse/services/autoCombo/modelFamily.ts` (`detectModelFamily`) classifies by model-id prefix for six families; `zai` is instead resolved by provider id (z.ai's hosted API serves the same `glm-*` model ids as every other GLM backend, so `auto/zai` means "route to my z.ai backend specifically" vs `auto/glm`'s "any connected GLM backend"). Reuses the existing `createVirtualAutoCombo` on-demand materialization path (no DB writes) and the `/v1/models` catalog advertising loop. Regression guard: `tests/unit/autoCombo/provider-family-combos.test.ts` (11).
+- **feat(proxy):** native **proxy-pool round-robin / egress IP rotation** ([#6365](https://github.com/diegosouzapw/OmniRoute/issues/6365)) — a scope (global / provider / account) can now hold **multiple** proxies as a pool with a rotation strategy, so outbound requests cycle their egress IP instead of pinning one proxy per scope. Migration `117_proxy_pool_rotation.sql` lifts the `UNIQUE(scope, scope_id)` constraint (rebuild via the canonical rename/copy/drop; existing single assignments become 1-element pools) and adds a `proxy_scope_rotation` companion table holding the per-scope strategy + a persisted monotonic round-robin cursor. Strategies: `round-robin` (default, monotonic cursor — never `Math.random`), `random`, and `sticky-per-N-min`. Resolution (`resolveProxyForScopeFromRegistry` / `resolveProxyForConnectionFromRegistry`) now fetches the alive, position-ordered candidate set (unchanged `PROXY_ALIVE_PREDICATE`) and applies the strategy; an empty / all-dead pool still returns `null` — the #6246 fail-closed guard is **untouched** (never falls through to direct egress). Backend + DB only; dashboard pool-builder UI is a follow-up. Regression guard: `tests/unit/proxy-pool-rotation-6365.test.ts` (8, incl. fail-closed + backward-compat).
+- **feat(providers):** end-to-end **tool/function calling on the native Gemini `/v1beta` endpoint** ([#6222](https://github.com/diegosouzapw/OmniRoute/issues/6222)) — both directions of the Gemini↔OpenAI conversion now preserve tool data (previously silently dropped). Request side: `convertGeminiToInternal` (extracted to its own testable module) maps `tools[].functionDeclarations` → OpenAI `tools`, prior `functionCall` parts → assistant `tool_calls`, and `functionResponse` parts → `tool`-role messages. Response side: `convertOpenAIResponseToGemini` emits `parts[].functionCall {name,args}` from `message.tool_calls`, and the streaming `openAIChunkToGeminiChunk` accumulates fragmented `tool_calls` deltas by index into complete `functionCall` parts. The non-Gemini client paths (Claude, OpenAI-Responses) already preserved tool calls — this closes the gap specific to the native Gemini surface. Regression guard: `tests/unit/v1beta-gemini-tool-calling-6222.test.ts` (6, incl. a streaming SSE round-trip).
+- **feat(providers):** copilot-m365-web **enterprise / work tier** support ([#6334](https://github.com/diegosouzapw/OmniRoute/issues/6334)) — mirrors the EDU-tier pattern (#6210): `M365ConnectionParams` gains an `agent` field, a new opt-in `M365_ENTERPRISE_OVERRIDES` preset (`agent=work`, `scenario=officeweb`, `licenseType=Premium`) applies via `providerSpecificData.tier="enterprise"` (alias `"work"`), and `agent` is also overridable directly via `providerSpecificData.agent`. `buildWsUrl` was hardcoding `agent="web"` (the one enterprise-distinguishing param with no override path), so a Premium work account handshook then returned an empty stream. The individual and EDU paths are untouched. Kilo's dup flag vs #6210 (EDU tier) was a false positive — different tier. Regression guard: `tests/unit/copilot-m365-enterprise-6334.test.ts` (7). End-to-end confirmation on a real Premium work account is a live-VPS validation follow-up (Hard Rule #18). (thanks @Forcerecon)
+- **feat(api):** standardized, provider-agnostic **`effort` + `thinking` request params** ([#6241](https://github.com/diegosouzapw/OmniRoute/issues/6241)) — a thin standardization layer over the existing mature per-provider reasoning plumbing (no provider mapper touched). `providerChatCompletionSchema` gains a canonical `effort` (reusing the shared `none/low/medium/high/xhigh` vocabulary — the UI tiers `extra`/`max` collapse onto `xhigh`) and a boolean `thinking`. A pure `normalizeReasoningRequest` (wired once in `src/sse/handlers/chat.ts`, before any reasoning field is read) folds them onto the fields the translators already consume (`reasoning_effort` / `reasoning.effort` / `thinking`), so they fan out to Anthropic / Gemini / xAI / Responses — an explicit client `reasoning_effort` / object-shaped `thinking` always wins (backward-compatible). `/models` additively exposes `supportsThinking` + `effort_tiers` so the frontend can render the toggles (UI component is a follow-up). Regression guard: `tests/unit/effort-thinking-standardization-6241.test.ts` (12). (thanks @Iammilansoni, @shabeer)
+- **feat(combo):** new **`pipeline` (sequential) combo strategy** ([#6297](https://github.com/diegosouzapw/OmniRoute/issues/6297)) — the 18th routing strategy runs targets **in order**, threading each step's output into the next step's input, with an optional per-step `prompt` (system instruction); only the final step's response is returned. Distinct from `fusion` (parallel fan-out + judge). Implemented as a self-contained `open-sse/services/pipeline.ts` (sibling to `fusion.ts`), dispatched from `combo.ts`; the step list reuses `combo.models` order and reads an optional `prompt` off each target (backward-compatible — ignored by every other strategy). Intermediate steps run non-streaming with tools stripped (complete prose to thread forward); the final step keeps the client's `stream` flag + tools. A failing/empty/unparseable intermediate step fails the whole pipeline explicitly via a sanitized error (never silently swallowed). Kilo's dup flag vs #563 was a false positive (that's model→chain selection; this is a sequential chain). Regression guard: `tests/unit/combo-pipeline-strategy.test.ts` (5). (thanks @ofekbetzalel)
+- **feat(ci):** `check:test-masking` now flags **inline-reimplemented prod conditions** ([#6348](https://github.com/diegosouzapw/OmniRoute/issues/6348)) — a new **report-only** subcheck (v2, 6A.10 family) catches the wrong-shape contract test: a test that recomputes the condition under test inline instead of importing/exercising the real function (the #6216 class, where `=== 500` → `>= 500` stayed green because the test re-implemented the branch). For each added/modified test file it warns when the file textually duplicates a ≥3-token conditional from a production file touched in the same PR **and** does not import the symbol/module owning it, via a pure, fixture-tested `findReimplementedConditions()` with an allowlist mirroring `assertReductionAllowlist`. Report-only for now (does not fail the gate) — to be promoted to blocking after a triage cycle. Regression guard: `tests/unit/check-test-masking.test.ts` (45).
+- **feat(sse):** per-connection routing override (native vs CLIProxyAPI) ([#6339](https://github.com/diegosouzapw/OmniRoute/issues/6339)) — the previously-dead `isCliproxyapiDeepModeEnabled` helper is now wired into `resolveExecutorWithProxy`: a single connection can opt itself into the CLIProxyAPI passthrough executor via `providerSpecificData.cliproxyapiMode="claude-native"`, with precedence **connection override > provider `upstream_proxy_config` mode > default**. `resolveExecutorWithProxy` now receives the resolved connection's `providerSpecificData` (threaded from `chatCore.ts`), so one connection can deep-route while the provider's other connections stay native — no DB schema change (the toggle rides in `providerSpecificData`). Also resolves the same-provider mixing ask in #6340. Regression guard: `tests/unit/chatcore-executor-proxy.test.ts` (9). (thanks @RaviTharuma)
+- **feat(dashboard):** "Add session cookie" modal now shows a prominent **"Open ‹host› →"** link to the provider's own site ([#6268](https://github.com/diegosouzapw/OmniRoute/issues/6268)) — every `-web` cookie-session provider (chatgpt-web, claude-web, gemini-web, kimi-web, lmarena, qwen-web, m365-copilot-web, …) renders a one-click external link (opening the provider's login/home page in a new tab) so operators no longer tab away to retype the URL mid-setup. The host resolves from a pure, unit-tested `resolveWebProviderHost()` (prefers `WEB_COOKIE_PROVIDERS[id].website`, falls back to the registry `baseUrl` origin); non-web providers render exactly as before. Kilo's dup flag vs #6265 (modal-too-small-on-1080p) was a false positive — distinct concern. Regression guard: `tests/unit/resolve-web-provider-host.test.ts` (5). (thanks @chirag127)
+- **feat(providers):** add **DigitalOcean** AI (serverless inference) as an OpenAI-compatible API-key provider ([#6373](https://github.com/diegosouzapw/OmniRoute/pull/6373)) — base `https://inference.do-ai.run/v1`, wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/digitalocean/`, `src/shared/constants/providers/apikey/inference-hosts.ts`). Regression guard: `tests/unit/digitalocean-provider.test.ts`. (thanks @newnol)
+- **feat(providers):** add **Huancheng Public API** (`hcnsec`) as an OpenAI-compatible regional provider ([#6410](https://github.com/diegosouzapw/OmniRoute/pull/6410)) — Xinjiang Huancheng Cybersecurity's public LLM platform (base `https://api.hcnsec.cn/v1`, free credits via daily check-ins), wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/hcnsec/`, `src/shared/constants/providers/apikey/regional.ts`). Regression guard: `tests/unit/hcnsec-provider.test.ts`. (thanks @UnrealAryan)
+- **feat(dashboard):** the web-session credential guide now shows an **"Open {host}" link** ([#6316](https://github.com/diegosouzapw/OmniRoute/pull/6316)) to the provider's sign-in site (derived from the provider `website` via `getProviderWebsiteHost`), so you can jump straight to the page where the cookie/session must be captured. Regression guard: `tests/unit/web-session-provider-link-6316.test.ts`. (thanks @jordansilly77-stack)
+- **feat(cerebras):** add the **Gemma 4 31B** model (`gemma-4-31b`) to the Cerebras registry + pricing table ([#6331](https://github.com/diegosouzapw/OmniRoute/pull/6331)). Regression guard extends `tests/unit/t28-model-catalog-updates.test.ts`. (thanks @backryun)
+- **feat(providers):** add **Yuanbao (web)** as a cookie-session provider ([#6196](https://github.com/diegosouzapw/OmniRoute/issues/6196)) — `yuanbao-web` (Tencent Yuanbao, `yuanbao.tencent.com`) with cookie-only auth (`hy_user`/`hy_token` + public agent id), SSE→OpenAI translation incl. `reasoning_content`, exposing DeepSeek V3/R1 + Hunyuan / Hunyuan-T1. Regression guard: `tests/unit/providers-yuanbao-web.test.ts`. `together-web` was **deferred** (no verifiable web-session endpoint — needs a captured request) and `huggingchat-web` **dropped** (the existing `huggingchat` already is a web-cookie provider). (thanks @chirag127)
+- **feat(providers):** route the built-in **agentrouter** through the dynamic Claude-Code wire image ([#6056](https://github.com/diegosouzapw/OmniRoute/issues/6056)) — a small static allow-set (`CC_WIRE_IMAGE_BUILTINS` in `open-sse/services/ccWireImageBuiltins.ts`), consulted by `isClaudeCodeCompatible` / `isClaudeCodeCompatibleProvider` / `applyFingerprint`, makes agentrouter adopt the CC wire-image headers + fingerprint **while guarding the CC baseUrl/auth branches** so it keeps its own registry `baseUrl` and `x-api-key` auth. Regression guard: `tests/unit/agentrouter-cc-wire-image.test.ts` (asserts the wire image is applied AND agentrouter's baseUrl/auth are preserved). Live WAF-acceptance against agentrouter.org is a VPS validation follow-up (Hard Rule #18).
+- **feat(providers):** **bulk-add API keys for Cloudflare Workers AI** ([#6174](https://github.com/diegosouzapw/OmniRoute/issues/6174)) — `cloudflare-ai` is removed from the bulk-add exclusion list and the bulk parser gains a 3-field `name|accountId|apiKey` mode; the bulk route now builds a **per-entry** `providerSpecificData` so each key carries its own `accountId` (fixing the previous shared-object reuse), and both the create + key-validation paths receive it. Regression guard: `tests/unit/bulk-api-key-parser-cloudflare.test.ts`. (thanks @muflifadla38)
+- **feat(dashboard):** routing/settings UX clarity ([#6147](https://github.com/diegosouzapw/OmniRoute/issues/6147)) — (1) weighted combos show the **effective routing share %** next to each weight when weights don't sum to 100 (`WeightTotalBar.tsx`); (2) the status widget's user-facing **"Cloud Sync" label is renamed** to "Remote Settings Sync" (`CloudSyncStatus.tsx`; internal ids/state untouched); (3) built-in providers gain an **opt-in advanced base-URL override** (`isBaseUrlOverrideEligibleProvider`, hidden behind an "Advanced" toggle, reusing the existing `providerSpecificData.baseUrl` persistence — not globally widened). Regression guard: `tests/unit/routing-settings-ux-6147.test.ts`.
+- **feat(combo):** add an option to **disable session stickiness**, per-combo or globally — round-robin / random combos can rotate to a different connection on every request instead of pinning a whole conversation to one connection by its first-message hash. Resolution precedence per-combo `config.disableSessionStickiness` → global `settings.disableSessionStickiness` → default `false` (preserves the #3825 prompt-cache/504 fix); gates **both** stickiness call sites in `open-sse/services/combo.ts`. Exposed as a global toggle (Combo Defaults) and a per-combo Inherit/on/off control. ([#6168](https://github.com/diegosouzapw/OmniRoute/issues/6168)) Regression guard: `tests/unit/combo-disable-session-stickiness.test.ts`. (thanks @RCrushMe)
+- **feat(docker):** add the `OMNIROUTE_NO_SUDO` env flag for root-less / user-namespaced deployments — the MITM cert-trust command path (`resolveSudoSpawn` in `src/mitm/systemCommands.ts`) now strips the leading `sudo` when the flag is truthy, in addition to the existing root / sudo-missing cases, so the Proxy Agent runs without `sudo` (the operator trusts the CA manually, e.g. via `NODE_EXTRA_CA_CERTS`). Argv-array `spawn` preserved — no shell interpolation (Hard Rule #13). ([#6122](https://github.com/diegosouzapw/OmniRoute/issues/6122)) Regression guard: `tests/unit/mitm-systemCommands-no-sudo.test.ts`. (thanks @powellnorma)
+- **feat(providers):** add **Requesty** as an OpenAI-compatible gateway provider (BYOK, base `https://router.requesty.ai/v1`, ~200 free requests/day) — wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/requesty/`, `src/shared/constants/providers/apikey/gateways.ts`). ([#6120](https://github.com/diegosouzapw/OmniRoute/issues/6120)) Regression guard: `tests/unit/requesty-provider.test.ts`. (thanks @chirag127)
+- **feat(dashboard):** add **configured-only / available-only filters** to the Free Provider Rankings page ([#6150](https://github.com/diegosouzapw/OmniRoute/issues/6150)) — hide providers you haven't configured, or whose connections are all rate-limited / out of quota, via server-side query params (`?configuredOnly` / `?availableOnly` on `GET /api/free-provider-rankings`) backed by a testable lib helper reusing the in-process connection state (no Redis). Both filters default off, so the default view is unchanged; this supersedes the earlier client-side "Configured Only" toggle (#6245) with an available-only dimension and unit-tested logic. Regression guard: `tests/unit/freeProviderRankings-filters.test.ts`.
+
+### 🔧 Bug Fixes
+
+- **fix(dashboard):** adding a second API key connection for the same provider no longer silently overwrites the first — the Add-API-key modal now derives a unique default connection name (`main`, then `main-2`, `main-3`, …) so the backend name-based upsert can't collide (#6499 — thanks @dilneiss).
+- **fix(compression):** the session-dedup engine now also deduplicates a large multi-line block repeated **within a single message** (intra-message dedup), not just across turns; the compression-preview API surfaces a `fallbackReason`, and the fusion panel reports how many models were rate-limited vs failed on total-panel failure (#6501 — thanks @chirag127).
+- **fix(compression):** a stacked-pipeline step naming an unregistered engine now surfaces a `validationErrors` entry instead of silently no-op'ing, so misconfigured pipelines are visible in the preview API (#6506 — thanks @chirag127).
+- **feat(usage):** add a **Codex reset-credit redemption** flow to the Provider Limits UI ([#6361](https://github.com/diegosouzapw/OmniRoute/pull/6361)) — a `useCodexResetCreditRedemption` hook + `/api/usage/codex-reset-credit` route + `codexResetCredits` lib let you redeem banked Codex reset credits from the quota card. Regression guard: `tests/unit/codex-reset-credits.test.ts`. (thanks @JxnLexn)
+- **feat(glm):** add **team-plan quota settings** for `glm-cn` connections ([#6351](https://github.com/diegosouzapw/OmniRoute/pull/6351)) — a dedicated `GlmTeamQuotaFields` form section (team quota id / limits) threaded through the Add/Edit connection modals, persisted via `providerSpecificData`, with the GLM usage service reading the team quota. Regression guards: `tests/unit/glm-team-quota.test.ts`, `provider-specific-data-schema.test.ts`. (thanks @hao3039032)
+- **feat(providers):** add **TinyFish** web-fetch/search support ([#6349](https://github.com/diegosouzapw/OmniRoute/pull/6349)) — a `tinyfish-fetch` executor + `/v1/web/fetch` route + MCP web-fetch tool, registered as a specialty-media provider with request-validation and a search-provider catalog entry. Regression guards: `tests/unit/executor-tinyfish-fetch.test.ts`, `web-fetch-handler.test.ts`, `mcp-web-fetch-tool.test.ts`, `provider-validation-tinyfish.test.ts`. (thanks @dtybnrj)
+- **fix(cli):** `omniroute launch-codex` now spawns `codex.cmd` through a shell on Windows (the npm `.cmd` shim is unresolvable by bare `spawn` → ENOENT), mirroring the qodercli Windows fix (#6263) ([#6312](https://github.com/diegosouzapw/OmniRoute/pull/6312)). Regression guard: `tests/unit/launch-codex-windows-spawn-6312.test.ts`. (thanks @swingtempo)
+- **fix(codex):** isolate the **Spark** quota from the shared Codex quota and stabilize the quota UI ordering / hydration so per-scope limits render consistently ([#6336](https://github.com/diegosouzapw/OmniRoute/pull/6336)). Regression guards: `tests/unit/codex-quota-selection-hydration.test.ts`, `provider-limits-ui.test.ts` + 3 more. (thanks @xz-dev)
+- **feat(api):** add a `hidePaidModels` setting that filters paid-only models out of the `/v1/models` catalog. Regression guard: `tests/unit/models-catalog-hide-paid.test.ts`. (thanks @chirag127)
+- **fix(api-manager):** the fallback model picker now preserves combos instead of dropping them when a primary model is unavailable ([#6443](https://github.com/diegosouzapw/OmniRoute/pull/6443)). Regression guard: `tests/unit/api-manager-page-static.test.ts`. (thanks @jmengit)
+- **fix(providers):** recoverable Antigravity / Cloud-Code (Gemini Code Assist) `403` responses ([#6452](https://github.com/diegosouzapw/OmniRoute/pull/6452)) are now classified as a retryable project-config error instead of a terminal account ban, so a fixable project/API-disabled 403 no longer forces a ~1-year cooldown / full OAuth reconnect. Regression guard: `tests/unit/errorclassifier-antigravity-403.test.ts`. (thanks @developerjillur)
+- **fix(mitm):** `sanitizeHeaders` now redacts `Set-Cookie` response headers so upstream session cookies never leak into logs / diagnostics ([#6451](https://github.com/diegosouzapw/OmniRoute/pull/6451)). Regression guard: `tests/unit/mitm-sanitize-headers.test.ts`. (thanks @developerjillur)
+- **fix(api):** `/api/compression/preview` now accepts `mode: "caveman"` and correctly handles stacked / zero-compression previews ([#6425](https://github.com/diegosouzapw/OmniRoute/issues/6425)). Regression guard: `tests/unit/api/compression-preview-caveman-and-stacked-6425.test.ts`. (thanks @chirag127)
+- **feat(providers):** add **Zed** hosted LLM aggregator as a native-app provider ([#6118](https://github.com/diegosouzapw/OmniRoute/pull/6118)) — OAuth sign-in via the Zed hosted flow, registered through the shared provider registry + executor. Regression guards: `tests/unit/zed-oauth-provider.test.ts`, `zed-import-utils.test.ts`, `zed-docker-detect.test.ts`, `mitm-handler-zed.test.ts`. VPS-validated via live operator login (Hard Rule #18).
+- **fix(oauth):** the Kiro SSO-cache auto-import now **preserves the IDC region** — cross-region Amazon Q / Kiro profiles imported from the SSO cache are no longer collapsed to the default region ([#6113](https://github.com/diegosouzapw/OmniRoute/pull/6113)). Regression guard: `tests/unit/kiro-auto-import-idc-2059.test.ts`. VPS-validated via live operator login (Hard Rule #18).
+- **fix(dashboard):** passthrough model aliases no longer collide when two namespaced model ids share a last segment (port from 9router#1850, [#6431](https://github.com/diegosouzapw/OmniRoute/pull/6431)). `enx/gpt-5.5` and `enx/codebuddy/gpt-5.5` both auto-generated the alias `gpt-5.5`, so the second model could never be added (the UI just alerted "alias already exists"). Aliases are now disambiguated deterministically — bare last segment when free, then parent-qualified (`codebuddy-gpt-5.5`), then a numeric suffix — while re-adding the exact same model id is still blocked. Regression guard: `tests/unit/passthrough-alias-1850.test.ts`. (thanks @arpicato)
+- **fix(translator):** preserve a Gemini `functionResponse` co-located with other parts (another `functionCall`, or trailing `text`) in the same content when translating **Gemini → OpenAI** ([#6376](https://github.com/diegosouzapw/OmniRoute/pull/6376)). `convertGeminiContent()` early-returned the tool message on the first `functionResponse` part, dropping any co-located parts; such contents are now pre-split (one tool message per `functionResponse`, emitted first, plus one message for the remaining parts). Regression guard: `tests/unit/gemini-to-openai-function-response.test.ts`. (thanks @warelik)
+- **fix(headroom):** detect a python interpreter managed by **mise / pyenv / asdf / conda** (port from 9router#2353, [#6382](https://github.com/diegosouzapw/OmniRoute/pull/6382)). Headroom's python probe (`src/lib/headroom/detect.ts`) searched a hardcoded `PATH`, but version managers expose their interpreters via shim dirs that only join `PATH` through interactive-shell activation — which the non-interactive server never runs, so a managed python (≥3.10) was invisible and Headroom reported it missing. The search path now prepends the well-known shim/bin dirs (`~/.local/share/mise/shims`, `~/.pyenv/shims`, `~/.asdf/shims`, `$CONDA_PREFIX/bin`, `~/.local/bin`, respecting `MISE_DATA_DIR`/`PYENV_ROOT`/`ASDF_DATA_DIR` when set), and a new `HEADROOM_PYTHON` env override lets operators point straight at their interpreter (mirroring `HEADROOM_URL`). Still shell-free (`execFileSync`). Regression guard: `tests/unit/headroom-detect.test.ts` (5). (thanks @loopyd)
+- **fix(executors):** strip the OpenAI-Codex/Claude-CLI `client_metadata` passthrough field for **NVIDIA** requests (port from 9router#1887, [#6411](https://github.com/diegosouzapw/OmniRoute/pull/6411)). NVIDIA's OpenAI-compatible wrapper rejects it with `400 Unsupported parameter`, the same class already handled for `cerebras`/`mistral`; `nvidia` (executor `default`) was missing from the strip allowlist so Codex/Claude-Code passthrough requests 400'd. Regression guard: `tests/unit/executor-default-strip-client-metadata.test.ts` (+nvidia case). (thanks @phidinhmanh)
+- **fix(translator):** strip the Claude-style `thinking` field for **NVIDIA `z-ai/glm-5.2`** (port from 9router#2023, [#6413](https://github.com/diegosouzapw/OmniRoute/pull/6413)). NVIDIA's OpenAI-compatible wrapper 400s on `thinking` (a Claude-format client routed here leaves a `thinking:{type:"adaptive"}`); the existing strip rule only dropped `reasoning`. Same class already handled for `minimax-m2.7`. Regression guard: `tests/unit/nvidia-minimax-thinking-strip.test.ts` (+glm-5.2 case). (thanks @phidinhmanh)
+- **fix(translator):** suppress the streamed `</think>` close marker for the **Antigravity IDE** client (port from 9router#1061, [#6415](https://github.com/diegosouzapw/OmniRoute/pull/6415)). On thinking-only turns Antigravity rendered a bare `</think>` as the sole visible content, tripping its loop-detection and wasting requests. Antigravity's UA (`vscode/<v> (Antigravity/<v>)`) is added to the marker-suppress allowlist (alongside OpenCode); Claude Code / Cursor still get the marker, and `x-omniroute-thinking-marker: on` force-restores it. Regression guard: `tests/unit/think-close-marker-suppress-5245.test.ts`. (thanks @abdofallah)
+- **fix(executors):** strip nested `reasoning_content` from messages for **Mistral** (port from 9router#1649, [#6417](https://github.com/diegosouzapw/OmniRoute/pull/6417)). Mistral's API returns `422 extra_forbidden` when an assistant message carries `reasoning_content` (replayed thinking from a prior turn, e.g. via the Codex `/responses` path); the generic top-level 400 field-downgrade retry never covered the nested per-message field. `DefaultExecutor` now strips it for provider `mistral` only, so DeepSeek (which requires replayed `reasoning_content`) is unaffected. Regression guard: `tests/unit/mistral-strip-reasoning-content-1649.test.ts`. (thanks @xxy9468615)
+- **fix(executors):** strip the `client_metadata` passthrough field on the **OpenCode** path (port from 9router#1442, [#6418](https://github.com/diegosouzapw/OmniRoute/pull/6418)). OpenCode upstreams (e.g. `kimi-k2.6` via opencode-go) reject it with `400 "Extra inputs are not permitted, field: 'client_metadata'"`; the DefaultExecutor strip only covered cerebras/mistral and `OpencodeExecutor` extends `BaseExecutor` directly, so nothing removed it there. Regression guard: `tests/unit/opencode-strip-client-metadata-1442.test.ts`. (thanks @yanpaing007)
+- **fix(executors):** inject the `reasoning_content` echo for the native **Moonshot Kimi** provider (port from 9router#1480, [#6419](https://github.com/diegosouzapw/OmniRoute/pull/6419)). Kimi (executor `default`) is a thinking-mode upstream that 400s with "reasoning_content must be passed back" when a prior assistant turn lacks it; the placeholder injection was only wired into the OpenCode meta-provider, so direct multi-turn Kimi conversations failed. Scoped to `kimi` (gateway-served models matching the thinking-model name pattern are unaffected). Regression guard: `tests/unit/kimi-native-reasoning-injected-1480.test.ts`. (thanks @2220258345)
+- **fix(executors):** recover from a strict gateway's `context_management: Extra inputs are not permitted` 400 (port from 9router#1468, [#6420](https://github.com/diegosouzapw/OmniRoute/pull/6420)). **Claude Code** always sends a top-level `context_management` field; strict anthropic-compatible gateways reject it. The dedicated context-editing 400-fallback only fired when OmniRoute's own `contextEditing` feature was enabled (default off), so a client-sent field passed through untouched and 400'd. `context_management` is now in the generic reactive field-strip list, so it's stripped-and-retried once regardless of the feature flag (with correct request re-signing for claude-compatible relays). Regression guard: `tests/unit/provider-field-strips.test.ts`. (thanks @ohahe52-dot)
+- **fix(network):** enable **RFC 8305 Happy Eyeballs** (`autoSelectFamily`) on the direct-egress undici dispatcher (port from 9router#1237, [#6423](https://github.com/diegosouzapw/OmniRoute/pull/6423)). When DNS returns both IPv6 (AAAA) and IPv4 (A) and the IPv6 route is broken (e.g. a NAT64 `64:ff9b::` prefix without routing), undici tried IPv6 first and hung until `ETIMEDOUT` (then a 502 + account lockout), even though `curl` reached the same host. The direct dispatcher now races both families and uses whichever connects first. Proxy paths pin family via `proxyTls` and are unaffected. Regression guard: `tests/unit/direct-dispatcher-pipelining-4580.test.ts`. (thanks @adentdk)
+- **fix(combo):** round-robin now advances the rotation pointer past the model that **actually served**, not the eagerly-scheduled one (port from 9router#948, [#6428](https://github.com/diegosouzapw/OmniRoute/pull/6428)). With `stickyLimit: 1` (true round-robin), when the scheduled model failed and a _different_ model served via fallback, the counter had already advanced +1 from the scheduled index — so the next request reused the fallback-served model, degrading round-robin into hot-spotting on whichever model was healthy. The pointer now advances to the served index + 1 (mirroring the sticky-limit>1 path). Session-stickiness (#3825) and distribution are preserved. Regression guard: `tests/unit/combo-rr-fallback-advance-948.test.ts`. (thanks @binsarjr)
+- **fix(sse):** a non-string `model` field is now rejected with a `400` before the resolver, instead of crashing downstream `.toLowerCase()`/`.split()` calls into an empty-body `500` that escapes the error sanitizer ([#6407](https://github.com/diegosouzapw/OmniRoute/issues/6407)). Regression guard: `tests/unit/chat-non-string-model-6407.test.ts`. (thanks @chirag127)
+- **fix(api):** unknown `/api/*` routes now return a JSON `404` (instead of the dashboard HTML shell) and scalar chat params (`model`/`temperature`/etc.) are validated **before** the provider lookup so malformed requests fail fast with a clear `400` ([#6424](https://github.com/diegosouzapw/OmniRoute/issues/6424), [#6412](https://github.com/diegosouzapw/OmniRoute/issues/6412)). Regression guards: `tests/unit/api/api-catchall-json-404.test.ts`, `tests/unit/chat-early-schema-validation-6412.test.ts`. (thanks @chirag127)
+- **fix(api):** `/v1/chat/completions` now rejects a non-JSON `Content-Type` with a `400` before parsing the body ([#6414](https://github.com/diegosouzapw/OmniRoute/issues/6414)). Regression guard: `tests/unit/v1-chat-completions-content-type-6414.test.ts`. (thanks @chirag127)
+- **fix(api):** the `X-OmniRoute-Compression` response header is now echoed on `/v1/chat/completions` and `/v1/completions` ([#6422](https://github.com/diegosouzapw/OmniRoute/issues/6422)). Regression guard: `tests/unit/compression-header-echo-6422.test.ts`. (thanks @chirag127)
+- **fix(api):** concurrent `GET /v1/models` requests are coalesced into a single catalog build ([#6408](https://github.com/diegosouzapw/OmniRoute/issues/6408)). Regression guard: `tests/unit/v1-models-concurrent-6408.test.ts`. (thanks @chirag127)
+- **fix(api):** `/v1/completions` now echoes the requested `body.model` in its JSON + streamed responses ([#6429](https://github.com/diegosouzapw/OmniRoute/pull/6429)). Regression guard: `tests/unit/completions-body-model-echo.test.ts`. (thanks @chirag127)
+- **fix(api):** env-var master keys now see the full `/v1/models` catalog ([#6406](https://github.com/diegosouzapw/OmniRoute/issues/6406)). Regression guard: `tests/unit/models-catalog-envkey-6406.test.ts`. (thanks @chirag127)
+- **fix(api):** non-streaming `/v1/completions` responses now echo `body.model` aligned with the `X-OmniRoute-Model` header ([#6426](https://github.com/diegosouzapw/OmniRoute/issues/6426)). Regression guard: `tests/unit/v1-completions-model-header-match-6426.test.ts`. (thanks @chirag127)
+- **fix(api):** unknown `/v1/*` routes now return a JSON `404 not_found` instead of the Next.js dashboard HTML shell ([#6405](https://github.com/diegosouzapw/OmniRoute/issues/6405)). Regression guard: `tests/unit/api/v1-catchall-json-404.test.ts`. (thanks @chirag127)
+- **fix(api):** the per-connection provider models route now degrades to the shipped catalog when a provider's `/models` endpoint answers with a redirect ([#6267](https://github.com/diegosouzapw/OmniRoute/issues/6267)) — a `qwen-web` import failed with a raw `Redirect blocked … (307)` 503. `safeOutboundFetch` throws `REDIRECT_BLOCKED` on the 307, `getSafeOutboundFetchErrorStatus` maps it to 503, and `buildDiscoveryErrorFallbackResponse` treated every 503 as a hard error — so the non-empty `getModelsByProviderId("qwen-web")` catalog was never surfaced. A models-endpoint redirect is not a fixable-config error (unlike `URL_GUARD_BLOCKED`/`INVALID_URL`, which stay hard errors), so it now falls back to the local/cached catalog before the 503 short-circuit. General fix — covers any config-driven provider that 307s. Regression guard: `tests/unit/provider-models-qwen-web-redirect-6267.test.ts`. (thanks @chirag127)
+- **fix(api):** the per-connection provider models route (MCP `list_models_catalog` + the dashboard import view) now merges USER-ADDED custom models into its response ([#6247](https://github.com/diegosouzapw/OmniRoute/issues/6247)) — custom models live in the `key_value` namespace `customModels`, which the live REST `/api/v1/models` already merges, but `src/app/api/providers/[id]/models/route.ts` never read `getCustomModels`, so custom models were dropped on both the discovery-success and local_catalog paths. They are now folded into the returned model list (deduped by id, stamped `owned_by: provider`), fixing MCP + the dashboard import view in one place. Regression guard: `tests/unit/provider-models-custom-merge-6247.test.ts`. (thanks @RCrushMe)
+
+- **fix(providers):** GitLab Duo tool-calling follow-up turns no longer fail upstream with `422 {"detail":"Validation error"}` (tokens 0/0, rejected pre-inference). The #6234 tool-result-feedback fix serialized the **entire** multi-turn conversation into GitLab's single-file `code_suggestions` (`small_file`) generation endpoint — folded history that turn-N sent as an oversized `current_file.content_above_cursor` **and** duplicated verbatim into `user_instruction`, tripping the AI-Gateway's `small_file` validation guard. The executor now **bounds** that prompt: it keeps system + latest user message + the most-recent tool round (dropping older turns), caps oversized tool results, and stops duplicating the full prompt into `user_instruction` (which now carries only the short latest user message) — while still feeding the most-recent tool result back so the agent continues (`open-sse/executors/gitlab.ts`, [#6220](https://github.com/diegosouzapw/OmniRoute/issues/6220)). The unit test covers the bounding logic; the upstream 422→200 clearing is VPS-only (Hard Rule #18). Regression guard: `tests/unit/gitlab-tool-exchange-bounded-6220.test.ts`.
+
+- **fix(i18n):** the provider-detail (`/dashboard/providers/[id]`) connection-status filter labels no longer render as `__MISSING__:All` / `__MISSING__:Active` / `__MISSING__:Error` / `__MISSING__:Banned` / `__MISSING__:CreditsExhausted` in non-English locales (notably pt-BR) ([#6290](https://github.com/diegosouzapw/OmniRoute/issues/6290)). Root cause was **not** the namespace mismatch the issue guessed — the `providers.filter*` keys resolve correctly in `en.json`; the debt lived in the locale mirrors (`src/i18n/messages/*.json`), where these five keys carried the `__MISSING__:` sync sentinel in ~15 locales and were absent entirely in ~26 others, so next-intl found the key and echoed the sentinel verbatim. All 40 non-English/-Chinese mirrors now ship real translations for the five `providers.filter*` labels. Regression guard: `tests/unit/i18n-provider-filter-keys-6290.test.ts`. (thanks @diegosouzapw)
+
+- **fix(providers):** the `copilot-m365-web` streaming executor now emits `debug`-level WebSocket diagnostics ([#6210](https://github.com/diegosouzapw/OmniRoute/issues/6210)) — the outbound WS URL (with the `access_token` **redacted** via `redactWsUrl()`), handshake success/failure, and each received SignalR frame's `type`/`target`. Previously the streaming path logged nothing, so an empty `content:null` response (the M365 Education / Starter tier symptom fixed in #6234) was undiagnosable even at `APP_LOG_LEVEL=debug`. The change is debug-level and side-effect-free — it does not alter streaming behavior or the frame parser, and the token never reaches the logs. Regression guard: `tests/unit/copilot-m365-web-logging-6210.test.ts` (thanks @qpeyba)
+
+- **fix(resilience):** a round-robin combo no longer returns `503 all upstream accounts are unavailable` when a compatibility-rejected target is actually healthy ([#6238](https://github.com/diegosouzapw/OmniRoute/issues/6238)). `filterTargetsByRequestCompatibility` drops request-incompatible targets (tool/vision/structured-output unsupported, or below the required context window) **before** any availability check runs, and its `compatible.length === 0` safety net only fired when _all_ targets were filtered — not when the kept targets later all turned out runtime-unavailable (circuit-open / cooldown / no credentials). So a combo could 503 while a compat-rejected-but-healthy provider sat unused. `handleRoundRobinCombo` now keeps the compat-rejected set and, when every compat-kept target was skipped without a single real attempt, probes those rejected targets as a **last-resort fallback tier** (via the new pure `open-sse/services/combo/comboCompatFallback.ts`) before crystallizing the 503. Regression guard: `tests/unit/combo-roundrobin-compat-fallback-6238.test.ts`. (thanks @ThongAccount)
+
+- **fix(startup):** best-effort self-heal for a corrupted Turbopack dev cache on Windows ([#6289](https://github.com/diegosouzapw/OmniRoute/issues/6289)). On Windows, `pnpm dev` can fail at startup when Turbopack `mmap`s a persistent-cache SST file and the OS refuses the mapping (`os error 1455` — "paging file too small"), which Turbopack surfaces as a misleading `Module not found: Can't resolve '@/shared/utils/machine'`. This is a **known upstream Turbopack cache-corruption bug — not our code**. The dev launcher (`scripts/dev/run-next.mjs`) now wraps `nextApp.prepare()` and, when it rejects with that signature (`isTurbopackCacheCorruption` in the new `scripts/dev/turbopackCacheHeal.mjs`), purges `.build/next/**/cache/turbopack` and retries **once** with a clear log. **Caveat — best-effort only:** the corruption often surfaces as a runtime overlay rather than a `prepare()` rejection, so this cannot always intercept it; the reliable remedy remains manually deleting the Turbopack cache dir. Regression guard: `tests/unit/turbopack-cache-heal-6289.test.ts`. (thanks @chirag127)
+
+- **fix(providers):** qodercli PAT auth no longer fails with `spawn qodercli ENOENT` on Windows ([#6263](https://github.com/diegosouzapw/OmniRoute/issues/6263)) — `spawnQoderCli` spawned the bare `qodercli` name with `shell:false` and an unenriched env, so the npm `.cmd` wrapper under `%APPDATA%\npm` (a user-PATH directory) was never resolved. It now resolves the absolute `.cmd`/`.exe` path through the existing `getCliRuntimeStatus("qoder")` resolver in `src/shared/services/cliRuntime.ts` (memoized), spawns with `shell` when the target is a `.cmd`/`.bat`, and uses the cliRuntime-enriched env (PATH + PATHEXT + APPDATA); the ENOENT error now lists the searched paths plus the `CLI_QODER_BIN` override. End-to-end spawn on a real Windows host is host-only (Hard Rule #18); the path-resolution logic is unit-tested. Regression guard: `tests/unit/qodercli-windows-resolve-6263.test.ts`. (thanks @chirag127)
+
+- **fix(sse):** the reasoning-token buffer no longer inflates **probe-sized `max_tokens`** ([#6274](https://github.com/diegosouzapw/OmniRoute/issues/6274)) — Claude Code's `/model` capability check sends `max_tokens: 1`, but for a thinking-capable model with a large output cap (e.g. `glm-5.2`) the #3587 headroom heuristic (`max(current + 1000, ceil(current * 1.5))`) rewrote it to `1001` and forwarded that upstream, wasting tokens on a request that was never a genuine reasoning budget. `resolveReasoningBufferedMaxTokens()` (`open-sse/services/reasoningTokenBuffer.ts`) now short-circuits and returns the caller's value verbatim when it is below the new `REASONING_BUFFER_MIN_TRIGGER` (256) threshold — a tiny explicit limit is a probe, not a reasoning request. Real budgets still receive the #3587 headroom unchanged, and the guard runs after the existing capability checks so unknown / non-reasoning models keep returning `null`. Regression guard: `tests/unit/reasoning-token-buffer-6274.test.ts`. (thanks @brightfiscalband)
+
+- **fix(cli):** `omniroute reset-password` now works as a real subcommand, and password resets over piped (non-TTY) stdin actually apply ([#6261](https://github.com/diegosouzapw/OmniRoute/issues/6261), [#6258](https://github.com/diegosouzapw/OmniRoute/issues/6258)). Two coupled defects: (1) **#6261** — `bin/omniroute.mjs` routed everything through Commander with only two pre-Commander bypasses (`--mcp`, `reset-encrypted-columns`), so `omniroute reset-password` was rejected as an unknown command; only the separate `omniroute-reset-password` bin worked, while the docs falsely advertised the subcommand (incl. a bogus "legacy alias still works"). A pre-Commander bypass mirroring `reset-encrypted-columns` now dynamically imports `bin/reset-password.mjs` (which self-executes) before Commander parses; the three doc lines were corrected. (2) **#6258** — `bin/reset-password.mjs` issued two sequential `rl.question` prompts; under piped stdin the second read never settled at EOF, so `main()` never reached `resetManagementPassword` and the reset was a silent no-op (both prompts printed, no success, password unchanged). The CLI now detects non-TTY stdin and reads it once (first line = password, second line = confirm if present, else reused), adds a `--password-stdin` flag (entire stdin is the password, no confirmation), and exits `0` explicitly so the success line always flushes; interactive TTY behavior is unchanged. Regression guard: `tests/unit/reset-password-cli-6261-6258.test.ts` (3). (thanks @chirag127)
+
+- **fix(db):** the mass-migration **safety abort** now tells the operator how to bypass it and stops flooding the log ([#6260](https://github.com/diegosouzapw/OmniRoute/issues/6260)) — after restoring a backup that wiped the migration tracking table, `runMigrations()` threw the abort on every downstream `ensureDbInitialized()`, re-logging the full banner 11+ times, and the message never mentioned the existing `OMNIROUTE_MAX_PENDING_MIGRATIONS` escape hatch. The abort text now appends a bypass hint (set `OMNIROUTE_MAX_PENDING_MIGRATIONS=0` in `server.env` / `DATA_DIR/.env`), and a new `MigrationSafetyAbortError` is memoized so repeated calls in the same process throw the same instance and emit a single concise line instead of the full cascade. Regression guard: `tests/unit/migration-safety-abort-6260.test.ts`. (thanks @chirag127)
+
+- **fix(auth):** importing a **distinct** Codex/ChatGPT OAuth `auth.json` is no longer falsely rejected as "already exists" when it belongs to a different user in the same workspace ([#6301](https://github.com/diegosouzapw/OmniRoute/issues/6301)). `findExistingCodexConnection` (in `src/lib/oauth/utils/codexAuthImport.ts`) deduped **only** on `providerSpecificData.workspaceId === accountId`, where `accountId` is the shared `chatgpt_account_id`/`tokens.account_id` — so two members of the same ChatGPT Team collapsed onto a single connection (409 `duplicate_account`). The id_token's `https://api.openai.com/auth` claim carries a per-user `chatgpt_user_id` alongside the workspace id (the device-flow path already persisted it as `chatgptUserId`, but the import path did not). Now `parseAndValidateCodexAuth` extracts `userId` (`chatgpt_user_id` → `user_id` → JWT `sub`) into `ParsedCodexAuth`, the create/update paths persist `chatgptUserId` in `providerSpecificData` (mirroring `codex.ts`), and dedup keys on `workspaceId` **AND** `chatgptUserId` — with a backward-compat fallback to legacy accountId-only matching when no stored connection for that workspace records a `chatgptUserId`, so genuinely-same accounts still dedup. Regression guard: `tests/unit/codex-auth-import-userid-dedup-6301.test.ts` (4). (thanks @anungma)
+
+- **fix(providers):** importing models for the **venice-web** provider no longer fails with a red "Provider venice-web does not support models listing" ([#6269](https://github.com/diegosouzapw/OmniRoute/issues/6269)). `venice-web` is a web-cookie provider with an executor but no upstream `/v1/models` endpoint and no registry `models`, so the models route fell through to the tail `400`. Mirroring the `jules`/`linkup-search`/`ollama-search` fix (#5569), it now ships a static local catalog entry in `src/lib/providers/staticModels.ts` — seeding the current Venice lineup (`venice-uncensored`, `llama-3.3-70b`, `qwen3-235b`, `qwen3-4b`, `deepseek-r1-671b`; Venice rotates its catalog, see docs.venice.ai/models/overview) — so the route returns `200` with `source:"local_catalog"`, `intentional:true`. Regression guard: `tests/unit/static-models-venice-web-6269.test.ts`. (thanks @chirag127)
+
+- **fix(api):** the specialty model catalogs (`/v1/embeddings`, `/v1/images`, `/v1/music`, `/v1/videos` model lists) are now derived from the **unified catalog filtered by a predicate** (`getSpecialtyModelsResponse`) instead of ad-hoc per-route logic, so they consistently respect active-credential visibility and stay in sync with the main catalog ([#6303](https://github.com/diegosouzapw/OmniRoute/pull/6303)). Regression guard: `tests/unit/specialty-model-catalog-routes.test.ts`. (thanks @makcimbx)
+- **fix(api):** the agent-bridge server route now resolves the MITM manager via a **dynamic `import("@/mitm/manager.runtime")`** so Turbopack does not statically pull the stub (or over-bundle the manager), and the agent-skills generator anchors its output base path with `path.join(process.cwd(), …)` so Turbopack's static analyzer stops tracing the whole project root ([#6329](https://github.com/diegosouzapw/OmniRoute/issues/6329), [#6366](https://github.com/diegosouzapw/OmniRoute/pull/6366)). Regression guard: `tests/unit/agent-bridge-server-route-dynamic-import.test.ts`. (thanks @Iammilansoni)
+- **fix(api):** internal probes (combo-test, cloud-sync verify) now pick a **management-scoped / allow-all API key** instead of naively grabbing `getApiKeys()[0]` — a restricted `self:usage` first row made the probe fail with "Model X is not allowed for this API key" even when the combo path was healthy (`pickApiKeyForInternalUse` in `src/lib/db/apiKeys.ts`). The API-manager model editor also falls back to `/api/models?all=true` when `/v1/models` is catalog-protected ([#6372](https://github.com/diegosouzapw/OmniRoute/pull/6372)). Regression guard: `tests/unit/pick-internal-api-key-6372.test.ts`. (thanks @jmengit)
+- **fix(live-ws):** the Live Dashboard WebSocket server now **rejects on bind failure** (e.g. `EADDRINUSE` when the API bridge already holds the port) instead of letting the error surface as an unhandled `error` event that crash-loops the process — the `error` listener is attached to `wss` (not `server`) and releases the EventBus subscription on a failed start ([#6324](https://github.com/diegosouzapw/OmniRoute/issues/6324)). Regression guard: `tests/unit/live-ws-eaddrinuse-6324.test.ts`. (thanks @vinayakkulkarni)
+- **fix(dashboard):** the Home provider-topology widget now trusts the live provider-metrics snapshot — it uses `topology.errorProvider` and live `activeRequests` directly instead of re-deriving state from a stale `lastErrorAt` or applying a frontend timeout filter, so the topology reflects real-time provider health ([#6322](https://github.com/diegosouzapw/OmniRoute/pull/6322)). Regression guard: `tests/unit/home-provider-topology-live-state.test.ts`. (thanks @xz-dev)
+- **fix(sse):** strip zero-width markers from streamed **tool-call arguments** — a follow-up to [#5857](https://github.com/diegosouzapw/OmniRoute/pull/5857). That PR removed injected zero-width joiners (U+200D) from streamed assistant text/reasoning but deliberately left tool-call argument JSON byte-exact. The request-side obfuscation (`open-sse/services/claudeCodeObfuscation.ts`) injects ZWJ into agent words — including the temp path inside the Bash tool description — and Claude models copy that verbatim into generated commands, which are delivered as tool-call arguments rather than assistant text. As a result the ZWJ survived and corrupted code blocks (e.g. a temp path rendered with an invisible joiner). Now `open-sse/handlers/responseSanitizer.ts` strips zero-width code points from tool-call argument strings at every emit site (OpenAI non-stream/stream chat `tool_calls` + legacy `function_call`, native Responses `function_call` items, the OpenAI→Responses conversion, and the native Responses streaming `response.function_call_arguments.delta/.done` events). Only zero-width code points are removed; JSON structure and all other bytes stay identical (no parse/restringify), so normal arguments remain byte-exact. Regression guard: 6 new cases in `tests/unit/response-sanitizer.test.ts` (suite 50/50).
+
+- **fix(nodejs):** the default app log path now resolves under `DATA_DIR` (`~/.omniroute/logs/application/app.log`) instead of `process.cwd()` ([#6197](https://github.com/diegosouzapw/OmniRoute/issues/6197)) — the globally-installed CLI runs from an arbitrary working directory, so anchoring the default to cwd made file logging silently write to (or no-op under) an unrelated directory, contradicting the documented `.env.example` default. `getAppLogFilePath()` now computes the default lazily via the pure `resolveDataDir()` resolver (honours a per-process `DATA_DIR`, no directory-creation side effect); an explicit `APP_LOG_FILE_PATH` still wins. Regression guard: `tests/unit/logenv-datadir-path-6197.test.ts` (3).
+
+- **fix(docker):** AgentBridge/`startMitm` no longer aborts in containers/headless when the Antigravity-default DNS step can't write `/etc/hosts` ([#6127](https://github.com/diegosouzapw/OmniRoute/issues/6127)), and the privileged command's stderr now reaches `app.log` instead of only a bare exit code hitting the toast ([#6198](https://github.com/diegosouzapw/OmniRoute/issues/6198)). The default DNS step (`addDNSEntry`) was called unguarded while cert install and the two sibling DNS steps were each best-effort — in the runtime Docker image (`USER node`, no `sudo`, read-only `/etc/hosts`) it threw `Command failed with code 1` out of `startMitmInternal` and killed the whole start, discarding the stderr. The three DNS steps are extracted into a best-effort `provisionDnsEntries()` where each failure is logged with the full `err` (stderr included, folded in by `systemCommands.ts`) and never aborts the start. Regression guard: `tests/unit/mitm-dns-graceful-degrade-6127.test.ts` (4).
+
+- **fix(providers):** copilot-m365-web now supports the M365 Education "Starter / OfficeWebIncludedCopilot" tier and no longer returns an empty `content:null` stream ([#6210](https://github.com/diegosouzapw/OmniRoute/issues/6210)). Two gaps: (1) `buildWsUrl()` hardcoded the individual-consumer scenario (`OfficeWebPaidConsumerCopilot`, `isEdu=false`) — the EDU tier is now opt-in via `providerSpecificData.tier="edu"`, emitting `scenario=OfficeWebIncludedCopilot`/`isEdu=true` (the individual path is unchanged); (2) the EDU/GPT-5.5 path streams deltas via `arguments[0].writeAtCursor` (incremental) instead of only `messages[].text` (accumulated snapshots), which the parser dropped — a new `accumulateBotContent()` folds both formats, with `type:2 item.result.message` as a last-resort fallback. Regression guard: `tests/unit/copilot-m365-edu-writeatcursor-6210.test.ts` (10). (thanks @qpeyba)
+
+- **fix(providers):** GitLab Duo executor now feeds tool results back into the prompt instead of looping ([#6220](https://github.com/diegosouzapw/OmniRoute/issues/6220)) — `buildPrompt()` branched only on `system`/`user` and took `userParts.at(-1)`, silently dropping the `assistant{tool_calls}` + `tool{result}` turns the client appended, so the reconstructed prompt was byte-identical to turn 1 and the model re-emitted the same `<tool>` call forever. When a tool exchange is present the full conversation is now serialized, folding each tool result back keyed by its `tool_call_id`; simple conversations keep the legacy shape. Complements the tool_call emission from [#6051](https://github.com/diegosouzapw/OmniRoute/issues/6051) (the `kilo-duplicate` label was a false positive — different, sequential defect). Regression guard: `tests/unit/gitlab-tool-result-feedback-6220.test.ts` (4).
+
+- **fix(providers):** opencode-go/opencode-zen can now synthesize the OpenCode CLI identity headers Cloudflare requires on VPS egress ([#5997](https://github.com/diegosouzapw/OmniRoute/issues/5997)) — on a datacenter VPS, `opencode.ai/zen/go/v1/chat/completions` 403s (HTML challenge) requests lacking CLI identity, while the reporter's control curl proved that `User-Agent: opencode-cli/1.0.0` + `x-opencode-client: cli` + `x-opencode-project: default` + fresh request/session UUIDs succeed. Opt-in via `OPENCODE_SYNTHESIZE_CLI_HEADERS=true` (values overridable via `OPENCODE_GO_USER_AGENT`/`OPENCODE_USER_AGENT`/`OPENCODE_CLIENT`/`OPENCODE_PROJECT`); it fills only headers the client did not already send. Kept **off by default** — the forward-only path is deliberate (fabricating a wrong value risks upstream rejection; a prior dedup regressed with `opencode/local`), so this replaces the fragile local header-injection shim without changing default behavior. Regression guard: `tests/unit/opencode-cli-headers-synthesis-5997.test.ts` (6). (thanks @aleksesipenko)
+- fix(resilience): sticky session affinity now evicts and fails over to another account when the pinned account is exhausted/unavailable (#6219)
+
+- fix(sse): Responses API passthrough now drops internal commentary-phase output before forwarding to clients (gated by RESPONSES_PASSTHROUGH_DROP_COMMENTARY, default on) (#6199)
+
+- **fix(sse):** tool-call function schemas with a root `type: null` are now coerced to `type: "object"` before dispatch ([#6359](https://github.com/diegosouzapw/OmniRoute/issues/6359)) — clients like the Codex app emit `parameters: { type: null, ... }` for some tools, which OpenAI-compatible upstreams reject with `400 Invalid schema for function '...': schema must be a JSON Schema of 'type: "object"', got 'type: null'`, failing the whole request. `toolSchemaSanitizer` already stripped the null; it now re-adds the mandatory root `"object"` type (and empty `properties`/open `additionalProperties` when absent). Combinator roots (`anyOf`/`oneOf`/`allOf`) and explicit root types are left untouched. Regression guard: 5 new cases in `tests/unit/tool-schema-sanitizer.test.mjs`.
+
+- **fix(docker):** AgentBridge no longer fails to start on npm/Electron/VPS installs with "MITM manager stub reached at runtime" ([#6344](https://github.com/diegosouzapw/OmniRoute/issues/6344)) — v3.8.45 flipped the production bundler default to Turbopack, but `next.config.mjs` aliased `@/mitm/manager` to its Docker-only degraded stub **unconditionally**. That was harmless while Docker (which sets the alias intentionally for #3390 graceful degradation) was the sole Turbopack consumer, but once every artifact built with Turbopack the stub shipped to all non-Docker users and `startMitm` threw on the first Agent-Bridge start. The alias is now opt-in via `OMNIROUTE_MITM_STUB=1` (set only by the Dockerfile) through the shared `scripts/build/mitm-stub-flag.mjs` helper; default builds bundle the real manager. Regression guard: `tests/unit/mitm-stub-alias-6344.test.mjs` (4).
+
+- **fix(proxy):** stop the v3.8.44 proxy regression that leaked the real IP and disabled healthy proxies ([#6246](https://github.com/diegosouzapw/OmniRoute/issues/6246)). Two coupled defects from the new health scheduler: (1) **IP leak** — when a proxy assigned to a connection was marked `inactive`, resolution fell through to a **direct** egress instead of blocking, exposing the operator's real IP; (2) **over-deactivation** — the sweep flipped a proxy to `inactive` on the **first** failed probe and counted our own 5s timeout / a probe-target `5xx` as the proxy's fault, so healthy paid proxies vanished from egress selection ("my proxies are not being used anymore"). Fix: the sweep decision is extracted into a pure, network-free `decideProxyHealthAction` (`src/lib/proxyHealth/decision.ts`) — by default the health check now **only counts/logs and never downgrades status** (a proxy is downgraded/removed only with `PROXY_AUTO_REMOVE=true`, after `PROXY_AUTO_REMOVE_AFTER` **consecutive** conclusive failures); probes are classified tri-state so an inconclusive result (our timeout, or a `5xx` from the probe target) never penalizes the proxy, and the probe timeout is raised 5s→15s. Separately, `safeResolveProxy` now **fails closed** via the existing policy: a connection whose assigned proxy is dead is blocked instead of leaking direct (`hasBlockingProxyAssignment`), honoring the explicit `proxy off` toggles and the `PROXY_FAIL_OPEN=true` opt-out. Existing proxies stuck `inactive` by the old behavior need a one-time manual re-activate (the operator owns proxy status). Regression guards: `tests/unit/proxy-health-decide-action-6246.test.ts`, `tests/unit/proxy-assigned-unavailable-6246.test.ts`.
+
+- **fix(proxy):** make "Test All" read-only and add bulk enable/disable ([#6246](https://github.com/diegosouzapw/OmniRoute/issues/6246)). Complements the core fail-closed / scheduler fix (#6296) with the two remaining reporter asks. (1) The **"Test All" button** (`POST /api/settings/proxies/auto-test`) used to flip a proxy to `inactive` on a failed reachability probe; since the egress selector excludes `inactive` proxies, a flaky probe (an unreachable `httpbin.org`, a proxy that blocks `HEAD`, or a slow paid proxy) silently disabled every proxy that failed — "Test All" is now **read-only by default** (only the operator sets a proxy active/inactive; opt back into the legacy test-and-set with `PROXY_HEALTH_AUTO_DEACTIVATE=true`). (2) Adds a **bulk enable/disable** proxies endpoint + toolbar action (`POST /api/settings/proxies/batch-activate`) so an operator can re-activate proxies in one click. Regression guard: `tests/unit/proxy-health-6246.test.ts`. (thanks @tenshiak)
+
+- **chatcore (tools): stop the default 128-tool cap from silently dropping opencode's `task`/MCP tools.** opencode (used as an MCP/agent host) sends a large tool list; when it exceeds the speculative `MAX_TOOLS_LIMIT` (128) default, `truncateToolList` did a blind `tools.slice(0, 128)`, dropping every tool past index 128 — including opencode's built-in `task` tool (subagent launch) and many MCP tools, so models routed through OmniRoute could no longer spawn subagents or reach part of their tools. The cap exists to avoid upstream `400`s for providers with real hard limits (e.g. grok-cli 200), so it is kept for those: detection of the opencode client (`isOpencodeClient` — any `x-opencode-*` header, or `opencode` in the user-agent) now only bypasses the **speculative 128 default**, never a known provider ceiling. Precedence is explicit — a proactive/detected provider limit always truncates (even for opencode); otherwise opencode forwards its full tool list; otherwise the unchanged 128 default applies to every other client. Refactors `getEffectiveToolLimit` into `getKnownToolLimit(provider) ?? DEFAULT_LIMIT` (byte-identical for existing callers) and fixes a cosmetic debug-log that reported the truncated count instead of the original. Regression guard: `tests/unit/tool-limit-detector.test.ts`.
+
+- **fix(mitm):** the macOS MITM-cert install check now matches the system keychain again. `security find-certificate -a -Z` prints the SHA-1 as a colon-less hex string, but the installed-check compared it against `getCertFingerprint()`'s colon-separated form, so the substring match never hit — the cert was reported as not-installed and re-prompted for the sudo install on every run. Fingerprints are now normalized (colons stripped, upper-cased) on both sides via the extracted `macCertOutputHasFingerprint` helper. Regression guard: `tests/unit/mitm-cert-mac-fingerprint.test.ts`. ([#6204](https://github.com/diegosouzapw/OmniRoute/pull/6204), closes [#6134](https://github.com/diegosouzapw/OmniRoute/issues/6134) — thanks @rianonehub)
+
+- **fix(api):** `/v1/messages/count_tokens` now counts `tool_use`, `tool_result` and `thinking` content blocks (and array-form `system` prompts) in the local-estimation path, instead of only `text`. Real agentic conversations keep ~95% of their tokens inside tool results; the previous estimate returned near-zero for them, which silently broke Claude Code's auto-compaction (context grew past the window with no compaction until the upstream API rejected the request). The real provider-side count path is unchanged. Regression guard: `tests/unit/messages-count-tokens-route.test.ts`. ([#6221](https://github.com/diegosouzapw/OmniRoute/pull/6221) — thanks @luweiCN)
+
+- **fix(antigravity):** strip a trailing assistant prefill turn for Vertex Claude models to avoid upstream 400s ([#6114](https://github.com/diegosouzapw/OmniRoute/pull/6114)). Regression guard: `tests/unit/antigravity-claude-prefill-strip.test.ts`. (thanks @anki1kr)
+
+- **fix(security):** the mutable cloud-agent routes (`/api/cloud/credentials/update`, `/api/cloud/models/alias`) now require management auth instead of being treated as public. They were classified as public API routes, so a request without management credentials could update stored cloud-agent credentials and model aliases. They are removed from the public-route set, classified as management routes in the authz pipeline, and gated by `requireManagementAuth`; cloud **read**/auth routes stay public. Regression guards: `tests/unit/cloud-write-auth.test.ts`, `tests/unit/authz/classify.test.ts`, `tests/unit/public-api-routes.test.ts`. ([#6233](https://github.com/diegosouzapw/OmniRoute/pull/6233) — thanks @vittoroliveira-dev)
+
+- **refactor(dashboard):** extract the onboarding-wizard "Open provider details" link target into a pure, unit-tested `buildProviderDetailsHref(connection)` helper. The wizard already routes by `connection.id` (the node UUID) rather than the provider category slug (#6144/#6145); this hardens that behavior behind a tested helper that guards a missing id/connection. Regression guard: `tests/unit/provider-onboarding-href.test.ts`. ([#6166](https://github.com/diegosouzapw/OmniRoute/pull/6166) — thanks @KooshaPari)
+
+- **fix(security):** the doubao synthetic device-id generator now derives its digits via an unbiased crypto-random draw (rejection sampling over `crypto.randomBytes()`) instead of a `% 10` reduction, closing a CodeQL `js/biased-cryptographic-random` finding.
+- **fix(agentSkills):** the GitHub-skills generator now resolves `outputDir` to an absolute path before writing, fixing a regression introduced by #6366 (relative-to-cwd base path) that could write generated skill files to the wrong directory.
+- **fix(security):** `/api/keys/{id}/devices` now checks the HTTP method before auth/validation, returning a `405` for non-GET/DELETE verbs instead of a misleading `401`/`500` (closes a `dast-smoke` QUERY-method finding).
+- **fix(quality):** clear the last 2 heavy quality-gate reds on the release tip (cycle pre-flight).
+- **fix(mitm):** the test suite and CI can never mutate the OS trust store — `OMNIROUTE_SKIP_SYSTEM_TRUST=1` is set globally for tests/CI so `installCert`/`uninstallCert`/`installTproxyCa` skip the privileged OS dispatch ([#6310](https://github.com/diegosouzapw/OmniRoute/pull/6310); full detail is under the [3.8.45] section below — this branch received it via the parallel-cycle sync-back).
+- **fix(api):** `POST /api/github-skills` now Zod-validates its request body; documented the new quality-gate env vars and pinned the merge-integrity GitHub Actions to a commit SHA.
+- **fix(skills):** generate the missing `omni-github-skills` registry entry and align the agent-skills catalog-count tests (follow-up to #6186).
+- **fix(quality):** clear the cycle's 11 net-new ESLint errors and make `validate-release-green` suppressions-aware.
+
+### 📝 Maintenance
+
+- **i18n(it):** add 118 missing Italian (`it`) translations (net-additive — no existing keys dropped, valid JSON), improving Italian UI coverage. ([#6212](https://github.com/diegosouzapw/OmniRoute/pull/6212) — thanks @serverless83)
+- **chore(providers):** remove deprecated **MiMo V2** model entries from the catalogs (xiaomi-mimo, opencode-go, zenmux-free, audio TTS) — the upstream V2 line is superseded by MiMo V2.5; drops `mimo-v2-tts`, `mimo-v2-pro`, `mimo-v2-omni`, `mimo-v2-flash`, `mimo-v2-flash-free` and realigns the provider-catalog tests. ([#6248](https://github.com/diegosouzapw/OmniRoute/pull/6248) — thanks @backryun)
+- **chore(release):** ~50 commits on this branch are v3.8.45 pre-flight/hardening fixes and CI-perf work that landed here via the parallel-cycle sync-back (`sync-next-cycle.mjs`, Hard Rule #21) **after** the `v3.8.45` git tag was cut, and are already fully documented under the **[3.8.45]** section below — listed here only so the per-cycle commit-coverage check (`npm run release:uncovered`) doesn't flag them as gaps. Provider/catalog/UX/backend fixes: #6041, #6078, #6108, #6135, #6148, #6149, #6154, #6158, #6161, #6162, #6163, #6164, #6165, #6170, #6177, #6178, #6181, #6186, #6187, #6191, #6193, #6194, #6195, #6200, #6205, #6208, #6209, #6211, #6213, #6223, #6224, #6225, #6226, #6227, #6228, #6229, #6230, #6235, #6291, #6292. CI/release-pipeline work: #6167, #6203, #6214, #6215, #6218, #6273, #6275, #6283, #6284, #6285, #6300, #6305.
+- **chore(release):** additional zero-ref release-cycle plumbing on this branch, kept out of `release:uncovered` on purpose (no `#N` in the commit subject to cite): opening the v3.8.46 cycle, opening/closing the v3.8.45 cycle, the finalized [3.8.45] CHANGELOG i18n sync-back to 42 mirrors, the v3.8.45 cognitive/cyclomatic and file-size drift rebaselines, ESLint stale-suppression pruning (4,273 → 4,233), and clearing test-masking/docs-all pre-flight reds for v3.8.45.
+
+### ⚡ Performance & Infrastructure
+
+- **perf(release-green):** the pre-flight validator (`scripts/quality/validate-release-green.mjs`) now runs its 4 slow suites (unit / vitest / integration / pack-artifact) **concurrently** via `Promise.all` — pre-flight wall time drops from ~the sum of the suites to ~the slowest one (~30min saved per round; Phase 0 was the nº1 bottleneck of the v3.8.45 release benchmark, 2h54 of 6h34 e2e). Guard: `tests/unit/validate-release-green.test.ts` ("runs the slow suites CONCURRENTLY"). ([#6319](https://github.com/diegosouzapw/OmniRoute/pull/6319))
+- **fix(ci):** `scripts/release/sync-next-cycle.mjs` — two defects found live in its first production run (v3.8.45 Phase 5): (1) the `git()` helper's default 1 MiB `maxBuffer` crashed with `ENOBUFS` on `git show origin/main:CHANGELOG.md` (the CHANGELOG alone is >1 MiB) — widened to 64 MiB; (2) the i18n resync only propagated the `[NEXT]` (TBD) section, leaving the just-shipped finalized section as "— TBD" in all 42 mirrors — it now also syncs `[prevVersion]` bounded by the heading below it (new exported pure helper `versionAfter`). Guards: +5 tests in `tests/unit/sync-next-cycle.test.ts` (8/8). ([#6327](https://github.com/diegosouzapw/OmniRoute/pull/6327))
+- **test(ci):** concurrency-sensitive flaky tests are **quarantined into a serial pass** (`tests/unit/serial/`, `--test-concurrency=1`, appended to every unit runner incl. sharded variants — the serial pass is sharded too so concurrent shard jobs never self-collide). Initial set: `glm-coding-plan-monthly-3580`, `quota-division-blocks`, `provider-health-autopilot`, `combo-health-autopilot` — the class behind the ~28min CI wedges/re-runs (two live 1h+ wedges cancelled during this PR's own validation). Discovery + TIA gates track the new glob; systemic root cause (async logger writing after teardown) tracked in [#6360](https://github.com/diegosouzapw/OmniRoute/issues/6360). Guard: `tests/unit/test-serial-quarantine.test.ts` (4). ([#6347](https://github.com/diegosouzapw/OmniRoute/pull/6347))
+
+### 🙌 Contributors
+
+Thanks to everyone whose work landed in v3.8.46:
+
+| Contributor                                                    | PRs / Issues           |
+| -------------------------------------------------------------- | ---------------------- |
+| [@2220258345](https://github.com/2220258345)                   | direct commit / report |
+| [@abdofallah](https://github.com/abdofallah)                   | direct commit / report |
+| [@adentdk](https://github.com/adentdk)                         | direct commit / report |
+| [@aleksesipenko](https://github.com/aleksesipenko)             | direct commit / report |
+| [@anki1kr](https://github.com/anki1kr)                         | direct commit / report |
+| [@anungma](https://github.com/anungma)                         | direct commit / report |
+| [@arpicato](https://github.com/arpicato)                       | direct commit / report |
+| [@backryun](https://github.com/backryun)                       | #6248                  |
+| [@binsarjr](https://github.com/binsarjr)                       | direct commit / report |
+| [@brightfiscalband](https://github.com/brightfiscalband)       | direct commit / report |
+| [@chirag127](https://github.com/chirag127)                     | #6501, #6506           |
+| [@developerjillur](https://github.com/developerjillur)         | direct commit / report |
+| [@dilneiss](https://github.com/dilneiss)                       | #6499                  |
+| [@dtybnrj](https://github.com/dtybnrj)                         | direct commit / report |
+| [@Forcerecon](https://github.com/Forcerecon)                   | direct commit / report |
+| [@hao3039032](https://github.com/hao3039032)                   | direct commit / report |
+| [@Iammilansoni](https://github.com/Iammilansoni)               | #6150, #6245           |
+| [@jmengit](https://github.com/jmengit)                         | direct commit / report |
+| [@jordansilly77-stack](https://github.com/jordansilly77-stack) | direct commit / report |
+| [@JxnLexn](https://github.com/JxnLexn)                         | direct commit / report |
+| [@KooshaPari](https://github.com/KooshaPari)                   | #6166                  |
+| [@loopyd](https://github.com/loopyd)                           | direct commit / report |
+| [@luweiCN](https://github.com/luweiCN)                         | #6221                  |
+| [@makcimbx](https://github.com/makcimbx)                       | direct commit / report |
+| [@muflifadla38](https://github.com/muflifadla38)               | direct commit / report |
+| [@newnol](https://github.com/newnol)                           | direct commit / report |
+| [@ofekbetzalel](https://github.com/ofekbetzalel)               | direct commit / report |
+| [@ohahe52-dot](https://github.com/ohahe52-dot)                 | direct commit / report |
+| [@phidinhmanh](https://github.com/phidinhmanh)                 | direct commit / report |
+| [@powellnorma](https://github.com/powellnorma)                 | direct commit / report |
+| [@qpeyba](https://github.com/qpeyba)                           | direct commit / report |
+| [@RaviTharuma](https://github.com/RaviTharuma)                 | direct commit / report |
+| [@RCrushMe](https://github.com/RCrushMe)                       | direct commit / report |
+| [@rianonehub](https://github.com/rianonehub)                   | #6134, #6204           |
+| [@serverless83](https://github.com/serverless83)               | #6212                  |
+| [@swingtempo](https://github.com/swingtempo)                   | direct commit / report |
+| [@tenshiak](https://github.com/tenshiak)                       | direct commit / report |
+| [@ThongAccount](https://github.com/ThongAccount)               | direct commit / report |
+| [@UnrealAryan](https://github.com/UnrealAryan)                 | direct commit / report |
+| [@vinayakkulkarni](https://github.com/vinayakkulkarni)         | direct commit / report |
+| [@vittoroliveira-dev](https://github.com/vittoroliveira-dev)   | #6233                  |
+| [@warelik](https://github.com/warelik)                         | direct commit / report |
+| [@xxy9468615](https://github.com/xxy9468615)                   | direct commit / report |
+| [@xz-dev](https://github.com/xz-dev)                           | direct commit / report |
+| [@yanpaing007](https://github.com/yanpaing007)                 | direct commit / report |
+| [@diegosouzapw](https://github.com/diegosouzapw)               | maintainer             |
+
+---
+
 ## [3.8.45] — 2026-07-06
 
 ### ✨ New Features

docs/i18n/vi/CHANGELOG.md

@@ -6,6 +6,213 @@
 
 ## [3.8.31] — 2026-06-20
 
+## [3.8.46] — TBD
+
+### ✨ New Features
+
+- **feat(sse):** **hide paid-only models from `auto/*` routing** when `hidePaidModels` is on ([#6512](https://github.com/diegosouzapw/OmniRoute/issues/6512)) — follow-up to #6328/#6495. PR #6495 hid paid-only models from the `GET /v1/models` listing, but `auto/*` combos (`auto/best-coding`, `auto/glm`, …) could still pick a paid-only backend into their candidate pool → a 402/403 at request time. `createVirtualAutoCombo` now filters the candidate pool through the new pure `open-sse/services/autoCombo/paidModelFilter.ts` (`filterPaidOnlyCandidates`), applying the same free-model predicate #6495 uses in `catalog.ts` (`providerHasFreeModels(provider) && isFreeModel(provider, {id})`) whenever `settings.hidePaidModels === true`. Applied before the category/tier/family narrowing, so it covers every `auto/*` combo; an all-paid pool degrades to the existing graceful empty-pool path. **Opt-in — default OFF leaves the pool unchanged** (identity). Regression guard: `tests/unit/autoCombo/paid-model-filter-6512.test.ts` (4, incl. the default-off identity guard).
+- **feat(sse):** **provider-family auto combos** — `auto/glm`, `auto/minimax`, `auto/mimo`, `auto/zai`, `auto/gemma`, `auto/llama`, `auto/gemini` ([#6453](https://github.com/diegosouzapw/OmniRoute/issues/6453)) — new routable ids that materialize an on-demand virtual combo spanning whatever installed backends currently expose that model family, degrading gracefully as backends rotate. A new pure `open-sse/services/autoCombo/modelFamily.ts` (`detectModelFamily`) classifies by model-id prefix for six families; `zai` is instead resolved by provider id (z.ai's hosted API serves the same `glm-*` model ids as every other GLM backend, so `auto/zai` means "route to my z.ai backend specifically" vs `auto/glm`'s "any connected GLM backend"). Reuses the existing `createVirtualAutoCombo` on-demand materialization path (no DB writes) and the `/v1/models` catalog advertising loop. Regression guard: `tests/unit/autoCombo/provider-family-combos.test.ts` (11).
+- **feat(proxy):** native **proxy-pool round-robin / egress IP rotation** ([#6365](https://github.com/diegosouzapw/OmniRoute/issues/6365)) — a scope (global / provider / account) can now hold **multiple** proxies as a pool with a rotation strategy, so outbound requests cycle their egress IP instead of pinning one proxy per scope. Migration `117_proxy_pool_rotation.sql` lifts the `UNIQUE(scope, scope_id)` constraint (rebuild via the canonical rename/copy/drop; existing single assignments become 1-element pools) and adds a `proxy_scope_rotation` companion table holding the per-scope strategy + a persisted monotonic round-robin cursor. Strategies: `round-robin` (default, monotonic cursor — never `Math.random`), `random`, and `sticky-per-N-min`. Resolution (`resolveProxyForScopeFromRegistry` / `resolveProxyForConnectionFromRegistry`) now fetches the alive, position-ordered candidate set (unchanged `PROXY_ALIVE_PREDICATE`) and applies the strategy; an empty / all-dead pool still returns `null` — the #6246 fail-closed guard is **untouched** (never falls through to direct egress). Backend + DB only; dashboard pool-builder UI is a follow-up. Regression guard: `tests/unit/proxy-pool-rotation-6365.test.ts` (8, incl. fail-closed + backward-compat).
+- **feat(providers):** end-to-end **tool/function calling on the native Gemini `/v1beta` endpoint** ([#6222](https://github.com/diegosouzapw/OmniRoute/issues/6222)) — both directions of the Gemini↔OpenAI conversion now preserve tool data (previously silently dropped). Request side: `convertGeminiToInternal` (extracted to its own testable module) maps `tools[].functionDeclarations` → OpenAI `tools`, prior `functionCall` parts → assistant `tool_calls`, and `functionResponse` parts → `tool`-role messages. Response side: `convertOpenAIResponseToGemini` emits `parts[].functionCall {name,args}` from `message.tool_calls`, and the streaming `openAIChunkToGeminiChunk` accumulates fragmented `tool_calls` deltas by index into complete `functionCall` parts. The non-Gemini client paths (Claude, OpenAI-Responses) already preserved tool calls — this closes the gap specific to the native Gemini surface. Regression guard: `tests/unit/v1beta-gemini-tool-calling-6222.test.ts` (6, incl. a streaming SSE round-trip).
+- **feat(providers):** copilot-m365-web **enterprise / work tier** support ([#6334](https://github.com/diegosouzapw/OmniRoute/issues/6334)) — mirrors the EDU-tier pattern (#6210): `M365ConnectionParams` gains an `agent` field, a new opt-in `M365_ENTERPRISE_OVERRIDES` preset (`agent=work`, `scenario=officeweb`, `licenseType=Premium`) applies via `providerSpecificData.tier="enterprise"` (alias `"work"`), and `agent` is also overridable directly via `providerSpecificData.agent`. `buildWsUrl` was hardcoding `agent="web"` (the one enterprise-distinguishing param with no override path), so a Premium work account handshook then returned an empty stream. The individual and EDU paths are untouched. Kilo's dup flag vs #6210 (EDU tier) was a false positive — different tier. Regression guard: `tests/unit/copilot-m365-enterprise-6334.test.ts` (7). End-to-end confirmation on a real Premium work account is a live-VPS validation follow-up (Hard Rule #18). (thanks @Forcerecon)
+- **feat(api):** standardized, provider-agnostic **`effort` + `thinking` request params** ([#6241](https://github.com/diegosouzapw/OmniRoute/issues/6241)) — a thin standardization layer over the existing mature per-provider reasoning plumbing (no provider mapper touched). `providerChatCompletionSchema` gains a canonical `effort` (reusing the shared `none/low/medium/high/xhigh` vocabulary — the UI tiers `extra`/`max` collapse onto `xhigh`) and a boolean `thinking`. A pure `normalizeReasoningRequest` (wired once in `src/sse/handlers/chat.ts`, before any reasoning field is read) folds them onto the fields the translators already consume (`reasoning_effort` / `reasoning.effort` / `thinking`), so they fan out to Anthropic / Gemini / xAI / Responses — an explicit client `reasoning_effort` / object-shaped `thinking` always wins (backward-compatible). `/models` additively exposes `supportsThinking` + `effort_tiers` so the frontend can render the toggles (UI component is a follow-up). Regression guard: `tests/unit/effort-thinking-standardization-6241.test.ts` (12). (thanks @Iammilansoni, @shabeer)
+- **feat(combo):** new **`pipeline` (sequential) combo strategy** ([#6297](https://github.com/diegosouzapw/OmniRoute/issues/6297)) — the 18th routing strategy runs targets **in order**, threading each step's output into the next step's input, with an optional per-step `prompt` (system instruction); only the final step's response is returned. Distinct from `fusion` (parallel fan-out + judge). Implemented as a self-contained `open-sse/services/pipeline.ts` (sibling to `fusion.ts`), dispatched from `combo.ts`; the step list reuses `combo.models` order and reads an optional `prompt` off each target (backward-compatible — ignored by every other strategy). Intermediate steps run non-streaming with tools stripped (complete prose to thread forward); the final step keeps the client's `stream` flag + tools. A failing/empty/unparseable intermediate step fails the whole pipeline explicitly via a sanitized error (never silently swallowed). Kilo's dup flag vs #563 was a false positive (that's model→chain selection; this is a sequential chain). Regression guard: `tests/unit/combo-pipeline-strategy.test.ts` (5). (thanks @ofekbetzalel)
+- **feat(ci):** `check:test-masking` now flags **inline-reimplemented prod conditions** ([#6348](https://github.com/diegosouzapw/OmniRoute/issues/6348)) — a new **report-only** subcheck (v2, 6A.10 family) catches the wrong-shape contract test: a test that recomputes the condition under test inline instead of importing/exercising the real function (the #6216 class, where `=== 500` → `>= 500` stayed green because the test re-implemented the branch). For each added/modified test file it warns when the file textually duplicates a ≥3-token conditional from a production file touched in the same PR **and** does not import the symbol/module owning it, via a pure, fixture-tested `findReimplementedConditions()` with an allowlist mirroring `assertReductionAllowlist`. Report-only for now (does not fail the gate) — to be promoted to blocking after a triage cycle. Regression guard: `tests/unit/check-test-masking.test.ts` (45).
+- **feat(sse):** per-connection routing override (native vs CLIProxyAPI) ([#6339](https://github.com/diegosouzapw/OmniRoute/issues/6339)) — the previously-dead `isCliproxyapiDeepModeEnabled` helper is now wired into `resolveExecutorWithProxy`: a single connection can opt itself into the CLIProxyAPI passthrough executor via `providerSpecificData.cliproxyapiMode="claude-native"`, with precedence **connection override > provider `upstream_proxy_config` mode > default**. `resolveExecutorWithProxy` now receives the resolved connection's `providerSpecificData` (threaded from `chatCore.ts`), so one connection can deep-route while the provider's other connections stay native — no DB schema change (the toggle rides in `providerSpecificData`). Also resolves the same-provider mixing ask in #6340. Regression guard: `tests/unit/chatcore-executor-proxy.test.ts` (9). (thanks @RaviTharuma)
+- **feat(dashboard):** "Add session cookie" modal now shows a prominent **"Open ‹host› →"** link to the provider's own site ([#6268](https://github.com/diegosouzapw/OmniRoute/issues/6268)) — every `-web` cookie-session provider (chatgpt-web, claude-web, gemini-web, kimi-web, lmarena, qwen-web, m365-copilot-web, …) renders a one-click external link (opening the provider's login/home page in a new tab) so operators no longer tab away to retype the URL mid-setup. The host resolves from a pure, unit-tested `resolveWebProviderHost()` (prefers `WEB_COOKIE_PROVIDERS[id].website`, falls back to the registry `baseUrl` origin); non-web providers render exactly as before. Kilo's dup flag vs #6265 (modal-too-small-on-1080p) was a false positive — distinct concern. Regression guard: `tests/unit/resolve-web-provider-host.test.ts` (5). (thanks @chirag127)
+- **feat(providers):** add **DigitalOcean** AI (serverless inference) as an OpenAI-compatible API-key provider ([#6373](https://github.com/diegosouzapw/OmniRoute/pull/6373)) — base `https://inference.do-ai.run/v1`, wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/digitalocean/`, `src/shared/constants/providers/apikey/inference-hosts.ts`). Regression guard: `tests/unit/digitalocean-provider.test.ts`. (thanks @newnol)
+- **feat(providers):** add **Huancheng Public API** (`hcnsec`) as an OpenAI-compatible regional provider ([#6410](https://github.com/diegosouzapw/OmniRoute/pull/6410)) — Xinjiang Huancheng Cybersecurity's public LLM platform (base `https://api.hcnsec.cn/v1`, free credits via daily check-ins), wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/hcnsec/`, `src/shared/constants/providers/apikey/regional.ts`). Regression guard: `tests/unit/hcnsec-provider.test.ts`. (thanks @UnrealAryan)
+- **feat(dashboard):** the web-session credential guide now shows an **"Open {host}" link** ([#6316](https://github.com/diegosouzapw/OmniRoute/pull/6316)) to the provider's sign-in site (derived from the provider `website` via `getProviderWebsiteHost`), so you can jump straight to the page where the cookie/session must be captured. Regression guard: `tests/unit/web-session-provider-link-6316.test.ts`. (thanks @jordansilly77-stack)
+- **feat(cerebras):** add the **Gemma 4 31B** model (`gemma-4-31b`) to the Cerebras registry + pricing table ([#6331](https://github.com/diegosouzapw/OmniRoute/pull/6331)). Regression guard extends `tests/unit/t28-model-catalog-updates.test.ts`. (thanks @backryun)
+- **feat(providers):** add **Yuanbao (web)** as a cookie-session provider ([#6196](https://github.com/diegosouzapw/OmniRoute/issues/6196)) — `yuanbao-web` (Tencent Yuanbao, `yuanbao.tencent.com`) with cookie-only auth (`hy_user`/`hy_token` + public agent id), SSE→OpenAI translation incl. `reasoning_content`, exposing DeepSeek V3/R1 + Hunyuan / Hunyuan-T1. Regression guard: `tests/unit/providers-yuanbao-web.test.ts`. `together-web` was **deferred** (no verifiable web-session endpoint — needs a captured request) and `huggingchat-web` **dropped** (the existing `huggingchat` already is a web-cookie provider). (thanks @chirag127)
+- **feat(providers):** route the built-in **agentrouter** through the dynamic Claude-Code wire image ([#6056](https://github.com/diegosouzapw/OmniRoute/issues/6056)) — a small static allow-set (`CC_WIRE_IMAGE_BUILTINS` in `open-sse/services/ccWireImageBuiltins.ts`), consulted by `isClaudeCodeCompatible` / `isClaudeCodeCompatibleProvider` / `applyFingerprint`, makes agentrouter adopt the CC wire-image headers + fingerprint **while guarding the CC baseUrl/auth branches** so it keeps its own registry `baseUrl` and `x-api-key` auth. Regression guard: `tests/unit/agentrouter-cc-wire-image.test.ts` (asserts the wire image is applied AND agentrouter's baseUrl/auth are preserved). Live WAF-acceptance against agentrouter.org is a VPS validation follow-up (Hard Rule #18).
+- **feat(providers):** **bulk-add API keys for Cloudflare Workers AI** ([#6174](https://github.com/diegosouzapw/OmniRoute/issues/6174)) — `cloudflare-ai` is removed from the bulk-add exclusion list and the bulk parser gains a 3-field `name|accountId|apiKey` mode; the bulk route now builds a **per-entry** `providerSpecificData` so each key carries its own `accountId` (fixing the previous shared-object reuse), and both the create + key-validation paths receive it. Regression guard: `tests/unit/bulk-api-key-parser-cloudflare.test.ts`. (thanks @muflifadla38)
+- **feat(dashboard):** routing/settings UX clarity ([#6147](https://github.com/diegosouzapw/OmniRoute/issues/6147)) — (1) weighted combos show the **effective routing share %** next to each weight when weights don't sum to 100 (`WeightTotalBar.tsx`); (2) the status widget's user-facing **"Cloud Sync" label is renamed** to "Remote Settings Sync" (`CloudSyncStatus.tsx`; internal ids/state untouched); (3) built-in providers gain an **opt-in advanced base-URL override** (`isBaseUrlOverrideEligibleProvider`, hidden behind an "Advanced" toggle, reusing the existing `providerSpecificData.baseUrl` persistence — not globally widened). Regression guard: `tests/unit/routing-settings-ux-6147.test.ts`.
+- **feat(combo):** add an option to **disable session stickiness**, per-combo or globally — round-robin / random combos can rotate to a different connection on every request instead of pinning a whole conversation to one connection by its first-message hash. Resolution precedence per-combo `config.disableSessionStickiness` → global `settings.disableSessionStickiness` → default `false` (preserves the #3825 prompt-cache/504 fix); gates **both** stickiness call sites in `open-sse/services/combo.ts`. Exposed as a global toggle (Combo Defaults) and a per-combo Inherit/on/off control. ([#6168](https://github.com/diegosouzapw/OmniRoute/issues/6168)) Regression guard: `tests/unit/combo-disable-session-stickiness.test.ts`. (thanks @RCrushMe)
+- **feat(docker):** add the `OMNIROUTE_NO_SUDO` env flag for root-less / user-namespaced deployments — the MITM cert-trust command path (`resolveSudoSpawn` in `src/mitm/systemCommands.ts`) now strips the leading `sudo` when the flag is truthy, in addition to the existing root / sudo-missing cases, so the Proxy Agent runs without `sudo` (the operator trusts the CA manually, e.g. via `NODE_EXTRA_CA_CERTS`). Argv-array `spawn` preserved — no shell interpolation (Hard Rule #13). ([#6122](https://github.com/diegosouzapw/OmniRoute/issues/6122)) Regression guard: `tests/unit/mitm-systemCommands-no-sudo.test.ts`. (thanks @powellnorma)
+- **feat(providers):** add **Requesty** as an OpenAI-compatible gateway provider (BYOK, base `https://router.requesty.ai/v1`, ~200 free requests/day) — wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/requesty/`, `src/shared/constants/providers/apikey/gateways.ts`). ([#6120](https://github.com/diegosouzapw/OmniRoute/issues/6120)) Regression guard: `tests/unit/requesty-provider.test.ts`. (thanks @chirag127)
+- **feat(dashboard):** add **configured-only / available-only filters** to the Free Provider Rankings page ([#6150](https://github.com/diegosouzapw/OmniRoute/issues/6150)) — hide providers you haven't configured, or whose connections are all rate-limited / out of quota, via server-side query params (`?configuredOnly` / `?availableOnly` on `GET /api/free-provider-rankings`) backed by a testable lib helper reusing the in-process connection state (no Redis). Both filters default off, so the default view is unchanged; this supersedes the earlier client-side "Configured Only" toggle (#6245) with an available-only dimension and unit-tested logic. Regression guard: `tests/unit/freeProviderRankings-filters.test.ts`.
+
+### 🔧 Bug Fixes
+
+- **fix(dashboard):** adding a second API key connection for the same provider no longer silently overwrites the first — the Add-API-key modal now derives a unique default connection name (`main`, then `main-2`, `main-3`, …) so the backend name-based upsert can't collide (#6499 — thanks @dilneiss).
+- **fix(compression):** the session-dedup engine now also deduplicates a large multi-line block repeated **within a single message** (intra-message dedup), not just across turns; the compression-preview API surfaces a `fallbackReason`, and the fusion panel reports how many models were rate-limited vs failed on total-panel failure (#6501 — thanks @chirag127).
+- **fix(compression):** a stacked-pipeline step naming an unregistered engine now surfaces a `validationErrors` entry instead of silently no-op'ing, so misconfigured pipelines are visible in the preview API (#6506 — thanks @chirag127).
+- **feat(usage):** add a **Codex reset-credit redemption** flow to the Provider Limits UI ([#6361](https://github.com/diegosouzapw/OmniRoute/pull/6361)) — a `useCodexResetCreditRedemption` hook + `/api/usage/codex-reset-credit` route + `codexResetCredits` lib let you redeem banked Codex reset credits from the quota card. Regression guard: `tests/unit/codex-reset-credits.test.ts`. (thanks @JxnLexn)
+- **feat(glm):** add **team-plan quota settings** for `glm-cn` connections ([#6351](https://github.com/diegosouzapw/OmniRoute/pull/6351)) — a dedicated `GlmTeamQuotaFields` form section (team quota id / limits) threaded through the Add/Edit connection modals, persisted via `providerSpecificData`, with the GLM usage service reading the team quota. Regression guards: `tests/unit/glm-team-quota.test.ts`, `provider-specific-data-schema.test.ts`. (thanks @hao3039032)
+- **feat(providers):** add **TinyFish** web-fetch/search support ([#6349](https://github.com/diegosouzapw/OmniRoute/pull/6349)) — a `tinyfish-fetch` executor + `/v1/web/fetch` route + MCP web-fetch tool, registered as a specialty-media provider with request-validation and a search-provider catalog entry. Regression guards: `tests/unit/executor-tinyfish-fetch.test.ts`, `web-fetch-handler.test.ts`, `mcp-web-fetch-tool.test.ts`, `provider-validation-tinyfish.test.ts`. (thanks @dtybnrj)
+- **fix(cli):** `omniroute launch-codex` now spawns `codex.cmd` through a shell on Windows (the npm `.cmd` shim is unresolvable by bare `spawn` → ENOENT), mirroring the qodercli Windows fix (#6263) ([#6312](https://github.com/diegosouzapw/OmniRoute/pull/6312)). Regression guard: `tests/unit/launch-codex-windows-spawn-6312.test.ts`. (thanks @swingtempo)
+- **fix(codex):** isolate the **Spark** quota from the shared Codex quota and stabilize the quota UI ordering / hydration so per-scope limits render consistently ([#6336](https://github.com/diegosouzapw/OmniRoute/pull/6336)). Regression guards: `tests/unit/codex-quota-selection-hydration.test.ts`, `provider-limits-ui.test.ts` + 3 more. (thanks @xz-dev)
+- **feat(api):** add a `hidePaidModels` setting that filters paid-only models out of the `/v1/models` catalog. Regression guard: `tests/unit/models-catalog-hide-paid.test.ts`. (thanks @chirag127)
+- **fix(api-manager):** the fallback model picker now preserves combos instead of dropping them when a primary model is unavailable ([#6443](https://github.com/diegosouzapw/OmniRoute/pull/6443)). Regression guard: `tests/unit/api-manager-page-static.test.ts`. (thanks @jmengit)
+- **fix(providers):** recoverable Antigravity / Cloud-Code (Gemini Code Assist) `403` responses ([#6452](https://github.com/diegosouzapw/OmniRoute/pull/6452)) are now classified as a retryable project-config error instead of a terminal account ban, so a fixable project/API-disabled 403 no longer forces a ~1-year cooldown / full OAuth reconnect. Regression guard: `tests/unit/errorclassifier-antigravity-403.test.ts`. (thanks @developerjillur)
+- **fix(mitm):** `sanitizeHeaders` now redacts `Set-Cookie` response headers so upstream session cookies never leak into logs / diagnostics ([#6451](https://github.com/diegosouzapw/OmniRoute/pull/6451)). Regression guard: `tests/unit/mitm-sanitize-headers.test.ts`. (thanks @developerjillur)
+- **fix(api):** `/api/compression/preview` now accepts `mode: "caveman"` and correctly handles stacked / zero-compression previews ([#6425](https://github.com/diegosouzapw/OmniRoute/issues/6425)). Regression guard: `tests/unit/api/compression-preview-caveman-and-stacked-6425.test.ts`. (thanks @chirag127)
+- **feat(providers):** add **Zed** hosted LLM aggregator as a native-app provider ([#6118](https://github.com/diegosouzapw/OmniRoute/pull/6118)) — OAuth sign-in via the Zed hosted flow, registered through the shared provider registry + executor. Regression guards: `tests/unit/zed-oauth-provider.test.ts`, `zed-import-utils.test.ts`, `zed-docker-detect.test.ts`, `mitm-handler-zed.test.ts`. VPS-validated via live operator login (Hard Rule #18).
+- **fix(oauth):** the Kiro SSO-cache auto-import now **preserves the IDC region** — cross-region Amazon Q / Kiro profiles imported from the SSO cache are no longer collapsed to the default region ([#6113](https://github.com/diegosouzapw/OmniRoute/pull/6113)). Regression guard: `tests/unit/kiro-auto-import-idc-2059.test.ts`. VPS-validated via live operator login (Hard Rule #18).
+- **fix(dashboard):** passthrough model aliases no longer collide when two namespaced model ids share a last segment (port from 9router#1850, [#6431](https://github.com/diegosouzapw/OmniRoute/pull/6431)). `enx/gpt-5.5` and `enx/codebuddy/gpt-5.5` both auto-generated the alias `gpt-5.5`, so the second model could never be added (the UI just alerted "alias already exists"). Aliases are now disambiguated deterministically — bare last segment when free, then parent-qualified (`codebuddy-gpt-5.5`), then a numeric suffix — while re-adding the exact same model id is still blocked. Regression guard: `tests/unit/passthrough-alias-1850.test.ts`. (thanks @arpicato)
+- **fix(translator):** preserve a Gemini `functionResponse` co-located with other parts (another `functionCall`, or trailing `text`) in the same content when translating **Gemini → OpenAI** ([#6376](https://github.com/diegosouzapw/OmniRoute/pull/6376)). `convertGeminiContent()` early-returned the tool message on the first `functionResponse` part, dropping any co-located parts; such contents are now pre-split (one tool message per `functionResponse`, emitted first, plus one message for the remaining parts). Regression guard: `tests/unit/gemini-to-openai-function-response.test.ts`. (thanks @warelik)
+- **fix(headroom):** detect a python interpreter managed by **mise / pyenv / asdf / conda** (port from 9router#2353, [#6382](https://github.com/diegosouzapw/OmniRoute/pull/6382)). Headroom's python probe (`src/lib/headroom/detect.ts`) searched a hardcoded `PATH`, but version managers expose their interpreters via shim dirs that only join `PATH` through interactive-shell activation — which the non-interactive server never runs, so a managed python (≥3.10) was invisible and Headroom reported it missing. The search path now prepends the well-known shim/bin dirs (`~/.local/share/mise/shims`, `~/.pyenv/shims`, `~/.asdf/shims`, `$CONDA_PREFIX/bin`, `~/.local/bin`, respecting `MISE_DATA_DIR`/`PYENV_ROOT`/`ASDF_DATA_DIR` when set), and a new `HEADROOM_PYTHON` env override lets operators point straight at their interpreter (mirroring `HEADROOM_URL`). Still shell-free (`execFileSync`). Regression guard: `tests/unit/headroom-detect.test.ts` (5). (thanks @loopyd)
+- **fix(executors):** strip the OpenAI-Codex/Claude-CLI `client_metadata` passthrough field for **NVIDIA** requests (port from 9router#1887, [#6411](https://github.com/diegosouzapw/OmniRoute/pull/6411)). NVIDIA's OpenAI-compatible wrapper rejects it with `400 Unsupported parameter`, the same class already handled for `cerebras`/`mistral`; `nvidia` (executor `default`) was missing from the strip allowlist so Codex/Claude-Code passthrough requests 400'd. Regression guard: `tests/unit/executor-default-strip-client-metadata.test.ts` (+nvidia case). (thanks @phidinhmanh)
+- **fix(translator):** strip the Claude-style `thinking` field for **NVIDIA `z-ai/glm-5.2`** (port from 9router#2023, [#6413](https://github.com/diegosouzapw/OmniRoute/pull/6413)). NVIDIA's OpenAI-compatible wrapper 400s on `thinking` (a Claude-format client routed here leaves a `thinking:{type:"adaptive"}`); the existing strip rule only dropped `reasoning`. Same class already handled for `minimax-m2.7`. Regression guard: `tests/unit/nvidia-minimax-thinking-strip.test.ts` (+glm-5.2 case). (thanks @phidinhmanh)
+- **fix(translator):** suppress the streamed `</think>` close marker for the **Antigravity IDE** client (port from 9router#1061, [#6415](https://github.com/diegosouzapw/OmniRoute/pull/6415)). On thinking-only turns Antigravity rendered a bare `</think>` as the sole visible content, tripping its loop-detection and wasting requests. Antigravity's UA (`vscode/<v> (Antigravity/<v>)`) is added to the marker-suppress allowlist (alongside OpenCode); Claude Code / Cursor still get the marker, and `x-omniroute-thinking-marker: on` force-restores it. Regression guard: `tests/unit/think-close-marker-suppress-5245.test.ts`. (thanks @abdofallah)
+- **fix(executors):** strip nested `reasoning_content` from messages for **Mistral** (port from 9router#1649, [#6417](https://github.com/diegosouzapw/OmniRoute/pull/6417)). Mistral's API returns `422 extra_forbidden` when an assistant message carries `reasoning_content` (replayed thinking from a prior turn, e.g. via the Codex `/responses` path); the generic top-level 400 field-downgrade retry never covered the nested per-message field. `DefaultExecutor` now strips it for provider `mistral` only, so DeepSeek (which requires replayed `reasoning_content`) is unaffected. Regression guard: `tests/unit/mistral-strip-reasoning-content-1649.test.ts`. (thanks @xxy9468615)
+- **fix(executors):** strip the `client_metadata` passthrough field on the **OpenCode** path (port from 9router#1442, [#6418](https://github.com/diegosouzapw/OmniRoute/pull/6418)). OpenCode upstreams (e.g. `kimi-k2.6` via opencode-go) reject it with `400 "Extra inputs are not permitted, field: 'client_metadata'"`; the DefaultExecutor strip only covered cerebras/mistral and `OpencodeExecutor` extends `BaseExecutor` directly, so nothing removed it there. Regression guard: `tests/unit/opencode-strip-client-metadata-1442.test.ts`. (thanks @yanpaing007)
+- **fix(executors):** inject the `reasoning_content` echo for the native **Moonshot Kimi** provider (port from 9router#1480, [#6419](https://github.com/diegosouzapw/OmniRoute/pull/6419)). Kimi (executor `default`) is a thinking-mode upstream that 400s with "reasoning_content must be passed back" when a prior assistant turn lacks it; the placeholder injection was only wired into the OpenCode meta-provider, so direct multi-turn Kimi conversations failed. Scoped to `kimi` (gateway-served models matching the thinking-model name pattern are unaffected). Regression guard: `tests/unit/kimi-native-reasoning-injected-1480.test.ts`. (thanks @2220258345)
+- **fix(executors):** recover from a strict gateway's `context_management: Extra inputs are not permitted` 400 (port from 9router#1468, [#6420](https://github.com/diegosouzapw/OmniRoute/pull/6420)). **Claude Code** always sends a top-level `context_management` field; strict anthropic-compatible gateways reject it. The dedicated context-editing 400-fallback only fired when OmniRoute's own `contextEditing` feature was enabled (default off), so a client-sent field passed through untouched and 400'd. `context_management` is now in the generic reactive field-strip list, so it's stripped-and-retried once regardless of the feature flag (with correct request re-signing for claude-compatible relays). Regression guard: `tests/unit/provider-field-strips.test.ts`. (thanks @ohahe52-dot)
+- **fix(network):** enable **RFC 8305 Happy Eyeballs** (`autoSelectFamily`) on the direct-egress undici dispatcher (port from 9router#1237, [#6423](https://github.com/diegosouzapw/OmniRoute/pull/6423)). When DNS returns both IPv6 (AAAA) and IPv4 (A) and the IPv6 route is broken (e.g. a NAT64 `64:ff9b::` prefix without routing), undici tried IPv6 first and hung until `ETIMEDOUT` (then a 502 + account lockout), even though `curl` reached the same host. The direct dispatcher now races both families and uses whichever connects first. Proxy paths pin family via `proxyTls` and are unaffected. Regression guard: `tests/unit/direct-dispatcher-pipelining-4580.test.ts`. (thanks @adentdk)
+- **fix(combo):** round-robin now advances the rotation pointer past the model that **actually served**, not the eagerly-scheduled one (port from 9router#948, [#6428](https://github.com/diegosouzapw/OmniRoute/pull/6428)). With `stickyLimit: 1` (true round-robin), when the scheduled model failed and a _different_ model served via fallback, the counter had already advanced +1 from the scheduled index — so the next request reused the fallback-served model, degrading round-robin into hot-spotting on whichever model was healthy. The pointer now advances to the served index + 1 (mirroring the sticky-limit>1 path). Session-stickiness (#3825) and distribution are preserved. Regression guard: `tests/unit/combo-rr-fallback-advance-948.test.ts`. (thanks @binsarjr)
+- **fix(sse):** a non-string `model` field is now rejected with a `400` before the resolver, instead of crashing downstream `.toLowerCase()`/`.split()` calls into an empty-body `500` that escapes the error sanitizer ([#6407](https://github.com/diegosouzapw/OmniRoute/issues/6407)). Regression guard: `tests/unit/chat-non-string-model-6407.test.ts`. (thanks @chirag127)
+- **fix(api):** unknown `/api/*` routes now return a JSON `404` (instead of the dashboard HTML shell) and scalar chat params (`model`/`temperature`/etc.) are validated **before** the provider lookup so malformed requests fail fast with a clear `400` ([#6424](https://github.com/diegosouzapw/OmniRoute/issues/6424), [#6412](https://github.com/diegosouzapw/OmniRoute/issues/6412)). Regression guards: `tests/unit/api/api-catchall-json-404.test.ts`, `tests/unit/chat-early-schema-validation-6412.test.ts`. (thanks @chirag127)
+- **fix(api):** `/v1/chat/completions` now rejects a non-JSON `Content-Type` with a `400` before parsing the body ([#6414](https://github.com/diegosouzapw/OmniRoute/issues/6414)). Regression guard: `tests/unit/v1-chat-completions-content-type-6414.test.ts`. (thanks @chirag127)
+- **fix(api):** the `X-OmniRoute-Compression` response header is now echoed on `/v1/chat/completions` and `/v1/completions` ([#6422](https://github.com/diegosouzapw/OmniRoute/issues/6422)). Regression guard: `tests/unit/compression-header-echo-6422.test.ts`. (thanks @chirag127)
+- **fix(api):** concurrent `GET /v1/models` requests are coalesced into a single catalog build ([#6408](https://github.com/diegosouzapw/OmniRoute/issues/6408)). Regression guard: `tests/unit/v1-models-concurrent-6408.test.ts`. (thanks @chirag127)
+- **fix(api):** `/v1/completions` now echoes the requested `body.model` in its JSON + streamed responses ([#6429](https://github.com/diegosouzapw/OmniRoute/pull/6429)). Regression guard: `tests/unit/completions-body-model-echo.test.ts`. (thanks @chirag127)
+- **fix(api):** env-var master keys now see the full `/v1/models` catalog ([#6406](https://github.com/diegosouzapw/OmniRoute/issues/6406)). Regression guard: `tests/unit/models-catalog-envkey-6406.test.ts`. (thanks @chirag127)
+- **fix(api):** non-streaming `/v1/completions` responses now echo `body.model` aligned with the `X-OmniRoute-Model` header ([#6426](https://github.com/diegosouzapw/OmniRoute/issues/6426)). Regression guard: `tests/unit/v1-completions-model-header-match-6426.test.ts`. (thanks @chirag127)
+- **fix(api):** unknown `/v1/*` routes now return a JSON `404 not_found` instead of the Next.js dashboard HTML shell ([#6405](https://github.com/diegosouzapw/OmniRoute/issues/6405)). Regression guard: `tests/unit/api/v1-catchall-json-404.test.ts`. (thanks @chirag127)
+- **fix(api):** the per-connection provider models route now degrades to the shipped catalog when a provider's `/models` endpoint answers with a redirect ([#6267](https://github.com/diegosouzapw/OmniRoute/issues/6267)) — a `qwen-web` import failed with a raw `Redirect blocked … (307)` 503. `safeOutboundFetch` throws `REDIRECT_BLOCKED` on the 307, `getSafeOutboundFetchErrorStatus` maps it to 503, and `buildDiscoveryErrorFallbackResponse` treated every 503 as a hard error — so the non-empty `getModelsByProviderId("qwen-web")` catalog was never surfaced. A models-endpoint redirect is not a fixable-config error (unlike `URL_GUARD_BLOCKED`/`INVALID_URL`, which stay hard errors), so it now falls back to the local/cached catalog before the 503 short-circuit. General fix — covers any config-driven provider that 307s. Regression guard: `tests/unit/provider-models-qwen-web-redirect-6267.test.ts`. (thanks @chirag127)
+- **fix(api):** the per-connection provider models route (MCP `list_models_catalog` + the dashboard import view) now merges USER-ADDED custom models into its response ([#6247](https://github.com/diegosouzapw/OmniRoute/issues/6247)) — custom models live in the `key_value` namespace `customModels`, which the live REST `/api/v1/models` already merges, but `src/app/api/providers/[id]/models/route.ts` never read `getCustomModels`, so custom models were dropped on both the discovery-success and local_catalog paths. They are now folded into the returned model list (deduped by id, stamped `owned_by: provider`), fixing MCP + the dashboard import view in one place. Regression guard: `tests/unit/provider-models-custom-merge-6247.test.ts`. (thanks @RCrushMe)
+
+- **fix(providers):** GitLab Duo tool-calling follow-up turns no longer fail upstream with `422 {"detail":"Validation error"}` (tokens 0/0, rejected pre-inference). The #6234 tool-result-feedback fix serialized the **entire** multi-turn conversation into GitLab's single-file `code_suggestions` (`small_file`) generation endpoint — folded history that turn-N sent as an oversized `current_file.content_above_cursor` **and** duplicated verbatim into `user_instruction`, tripping the AI-Gateway's `small_file` validation guard. The executor now **bounds** that prompt: it keeps system + latest user message + the most-recent tool round (dropping older turns), caps oversized tool results, and stops duplicating the full prompt into `user_instruction` (which now carries only the short latest user message) — while still feeding the most-recent tool result back so the agent continues (`open-sse/executors/gitlab.ts`, [#6220](https://github.com/diegosouzapw/OmniRoute/issues/6220)). The unit test covers the bounding logic; the upstream 422→200 clearing is VPS-only (Hard Rule #18). Regression guard: `tests/unit/gitlab-tool-exchange-bounded-6220.test.ts`.
+
+- **fix(i18n):** the provider-detail (`/dashboard/providers/[id]`) connection-status filter labels no longer render as `__MISSING__:All` / `__MISSING__:Active` / `__MISSING__:Error` / `__MISSING__:Banned` / `__MISSING__:CreditsExhausted` in non-English locales (notably pt-BR) ([#6290](https://github.com/diegosouzapw/OmniRoute/issues/6290)). Root cause was **not** the namespace mismatch the issue guessed — the `providers.filter*` keys resolve correctly in `en.json`; the debt lived in the locale mirrors (`src/i18n/messages/*.json`), where these five keys carried the `__MISSING__:` sync sentinel in ~15 locales and were absent entirely in ~26 others, so next-intl found the key and echoed the sentinel verbatim. All 40 non-English/-Chinese mirrors now ship real translations for the five `providers.filter*` labels. Regression guard: `tests/unit/i18n-provider-filter-keys-6290.test.ts`. (thanks @diegosouzapw)
+
+- **fix(providers):** the `copilot-m365-web` streaming executor now emits `debug`-level WebSocket diagnostics ([#6210](https://github.com/diegosouzapw/OmniRoute/issues/6210)) — the outbound WS URL (with the `access_token` **redacted** via `redactWsUrl()`), handshake success/failure, and each received SignalR frame's `type`/`target`. Previously the streaming path logged nothing, so an empty `content:null` response (the M365 Education / Starter tier symptom fixed in #6234) was undiagnosable even at `APP_LOG_LEVEL=debug`. The change is debug-level and side-effect-free — it does not alter streaming behavior or the frame parser, and the token never reaches the logs. Regression guard: `tests/unit/copilot-m365-web-logging-6210.test.ts` (thanks @qpeyba)
+
+- **fix(resilience):** a round-robin combo no longer returns `503 all upstream accounts are unavailable` when a compatibility-rejected target is actually healthy ([#6238](https://github.com/diegosouzapw/OmniRoute/issues/6238)). `filterTargetsByRequestCompatibility` drops request-incompatible targets (tool/vision/structured-output unsupported, or below the required context window) **before** any availability check runs, and its `compatible.length === 0` safety net only fired when _all_ targets were filtered — not when the kept targets later all turned out runtime-unavailable (circuit-open / cooldown / no credentials). So a combo could 503 while a compat-rejected-but-healthy provider sat unused. `handleRoundRobinCombo` now keeps the compat-rejected set and, when every compat-kept target was skipped without a single real attempt, probes those rejected targets as a **last-resort fallback tier** (via the new pure `open-sse/services/combo/comboCompatFallback.ts`) before crystallizing the 503. Regression guard: `tests/unit/combo-roundrobin-compat-fallback-6238.test.ts`. (thanks @ThongAccount)
+
+- **fix(startup):** best-effort self-heal for a corrupted Turbopack dev cache on Windows ([#6289](https://github.com/diegosouzapw/OmniRoute/issues/6289)). On Windows, `pnpm dev` can fail at startup when Turbopack `mmap`s a persistent-cache SST file and the OS refuses the mapping (`os error 1455` — "paging file too small"), which Turbopack surfaces as a misleading `Module not found: Can't resolve '@/shared/utils/machine'`. This is a **known upstream Turbopack cache-corruption bug — not our code**. The dev launcher (`scripts/dev/run-next.mjs`) now wraps `nextApp.prepare()` and, when it rejects with that signature (`isTurbopackCacheCorruption` in the new `scripts/dev/turbopackCacheHeal.mjs`), purges `.build/next/**/cache/turbopack` and retries **once** with a clear log. **Caveat — best-effort only:** the corruption often surfaces as a runtime overlay rather than a `prepare()` rejection, so this cannot always intercept it; the reliable remedy remains manually deleting the Turbopack cache dir. Regression guard: `tests/unit/turbopack-cache-heal-6289.test.ts`. (thanks @chirag127)
+
+- **fix(providers):** qodercli PAT auth no longer fails with `spawn qodercli ENOENT` on Windows ([#6263](https://github.com/diegosouzapw/OmniRoute/issues/6263)) — `spawnQoderCli` spawned the bare `qodercli` name with `shell:false` and an unenriched env, so the npm `.cmd` wrapper under `%APPDATA%\npm` (a user-PATH directory) was never resolved. It now resolves the absolute `.cmd`/`.exe` path through the existing `getCliRuntimeStatus("qoder")` resolver in `src/shared/services/cliRuntime.ts` (memoized), spawns with `shell` when the target is a `.cmd`/`.bat`, and uses the cliRuntime-enriched env (PATH + PATHEXT + APPDATA); the ENOENT error now lists the searched paths plus the `CLI_QODER_BIN` override. End-to-end spawn on a real Windows host is host-only (Hard Rule #18); the path-resolution logic is unit-tested. Regression guard: `tests/unit/qodercli-windows-resolve-6263.test.ts`. (thanks @chirag127)
+
+- **fix(sse):** the reasoning-token buffer no longer inflates **probe-sized `max_tokens`** ([#6274](https://github.com/diegosouzapw/OmniRoute/issues/6274)) — Claude Code's `/model` capability check sends `max_tokens: 1`, but for a thinking-capable model with a large output cap (e.g. `glm-5.2`) the #3587 headroom heuristic (`max(current + 1000, ceil(current * 1.5))`) rewrote it to `1001` and forwarded that upstream, wasting tokens on a request that was never a genuine reasoning budget. `resolveReasoningBufferedMaxTokens()` (`open-sse/services/reasoningTokenBuffer.ts`) now short-circuits and returns the caller's value verbatim when it is below the new `REASONING_BUFFER_MIN_TRIGGER` (256) threshold — a tiny explicit limit is a probe, not a reasoning request. Real budgets still receive the #3587 headroom unchanged, and the guard runs after the existing capability checks so unknown / non-reasoning models keep returning `null`. Regression guard: `tests/unit/reasoning-token-buffer-6274.test.ts`. (thanks @brightfiscalband)
+
+- **fix(cli):** `omniroute reset-password` now works as a real subcommand, and password resets over piped (non-TTY) stdin actually apply ([#6261](https://github.com/diegosouzapw/OmniRoute/issues/6261), [#6258](https://github.com/diegosouzapw/OmniRoute/issues/6258)). Two coupled defects: (1) **#6261** — `bin/omniroute.mjs` routed everything through Commander with only two pre-Commander bypasses (`--mcp`, `reset-encrypted-columns`), so `omniroute reset-password` was rejected as an unknown command; only the separate `omniroute-reset-password` bin worked, while the docs falsely advertised the subcommand (incl. a bogus "legacy alias still works"). A pre-Commander bypass mirroring `reset-encrypted-columns` now dynamically imports `bin/reset-password.mjs` (which self-executes) before Commander parses; the three doc lines were corrected. (2) **#6258** — `bin/reset-password.mjs` issued two sequential `rl.question` prompts; under piped stdin the second read never settled at EOF, so `main()` never reached `resetManagementPassword` and the reset was a silent no-op (both prompts printed, no success, password unchanged). The CLI now detects non-TTY stdin and reads it once (first line = password, second line = confirm if present, else reused), adds a `--password-stdin` flag (entire stdin is the password, no confirmation), and exits `0` explicitly so the success line always flushes; interactive TTY behavior is unchanged. Regression guard: `tests/unit/reset-password-cli-6261-6258.test.ts` (3). (thanks @chirag127)
+
+- **fix(db):** the mass-migration **safety abort** now tells the operator how to bypass it and stops flooding the log ([#6260](https://github.com/diegosouzapw/OmniRoute/issues/6260)) — after restoring a backup that wiped the migration tracking table, `runMigrations()` threw the abort on every downstream `ensureDbInitialized()`, re-logging the full banner 11+ times, and the message never mentioned the existing `OMNIROUTE_MAX_PENDING_MIGRATIONS` escape hatch. The abort text now appends a bypass hint (set `OMNIROUTE_MAX_PENDING_MIGRATIONS=0` in `server.env` / `DATA_DIR/.env`), and a new `MigrationSafetyAbortError` is memoized so repeated calls in the same process throw the same instance and emit a single concise line instead of the full cascade. Regression guard: `tests/unit/migration-safety-abort-6260.test.ts`. (thanks @chirag127)
+
+- **fix(auth):** importing a **distinct** Codex/ChatGPT OAuth `auth.json` is no longer falsely rejected as "already exists" when it belongs to a different user in the same workspace ([#6301](https://github.com/diegosouzapw/OmniRoute/issues/6301)). `findExistingCodexConnection` (in `src/lib/oauth/utils/codexAuthImport.ts`) deduped **only** on `providerSpecificData.workspaceId === accountId`, where `accountId` is the shared `chatgpt_account_id`/`tokens.account_id` — so two members of the same ChatGPT Team collapsed onto a single connection (409 `duplicate_account`). The id_token's `https://api.openai.com/auth` claim carries a per-user `chatgpt_user_id` alongside the workspace id (the device-flow path already persisted it as `chatgptUserId`, but the import path did not). Now `parseAndValidateCodexAuth` extracts `userId` (`chatgpt_user_id` → `user_id` → JWT `sub`) into `ParsedCodexAuth`, the create/update paths persist `chatgptUserId` in `providerSpecificData` (mirroring `codex.ts`), and dedup keys on `workspaceId` **AND** `chatgptUserId` — with a backward-compat fallback to legacy accountId-only matching when no stored connection for that workspace records a `chatgptUserId`, so genuinely-same accounts still dedup. Regression guard: `tests/unit/codex-auth-import-userid-dedup-6301.test.ts` (4). (thanks @anungma)
+
+- **fix(providers):** importing models for the **venice-web** provider no longer fails with a red "Provider venice-web does not support models listing" ([#6269](https://github.com/diegosouzapw/OmniRoute/issues/6269)). `venice-web` is a web-cookie provider with an executor but no upstream `/v1/models` endpoint and no registry `models`, so the models route fell through to the tail `400`. Mirroring the `jules`/`linkup-search`/`ollama-search` fix (#5569), it now ships a static local catalog entry in `src/lib/providers/staticModels.ts` — seeding the current Venice lineup (`venice-uncensored`, `llama-3.3-70b`, `qwen3-235b`, `qwen3-4b`, `deepseek-r1-671b`; Venice rotates its catalog, see docs.venice.ai/models/overview) — so the route returns `200` with `source:"local_catalog"`, `intentional:true`. Regression guard: `tests/unit/static-models-venice-web-6269.test.ts`. (thanks @chirag127)
+
+- **fix(api):** the specialty model catalogs (`/v1/embeddings`, `/v1/images`, `/v1/music`, `/v1/videos` model lists) are now derived from the **unified catalog filtered by a predicate** (`getSpecialtyModelsResponse`) instead of ad-hoc per-route logic, so they consistently respect active-credential visibility and stay in sync with the main catalog ([#6303](https://github.com/diegosouzapw/OmniRoute/pull/6303)). Regression guard: `tests/unit/specialty-model-catalog-routes.test.ts`. (thanks @makcimbx)
+- **fix(api):** the agent-bridge server route now resolves the MITM manager via a **dynamic `import("@/mitm/manager.runtime")`** so Turbopack does not statically pull the stub (or over-bundle the manager), and the agent-skills generator anchors its output base path with `path.join(process.cwd(), …)` so Turbopack's static analyzer stops tracing the whole project root ([#6329](https://github.com/diegosouzapw/OmniRoute/issues/6329), [#6366](https://github.com/diegosouzapw/OmniRoute/pull/6366)). Regression guard: `tests/unit/agent-bridge-server-route-dynamic-import.test.ts`. (thanks @Iammilansoni)
+- **fix(api):** internal probes (combo-test, cloud-sync verify) now pick a **management-scoped / allow-all API key** instead of naively grabbing `getApiKeys()[0]` — a restricted `self:usage` first row made the probe fail with "Model X is not allowed for this API key" even when the combo path was healthy (`pickApiKeyForInternalUse` in `src/lib/db/apiKeys.ts`). The API-manager model editor also falls back to `/api/models?all=true` when `/v1/models` is catalog-protected ([#6372](https://github.com/diegosouzapw/OmniRoute/pull/6372)). Regression guard: `tests/unit/pick-internal-api-key-6372.test.ts`. (thanks @jmengit)
+- **fix(live-ws):** the Live Dashboard WebSocket server now **rejects on bind failure** (e.g. `EADDRINUSE` when the API bridge already holds the port) instead of letting the error surface as an unhandled `error` event that crash-loops the process — the `error` listener is attached to `wss` (not `server`) and releases the EventBus subscription on a failed start ([#6324](https://github.com/diegosouzapw/OmniRoute/issues/6324)). Regression guard: `tests/unit/live-ws-eaddrinuse-6324.test.ts`. (thanks @vinayakkulkarni)
+- **fix(dashboard):** the Home provider-topology widget now trusts the live provider-metrics snapshot — it uses `topology.errorProvider` and live `activeRequests` directly instead of re-deriving state from a stale `lastErrorAt` or applying a frontend timeout filter, so the topology reflects real-time provider health ([#6322](https://github.com/diegosouzapw/OmniRoute/pull/6322)). Regression guard: `tests/unit/home-provider-topology-live-state.test.ts`. (thanks @xz-dev)
+- **fix(sse):** strip zero-width markers from streamed **tool-call arguments** — a follow-up to [#5857](https://github.com/diegosouzapw/OmniRoute/pull/5857). That PR removed injected zero-width joiners (U+200D) from streamed assistant text/reasoning but deliberately left tool-call argument JSON byte-exact. The request-side obfuscation (`open-sse/services/claudeCodeObfuscation.ts`) injects ZWJ into agent words — including the temp path inside the Bash tool description — and Claude models copy that verbatim into generated commands, which are delivered as tool-call arguments rather than assistant text. As a result the ZWJ survived and corrupted code blocks (e.g. a temp path rendered with an invisible joiner). Now `open-sse/handlers/responseSanitizer.ts` strips zero-width code points from tool-call argument strings at every emit site (OpenAI non-stream/stream chat `tool_calls` + legacy `function_call`, native Responses `function_call` items, the OpenAI→Responses conversion, and the native Responses streaming `response.function_call_arguments.delta/.done` events). Only zero-width code points are removed; JSON structure and all other bytes stay identical (no parse/restringify), so normal arguments remain byte-exact. Regression guard: 6 new cases in `tests/unit/response-sanitizer.test.ts` (suite 50/50).
+
+- **fix(nodejs):** the default app log path now resolves under `DATA_DIR` (`~/.omniroute/logs/application/app.log`) instead of `process.cwd()` ([#6197](https://github.com/diegosouzapw/OmniRoute/issues/6197)) — the globally-installed CLI runs from an arbitrary working directory, so anchoring the default to cwd made file logging silently write to (or no-op under) an unrelated directory, contradicting the documented `.env.example` default. `getAppLogFilePath()` now computes the default lazily via the pure `resolveDataDir()` resolver (honours a per-process `DATA_DIR`, no directory-creation side effect); an explicit `APP_LOG_FILE_PATH` still wins. Regression guard: `tests/unit/logenv-datadir-path-6197.test.ts` (3).
+
+- **fix(docker):** AgentBridge/`startMitm` no longer aborts in containers/headless when the Antigravity-default DNS step can't write `/etc/hosts` ([#6127](https://github.com/diegosouzapw/OmniRoute/issues/6127)), and the privileged command's stderr now reaches `app.log` instead of only a bare exit code hitting the toast ([#6198](https://github.com/diegosouzapw/OmniRoute/issues/6198)). The default DNS step (`addDNSEntry`) was called unguarded while cert install and the two sibling DNS steps were each best-effort — in the runtime Docker image (`USER node`, no `sudo`, read-only `/etc/hosts`) it threw `Command failed with code 1` out of `startMitmInternal` and killed the whole start, discarding the stderr. The three DNS steps are extracted into a best-effort `provisionDnsEntries()` where each failure is logged with the full `err` (stderr included, folded in by `systemCommands.ts`) and never aborts the start. Regression guard: `tests/unit/mitm-dns-graceful-degrade-6127.test.ts` (4).
+
+- **fix(providers):** copilot-m365-web now supports the M365 Education "Starter / OfficeWebIncludedCopilot" tier and no longer returns an empty `content:null` stream ([#6210](https://github.com/diegosouzapw/OmniRoute/issues/6210)). Two gaps: (1) `buildWsUrl()` hardcoded the individual-consumer scenario (`OfficeWebPaidConsumerCopilot`, `isEdu=false`) — the EDU tier is now opt-in via `providerSpecificData.tier="edu"`, emitting `scenario=OfficeWebIncludedCopilot`/`isEdu=true` (the individual path is unchanged); (2) the EDU/GPT-5.5 path streams deltas via `arguments[0].writeAtCursor` (incremental) instead of only `messages[].text` (accumulated snapshots), which the parser dropped — a new `accumulateBotContent()` folds both formats, with `type:2 item.result.message` as a last-resort fallback. Regression guard: `tests/unit/copilot-m365-edu-writeatcursor-6210.test.ts` (10). (thanks @qpeyba)
+
+- **fix(providers):** GitLab Duo executor now feeds tool results back into the prompt instead of looping ([#6220](https://github.com/diegosouzapw/OmniRoute/issues/6220)) — `buildPrompt()` branched only on `system`/`user` and took `userParts.at(-1)`, silently dropping the `assistant{tool_calls}` + `tool{result}` turns the client appended, so the reconstructed prompt was byte-identical to turn 1 and the model re-emitted the same `<tool>` call forever. When a tool exchange is present the full conversation is now serialized, folding each tool result back keyed by its `tool_call_id`; simple conversations keep the legacy shape. Complements the tool_call emission from [#6051](https://github.com/diegosouzapw/OmniRoute/issues/6051) (the `kilo-duplicate` label was a false positive — different, sequential defect). Regression guard: `tests/unit/gitlab-tool-result-feedback-6220.test.ts` (4).
+
+- **fix(providers):** opencode-go/opencode-zen can now synthesize the OpenCode CLI identity headers Cloudflare requires on VPS egress ([#5997](https://github.com/diegosouzapw/OmniRoute/issues/5997)) — on a datacenter VPS, `opencode.ai/zen/go/v1/chat/completions` 403s (HTML challenge) requests lacking CLI identity, while the reporter's control curl proved that `User-Agent: opencode-cli/1.0.0` + `x-opencode-client: cli` + `x-opencode-project: default` + fresh request/session UUIDs succeed. Opt-in via `OPENCODE_SYNTHESIZE_CLI_HEADERS=true` (values overridable via `OPENCODE_GO_USER_AGENT`/`OPENCODE_USER_AGENT`/`OPENCODE_CLIENT`/`OPENCODE_PROJECT`); it fills only headers the client did not already send. Kept **off by default** — the forward-only path is deliberate (fabricating a wrong value risks upstream rejection; a prior dedup regressed with `opencode/local`), so this replaces the fragile local header-injection shim without changing default behavior. Regression guard: `tests/unit/opencode-cli-headers-synthesis-5997.test.ts` (6). (thanks @aleksesipenko)
+- fix(resilience): sticky session affinity now evicts and fails over to another account when the pinned account is exhausted/unavailable (#6219)
+
+- fix(sse): Responses API passthrough now drops internal commentary-phase output before forwarding to clients (gated by RESPONSES_PASSTHROUGH_DROP_COMMENTARY, default on) (#6199)
+
+- **fix(sse):** tool-call function schemas with a root `type: null` are now coerced to `type: "object"` before dispatch ([#6359](https://github.com/diegosouzapw/OmniRoute/issues/6359)) — clients like the Codex app emit `parameters: { type: null, ... }` for some tools, which OpenAI-compatible upstreams reject with `400 Invalid schema for function '...': schema must be a JSON Schema of 'type: "object"', got 'type: null'`, failing the whole request. `toolSchemaSanitizer` already stripped the null; it now re-adds the mandatory root `"object"` type (and empty `properties`/open `additionalProperties` when absent). Combinator roots (`anyOf`/`oneOf`/`allOf`) and explicit root types are left untouched. Regression guard: 5 new cases in `tests/unit/tool-schema-sanitizer.test.mjs`.
+
+- **fix(docker):** AgentBridge no longer fails to start on npm/Electron/VPS installs with "MITM manager stub reached at runtime" ([#6344](https://github.com/diegosouzapw/OmniRoute/issues/6344)) — v3.8.45 flipped the production bundler default to Turbopack, but `next.config.mjs` aliased `@/mitm/manager` to its Docker-only degraded stub **unconditionally**. That was harmless while Docker (which sets the alias intentionally for #3390 graceful degradation) was the sole Turbopack consumer, but once every artifact built with Turbopack the stub shipped to all non-Docker users and `startMitm` threw on the first Agent-Bridge start. The alias is now opt-in via `OMNIROUTE_MITM_STUB=1` (set only by the Dockerfile) through the shared `scripts/build/mitm-stub-flag.mjs` helper; default builds bundle the real manager. Regression guard: `tests/unit/mitm-stub-alias-6344.test.mjs` (4).
+
+- **fix(proxy):** stop the v3.8.44 proxy regression that leaked the real IP and disabled healthy proxies ([#6246](https://github.com/diegosouzapw/OmniRoute/issues/6246)). Two coupled defects from the new health scheduler: (1) **IP leak** — when a proxy assigned to a connection was marked `inactive`, resolution fell through to a **direct** egress instead of blocking, exposing the operator's real IP; (2) **over-deactivation** — the sweep flipped a proxy to `inactive` on the **first** failed probe and counted our own 5s timeout / a probe-target `5xx` as the proxy's fault, so healthy paid proxies vanished from egress selection ("my proxies are not being used anymore"). Fix: the sweep decision is extracted into a pure, network-free `decideProxyHealthAction` (`src/lib/proxyHealth/decision.ts`) — by default the health check now **only counts/logs and never downgrades status** (a proxy is downgraded/removed only with `PROXY_AUTO_REMOVE=true`, after `PROXY_AUTO_REMOVE_AFTER` **consecutive** conclusive failures); probes are classified tri-state so an inconclusive result (our timeout, or a `5xx` from the probe target) never penalizes the proxy, and the probe timeout is raised 5s→15s. Separately, `safeResolveProxy` now **fails closed** via the existing policy: a connection whose assigned proxy is dead is blocked instead of leaking direct (`hasBlockingProxyAssignment`), honoring the explicit `proxy off` toggles and the `PROXY_FAIL_OPEN=true` opt-out. Existing proxies stuck `inactive` by the old behavior need a one-time manual re-activate (the operator owns proxy status). Regression guards: `tests/unit/proxy-health-decide-action-6246.test.ts`, `tests/unit/proxy-assigned-unavailable-6246.test.ts`.
+
+- **fix(proxy):** make "Test All" read-only and add bulk enable/disable ([#6246](https://github.com/diegosouzapw/OmniRoute/issues/6246)). Complements the core fail-closed / scheduler fix (#6296) with the two remaining reporter asks. (1) The **"Test All" button** (`POST /api/settings/proxies/auto-test`) used to flip a proxy to `inactive` on a failed reachability probe; since the egress selector excludes `inactive` proxies, a flaky probe (an unreachable `httpbin.org`, a proxy that blocks `HEAD`, or a slow paid proxy) silently disabled every proxy that failed — "Test All" is now **read-only by default** (only the operator sets a proxy active/inactive; opt back into the legacy test-and-set with `PROXY_HEALTH_AUTO_DEACTIVATE=true`). (2) Adds a **bulk enable/disable** proxies endpoint + toolbar action (`POST /api/settings/proxies/batch-activate`) so an operator can re-activate proxies in one click. Regression guard: `tests/unit/proxy-health-6246.test.ts`. (thanks @tenshiak)
+
+- **chatcore (tools): stop the default 128-tool cap from silently dropping opencode's `task`/MCP tools.** opencode (used as an MCP/agent host) sends a large tool list; when it exceeds the speculative `MAX_TOOLS_LIMIT` (128) default, `truncateToolList` did a blind `tools.slice(0, 128)`, dropping every tool past index 128 — including opencode's built-in `task` tool (subagent launch) and many MCP tools, so models routed through OmniRoute could no longer spawn subagents or reach part of their tools. The cap exists to avoid upstream `400`s for providers with real hard limits (e.g. grok-cli 200), so it is kept for those: detection of the opencode client (`isOpencodeClient` — any `x-opencode-*` header, or `opencode` in the user-agent) now only bypasses the **speculative 128 default**, never a known provider ceiling. Precedence is explicit — a proactive/detected provider limit always truncates (even for opencode); otherwise opencode forwards its full tool list; otherwise the unchanged 128 default applies to every other client. Refactors `getEffectiveToolLimit` into `getKnownToolLimit(provider) ?? DEFAULT_LIMIT` (byte-identical for existing callers) and fixes a cosmetic debug-log that reported the truncated count instead of the original. Regression guard: `tests/unit/tool-limit-detector.test.ts`.
+
+- **fix(mitm):** the macOS MITM-cert install check now matches the system keychain again. `security find-certificate -a -Z` prints the SHA-1 as a colon-less hex string, but the installed-check compared it against `getCertFingerprint()`'s colon-separated form, so the substring match never hit — the cert was reported as not-installed and re-prompted for the sudo install on every run. Fingerprints are now normalized (colons stripped, upper-cased) on both sides via the extracted `macCertOutputHasFingerprint` helper. Regression guard: `tests/unit/mitm-cert-mac-fingerprint.test.ts`. ([#6204](https://github.com/diegosouzapw/OmniRoute/pull/6204), closes [#6134](https://github.com/diegosouzapw/OmniRoute/issues/6134) — thanks @rianonehub)
+
+- **fix(api):** `/v1/messages/count_tokens` now counts `tool_use`, `tool_result` and `thinking` content blocks (and array-form `system` prompts) in the local-estimation path, instead of only `text`. Real agentic conversations keep ~95% of their tokens inside tool results; the previous estimate returned near-zero for them, which silently broke Claude Code's auto-compaction (context grew past the window with no compaction until the upstream API rejected the request). The real provider-side count path is unchanged. Regression guard: `tests/unit/messages-count-tokens-route.test.ts`. ([#6221](https://github.com/diegosouzapw/OmniRoute/pull/6221) — thanks @luweiCN)
+
+- **fix(antigravity):** strip a trailing assistant prefill turn for Vertex Claude models to avoid upstream 400s ([#6114](https://github.com/diegosouzapw/OmniRoute/pull/6114)). Regression guard: `tests/unit/antigravity-claude-prefill-strip.test.ts`. (thanks @anki1kr)
+
+- **fix(security):** the mutable cloud-agent routes (`/api/cloud/credentials/update`, `/api/cloud/models/alias`) now require management auth instead of being treated as public. They were classified as public API routes, so a request without management credentials could update stored cloud-agent credentials and model aliases. They are removed from the public-route set, classified as management routes in the authz pipeline, and gated by `requireManagementAuth`; cloud **read**/auth routes stay public. Regression guards: `tests/unit/cloud-write-auth.test.ts`, `tests/unit/authz/classify.test.ts`, `tests/unit/public-api-routes.test.ts`. ([#6233](https://github.com/diegosouzapw/OmniRoute/pull/6233) — thanks @vittoroliveira-dev)
+
+- **refactor(dashboard):** extract the onboarding-wizard "Open provider details" link target into a pure, unit-tested `buildProviderDetailsHref(connection)` helper. The wizard already routes by `connection.id` (the node UUID) rather than the provider category slug (#6144/#6145); this hardens that behavior behind a tested helper that guards a missing id/connection. Regression guard: `tests/unit/provider-onboarding-href.test.ts`. ([#6166](https://github.com/diegosouzapw/OmniRoute/pull/6166) — thanks @KooshaPari)
+
+- **fix(security):** the doubao synthetic device-id generator now derives its digits via an unbiased crypto-random draw (rejection sampling over `crypto.randomBytes()`) instead of a `% 10` reduction, closing a CodeQL `js/biased-cryptographic-random` finding.
+- **fix(agentSkills):** the GitHub-skills generator now resolves `outputDir` to an absolute path before writing, fixing a regression introduced by #6366 (relative-to-cwd base path) that could write generated skill files to the wrong directory.
+- **fix(security):** `/api/keys/{id}/devices` now checks the HTTP method before auth/validation, returning a `405` for non-GET/DELETE verbs instead of a misleading `401`/`500` (closes a `dast-smoke` QUERY-method finding).
+- **fix(quality):** clear the last 2 heavy quality-gate reds on the release tip (cycle pre-flight).
+- **fix(mitm):** the test suite and CI can never mutate the OS trust store — `OMNIROUTE_SKIP_SYSTEM_TRUST=1` is set globally for tests/CI so `installCert`/`uninstallCert`/`installTproxyCa` skip the privileged OS dispatch ([#6310](https://github.com/diegosouzapw/OmniRoute/pull/6310); full detail is under the [3.8.45] section below — this branch received it via the parallel-cycle sync-back).
+- **fix(api):** `POST /api/github-skills` now Zod-validates its request body; documented the new quality-gate env vars and pinned the merge-integrity GitHub Actions to a commit SHA.
+- **fix(skills):** generate the missing `omni-github-skills` registry entry and align the agent-skills catalog-count tests (follow-up to #6186).
+- **fix(quality):** clear the cycle's 11 net-new ESLint errors and make `validate-release-green` suppressions-aware.
+
+### 📝 Maintenance
+
+- **i18n(it):** add 118 missing Italian (`it`) translations (net-additive — no existing keys dropped, valid JSON), improving Italian UI coverage. ([#6212](https://github.com/diegosouzapw/OmniRoute/pull/6212) — thanks @serverless83)
+- **chore(providers):** remove deprecated **MiMo V2** model entries from the catalogs (xiaomi-mimo, opencode-go, zenmux-free, audio TTS) — the upstream V2 line is superseded by MiMo V2.5; drops `mimo-v2-tts`, `mimo-v2-pro`, `mimo-v2-omni`, `mimo-v2-flash`, `mimo-v2-flash-free` and realigns the provider-catalog tests. ([#6248](https://github.com/diegosouzapw/OmniRoute/pull/6248) — thanks @backryun)
+- **chore(release):** ~50 commits on this branch are v3.8.45 pre-flight/hardening fixes and CI-perf work that landed here via the parallel-cycle sync-back (`sync-next-cycle.mjs`, Hard Rule #21) **after** the `v3.8.45` git tag was cut, and are already fully documented under the **[3.8.45]** section below — listed here only so the per-cycle commit-coverage check (`npm run release:uncovered`) doesn't flag them as gaps. Provider/catalog/UX/backend fixes: #6041, #6078, #6108, #6135, #6148, #6149, #6154, #6158, #6161, #6162, #6163, #6164, #6165, #6170, #6177, #6178, #6181, #6186, #6187, #6191, #6193, #6194, #6195, #6200, #6205, #6208, #6209, #6211, #6213, #6223, #6224, #6225, #6226, #6227, #6228, #6229, #6230, #6235, #6291, #6292. CI/release-pipeline work: #6167, #6203, #6214, #6215, #6218, #6273, #6275, #6283, #6284, #6285, #6300, #6305.
+- **chore(release):** additional zero-ref release-cycle plumbing on this branch, kept out of `release:uncovered` on purpose (no `#N` in the commit subject to cite): opening the v3.8.46 cycle, opening/closing the v3.8.45 cycle, the finalized [3.8.45] CHANGELOG i18n sync-back to 42 mirrors, the v3.8.45 cognitive/cyclomatic and file-size drift rebaselines, ESLint stale-suppression pruning (4,273 → 4,233), and clearing test-masking/docs-all pre-flight reds for v3.8.45.
+
+### ⚡ Performance & Infrastructure
+
+- **perf(release-green):** the pre-flight validator (`scripts/quality/validate-release-green.mjs`) now runs its 4 slow suites (unit / vitest / integration / pack-artifact) **concurrently** via `Promise.all` — pre-flight wall time drops from ~the sum of the suites to ~the slowest one (~30min saved per round; Phase 0 was the nº1 bottleneck of the v3.8.45 release benchmark, 2h54 of 6h34 e2e). Guard: `tests/unit/validate-release-green.test.ts` ("runs the slow suites CONCURRENTLY"). ([#6319](https://github.com/diegosouzapw/OmniRoute/pull/6319))
+- **fix(ci):** `scripts/release/sync-next-cycle.mjs` — two defects found live in its first production run (v3.8.45 Phase 5): (1) the `git()` helper's default 1 MiB `maxBuffer` crashed with `ENOBUFS` on `git show origin/main:CHANGELOG.md` (the CHANGELOG alone is >1 MiB) — widened to 64 MiB; (2) the i18n resync only propagated the `[NEXT]` (TBD) section, leaving the just-shipped finalized section as "— TBD" in all 42 mirrors — it now also syncs `[prevVersion]` bounded by the heading below it (new exported pure helper `versionAfter`). Guards: +5 tests in `tests/unit/sync-next-cycle.test.ts` (8/8). ([#6327](https://github.com/diegosouzapw/OmniRoute/pull/6327))
+- **test(ci):** concurrency-sensitive flaky tests are **quarantined into a serial pass** (`tests/unit/serial/`, `--test-concurrency=1`, appended to every unit runner incl. sharded variants — the serial pass is sharded too so concurrent shard jobs never self-collide). Initial set: `glm-coding-plan-monthly-3580`, `quota-division-blocks`, `provider-health-autopilot`, `combo-health-autopilot` — the class behind the ~28min CI wedges/re-runs (two live 1h+ wedges cancelled during this PR's own validation). Discovery + TIA gates track the new glob; systemic root cause (async logger writing after teardown) tracked in [#6360](https://github.com/diegosouzapw/OmniRoute/issues/6360). Guard: `tests/unit/test-serial-quarantine.test.ts` (4). ([#6347](https://github.com/diegosouzapw/OmniRoute/pull/6347))
+
+### 🙌 Contributors
+
+Thanks to everyone whose work landed in v3.8.46:
+
+| Contributor                                                    | PRs / Issues           |
+| -------------------------------------------------------------- | ---------------------- |
+| [@2220258345](https://github.com/2220258345)                   | direct commit / report |
+| [@abdofallah](https://github.com/abdofallah)                   | direct commit / report |
+| [@adentdk](https://github.com/adentdk)                         | direct commit / report |
+| [@aleksesipenko](https://github.com/aleksesipenko)             | direct commit / report |
+| [@anki1kr](https://github.com/anki1kr)                         | direct commit / report |
+| [@anungma](https://github.com/anungma)                         | direct commit / report |
+| [@arpicato](https://github.com/arpicato)                       | direct commit / report |
+| [@backryun](https://github.com/backryun)                       | #6248                  |
+| [@binsarjr](https://github.com/binsarjr)                       | direct commit / report |
+| [@brightfiscalband](https://github.com/brightfiscalband)       | direct commit / report |
+| [@chirag127](https://github.com/chirag127)                     | #6501, #6506           |
+| [@developerjillur](https://github.com/developerjillur)         | direct commit / report |
+| [@dilneiss](https://github.com/dilneiss)                       | #6499                  |
+| [@dtybnrj](https://github.com/dtybnrj)                         | direct commit / report |
+| [@Forcerecon](https://github.com/Forcerecon)                   | direct commit / report |
+| [@hao3039032](https://github.com/hao3039032)                   | direct commit / report |
+| [@Iammilansoni](https://github.com/Iammilansoni)               | #6150, #6245           |
+| [@jmengit](https://github.com/jmengit)                         | direct commit / report |
+| [@jordansilly77-stack](https://github.com/jordansilly77-stack) | direct commit / report |
+| [@JxnLexn](https://github.com/JxnLexn)                         | direct commit / report |
+| [@KooshaPari](https://github.com/KooshaPari)                   | #6166                  |
+| [@loopyd](https://github.com/loopyd)                           | direct commit / report |
+| [@luweiCN](https://github.com/luweiCN)                         | #6221                  |
+| [@makcimbx](https://github.com/makcimbx)                       | direct commit / report |
+| [@muflifadla38](https://github.com/muflifadla38)               | direct commit / report |
+| [@newnol](https://github.com/newnol)                           | direct commit / report |
+| [@ofekbetzalel](https://github.com/ofekbetzalel)               | direct commit / report |
+| [@ohahe52-dot](https://github.com/ohahe52-dot)                 | direct commit / report |
+| [@phidinhmanh](https://github.com/phidinhmanh)                 | direct commit / report |
+| [@powellnorma](https://github.com/powellnorma)                 | direct commit / report |
+| [@qpeyba](https://github.com/qpeyba)                           | direct commit / report |
+| [@RaviTharuma](https://github.com/RaviTharuma)                 | direct commit / report |
+| [@RCrushMe](https://github.com/RCrushMe)                       | direct commit / report |
+| [@rianonehub](https://github.com/rianonehub)                   | #6134, #6204           |
+| [@serverless83](https://github.com/serverless83)               | #6212                  |
+| [@swingtempo](https://github.com/swingtempo)                   | direct commit / report |
+| [@tenshiak](https://github.com/tenshiak)                       | direct commit / report |
+| [@ThongAccount](https://github.com/ThongAccount)               | direct commit / report |
+| [@UnrealAryan](https://github.com/UnrealAryan)                 | direct commit / report |
+| [@vinayakkulkarni](https://github.com/vinayakkulkarni)         | direct commit / report |
+| [@vittoroliveira-dev](https://github.com/vittoroliveira-dev)   | #6233                  |
+| [@warelik](https://github.com/warelik)                         | direct commit / report |
+| [@xxy9468615](https://github.com/xxy9468615)                   | direct commit / report |
+| [@xz-dev](https://github.com/xz-dev)                           | direct commit / report |
+| [@yanpaing007](https://github.com/yanpaing007)                 | direct commit / report |
+| [@diegosouzapw](https://github.com/diegosouzapw)               | maintainer             |
+
+---
+
 ## [3.8.45] — 2026-07-06
 
 ### ✨ New Features

docs/i18n/zh-CN/CHANGELOG.md

@@ -6,6 +6,213 @@
 
 ## [3.8.31] — 2026-06-20
 
+## [3.8.46] — TBD
+
+### ✨ New Features
+
+- **feat(sse):** **hide paid-only models from `auto/*` routing** when `hidePaidModels` is on ([#6512](https://github.com/diegosouzapw/OmniRoute/issues/6512)) — follow-up to #6328/#6495. PR #6495 hid paid-only models from the `GET /v1/models` listing, but `auto/*` combos (`auto/best-coding`, `auto/glm`, …) could still pick a paid-only backend into their candidate pool → a 402/403 at request time. `createVirtualAutoCombo` now filters the candidate pool through the new pure `open-sse/services/autoCombo/paidModelFilter.ts` (`filterPaidOnlyCandidates`), applying the same free-model predicate #6495 uses in `catalog.ts` (`providerHasFreeModels(provider) && isFreeModel(provider, {id})`) whenever `settings.hidePaidModels === true`. Applied before the category/tier/family narrowing, so it covers every `auto/*` combo; an all-paid pool degrades to the existing graceful empty-pool path. **Opt-in — default OFF leaves the pool unchanged** (identity). Regression guard: `tests/unit/autoCombo/paid-model-filter-6512.test.ts` (4, incl. the default-off identity guard).
+- **feat(sse):** **provider-family auto combos** — `auto/glm`, `auto/minimax`, `auto/mimo`, `auto/zai`, `auto/gemma`, `auto/llama`, `auto/gemini` ([#6453](https://github.com/diegosouzapw/OmniRoute/issues/6453)) — new routable ids that materialize an on-demand virtual combo spanning whatever installed backends currently expose that model family, degrading gracefully as backends rotate. A new pure `open-sse/services/autoCombo/modelFamily.ts` (`detectModelFamily`) classifies by model-id prefix for six families; `zai` is instead resolved by provider id (z.ai's hosted API serves the same `glm-*` model ids as every other GLM backend, so `auto/zai` means "route to my z.ai backend specifically" vs `auto/glm`'s "any connected GLM backend"). Reuses the existing `createVirtualAutoCombo` on-demand materialization path (no DB writes) and the `/v1/models` catalog advertising loop. Regression guard: `tests/unit/autoCombo/provider-family-combos.test.ts` (11).
+- **feat(proxy):** native **proxy-pool round-robin / egress IP rotation** ([#6365](https://github.com/diegosouzapw/OmniRoute/issues/6365)) — a scope (global / provider / account) can now hold **multiple** proxies as a pool with a rotation strategy, so outbound requests cycle their egress IP instead of pinning one proxy per scope. Migration `117_proxy_pool_rotation.sql` lifts the `UNIQUE(scope, scope_id)` constraint (rebuild via the canonical rename/copy/drop; existing single assignments become 1-element pools) and adds a `proxy_scope_rotation` companion table holding the per-scope strategy + a persisted monotonic round-robin cursor. Strategies: `round-robin` (default, monotonic cursor — never `Math.random`), `random`, and `sticky-per-N-min`. Resolution (`resolveProxyForScopeFromRegistry` / `resolveProxyForConnectionFromRegistry`) now fetches the alive, position-ordered candidate set (unchanged `PROXY_ALIVE_PREDICATE`) and applies the strategy; an empty / all-dead pool still returns `null` — the #6246 fail-closed guard is **untouched** (never falls through to direct egress). Backend + DB only; dashboard pool-builder UI is a follow-up. Regression guard: `tests/unit/proxy-pool-rotation-6365.test.ts` (8, incl. fail-closed + backward-compat).
+- **feat(providers):** end-to-end **tool/function calling on the native Gemini `/v1beta` endpoint** ([#6222](https://github.com/diegosouzapw/OmniRoute/issues/6222)) — both directions of the Gemini↔OpenAI conversion now preserve tool data (previously silently dropped). Request side: `convertGeminiToInternal` (extracted to its own testable module) maps `tools[].functionDeclarations` → OpenAI `tools`, prior `functionCall` parts → assistant `tool_calls`, and `functionResponse` parts → `tool`-role messages. Response side: `convertOpenAIResponseToGemini` emits `parts[].functionCall {name,args}` from `message.tool_calls`, and the streaming `openAIChunkToGeminiChunk` accumulates fragmented `tool_calls` deltas by index into complete `functionCall` parts. The non-Gemini client paths (Claude, OpenAI-Responses) already preserved tool calls — this closes the gap specific to the native Gemini surface. Regression guard: `tests/unit/v1beta-gemini-tool-calling-6222.test.ts` (6, incl. a streaming SSE round-trip).
+- **feat(providers):** copilot-m365-web **enterprise / work tier** support ([#6334](https://github.com/diegosouzapw/OmniRoute/issues/6334)) — mirrors the EDU-tier pattern (#6210): `M365ConnectionParams` gains an `agent` field, a new opt-in `M365_ENTERPRISE_OVERRIDES` preset (`agent=work`, `scenario=officeweb`, `licenseType=Premium`) applies via `providerSpecificData.tier="enterprise"` (alias `"work"`), and `agent` is also overridable directly via `providerSpecificData.agent`. `buildWsUrl` was hardcoding `agent="web"` (the one enterprise-distinguishing param with no override path), so a Premium work account handshook then returned an empty stream. The individual and EDU paths are untouched. Kilo's dup flag vs #6210 (EDU tier) was a false positive — different tier. Regression guard: `tests/unit/copilot-m365-enterprise-6334.test.ts` (7). End-to-end confirmation on a real Premium work account is a live-VPS validation follow-up (Hard Rule #18). (thanks @Forcerecon)
+- **feat(api):** standardized, provider-agnostic **`effort` + `thinking` request params** ([#6241](https://github.com/diegosouzapw/OmniRoute/issues/6241)) — a thin standardization layer over the existing mature per-provider reasoning plumbing (no provider mapper touched). `providerChatCompletionSchema` gains a canonical `effort` (reusing the shared `none/low/medium/high/xhigh` vocabulary — the UI tiers `extra`/`max` collapse onto `xhigh`) and a boolean `thinking`. A pure `normalizeReasoningRequest` (wired once in `src/sse/handlers/chat.ts`, before any reasoning field is read) folds them onto the fields the translators already consume (`reasoning_effort` / `reasoning.effort` / `thinking`), so they fan out to Anthropic / Gemini / xAI / Responses — an explicit client `reasoning_effort` / object-shaped `thinking` always wins (backward-compatible). `/models` additively exposes `supportsThinking` + `effort_tiers` so the frontend can render the toggles (UI component is a follow-up). Regression guard: `tests/unit/effort-thinking-standardization-6241.test.ts` (12). (thanks @Iammilansoni, @shabeer)
+- **feat(combo):** new **`pipeline` (sequential) combo strategy** ([#6297](https://github.com/diegosouzapw/OmniRoute/issues/6297)) — the 18th routing strategy runs targets **in order**, threading each step's output into the next step's input, with an optional per-step `prompt` (system instruction); only the final step's response is returned. Distinct from `fusion` (parallel fan-out + judge). Implemented as a self-contained `open-sse/services/pipeline.ts` (sibling to `fusion.ts`), dispatched from `combo.ts`; the step list reuses `combo.models` order and reads an optional `prompt` off each target (backward-compatible — ignored by every other strategy). Intermediate steps run non-streaming with tools stripped (complete prose to thread forward); the final step keeps the client's `stream` flag + tools. A failing/empty/unparseable intermediate step fails the whole pipeline explicitly via a sanitized error (never silently swallowed). Kilo's dup flag vs #563 was a false positive (that's model→chain selection; this is a sequential chain). Regression guard: `tests/unit/combo-pipeline-strategy.test.ts` (5). (thanks @ofekbetzalel)
+- **feat(ci):** `check:test-masking` now flags **inline-reimplemented prod conditions** ([#6348](https://github.com/diegosouzapw/OmniRoute/issues/6348)) — a new **report-only** subcheck (v2, 6A.10 family) catches the wrong-shape contract test: a test that recomputes the condition under test inline instead of importing/exercising the real function (the #6216 class, where `=== 500` → `>= 500` stayed green because the test re-implemented the branch). For each added/modified test file it warns when the file textually duplicates a ≥3-token conditional from a production file touched in the same PR **and** does not import the symbol/module owning it, via a pure, fixture-tested `findReimplementedConditions()` with an allowlist mirroring `assertReductionAllowlist`. Report-only for now (does not fail the gate) — to be promoted to blocking after a triage cycle. Regression guard: `tests/unit/check-test-masking.test.ts` (45).
+- **feat(sse):** per-connection routing override (native vs CLIProxyAPI) ([#6339](https://github.com/diegosouzapw/OmniRoute/issues/6339)) — the previously-dead `isCliproxyapiDeepModeEnabled` helper is now wired into `resolveExecutorWithProxy`: a single connection can opt itself into the CLIProxyAPI passthrough executor via `providerSpecificData.cliproxyapiMode="claude-native"`, with precedence **connection override > provider `upstream_proxy_config` mode > default**. `resolveExecutorWithProxy` now receives the resolved connection's `providerSpecificData` (threaded from `chatCore.ts`), so one connection can deep-route while the provider's other connections stay native — no DB schema change (the toggle rides in `providerSpecificData`). Also resolves the same-provider mixing ask in #6340. Regression guard: `tests/unit/chatcore-executor-proxy.test.ts` (9). (thanks @RaviTharuma)
+- **feat(dashboard):** "Add session cookie" modal now shows a prominent **"Open ‹host› →"** link to the provider's own site ([#6268](https://github.com/diegosouzapw/OmniRoute/issues/6268)) — every `-web` cookie-session provider (chatgpt-web, claude-web, gemini-web, kimi-web, lmarena, qwen-web, m365-copilot-web, …) renders a one-click external link (opening the provider's login/home page in a new tab) so operators no longer tab away to retype the URL mid-setup. The host resolves from a pure, unit-tested `resolveWebProviderHost()` (prefers `WEB_COOKIE_PROVIDERS[id].website`, falls back to the registry `baseUrl` origin); non-web providers render exactly as before. Kilo's dup flag vs #6265 (modal-too-small-on-1080p) was a false positive — distinct concern. Regression guard: `tests/unit/resolve-web-provider-host.test.ts` (5). (thanks @chirag127)
+- **feat(providers):** add **DigitalOcean** AI (serverless inference) as an OpenAI-compatible API-key provider ([#6373](https://github.com/diegosouzapw/OmniRoute/pull/6373)) — base `https://inference.do-ai.run/v1`, wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/digitalocean/`, `src/shared/constants/providers/apikey/inference-hosts.ts`). Regression guard: `tests/unit/digitalocean-provider.test.ts`. (thanks @newnol)
+- **feat(providers):** add **Huancheng Public API** (`hcnsec`) as an OpenAI-compatible regional provider ([#6410](https://github.com/diegosouzapw/OmniRoute/pull/6410)) — Xinjiang Huancheng Cybersecurity's public LLM platform (base `https://api.hcnsec.cn/v1`, free credits via daily check-ins), wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/hcnsec/`, `src/shared/constants/providers/apikey/regional.ts`). Regression guard: `tests/unit/hcnsec-provider.test.ts`. (thanks @UnrealAryan)
+- **feat(dashboard):** the web-session credential guide now shows an **"Open {host}" link** ([#6316](https://github.com/diegosouzapw/OmniRoute/pull/6316)) to the provider's sign-in site (derived from the provider `website` via `getProviderWebsiteHost`), so you can jump straight to the page where the cookie/session must be captured. Regression guard: `tests/unit/web-session-provider-link-6316.test.ts`. (thanks @jordansilly77-stack)
+- **feat(cerebras):** add the **Gemma 4 31B** model (`gemma-4-31b`) to the Cerebras registry + pricing table ([#6331](https://github.com/diegosouzapw/OmniRoute/pull/6331)). Regression guard extends `tests/unit/t28-model-catalog-updates.test.ts`. (thanks @backryun)
+- **feat(providers):** add **Yuanbao (web)** as a cookie-session provider ([#6196](https://github.com/diegosouzapw/OmniRoute/issues/6196)) — `yuanbao-web` (Tencent Yuanbao, `yuanbao.tencent.com`) with cookie-only auth (`hy_user`/`hy_token` + public agent id), SSE→OpenAI translation incl. `reasoning_content`, exposing DeepSeek V3/R1 + Hunyuan / Hunyuan-T1. Regression guard: `tests/unit/providers-yuanbao-web.test.ts`. `together-web` was **deferred** (no verifiable web-session endpoint — needs a captured request) and `huggingchat-web` **dropped** (the existing `huggingchat` already is a web-cookie provider). (thanks @chirag127)
+- **feat(providers):** route the built-in **agentrouter** through the dynamic Claude-Code wire image ([#6056](https://github.com/diegosouzapw/OmniRoute/issues/6056)) — a small static allow-set (`CC_WIRE_IMAGE_BUILTINS` in `open-sse/services/ccWireImageBuiltins.ts`), consulted by `isClaudeCodeCompatible` / `isClaudeCodeCompatibleProvider` / `applyFingerprint`, makes agentrouter adopt the CC wire-image headers + fingerprint **while guarding the CC baseUrl/auth branches** so it keeps its own registry `baseUrl` and `x-api-key` auth. Regression guard: `tests/unit/agentrouter-cc-wire-image.test.ts` (asserts the wire image is applied AND agentrouter's baseUrl/auth are preserved). Live WAF-acceptance against agentrouter.org is a VPS validation follow-up (Hard Rule #18).
+- **feat(providers):** **bulk-add API keys for Cloudflare Workers AI** ([#6174](https://github.com/diegosouzapw/OmniRoute/issues/6174)) — `cloudflare-ai` is removed from the bulk-add exclusion list and the bulk parser gains a 3-field `name|accountId|apiKey` mode; the bulk route now builds a **per-entry** `providerSpecificData` so each key carries its own `accountId` (fixing the previous shared-object reuse), and both the create + key-validation paths receive it. Regression guard: `tests/unit/bulk-api-key-parser-cloudflare.test.ts`. (thanks @muflifadla38)
+- **feat(dashboard):** routing/settings UX clarity ([#6147](https://github.com/diegosouzapw/OmniRoute/issues/6147)) — (1) weighted combos show the **effective routing share %** next to each weight when weights don't sum to 100 (`WeightTotalBar.tsx`); (2) the status widget's user-facing **"Cloud Sync" label is renamed** to "Remote Settings Sync" (`CloudSyncStatus.tsx`; internal ids/state untouched); (3) built-in providers gain an **opt-in advanced base-URL override** (`isBaseUrlOverrideEligibleProvider`, hidden behind an "Advanced" toggle, reusing the existing `providerSpecificData.baseUrl` persistence — not globally widened). Regression guard: `tests/unit/routing-settings-ux-6147.test.ts`.
+- **feat(combo):** add an option to **disable session stickiness**, per-combo or globally — round-robin / random combos can rotate to a different connection on every request instead of pinning a whole conversation to one connection by its first-message hash. Resolution precedence per-combo `config.disableSessionStickiness` → global `settings.disableSessionStickiness` → default `false` (preserves the #3825 prompt-cache/504 fix); gates **both** stickiness call sites in `open-sse/services/combo.ts`. Exposed as a global toggle (Combo Defaults) and a per-combo Inherit/on/off control. ([#6168](https://github.com/diegosouzapw/OmniRoute/issues/6168)) Regression guard: `tests/unit/combo-disable-session-stickiness.test.ts`. (thanks @RCrushMe)
+- **feat(docker):** add the `OMNIROUTE_NO_SUDO` env flag for root-less / user-namespaced deployments — the MITM cert-trust command path (`resolveSudoSpawn` in `src/mitm/systemCommands.ts`) now strips the leading `sudo` when the flag is truthy, in addition to the existing root / sudo-missing cases, so the Proxy Agent runs without `sudo` (the operator trusts the CA manually, e.g. via `NODE_EXTRA_CA_CERTS`). Argv-array `spawn` preserved — no shell interpolation (Hard Rule #13). ([#6122](https://github.com/diegosouzapw/OmniRoute/issues/6122)) Regression guard: `tests/unit/mitm-systemCommands-no-sudo.test.ts`. (thanks @powellnorma)
+- **feat(providers):** add **Requesty** as an OpenAI-compatible gateway provider (BYOK, base `https://router.requesty.ai/v1`, ~200 free requests/day) — wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/requesty/`, `src/shared/constants/providers/apikey/gateways.ts`). ([#6120](https://github.com/diegosouzapw/OmniRoute/issues/6120)) Regression guard: `tests/unit/requesty-provider.test.ts`. (thanks @chirag127)
+- **feat(dashboard):** add **configured-only / available-only filters** to the Free Provider Rankings page ([#6150](https://github.com/diegosouzapw/OmniRoute/issues/6150)) — hide providers you haven't configured, or whose connections are all rate-limited / out of quota, via server-side query params (`?configuredOnly` / `?availableOnly` on `GET /api/free-provider-rankings`) backed by a testable lib helper reusing the in-process connection state (no Redis). Both filters default off, so the default view is unchanged; this supersedes the earlier client-side "Configured Only" toggle (#6245) with an available-only dimension and unit-tested logic. Regression guard: `tests/unit/freeProviderRankings-filters.test.ts`.
+
+### 🔧 Bug Fixes
+
+- **fix(dashboard):** adding a second API key connection for the same provider no longer silently overwrites the first — the Add-API-key modal now derives a unique default connection name (`main`, then `main-2`, `main-3`, …) so the backend name-based upsert can't collide (#6499 — thanks @dilneiss).
+- **fix(compression):** the session-dedup engine now also deduplicates a large multi-line block repeated **within a single message** (intra-message dedup), not just across turns; the compression-preview API surfaces a `fallbackReason`, and the fusion panel reports how many models were rate-limited vs failed on total-panel failure (#6501 — thanks @chirag127).
+- **fix(compression):** a stacked-pipeline step naming an unregistered engine now surfaces a `validationErrors` entry instead of silently no-op'ing, so misconfigured pipelines are visible in the preview API (#6506 — thanks @chirag127).
+- **feat(usage):** add a **Codex reset-credit redemption** flow to the Provider Limits UI ([#6361](https://github.com/diegosouzapw/OmniRoute/pull/6361)) — a `useCodexResetCreditRedemption` hook + `/api/usage/codex-reset-credit` route + `codexResetCredits` lib let you redeem banked Codex reset credits from the quota card. Regression guard: `tests/unit/codex-reset-credits.test.ts`. (thanks @JxnLexn)
+- **feat(glm):** add **team-plan quota settings** for `glm-cn` connections ([#6351](https://github.com/diegosouzapw/OmniRoute/pull/6351)) — a dedicated `GlmTeamQuotaFields` form section (team quota id / limits) threaded through the Add/Edit connection modals, persisted via `providerSpecificData`, with the GLM usage service reading the team quota. Regression guards: `tests/unit/glm-team-quota.test.ts`, `provider-specific-data-schema.test.ts`. (thanks @hao3039032)
+- **feat(providers):** add **TinyFish** web-fetch/search support ([#6349](https://github.com/diegosouzapw/OmniRoute/pull/6349)) — a `tinyfish-fetch` executor + `/v1/web/fetch` route + MCP web-fetch tool, registered as a specialty-media provider with request-validation and a search-provider catalog entry. Regression guards: `tests/unit/executor-tinyfish-fetch.test.ts`, `web-fetch-handler.test.ts`, `mcp-web-fetch-tool.test.ts`, `provider-validation-tinyfish.test.ts`. (thanks @dtybnrj)
+- **fix(cli):** `omniroute launch-codex` now spawns `codex.cmd` through a shell on Windows (the npm `.cmd` shim is unresolvable by bare `spawn` → ENOENT), mirroring the qodercli Windows fix (#6263) ([#6312](https://github.com/diegosouzapw/OmniRoute/pull/6312)). Regression guard: `tests/unit/launch-codex-windows-spawn-6312.test.ts`. (thanks @swingtempo)
+- **fix(codex):** isolate the **Spark** quota from the shared Codex quota and stabilize the quota UI ordering / hydration so per-scope limits render consistently ([#6336](https://github.com/diegosouzapw/OmniRoute/pull/6336)). Regression guards: `tests/unit/codex-quota-selection-hydration.test.ts`, `provider-limits-ui.test.ts` + 3 more. (thanks @xz-dev)
+- **feat(api):** add a `hidePaidModels` setting that filters paid-only models out of the `/v1/models` catalog. Regression guard: `tests/unit/models-catalog-hide-paid.test.ts`. (thanks @chirag127)
+- **fix(api-manager):** the fallback model picker now preserves combos instead of dropping them when a primary model is unavailable ([#6443](https://github.com/diegosouzapw/OmniRoute/pull/6443)). Regression guard: `tests/unit/api-manager-page-static.test.ts`. (thanks @jmengit)
+- **fix(providers):** recoverable Antigravity / Cloud-Code (Gemini Code Assist) `403` responses ([#6452](https://github.com/diegosouzapw/OmniRoute/pull/6452)) are now classified as a retryable project-config error instead of a terminal account ban, so a fixable project/API-disabled 403 no longer forces a ~1-year cooldown / full OAuth reconnect. Regression guard: `tests/unit/errorclassifier-antigravity-403.test.ts`. (thanks @developerjillur)
+- **fix(mitm):** `sanitizeHeaders` now redacts `Set-Cookie` response headers so upstream session cookies never leak into logs / diagnostics ([#6451](https://github.com/diegosouzapw/OmniRoute/pull/6451)). Regression guard: `tests/unit/mitm-sanitize-headers.test.ts`. (thanks @developerjillur)
+- **fix(api):** `/api/compression/preview` now accepts `mode: "caveman"` and correctly handles stacked / zero-compression previews ([#6425](https://github.com/diegosouzapw/OmniRoute/issues/6425)). Regression guard: `tests/unit/api/compression-preview-caveman-and-stacked-6425.test.ts`. (thanks @chirag127)
+- **feat(providers):** add **Zed** hosted LLM aggregator as a native-app provider ([#6118](https://github.com/diegosouzapw/OmniRoute/pull/6118)) — OAuth sign-in via the Zed hosted flow, registered through the shared provider registry + executor. Regression guards: `tests/unit/zed-oauth-provider.test.ts`, `zed-import-utils.test.ts`, `zed-docker-detect.test.ts`, `mitm-handler-zed.test.ts`. VPS-validated via live operator login (Hard Rule #18).
+- **fix(oauth):** the Kiro SSO-cache auto-import now **preserves the IDC region** — cross-region Amazon Q / Kiro profiles imported from the SSO cache are no longer collapsed to the default region ([#6113](https://github.com/diegosouzapw/OmniRoute/pull/6113)). Regression guard: `tests/unit/kiro-auto-import-idc-2059.test.ts`. VPS-validated via live operator login (Hard Rule #18).
+- **fix(dashboard):** passthrough model aliases no longer collide when two namespaced model ids share a last segment (port from 9router#1850, [#6431](https://github.com/diegosouzapw/OmniRoute/pull/6431)). `enx/gpt-5.5` and `enx/codebuddy/gpt-5.5` both auto-generated the alias `gpt-5.5`, so the second model could never be added (the UI just alerted "alias already exists"). Aliases are now disambiguated deterministically — bare last segment when free, then parent-qualified (`codebuddy-gpt-5.5`), then a numeric suffix — while re-adding the exact same model id is still blocked. Regression guard: `tests/unit/passthrough-alias-1850.test.ts`. (thanks @arpicato)
+- **fix(translator):** preserve a Gemini `functionResponse` co-located with other parts (another `functionCall`, or trailing `text`) in the same content when translating **Gemini → OpenAI** ([#6376](https://github.com/diegosouzapw/OmniRoute/pull/6376)). `convertGeminiContent()` early-returned the tool message on the first `functionResponse` part, dropping any co-located parts; such contents are now pre-split (one tool message per `functionResponse`, emitted first, plus one message for the remaining parts). Regression guard: `tests/unit/gemini-to-openai-function-response.test.ts`. (thanks @warelik)
+- **fix(headroom):** detect a python interpreter managed by **mise / pyenv / asdf / conda** (port from 9router#2353, [#6382](https://github.com/diegosouzapw/OmniRoute/pull/6382)). Headroom's python probe (`src/lib/headroom/detect.ts`) searched a hardcoded `PATH`, but version managers expose their interpreters via shim dirs that only join `PATH` through interactive-shell activation — which the non-interactive server never runs, so a managed python (≥3.10) was invisible and Headroom reported it missing. The search path now prepends the well-known shim/bin dirs (`~/.local/share/mise/shims`, `~/.pyenv/shims`, `~/.asdf/shims`, `$CONDA_PREFIX/bin`, `~/.local/bin`, respecting `MISE_DATA_DIR`/`PYENV_ROOT`/`ASDF_DATA_DIR` when set), and a new `HEADROOM_PYTHON` env override lets operators point straight at their interpreter (mirroring `HEADROOM_URL`). Still shell-free (`execFileSync`). Regression guard: `tests/unit/headroom-detect.test.ts` (5). (thanks @loopyd)
+- **fix(executors):** strip the OpenAI-Codex/Claude-CLI `client_metadata` passthrough field for **NVIDIA** requests (port from 9router#1887, [#6411](https://github.com/diegosouzapw/OmniRoute/pull/6411)). NVIDIA's OpenAI-compatible wrapper rejects it with `400 Unsupported parameter`, the same class already handled for `cerebras`/`mistral`; `nvidia` (executor `default`) was missing from the strip allowlist so Codex/Claude-Code passthrough requests 400'd. Regression guard: `tests/unit/executor-default-strip-client-metadata.test.ts` (+nvidia case). (thanks @phidinhmanh)
+- **fix(translator):** strip the Claude-style `thinking` field for **NVIDIA `z-ai/glm-5.2`** (port from 9router#2023, [#6413](https://github.com/diegosouzapw/OmniRoute/pull/6413)). NVIDIA's OpenAI-compatible wrapper 400s on `thinking` (a Claude-format client routed here leaves a `thinking:{type:"adaptive"}`); the existing strip rule only dropped `reasoning`. Same class already handled for `minimax-m2.7`. Regression guard: `tests/unit/nvidia-minimax-thinking-strip.test.ts` (+glm-5.2 case). (thanks @phidinhmanh)
+- **fix(translator):** suppress the streamed `</think>` close marker for the **Antigravity IDE** client (port from 9router#1061, [#6415](https://github.com/diegosouzapw/OmniRoute/pull/6415)). On thinking-only turns Antigravity rendered a bare `</think>` as the sole visible content, tripping its loop-detection and wasting requests. Antigravity's UA (`vscode/<v> (Antigravity/<v>)`) is added to the marker-suppress allowlist (alongside OpenCode); Claude Code / Cursor still get the marker, and `x-omniroute-thinking-marker: on` force-restores it. Regression guard: `tests/unit/think-close-marker-suppress-5245.test.ts`. (thanks @abdofallah)
+- **fix(executors):** strip nested `reasoning_content` from messages for **Mistral** (port from 9router#1649, [#6417](https://github.com/diegosouzapw/OmniRoute/pull/6417)). Mistral's API returns `422 extra_forbidden` when an assistant message carries `reasoning_content` (replayed thinking from a prior turn, e.g. via the Codex `/responses` path); the generic top-level 400 field-downgrade retry never covered the nested per-message field. `DefaultExecutor` now strips it for provider `mistral` only, so DeepSeek (which requires replayed `reasoning_content`) is unaffected. Regression guard: `tests/unit/mistral-strip-reasoning-content-1649.test.ts`. (thanks @xxy9468615)
+- **fix(executors):** strip the `client_metadata` passthrough field on the **OpenCode** path (port from 9router#1442, [#6418](https://github.com/diegosouzapw/OmniRoute/pull/6418)). OpenCode upstreams (e.g. `kimi-k2.6` via opencode-go) reject it with `400 "Extra inputs are not permitted, field: 'client_metadata'"`; the DefaultExecutor strip only covered cerebras/mistral and `OpencodeExecutor` extends `BaseExecutor` directly, so nothing removed it there. Regression guard: `tests/unit/opencode-strip-client-metadata-1442.test.ts`. (thanks @yanpaing007)
+- **fix(executors):** inject the `reasoning_content` echo for the native **Moonshot Kimi** provider (port from 9router#1480, [#6419](https://github.com/diegosouzapw/OmniRoute/pull/6419)). Kimi (executor `default`) is a thinking-mode upstream that 400s with "reasoning_content must be passed back" when a prior assistant turn lacks it; the placeholder injection was only wired into the OpenCode meta-provider, so direct multi-turn Kimi conversations failed. Scoped to `kimi` (gateway-served models matching the thinking-model name pattern are unaffected). Regression guard: `tests/unit/kimi-native-reasoning-injected-1480.test.ts`. (thanks @2220258345)
+- **fix(executors):** recover from a strict gateway's `context_management: Extra inputs are not permitted` 400 (port from 9router#1468, [#6420](https://github.com/diegosouzapw/OmniRoute/pull/6420)). **Claude Code** always sends a top-level `context_management` field; strict anthropic-compatible gateways reject it. The dedicated context-editing 400-fallback only fired when OmniRoute's own `contextEditing` feature was enabled (default off), so a client-sent field passed through untouched and 400'd. `context_management` is now in the generic reactive field-strip list, so it's stripped-and-retried once regardless of the feature flag (with correct request re-signing for claude-compatible relays). Regression guard: `tests/unit/provider-field-strips.test.ts`. (thanks @ohahe52-dot)
+- **fix(network):** enable **RFC 8305 Happy Eyeballs** (`autoSelectFamily`) on the direct-egress undici dispatcher (port from 9router#1237, [#6423](https://github.com/diegosouzapw/OmniRoute/pull/6423)). When DNS returns both IPv6 (AAAA) and IPv4 (A) and the IPv6 route is broken (e.g. a NAT64 `64:ff9b::` prefix without routing), undici tried IPv6 first and hung until `ETIMEDOUT` (then a 502 + account lockout), even though `curl` reached the same host. The direct dispatcher now races both families and uses whichever connects first. Proxy paths pin family via `proxyTls` and are unaffected. Regression guard: `tests/unit/direct-dispatcher-pipelining-4580.test.ts`. (thanks @adentdk)
+- **fix(combo):** round-robin now advances the rotation pointer past the model that **actually served**, not the eagerly-scheduled one (port from 9router#948, [#6428](https://github.com/diegosouzapw/OmniRoute/pull/6428)). With `stickyLimit: 1` (true round-robin), when the scheduled model failed and a _different_ model served via fallback, the counter had already advanced +1 from the scheduled index — so the next request reused the fallback-served model, degrading round-robin into hot-spotting on whichever model was healthy. The pointer now advances to the served index + 1 (mirroring the sticky-limit>1 path). Session-stickiness (#3825) and distribution are preserved. Regression guard: `tests/unit/combo-rr-fallback-advance-948.test.ts`. (thanks @binsarjr)
+- **fix(sse):** a non-string `model` field is now rejected with a `400` before the resolver, instead of crashing downstream `.toLowerCase()`/`.split()` calls into an empty-body `500` that escapes the error sanitizer ([#6407](https://github.com/diegosouzapw/OmniRoute/issues/6407)). Regression guard: `tests/unit/chat-non-string-model-6407.test.ts`. (thanks @chirag127)
+- **fix(api):** unknown `/api/*` routes now return a JSON `404` (instead of the dashboard HTML shell) and scalar chat params (`model`/`temperature`/etc.) are validated **before** the provider lookup so malformed requests fail fast with a clear `400` ([#6424](https://github.com/diegosouzapw/OmniRoute/issues/6424), [#6412](https://github.com/diegosouzapw/OmniRoute/issues/6412)). Regression guards: `tests/unit/api/api-catchall-json-404.test.ts`, `tests/unit/chat-early-schema-validation-6412.test.ts`. (thanks @chirag127)
+- **fix(api):** `/v1/chat/completions` now rejects a non-JSON `Content-Type` with a `400` before parsing the body ([#6414](https://github.com/diegosouzapw/OmniRoute/issues/6414)). Regression guard: `tests/unit/v1-chat-completions-content-type-6414.test.ts`. (thanks @chirag127)
+- **fix(api):** the `X-OmniRoute-Compression` response header is now echoed on `/v1/chat/completions` and `/v1/completions` ([#6422](https://github.com/diegosouzapw/OmniRoute/issues/6422)). Regression guard: `tests/unit/compression-header-echo-6422.test.ts`. (thanks @chirag127)
+- **fix(api):** concurrent `GET /v1/models` requests are coalesced into a single catalog build ([#6408](https://github.com/diegosouzapw/OmniRoute/issues/6408)). Regression guard: `tests/unit/v1-models-concurrent-6408.test.ts`. (thanks @chirag127)
+- **fix(api):** `/v1/completions` now echoes the requested `body.model` in its JSON + streamed responses ([#6429](https://github.com/diegosouzapw/OmniRoute/pull/6429)). Regression guard: `tests/unit/completions-body-model-echo.test.ts`. (thanks @chirag127)
+- **fix(api):** env-var master keys now see the full `/v1/models` catalog ([#6406](https://github.com/diegosouzapw/OmniRoute/issues/6406)). Regression guard: `tests/unit/models-catalog-envkey-6406.test.ts`. (thanks @chirag127)
+- **fix(api):** non-streaming `/v1/completions` responses now echo `body.model` aligned with the `X-OmniRoute-Model` header ([#6426](https://github.com/diegosouzapw/OmniRoute/issues/6426)). Regression guard: `tests/unit/v1-completions-model-header-match-6426.test.ts`. (thanks @chirag127)
+- **fix(api):** unknown `/v1/*` routes now return a JSON `404 not_found` instead of the Next.js dashboard HTML shell ([#6405](https://github.com/diegosouzapw/OmniRoute/issues/6405)). Regression guard: `tests/unit/api/v1-catchall-json-404.test.ts`. (thanks @chirag127)
+- **fix(api):** the per-connection provider models route now degrades to the shipped catalog when a provider's `/models` endpoint answers with a redirect ([#6267](https://github.com/diegosouzapw/OmniRoute/issues/6267)) — a `qwen-web` import failed with a raw `Redirect blocked … (307)` 503. `safeOutboundFetch` throws `REDIRECT_BLOCKED` on the 307, `getSafeOutboundFetchErrorStatus` maps it to 503, and `buildDiscoveryErrorFallbackResponse` treated every 503 as a hard error — so the non-empty `getModelsByProviderId("qwen-web")` catalog was never surfaced. A models-endpoint redirect is not a fixable-config error (unlike `URL_GUARD_BLOCKED`/`INVALID_URL`, which stay hard errors), so it now falls back to the local/cached catalog before the 503 short-circuit. General fix — covers any config-driven provider that 307s. Regression guard: `tests/unit/provider-models-qwen-web-redirect-6267.test.ts`. (thanks @chirag127)
+- **fix(api):** the per-connection provider models route (MCP `list_models_catalog` + the dashboard import view) now merges USER-ADDED custom models into its response ([#6247](https://github.com/diegosouzapw/OmniRoute/issues/6247)) — custom models live in the `key_value` namespace `customModels`, which the live REST `/api/v1/models` already merges, but `src/app/api/providers/[id]/models/route.ts` never read `getCustomModels`, so custom models were dropped on both the discovery-success and local_catalog paths. They are now folded into the returned model list (deduped by id, stamped `owned_by: provider`), fixing MCP + the dashboard import view in one place. Regression guard: `tests/unit/provider-models-custom-merge-6247.test.ts`. (thanks @RCrushMe)
+
+- **fix(providers):** GitLab Duo tool-calling follow-up turns no longer fail upstream with `422 {"detail":"Validation error"}` (tokens 0/0, rejected pre-inference). The #6234 tool-result-feedback fix serialized the **entire** multi-turn conversation into GitLab's single-file `code_suggestions` (`small_file`) generation endpoint — folded history that turn-N sent as an oversized `current_file.content_above_cursor` **and** duplicated verbatim into `user_instruction`, tripping the AI-Gateway's `small_file` validation guard. The executor now **bounds** that prompt: it keeps system + latest user message + the most-recent tool round (dropping older turns), caps oversized tool results, and stops duplicating the full prompt into `user_instruction` (which now carries only the short latest user message) — while still feeding the most-recent tool result back so the agent continues (`open-sse/executors/gitlab.ts`, [#6220](https://github.com/diegosouzapw/OmniRoute/issues/6220)). The unit test covers the bounding logic; the upstream 422→200 clearing is VPS-only (Hard Rule #18). Regression guard: `tests/unit/gitlab-tool-exchange-bounded-6220.test.ts`.
+
+- **fix(i18n):** the provider-detail (`/dashboard/providers/[id]`) connection-status filter labels no longer render as `__MISSING__:All` / `__MISSING__:Active` / `__MISSING__:Error` / `__MISSING__:Banned` / `__MISSING__:CreditsExhausted` in non-English locales (notably pt-BR) ([#6290](https://github.com/diegosouzapw/OmniRoute/issues/6290)). Root cause was **not** the namespace mismatch the issue guessed — the `providers.filter*` keys resolve correctly in `en.json`; the debt lived in the locale mirrors (`src/i18n/messages/*.json`), where these five keys carried the `__MISSING__:` sync sentinel in ~15 locales and were absent entirely in ~26 others, so next-intl found the key and echoed the sentinel verbatim. All 40 non-English/-Chinese mirrors now ship real translations for the five `providers.filter*` labels. Regression guard: `tests/unit/i18n-provider-filter-keys-6290.test.ts`. (thanks @diegosouzapw)
+
+- **fix(providers):** the `copilot-m365-web` streaming executor now emits `debug`-level WebSocket diagnostics ([#6210](https://github.com/diegosouzapw/OmniRoute/issues/6210)) — the outbound WS URL (with the `access_token` **redacted** via `redactWsUrl()`), handshake success/failure, and each received SignalR frame's `type`/`target`. Previously the streaming path logged nothing, so an empty `content:null` response (the M365 Education / Starter tier symptom fixed in #6234) was undiagnosable even at `APP_LOG_LEVEL=debug`. The change is debug-level and side-effect-free — it does not alter streaming behavior or the frame parser, and the token never reaches the logs. Regression guard: `tests/unit/copilot-m365-web-logging-6210.test.ts` (thanks @qpeyba)
+
+- **fix(resilience):** a round-robin combo no longer returns `503 all upstream accounts are unavailable` when a compatibility-rejected target is actually healthy ([#6238](https://github.com/diegosouzapw/OmniRoute/issues/6238)). `filterTargetsByRequestCompatibility` drops request-incompatible targets (tool/vision/structured-output unsupported, or below the required context window) **before** any availability check runs, and its `compatible.length === 0` safety net only fired when _all_ targets were filtered — not when the kept targets later all turned out runtime-unavailable (circuit-open / cooldown / no credentials). So a combo could 503 while a compat-rejected-but-healthy provider sat unused. `handleRoundRobinCombo` now keeps the compat-rejected set and, when every compat-kept target was skipped without a single real attempt, probes those rejected targets as a **last-resort fallback tier** (via the new pure `open-sse/services/combo/comboCompatFallback.ts`) before crystallizing the 503. Regression guard: `tests/unit/combo-roundrobin-compat-fallback-6238.test.ts`. (thanks @ThongAccount)
+
+- **fix(startup):** best-effort self-heal for a corrupted Turbopack dev cache on Windows ([#6289](https://github.com/diegosouzapw/OmniRoute/issues/6289)). On Windows, `pnpm dev` can fail at startup when Turbopack `mmap`s a persistent-cache SST file and the OS refuses the mapping (`os error 1455` — "paging file too small"), which Turbopack surfaces as a misleading `Module not found: Can't resolve '@/shared/utils/machine'`. This is a **known upstream Turbopack cache-corruption bug — not our code**. The dev launcher (`scripts/dev/run-next.mjs`) now wraps `nextApp.prepare()` and, when it rejects with that signature (`isTurbopackCacheCorruption` in the new `scripts/dev/turbopackCacheHeal.mjs`), purges `.build/next/**/cache/turbopack` and retries **once** with a clear log. **Caveat — best-effort only:** the corruption often surfaces as a runtime overlay rather than a `prepare()` rejection, so this cannot always intercept it; the reliable remedy remains manually deleting the Turbopack cache dir. Regression guard: `tests/unit/turbopack-cache-heal-6289.test.ts`. (thanks @chirag127)
+
+- **fix(providers):** qodercli PAT auth no longer fails with `spawn qodercli ENOENT` on Windows ([#6263](https://github.com/diegosouzapw/OmniRoute/issues/6263)) — `spawnQoderCli` spawned the bare `qodercli` name with `shell:false` and an unenriched env, so the npm `.cmd` wrapper under `%APPDATA%\npm` (a user-PATH directory) was never resolved. It now resolves the absolute `.cmd`/`.exe` path through the existing `getCliRuntimeStatus("qoder")` resolver in `src/shared/services/cliRuntime.ts` (memoized), spawns with `shell` when the target is a `.cmd`/`.bat`, and uses the cliRuntime-enriched env (PATH + PATHEXT + APPDATA); the ENOENT error now lists the searched paths plus the `CLI_QODER_BIN` override. End-to-end spawn on a real Windows host is host-only (Hard Rule #18); the path-resolution logic is unit-tested. Regression guard: `tests/unit/qodercli-windows-resolve-6263.test.ts`. (thanks @chirag127)
+
+- **fix(sse):** the reasoning-token buffer no longer inflates **probe-sized `max_tokens`** ([#6274](https://github.com/diegosouzapw/OmniRoute/issues/6274)) — Claude Code's `/model` capability check sends `max_tokens: 1`, but for a thinking-capable model with a large output cap (e.g. `glm-5.2`) the #3587 headroom heuristic (`max(current + 1000, ceil(current * 1.5))`) rewrote it to `1001` and forwarded that upstream, wasting tokens on a request that was never a genuine reasoning budget. `resolveReasoningBufferedMaxTokens()` (`open-sse/services/reasoningTokenBuffer.ts`) now short-circuits and returns the caller's value verbatim when it is below the new `REASONING_BUFFER_MIN_TRIGGER` (256) threshold — a tiny explicit limit is a probe, not a reasoning request. Real budgets still receive the #3587 headroom unchanged, and the guard runs after the existing capability checks so unknown / non-reasoning models keep returning `null`. Regression guard: `tests/unit/reasoning-token-buffer-6274.test.ts`. (thanks @brightfiscalband)
+
+- **fix(cli):** `omniroute reset-password` now works as a real subcommand, and password resets over piped (non-TTY) stdin actually apply ([#6261](https://github.com/diegosouzapw/OmniRoute/issues/6261), [#6258](https://github.com/diegosouzapw/OmniRoute/issues/6258)). Two coupled defects: (1) **#6261** — `bin/omniroute.mjs` routed everything through Commander with only two pre-Commander bypasses (`--mcp`, `reset-encrypted-columns`), so `omniroute reset-password` was rejected as an unknown command; only the separate `omniroute-reset-password` bin worked, while the docs falsely advertised the subcommand (incl. a bogus "legacy alias still works"). A pre-Commander bypass mirroring `reset-encrypted-columns` now dynamically imports `bin/reset-password.mjs` (which self-executes) before Commander parses; the three doc lines were corrected. (2) **#6258** — `bin/reset-password.mjs` issued two sequential `rl.question` prompts; under piped stdin the second read never settled at EOF, so `main()` never reached `resetManagementPassword` and the reset was a silent no-op (both prompts printed, no success, password unchanged). The CLI now detects non-TTY stdin and reads it once (first line = password, second line = confirm if present, else reused), adds a `--password-stdin` flag (entire stdin is the password, no confirmation), and exits `0` explicitly so the success line always flushes; interactive TTY behavior is unchanged. Regression guard: `tests/unit/reset-password-cli-6261-6258.test.ts` (3). (thanks @chirag127)
+
+- **fix(db):** the mass-migration **safety abort** now tells the operator how to bypass it and stops flooding the log ([#6260](https://github.com/diegosouzapw/OmniRoute/issues/6260)) — after restoring a backup that wiped the migration tracking table, `runMigrations()` threw the abort on every downstream `ensureDbInitialized()`, re-logging the full banner 11+ times, and the message never mentioned the existing `OMNIROUTE_MAX_PENDING_MIGRATIONS` escape hatch. The abort text now appends a bypass hint (set `OMNIROUTE_MAX_PENDING_MIGRATIONS=0` in `server.env` / `DATA_DIR/.env`), and a new `MigrationSafetyAbortError` is memoized so repeated calls in the same process throw the same instance and emit a single concise line instead of the full cascade. Regression guard: `tests/unit/migration-safety-abort-6260.test.ts`. (thanks @chirag127)
+
+- **fix(auth):** importing a **distinct** Codex/ChatGPT OAuth `auth.json` is no longer falsely rejected as "already exists" when it belongs to a different user in the same workspace ([#6301](https://github.com/diegosouzapw/OmniRoute/issues/6301)). `findExistingCodexConnection` (in `src/lib/oauth/utils/codexAuthImport.ts`) deduped **only** on `providerSpecificData.workspaceId === accountId`, where `accountId` is the shared `chatgpt_account_id`/`tokens.account_id` — so two members of the same ChatGPT Team collapsed onto a single connection (409 `duplicate_account`). The id_token's `https://api.openai.com/auth` claim carries a per-user `chatgpt_user_id` alongside the workspace id (the device-flow path already persisted it as `chatgptUserId`, but the import path did not). Now `parseAndValidateCodexAuth` extracts `userId` (`chatgpt_user_id` → `user_id` → JWT `sub`) into `ParsedCodexAuth`, the create/update paths persist `chatgptUserId` in `providerSpecificData` (mirroring `codex.ts`), and dedup keys on `workspaceId` **AND** `chatgptUserId` — with a backward-compat fallback to legacy accountId-only matching when no stored connection for that workspace records a `chatgptUserId`, so genuinely-same accounts still dedup. Regression guard: `tests/unit/codex-auth-import-userid-dedup-6301.test.ts` (4). (thanks @anungma)
+
+- **fix(providers):** importing models for the **venice-web** provider no longer fails with a red "Provider venice-web does not support models listing" ([#6269](https://github.com/diegosouzapw/OmniRoute/issues/6269)). `venice-web` is a web-cookie provider with an executor but no upstream `/v1/models` endpoint and no registry `models`, so the models route fell through to the tail `400`. Mirroring the `jules`/`linkup-search`/`ollama-search` fix (#5569), it now ships a static local catalog entry in `src/lib/providers/staticModels.ts` — seeding the current Venice lineup (`venice-uncensored`, `llama-3.3-70b`, `qwen3-235b`, `qwen3-4b`, `deepseek-r1-671b`; Venice rotates its catalog, see docs.venice.ai/models/overview) — so the route returns `200` with `source:"local_catalog"`, `intentional:true`. Regression guard: `tests/unit/static-models-venice-web-6269.test.ts`. (thanks @chirag127)
+
+- **fix(api):** the specialty model catalogs (`/v1/embeddings`, `/v1/images`, `/v1/music`, `/v1/videos` model lists) are now derived from the **unified catalog filtered by a predicate** (`getSpecialtyModelsResponse`) instead of ad-hoc per-route logic, so they consistently respect active-credential visibility and stay in sync with the main catalog ([#6303](https://github.com/diegosouzapw/OmniRoute/pull/6303)). Regression guard: `tests/unit/specialty-model-catalog-routes.test.ts`. (thanks @makcimbx)
+- **fix(api):** the agent-bridge server route now resolves the MITM manager via a **dynamic `import("@/mitm/manager.runtime")`** so Turbopack does not statically pull the stub (or over-bundle the manager), and the agent-skills generator anchors its output base path with `path.join(process.cwd(), …)` so Turbopack's static analyzer stops tracing the whole project root ([#6329](https://github.com/diegosouzapw/OmniRoute/issues/6329), [#6366](https://github.com/diegosouzapw/OmniRoute/pull/6366)). Regression guard: `tests/unit/agent-bridge-server-route-dynamic-import.test.ts`. (thanks @Iammilansoni)
+- **fix(api):** internal probes (combo-test, cloud-sync verify) now pick a **management-scoped / allow-all API key** instead of naively grabbing `getApiKeys()[0]` — a restricted `self:usage` first row made the probe fail with "Model X is not allowed for this API key" even when the combo path was healthy (`pickApiKeyForInternalUse` in `src/lib/db/apiKeys.ts`). The API-manager model editor also falls back to `/api/models?all=true` when `/v1/models` is catalog-protected ([#6372](https://github.com/diegosouzapw/OmniRoute/pull/6372)). Regression guard: `tests/unit/pick-internal-api-key-6372.test.ts`. (thanks @jmengit)
+- **fix(live-ws):** the Live Dashboard WebSocket server now **rejects on bind failure** (e.g. `EADDRINUSE` when the API bridge already holds the port) instead of letting the error surface as an unhandled `error` event that crash-loops the process — the `error` listener is attached to `wss` (not `server`) and releases the EventBus subscription on a failed start ([#6324](https://github.com/diegosouzapw/OmniRoute/issues/6324)). Regression guard: `tests/unit/live-ws-eaddrinuse-6324.test.ts`. (thanks @vinayakkulkarni)
+- **fix(dashboard):** the Home provider-topology widget now trusts the live provider-metrics snapshot — it uses `topology.errorProvider` and live `activeRequests` directly instead of re-deriving state from a stale `lastErrorAt` or applying a frontend timeout filter, so the topology reflects real-time provider health ([#6322](https://github.com/diegosouzapw/OmniRoute/pull/6322)). Regression guard: `tests/unit/home-provider-topology-live-state.test.ts`. (thanks @xz-dev)
+- **fix(sse):** strip zero-width markers from streamed **tool-call arguments** — a follow-up to [#5857](https://github.com/diegosouzapw/OmniRoute/pull/5857). That PR removed injected zero-width joiners (U+200D) from streamed assistant text/reasoning but deliberately left tool-call argument JSON byte-exact. The request-side obfuscation (`open-sse/services/claudeCodeObfuscation.ts`) injects ZWJ into agent words — including the temp path inside the Bash tool description — and Claude models copy that verbatim into generated commands, which are delivered as tool-call arguments rather than assistant text. As a result the ZWJ survived and corrupted code blocks (e.g. a temp path rendered with an invisible joiner). Now `open-sse/handlers/responseSanitizer.ts` strips zero-width code points from tool-call argument strings at every emit site (OpenAI non-stream/stream chat `tool_calls` + legacy `function_call`, native Responses `function_call` items, the OpenAI→Responses conversion, and the native Responses streaming `response.function_call_arguments.delta/.done` events). Only zero-width code points are removed; JSON structure and all other bytes stay identical (no parse/restringify), so normal arguments remain byte-exact. Regression guard: 6 new cases in `tests/unit/response-sanitizer.test.ts` (suite 50/50).
+
+- **fix(nodejs):** the default app log path now resolves under `DATA_DIR` (`~/.omniroute/logs/application/app.log`) instead of `process.cwd()` ([#6197](https://github.com/diegosouzapw/OmniRoute/issues/6197)) — the globally-installed CLI runs from an arbitrary working directory, so anchoring the default to cwd made file logging silently write to (or no-op under) an unrelated directory, contradicting the documented `.env.example` default. `getAppLogFilePath()` now computes the default lazily via the pure `resolveDataDir()` resolver (honours a per-process `DATA_DIR`, no directory-creation side effect); an explicit `APP_LOG_FILE_PATH` still wins. Regression guard: `tests/unit/logenv-datadir-path-6197.test.ts` (3).
+
+- **fix(docker):** AgentBridge/`startMitm` no longer aborts in containers/headless when the Antigravity-default DNS step can't write `/etc/hosts` ([#6127](https://github.com/diegosouzapw/OmniRoute/issues/6127)), and the privileged command's stderr now reaches `app.log` instead of only a bare exit code hitting the toast ([#6198](https://github.com/diegosouzapw/OmniRoute/issues/6198)). The default DNS step (`addDNSEntry`) was called unguarded while cert install and the two sibling DNS steps were each best-effort — in the runtime Docker image (`USER node`, no `sudo`, read-only `/etc/hosts`) it threw `Command failed with code 1` out of `startMitmInternal` and killed the whole start, discarding the stderr. The three DNS steps are extracted into a best-effort `provisionDnsEntries()` where each failure is logged with the full `err` (stderr included, folded in by `systemCommands.ts`) and never aborts the start. Regression guard: `tests/unit/mitm-dns-graceful-degrade-6127.test.ts` (4).
+
+- **fix(providers):** copilot-m365-web now supports the M365 Education "Starter / OfficeWebIncludedCopilot" tier and no longer returns an empty `content:null` stream ([#6210](https://github.com/diegosouzapw/OmniRoute/issues/6210)). Two gaps: (1) `buildWsUrl()` hardcoded the individual-consumer scenario (`OfficeWebPaidConsumerCopilot`, `isEdu=false`) — the EDU tier is now opt-in via `providerSpecificData.tier="edu"`, emitting `scenario=OfficeWebIncludedCopilot`/`isEdu=true` (the individual path is unchanged); (2) the EDU/GPT-5.5 path streams deltas via `arguments[0].writeAtCursor` (incremental) instead of only `messages[].text` (accumulated snapshots), which the parser dropped — a new `accumulateBotContent()` folds both formats, with `type:2 item.result.message` as a last-resort fallback. Regression guard: `tests/unit/copilot-m365-edu-writeatcursor-6210.test.ts` (10). (thanks @qpeyba)
+
+- **fix(providers):** GitLab Duo executor now feeds tool results back into the prompt instead of looping ([#6220](https://github.com/diegosouzapw/OmniRoute/issues/6220)) — `buildPrompt()` branched only on `system`/`user` and took `userParts.at(-1)`, silently dropping the `assistant{tool_calls}` + `tool{result}` turns the client appended, so the reconstructed prompt was byte-identical to turn 1 and the model re-emitted the same `<tool>` call forever. When a tool exchange is present the full conversation is now serialized, folding each tool result back keyed by its `tool_call_id`; simple conversations keep the legacy shape. Complements the tool_call emission from [#6051](https://github.com/diegosouzapw/OmniRoute/issues/6051) (the `kilo-duplicate` label was a false positive — different, sequential defect). Regression guard: `tests/unit/gitlab-tool-result-feedback-6220.test.ts` (4).
+
+- **fix(providers):** opencode-go/opencode-zen can now synthesize the OpenCode CLI identity headers Cloudflare requires on VPS egress ([#5997](https://github.com/diegosouzapw/OmniRoute/issues/5997)) — on a datacenter VPS, `opencode.ai/zen/go/v1/chat/completions` 403s (HTML challenge) requests lacking CLI identity, while the reporter's control curl proved that `User-Agent: opencode-cli/1.0.0` + `x-opencode-client: cli` + `x-opencode-project: default` + fresh request/session UUIDs succeed. Opt-in via `OPENCODE_SYNTHESIZE_CLI_HEADERS=true` (values overridable via `OPENCODE_GO_USER_AGENT`/`OPENCODE_USER_AGENT`/`OPENCODE_CLIENT`/`OPENCODE_PROJECT`); it fills only headers the client did not already send. Kept **off by default** — the forward-only path is deliberate (fabricating a wrong value risks upstream rejection; a prior dedup regressed with `opencode/local`), so this replaces the fragile local header-injection shim without changing default behavior. Regression guard: `tests/unit/opencode-cli-headers-synthesis-5997.test.ts` (6). (thanks @aleksesipenko)
+- fix(resilience): sticky session affinity now evicts and fails over to another account when the pinned account is exhausted/unavailable (#6219)
+
+- fix(sse): Responses API passthrough now drops internal commentary-phase output before forwarding to clients (gated by RESPONSES_PASSTHROUGH_DROP_COMMENTARY, default on) (#6199)
+
+- **fix(sse):** tool-call function schemas with a root `type: null` are now coerced to `type: "object"` before dispatch ([#6359](https://github.com/diegosouzapw/OmniRoute/issues/6359)) — clients like the Codex app emit `parameters: { type: null, ... }` for some tools, which OpenAI-compatible upstreams reject with `400 Invalid schema for function '...': schema must be a JSON Schema of 'type: "object"', got 'type: null'`, failing the whole request. `toolSchemaSanitizer` already stripped the null; it now re-adds the mandatory root `"object"` type (and empty `properties`/open `additionalProperties` when absent). Combinator roots (`anyOf`/`oneOf`/`allOf`) and explicit root types are left untouched. Regression guard: 5 new cases in `tests/unit/tool-schema-sanitizer.test.mjs`.
+
+- **fix(docker):** AgentBridge no longer fails to start on npm/Electron/VPS installs with "MITM manager stub reached at runtime" ([#6344](https://github.com/diegosouzapw/OmniRoute/issues/6344)) — v3.8.45 flipped the production bundler default to Turbopack, but `next.config.mjs` aliased `@/mitm/manager` to its Docker-only degraded stub **unconditionally**. That was harmless while Docker (which sets the alias intentionally for #3390 graceful degradation) was the sole Turbopack consumer, but once every artifact built with Turbopack the stub shipped to all non-Docker users and `startMitm` threw on the first Agent-Bridge start. The alias is now opt-in via `OMNIROUTE_MITM_STUB=1` (set only by the Dockerfile) through the shared `scripts/build/mitm-stub-flag.mjs` helper; default builds bundle the real manager. Regression guard: `tests/unit/mitm-stub-alias-6344.test.mjs` (4).
+
+- **fix(proxy):** stop the v3.8.44 proxy regression that leaked the real IP and disabled healthy proxies ([#6246](https://github.com/diegosouzapw/OmniRoute/issues/6246)). Two coupled defects from the new health scheduler: (1) **IP leak** — when a proxy assigned to a connection was marked `inactive`, resolution fell through to a **direct** egress instead of blocking, exposing the operator's real IP; (2) **over-deactivation** — the sweep flipped a proxy to `inactive` on the **first** failed probe and counted our own 5s timeout / a probe-target `5xx` as the proxy's fault, so healthy paid proxies vanished from egress selection ("my proxies are not being used anymore"). Fix: the sweep decision is extracted into a pure, network-free `decideProxyHealthAction` (`src/lib/proxyHealth/decision.ts`) — by default the health check now **only counts/logs and never downgrades status** (a proxy is downgraded/removed only with `PROXY_AUTO_REMOVE=true`, after `PROXY_AUTO_REMOVE_AFTER` **consecutive** conclusive failures); probes are classified tri-state so an inconclusive result (our timeout, or a `5xx` from the probe target) never penalizes the proxy, and the probe timeout is raised 5s→15s. Separately, `safeResolveProxy` now **fails closed** via the existing policy: a connection whose assigned proxy is dead is blocked instead of leaking direct (`hasBlockingProxyAssignment`), honoring the explicit `proxy off` toggles and the `PROXY_FAIL_OPEN=true` opt-out. Existing proxies stuck `inactive` by the old behavior need a one-time manual re-activate (the operator owns proxy status). Regression guards: `tests/unit/proxy-health-decide-action-6246.test.ts`, `tests/unit/proxy-assigned-unavailable-6246.test.ts`.
+
+- **fix(proxy):** make "Test All" read-only and add bulk enable/disable ([#6246](https://github.com/diegosouzapw/OmniRoute/issues/6246)). Complements the core fail-closed / scheduler fix (#6296) with the two remaining reporter asks. (1) The **"Test All" button** (`POST /api/settings/proxies/auto-test`) used to flip a proxy to `inactive` on a failed reachability probe; since the egress selector excludes `inactive` proxies, a flaky probe (an unreachable `httpbin.org`, a proxy that blocks `HEAD`, or a slow paid proxy) silently disabled every proxy that failed — "Test All" is now **read-only by default** (only the operator sets a proxy active/inactive; opt back into the legacy test-and-set with `PROXY_HEALTH_AUTO_DEACTIVATE=true`). (2) Adds a **bulk enable/disable** proxies endpoint + toolbar action (`POST /api/settings/proxies/batch-activate`) so an operator can re-activate proxies in one click. Regression guard: `tests/unit/proxy-health-6246.test.ts`. (thanks @tenshiak)
+
+- **chatcore (tools): stop the default 128-tool cap from silently dropping opencode's `task`/MCP tools.** opencode (used as an MCP/agent host) sends a large tool list; when it exceeds the speculative `MAX_TOOLS_LIMIT` (128) default, `truncateToolList` did a blind `tools.slice(0, 128)`, dropping every tool past index 128 — including opencode's built-in `task` tool (subagent launch) and many MCP tools, so models routed through OmniRoute could no longer spawn subagents or reach part of their tools. The cap exists to avoid upstream `400`s for providers with real hard limits (e.g. grok-cli 200), so it is kept for those: detection of the opencode client (`isOpencodeClient` — any `x-opencode-*` header, or `opencode` in the user-agent) now only bypasses the **speculative 128 default**, never a known provider ceiling. Precedence is explicit — a proactive/detected provider limit always truncates (even for opencode); otherwise opencode forwards its full tool list; otherwise the unchanged 128 default applies to every other client. Refactors `getEffectiveToolLimit` into `getKnownToolLimit(provider) ?? DEFAULT_LIMIT` (byte-identical for existing callers) and fixes a cosmetic debug-log that reported the truncated count instead of the original. Regression guard: `tests/unit/tool-limit-detector.test.ts`.
+
+- **fix(mitm):** the macOS MITM-cert install check now matches the system keychain again. `security find-certificate -a -Z` prints the SHA-1 as a colon-less hex string, but the installed-check compared it against `getCertFingerprint()`'s colon-separated form, so the substring match never hit — the cert was reported as not-installed and re-prompted for the sudo install on every run. Fingerprints are now normalized (colons stripped, upper-cased) on both sides via the extracted `macCertOutputHasFingerprint` helper. Regression guard: `tests/unit/mitm-cert-mac-fingerprint.test.ts`. ([#6204](https://github.com/diegosouzapw/OmniRoute/pull/6204), closes [#6134](https://github.com/diegosouzapw/OmniRoute/issues/6134) — thanks @rianonehub)
+
+- **fix(api):** `/v1/messages/count_tokens` now counts `tool_use`, `tool_result` and `thinking` content blocks (and array-form `system` prompts) in the local-estimation path, instead of only `text`. Real agentic conversations keep ~95% of their tokens inside tool results; the previous estimate returned near-zero for them, which silently broke Claude Code's auto-compaction (context grew past the window with no compaction until the upstream API rejected the request). The real provider-side count path is unchanged. Regression guard: `tests/unit/messages-count-tokens-route.test.ts`. ([#6221](https://github.com/diegosouzapw/OmniRoute/pull/6221) — thanks @luweiCN)
+
+- **fix(antigravity):** strip a trailing assistant prefill turn for Vertex Claude models to avoid upstream 400s ([#6114](https://github.com/diegosouzapw/OmniRoute/pull/6114)). Regression guard: `tests/unit/antigravity-claude-prefill-strip.test.ts`. (thanks @anki1kr)
+
+- **fix(security):** the mutable cloud-agent routes (`/api/cloud/credentials/update`, `/api/cloud/models/alias`) now require management auth instead of being treated as public. They were classified as public API routes, so a request without management credentials could update stored cloud-agent credentials and model aliases. They are removed from the public-route set, classified as management routes in the authz pipeline, and gated by `requireManagementAuth`; cloud **read**/auth routes stay public. Regression guards: `tests/unit/cloud-write-auth.test.ts`, `tests/unit/authz/classify.test.ts`, `tests/unit/public-api-routes.test.ts`. ([#6233](https://github.com/diegosouzapw/OmniRoute/pull/6233) — thanks @vittoroliveira-dev)
+
+- **refactor(dashboard):** extract the onboarding-wizard "Open provider details" link target into a pure, unit-tested `buildProviderDetailsHref(connection)` helper. The wizard already routes by `connection.id` (the node UUID) rather than the provider category slug (#6144/#6145); this hardens that behavior behind a tested helper that guards a missing id/connection. Regression guard: `tests/unit/provider-onboarding-href.test.ts`. ([#6166](https://github.com/diegosouzapw/OmniRoute/pull/6166) — thanks @KooshaPari)
+
+- **fix(security):** the doubao synthetic device-id generator now derives its digits via an unbiased crypto-random draw (rejection sampling over `crypto.randomBytes()`) instead of a `% 10` reduction, closing a CodeQL `js/biased-cryptographic-random` finding.
+- **fix(agentSkills):** the GitHub-skills generator now resolves `outputDir` to an absolute path before writing, fixing a regression introduced by #6366 (relative-to-cwd base path) that could write generated skill files to the wrong directory.
+- **fix(security):** `/api/keys/{id}/devices` now checks the HTTP method before auth/validation, returning a `405` for non-GET/DELETE verbs instead of a misleading `401`/`500` (closes a `dast-smoke` QUERY-method finding).
+- **fix(quality):** clear the last 2 heavy quality-gate reds on the release tip (cycle pre-flight).
+- **fix(mitm):** the test suite and CI can never mutate the OS trust store — `OMNIROUTE_SKIP_SYSTEM_TRUST=1` is set globally for tests/CI so `installCert`/`uninstallCert`/`installTproxyCa` skip the privileged OS dispatch ([#6310](https://github.com/diegosouzapw/OmniRoute/pull/6310); full detail is under the [3.8.45] section below — this branch received it via the parallel-cycle sync-back).
+- **fix(api):** `POST /api/github-skills` now Zod-validates its request body; documented the new quality-gate env vars and pinned the merge-integrity GitHub Actions to a commit SHA.
+- **fix(skills):** generate the missing `omni-github-skills` registry entry and align the agent-skills catalog-count tests (follow-up to #6186).
+- **fix(quality):** clear the cycle's 11 net-new ESLint errors and make `validate-release-green` suppressions-aware.
+
+### 📝 Maintenance
+
+- **i18n(it):** add 118 missing Italian (`it`) translations (net-additive — no existing keys dropped, valid JSON), improving Italian UI coverage. ([#6212](https://github.com/diegosouzapw/OmniRoute/pull/6212) — thanks @serverless83)
+- **chore(providers):** remove deprecated **MiMo V2** model entries from the catalogs (xiaomi-mimo, opencode-go, zenmux-free, audio TTS) — the upstream V2 line is superseded by MiMo V2.5; drops `mimo-v2-tts`, `mimo-v2-pro`, `mimo-v2-omni`, `mimo-v2-flash`, `mimo-v2-flash-free` and realigns the provider-catalog tests. ([#6248](https://github.com/diegosouzapw/OmniRoute/pull/6248) — thanks @backryun)
+- **chore(release):** ~50 commits on this branch are v3.8.45 pre-flight/hardening fixes and CI-perf work that landed here via the parallel-cycle sync-back (`sync-next-cycle.mjs`, Hard Rule #21) **after** the `v3.8.45` git tag was cut, and are already fully documented under the **[3.8.45]** section below — listed here only so the per-cycle commit-coverage check (`npm run release:uncovered`) doesn't flag them as gaps. Provider/catalog/UX/backend fixes: #6041, #6078, #6108, #6135, #6148, #6149, #6154, #6158, #6161, #6162, #6163, #6164, #6165, #6170, #6177, #6178, #6181, #6186, #6187, #6191, #6193, #6194, #6195, #6200, #6205, #6208, #6209, #6211, #6213, #6223, #6224, #6225, #6226, #6227, #6228, #6229, #6230, #6235, #6291, #6292. CI/release-pipeline work: #6167, #6203, #6214, #6215, #6218, #6273, #6275, #6283, #6284, #6285, #6300, #6305.
+- **chore(release):** additional zero-ref release-cycle plumbing on this branch, kept out of `release:uncovered` on purpose (no `#N` in the commit subject to cite): opening the v3.8.46 cycle, opening/closing the v3.8.45 cycle, the finalized [3.8.45] CHANGELOG i18n sync-back to 42 mirrors, the v3.8.45 cognitive/cyclomatic and file-size drift rebaselines, ESLint stale-suppression pruning (4,273 → 4,233), and clearing test-masking/docs-all pre-flight reds for v3.8.45.
+
+### ⚡ Performance & Infrastructure
+
+- **perf(release-green):** the pre-flight validator (`scripts/quality/validate-release-green.mjs`) now runs its 4 slow suites (unit / vitest / integration / pack-artifact) **concurrently** via `Promise.all` — pre-flight wall time drops from ~the sum of the suites to ~the slowest one (~30min saved per round; Phase 0 was the nº1 bottleneck of the v3.8.45 release benchmark, 2h54 of 6h34 e2e). Guard: `tests/unit/validate-release-green.test.ts` ("runs the slow suites CONCURRENTLY"). ([#6319](https://github.com/diegosouzapw/OmniRoute/pull/6319))
+- **fix(ci):** `scripts/release/sync-next-cycle.mjs` — two defects found live in its first production run (v3.8.45 Phase 5): (1) the `git()` helper's default 1 MiB `maxBuffer` crashed with `ENOBUFS` on `git show origin/main:CHANGELOG.md` (the CHANGELOG alone is >1 MiB) — widened to 64 MiB; (2) the i18n resync only propagated the `[NEXT]` (TBD) section, leaving the just-shipped finalized section as "— TBD" in all 42 mirrors — it now also syncs `[prevVersion]` bounded by the heading below it (new exported pure helper `versionAfter`). Guards: +5 tests in `tests/unit/sync-next-cycle.test.ts` (8/8). ([#6327](https://github.com/diegosouzapw/OmniRoute/pull/6327))
+- **test(ci):** concurrency-sensitive flaky tests are **quarantined into a serial pass** (`tests/unit/serial/`, `--test-concurrency=1`, appended to every unit runner incl. sharded variants — the serial pass is sharded too so concurrent shard jobs never self-collide). Initial set: `glm-coding-plan-monthly-3580`, `quota-division-blocks`, `provider-health-autopilot`, `combo-health-autopilot` — the class behind the ~28min CI wedges/re-runs (two live 1h+ wedges cancelled during this PR's own validation). Discovery + TIA gates track the new glob; systemic root cause (async logger writing after teardown) tracked in [#6360](https://github.com/diegosouzapw/OmniRoute/issues/6360). Guard: `tests/unit/test-serial-quarantine.test.ts` (4). ([#6347](https://github.com/diegosouzapw/OmniRoute/pull/6347))
+
+### 🙌 Contributors
+
+Thanks to everyone whose work landed in v3.8.46:
+
+| Contributor                                                    | PRs / Issues           |
+| -------------------------------------------------------------- | ---------------------- |
+| [@2220258345](https://github.com/2220258345)                   | direct commit / report |
+| [@abdofallah](https://github.com/abdofallah)                   | direct commit / report |
+| [@adentdk](https://github.com/adentdk)                         | direct commit / report |
+| [@aleksesipenko](https://github.com/aleksesipenko)             | direct commit / report |
+| [@anki1kr](https://github.com/anki1kr)                         | direct commit / report |
+| [@anungma](https://github.com/anungma)                         | direct commit / report |
+| [@arpicato](https://github.com/arpicato)                       | direct commit / report |
+| [@backryun](https://github.com/backryun)                       | #6248                  |
+| [@binsarjr](https://github.com/binsarjr)                       | direct commit / report |
+| [@brightfiscalband](https://github.com/brightfiscalband)       | direct commit / report |
+| [@chirag127](https://github.com/chirag127)                     | #6501, #6506           |
+| [@developerjillur](https://github.com/developerjillur)         | direct commit / report |
+| [@dilneiss](https://github.com/dilneiss)                       | #6499                  |
+| [@dtybnrj](https://github.com/dtybnrj)                         | direct commit / report |
+| [@Forcerecon](https://github.com/Forcerecon)                   | direct commit / report |
+| [@hao3039032](https://github.com/hao3039032)                   | direct commit / report |
+| [@Iammilansoni](https://github.com/Iammilansoni)               | #6150, #6245           |
+| [@jmengit](https://github.com/jmengit)                         | direct commit / report |
+| [@jordansilly77-stack](https://github.com/jordansilly77-stack) | direct commit / report |
+| [@JxnLexn](https://github.com/JxnLexn)                         | direct commit / report |
+| [@KooshaPari](https://github.com/KooshaPari)                   | #6166                  |
+| [@loopyd](https://github.com/loopyd)                           | direct commit / report |
+| [@luweiCN](https://github.com/luweiCN)                         | #6221                  |
+| [@makcimbx](https://github.com/makcimbx)                       | direct commit / report |
+| [@muflifadla38](https://github.com/muflifadla38)               | direct commit / report |
+| [@newnol](https://github.com/newnol)                           | direct commit / report |
+| [@ofekbetzalel](https://github.com/ofekbetzalel)               | direct commit / report |
+| [@ohahe52-dot](https://github.com/ohahe52-dot)                 | direct commit / report |
+| [@phidinhmanh](https://github.com/phidinhmanh)                 | direct commit / report |
+| [@powellnorma](https://github.com/powellnorma)                 | direct commit / report |
+| [@qpeyba](https://github.com/qpeyba)                           | direct commit / report |
+| [@RaviTharuma](https://github.com/RaviTharuma)                 | direct commit / report |
+| [@RCrushMe](https://github.com/RCrushMe)                       | direct commit / report |
+| [@rianonehub](https://github.com/rianonehub)                   | #6134, #6204           |
+| [@serverless83](https://github.com/serverless83)               | #6212                  |
+| [@swingtempo](https://github.com/swingtempo)                   | direct commit / report |
+| [@tenshiak](https://github.com/tenshiak)                       | direct commit / report |
+| [@ThongAccount](https://github.com/ThongAccount)               | direct commit / report |
+| [@UnrealAryan](https://github.com/UnrealAryan)                 | direct commit / report |
+| [@vinayakkulkarni](https://github.com/vinayakkulkarni)         | direct commit / report |
+| [@vittoroliveira-dev](https://github.com/vittoroliveira-dev)   | #6233                  |
+| [@warelik](https://github.com/warelik)                         | direct commit / report |
+| [@xxy9468615](https://github.com/xxy9468615)                   | direct commit / report |
+| [@xz-dev](https://github.com/xz-dev)                           | direct commit / report |
+| [@yanpaing007](https://github.com/yanpaing007)                 | direct commit / report |
+| [@diegosouzapw](https://github.com/diegosouzapw)               | maintainer             |
+
+---
+
 ## [3.8.45] — 2026-07-06
 
 ### ✨ New Features

docs/i18n/zh-TW/CHANGELOG.md

@@ -6,6 +6,213 @@
 
 ## [3.8.31] — 2026-06-20
 
+## [3.8.46] — TBD
+
+### ✨ New Features
+
+- **feat(sse):** **hide paid-only models from `auto/*` routing** when `hidePaidModels` is on ([#6512](https://github.com/diegosouzapw/OmniRoute/issues/6512)) — follow-up to #6328/#6495. PR #6495 hid paid-only models from the `GET /v1/models` listing, but `auto/*` combos (`auto/best-coding`, `auto/glm`, …) could still pick a paid-only backend into their candidate pool → a 402/403 at request time. `createVirtualAutoCombo` now filters the candidate pool through the new pure `open-sse/services/autoCombo/paidModelFilter.ts` (`filterPaidOnlyCandidates`), applying the same free-model predicate #6495 uses in `catalog.ts` (`providerHasFreeModels(provider) && isFreeModel(provider, {id})`) whenever `settings.hidePaidModels === true`. Applied before the category/tier/family narrowing, so it covers every `auto/*` combo; an all-paid pool degrades to the existing graceful empty-pool path. **Opt-in — default OFF leaves the pool unchanged** (identity). Regression guard: `tests/unit/autoCombo/paid-model-filter-6512.test.ts` (4, incl. the default-off identity guard).
+- **feat(sse):** **provider-family auto combos** — `auto/glm`, `auto/minimax`, `auto/mimo`, `auto/zai`, `auto/gemma`, `auto/llama`, `auto/gemini` ([#6453](https://github.com/diegosouzapw/OmniRoute/issues/6453)) — new routable ids that materialize an on-demand virtual combo spanning whatever installed backends currently expose that model family, degrading gracefully as backends rotate. A new pure `open-sse/services/autoCombo/modelFamily.ts` (`detectModelFamily`) classifies by model-id prefix for six families; `zai` is instead resolved by provider id (z.ai's hosted API serves the same `glm-*` model ids as every other GLM backend, so `auto/zai` means "route to my z.ai backend specifically" vs `auto/glm`'s "any connected GLM backend"). Reuses the existing `createVirtualAutoCombo` on-demand materialization path (no DB writes) and the `/v1/models` catalog advertising loop. Regression guard: `tests/unit/autoCombo/provider-family-combos.test.ts` (11).
+- **feat(proxy):** native **proxy-pool round-robin / egress IP rotation** ([#6365](https://github.com/diegosouzapw/OmniRoute/issues/6365)) — a scope (global / provider / account) can now hold **multiple** proxies as a pool with a rotation strategy, so outbound requests cycle their egress IP instead of pinning one proxy per scope. Migration `117_proxy_pool_rotation.sql` lifts the `UNIQUE(scope, scope_id)` constraint (rebuild via the canonical rename/copy/drop; existing single assignments become 1-element pools) and adds a `proxy_scope_rotation` companion table holding the per-scope strategy + a persisted monotonic round-robin cursor. Strategies: `round-robin` (default, monotonic cursor — never `Math.random`), `random`, and `sticky-per-N-min`. Resolution (`resolveProxyForScopeFromRegistry` / `resolveProxyForConnectionFromRegistry`) now fetches the alive, position-ordered candidate set (unchanged `PROXY_ALIVE_PREDICATE`) and applies the strategy; an empty / all-dead pool still returns `null` — the #6246 fail-closed guard is **untouched** (never falls through to direct egress). Backend + DB only; dashboard pool-builder UI is a follow-up. Regression guard: `tests/unit/proxy-pool-rotation-6365.test.ts` (8, incl. fail-closed + backward-compat).
+- **feat(providers):** end-to-end **tool/function calling on the native Gemini `/v1beta` endpoint** ([#6222](https://github.com/diegosouzapw/OmniRoute/issues/6222)) — both directions of the Gemini↔OpenAI conversion now preserve tool data (previously silently dropped). Request side: `convertGeminiToInternal` (extracted to its own testable module) maps `tools[].functionDeclarations` → OpenAI `tools`, prior `functionCall` parts → assistant `tool_calls`, and `functionResponse` parts → `tool`-role messages. Response side: `convertOpenAIResponseToGemini` emits `parts[].functionCall {name,args}` from `message.tool_calls`, and the streaming `openAIChunkToGeminiChunk` accumulates fragmented `tool_calls` deltas by index into complete `functionCall` parts. The non-Gemini client paths (Claude, OpenAI-Responses) already preserved tool calls — this closes the gap specific to the native Gemini surface. Regression guard: `tests/unit/v1beta-gemini-tool-calling-6222.test.ts` (6, incl. a streaming SSE round-trip).
+- **feat(providers):** copilot-m365-web **enterprise / work tier** support ([#6334](https://github.com/diegosouzapw/OmniRoute/issues/6334)) — mirrors the EDU-tier pattern (#6210): `M365ConnectionParams` gains an `agent` field, a new opt-in `M365_ENTERPRISE_OVERRIDES` preset (`agent=work`, `scenario=officeweb`, `licenseType=Premium`) applies via `providerSpecificData.tier="enterprise"` (alias `"work"`), and `agent` is also overridable directly via `providerSpecificData.agent`. `buildWsUrl` was hardcoding `agent="web"` (the one enterprise-distinguishing param with no override path), so a Premium work account handshook then returned an empty stream. The individual and EDU paths are untouched. Kilo's dup flag vs #6210 (EDU tier) was a false positive — different tier. Regression guard: `tests/unit/copilot-m365-enterprise-6334.test.ts` (7). End-to-end confirmation on a real Premium work account is a live-VPS validation follow-up (Hard Rule #18). (thanks @Forcerecon)
+- **feat(api):** standardized, provider-agnostic **`effort` + `thinking` request params** ([#6241](https://github.com/diegosouzapw/OmniRoute/issues/6241)) — a thin standardization layer over the existing mature per-provider reasoning plumbing (no provider mapper touched). `providerChatCompletionSchema` gains a canonical `effort` (reusing the shared `none/low/medium/high/xhigh` vocabulary — the UI tiers `extra`/`max` collapse onto `xhigh`) and a boolean `thinking`. A pure `normalizeReasoningRequest` (wired once in `src/sse/handlers/chat.ts`, before any reasoning field is read) folds them onto the fields the translators already consume (`reasoning_effort` / `reasoning.effort` / `thinking`), so they fan out to Anthropic / Gemini / xAI / Responses — an explicit client `reasoning_effort` / object-shaped `thinking` always wins (backward-compatible). `/models` additively exposes `supportsThinking` + `effort_tiers` so the frontend can render the toggles (UI component is a follow-up). Regression guard: `tests/unit/effort-thinking-standardization-6241.test.ts` (12). (thanks @Iammilansoni, @shabeer)
+- **feat(combo):** new **`pipeline` (sequential) combo strategy** ([#6297](https://github.com/diegosouzapw/OmniRoute/issues/6297)) — the 18th routing strategy runs targets **in order**, threading each step's output into the next step's input, with an optional per-step `prompt` (system instruction); only the final step's response is returned. Distinct from `fusion` (parallel fan-out + judge). Implemented as a self-contained `open-sse/services/pipeline.ts` (sibling to `fusion.ts`), dispatched from `combo.ts`; the step list reuses `combo.models` order and reads an optional `prompt` off each target (backward-compatible — ignored by every other strategy). Intermediate steps run non-streaming with tools stripped (complete prose to thread forward); the final step keeps the client's `stream` flag + tools. A failing/empty/unparseable intermediate step fails the whole pipeline explicitly via a sanitized error (never silently swallowed). Kilo's dup flag vs #563 was a false positive (that's model→chain selection; this is a sequential chain). Regression guard: `tests/unit/combo-pipeline-strategy.test.ts` (5). (thanks @ofekbetzalel)
+- **feat(ci):** `check:test-masking` now flags **inline-reimplemented prod conditions** ([#6348](https://github.com/diegosouzapw/OmniRoute/issues/6348)) — a new **report-only** subcheck (v2, 6A.10 family) catches the wrong-shape contract test: a test that recomputes the condition under test inline instead of importing/exercising the real function (the #6216 class, where `=== 500` → `>= 500` stayed green because the test re-implemented the branch). For each added/modified test file it warns when the file textually duplicates a ≥3-token conditional from a production file touched in the same PR **and** does not import the symbol/module owning it, via a pure, fixture-tested `findReimplementedConditions()` with an allowlist mirroring `assertReductionAllowlist`. Report-only for now (does not fail the gate) — to be promoted to blocking after a triage cycle. Regression guard: `tests/unit/check-test-masking.test.ts` (45).
+- **feat(sse):** per-connection routing override (native vs CLIProxyAPI) ([#6339](https://github.com/diegosouzapw/OmniRoute/issues/6339)) — the previously-dead `isCliproxyapiDeepModeEnabled` helper is now wired into `resolveExecutorWithProxy`: a single connection can opt itself into the CLIProxyAPI passthrough executor via `providerSpecificData.cliproxyapiMode="claude-native"`, with precedence **connection override > provider `upstream_proxy_config` mode > default**. `resolveExecutorWithProxy` now receives the resolved connection's `providerSpecificData` (threaded from `chatCore.ts`), so one connection can deep-route while the provider's other connections stay native — no DB schema change (the toggle rides in `providerSpecificData`). Also resolves the same-provider mixing ask in #6340. Regression guard: `tests/unit/chatcore-executor-proxy.test.ts` (9). (thanks @RaviTharuma)
+- **feat(dashboard):** "Add session cookie" modal now shows a prominent **"Open ‹host› →"** link to the provider's own site ([#6268](https://github.com/diegosouzapw/OmniRoute/issues/6268)) — every `-web` cookie-session provider (chatgpt-web, claude-web, gemini-web, kimi-web, lmarena, qwen-web, m365-copilot-web, …) renders a one-click external link (opening the provider's login/home page in a new tab) so operators no longer tab away to retype the URL mid-setup. The host resolves from a pure, unit-tested `resolveWebProviderHost()` (prefers `WEB_COOKIE_PROVIDERS[id].website`, falls back to the registry `baseUrl` origin); non-web providers render exactly as before. Kilo's dup flag vs #6265 (modal-too-small-on-1080p) was a false positive — distinct concern. Regression guard: `tests/unit/resolve-web-provider-host.test.ts` (5). (thanks @chirag127)
+- **feat(providers):** add **DigitalOcean** AI (serverless inference) as an OpenAI-compatible API-key provider ([#6373](https://github.com/diegosouzapw/OmniRoute/pull/6373)) — base `https://inference.do-ai.run/v1`, wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/digitalocean/`, `src/shared/constants/providers/apikey/inference-hosts.ts`). Regression guard: `tests/unit/digitalocean-provider.test.ts`. (thanks @newnol)
+- **feat(providers):** add **Huancheng Public API** (`hcnsec`) as an OpenAI-compatible regional provider ([#6410](https://github.com/diegosouzapw/OmniRoute/pull/6410)) — Xinjiang Huancheng Cybersecurity's public LLM platform (base `https://api.hcnsec.cn/v1`, free credits via daily check-ins), wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/hcnsec/`, `src/shared/constants/providers/apikey/regional.ts`). Regression guard: `tests/unit/hcnsec-provider.test.ts`. (thanks @UnrealAryan)
+- **feat(dashboard):** the web-session credential guide now shows an **"Open {host}" link** ([#6316](https://github.com/diegosouzapw/OmniRoute/pull/6316)) to the provider's sign-in site (derived from the provider `website` via `getProviderWebsiteHost`), so you can jump straight to the page where the cookie/session must be captured. Regression guard: `tests/unit/web-session-provider-link-6316.test.ts`. (thanks @jordansilly77-stack)
+- **feat(cerebras):** add the **Gemma 4 31B** model (`gemma-4-31b`) to the Cerebras registry + pricing table ([#6331](https://github.com/diegosouzapw/OmniRoute/pull/6331)). Regression guard extends `tests/unit/t28-model-catalog-updates.test.ts`. (thanks @backryun)
+- **feat(providers):** add **Yuanbao (web)** as a cookie-session provider ([#6196](https://github.com/diegosouzapw/OmniRoute/issues/6196)) — `yuanbao-web` (Tencent Yuanbao, `yuanbao.tencent.com`) with cookie-only auth (`hy_user`/`hy_token` + public agent id), SSE→OpenAI translation incl. `reasoning_content`, exposing DeepSeek V3/R1 + Hunyuan / Hunyuan-T1. Regression guard: `tests/unit/providers-yuanbao-web.test.ts`. `together-web` was **deferred** (no verifiable web-session endpoint — needs a captured request) and `huggingchat-web` **dropped** (the existing `huggingchat` already is a web-cookie provider). (thanks @chirag127)
+- **feat(providers):** route the built-in **agentrouter** through the dynamic Claude-Code wire image ([#6056](https://github.com/diegosouzapw/OmniRoute/issues/6056)) — a small static allow-set (`CC_WIRE_IMAGE_BUILTINS` in `open-sse/services/ccWireImageBuiltins.ts`), consulted by `isClaudeCodeCompatible` / `isClaudeCodeCompatibleProvider` / `applyFingerprint`, makes agentrouter adopt the CC wire-image headers + fingerprint **while guarding the CC baseUrl/auth branches** so it keeps its own registry `baseUrl` and `x-api-key` auth. Regression guard: `tests/unit/agentrouter-cc-wire-image.test.ts` (asserts the wire image is applied AND agentrouter's baseUrl/auth are preserved). Live WAF-acceptance against agentrouter.org is a VPS validation follow-up (Hard Rule #18).
+- **feat(providers):** **bulk-add API keys for Cloudflare Workers AI** ([#6174](https://github.com/diegosouzapw/OmniRoute/issues/6174)) — `cloudflare-ai` is removed from the bulk-add exclusion list and the bulk parser gains a 3-field `name|accountId|apiKey` mode; the bulk route now builds a **per-entry** `providerSpecificData` so each key carries its own `accountId` (fixing the previous shared-object reuse), and both the create + key-validation paths receive it. Regression guard: `tests/unit/bulk-api-key-parser-cloudflare.test.ts`. (thanks @muflifadla38)
+- **feat(dashboard):** routing/settings UX clarity ([#6147](https://github.com/diegosouzapw/OmniRoute/issues/6147)) — (1) weighted combos show the **effective routing share %** next to each weight when weights don't sum to 100 (`WeightTotalBar.tsx`); (2) the status widget's user-facing **"Cloud Sync" label is renamed** to "Remote Settings Sync" (`CloudSyncStatus.tsx`; internal ids/state untouched); (3) built-in providers gain an **opt-in advanced base-URL override** (`isBaseUrlOverrideEligibleProvider`, hidden behind an "Advanced" toggle, reusing the existing `providerSpecificData.baseUrl` persistence — not globally widened). Regression guard: `tests/unit/routing-settings-ux-6147.test.ts`.
+- **feat(combo):** add an option to **disable session stickiness**, per-combo or globally — round-robin / random combos can rotate to a different connection on every request instead of pinning a whole conversation to one connection by its first-message hash. Resolution precedence per-combo `config.disableSessionStickiness` → global `settings.disableSessionStickiness` → default `false` (preserves the #3825 prompt-cache/504 fix); gates **both** stickiness call sites in `open-sse/services/combo.ts`. Exposed as a global toggle (Combo Defaults) and a per-combo Inherit/on/off control. ([#6168](https://github.com/diegosouzapw/OmniRoute/issues/6168)) Regression guard: `tests/unit/combo-disable-session-stickiness.test.ts`. (thanks @RCrushMe)
+- **feat(docker):** add the `OMNIROUTE_NO_SUDO` env flag for root-less / user-namespaced deployments — the MITM cert-trust command path (`resolveSudoSpawn` in `src/mitm/systemCommands.ts`) now strips the leading `sudo` when the flag is truthy, in addition to the existing root / sudo-missing cases, so the Proxy Agent runs without `sudo` (the operator trusts the CA manually, e.g. via `NODE_EXTRA_CA_CERTS`). Argv-array `spawn` preserved — no shell interpolation (Hard Rule #13). ([#6122](https://github.com/diegosouzapw/OmniRoute/issues/6122)) Regression guard: `tests/unit/mitm-systemCommands-no-sudo.test.ts`. (thanks @powellnorma)
+- **feat(providers):** add **Requesty** as an OpenAI-compatible gateway provider (BYOK, base `https://router.requesty.ai/v1`, ~200 free requests/day) — wired through the shared OpenAI-compatible registry with full model passthrough (`open-sse/config/providers/registry/requesty/`, `src/shared/constants/providers/apikey/gateways.ts`). ([#6120](https://github.com/diegosouzapw/OmniRoute/issues/6120)) Regression guard: `tests/unit/requesty-provider.test.ts`. (thanks @chirag127)
+- **feat(dashboard):** add **configured-only / available-only filters** to the Free Provider Rankings page ([#6150](https://github.com/diegosouzapw/OmniRoute/issues/6150)) — hide providers you haven't configured, or whose connections are all rate-limited / out of quota, via server-side query params (`?configuredOnly` / `?availableOnly` on `GET /api/free-provider-rankings`) backed by a testable lib helper reusing the in-process connection state (no Redis). Both filters default off, so the default view is unchanged; this supersedes the earlier client-side "Configured Only" toggle (#6245) with an available-only dimension and unit-tested logic. Regression guard: `tests/unit/freeProviderRankings-filters.test.ts`.
+
+### 🔧 Bug Fixes
+
+- **fix(dashboard):** adding a second API key connection for the same provider no longer silently overwrites the first — the Add-API-key modal now derives a unique default connection name (`main`, then `main-2`, `main-3`, …) so the backend name-based upsert can't collide (#6499 — thanks @dilneiss).
+- **fix(compression):** the session-dedup engine now also deduplicates a large multi-line block repeated **within a single message** (intra-message dedup), not just across turns; the compression-preview API surfaces a `fallbackReason`, and the fusion panel reports how many models were rate-limited vs failed on total-panel failure (#6501 — thanks @chirag127).
+- **fix(compression):** a stacked-pipeline step naming an unregistered engine now surfaces a `validationErrors` entry instead of silently no-op'ing, so misconfigured pipelines are visible in the preview API (#6506 — thanks @chirag127).
+- **feat(usage):** add a **Codex reset-credit redemption** flow to the Provider Limits UI ([#6361](https://github.com/diegosouzapw/OmniRoute/pull/6361)) — a `useCodexResetCreditRedemption` hook + `/api/usage/codex-reset-credit` route + `codexResetCredits` lib let you redeem banked Codex reset credits from the quota card. Regression guard: `tests/unit/codex-reset-credits.test.ts`. (thanks @JxnLexn)
+- **feat(glm):** add **team-plan quota settings** for `glm-cn` connections ([#6351](https://github.com/diegosouzapw/OmniRoute/pull/6351)) — a dedicated `GlmTeamQuotaFields` form section (team quota id / limits) threaded through the Add/Edit connection modals, persisted via `providerSpecificData`, with the GLM usage service reading the team quota. Regression guards: `tests/unit/glm-team-quota.test.ts`, `provider-specific-data-schema.test.ts`. (thanks @hao3039032)
+- **feat(providers):** add **TinyFish** web-fetch/search support ([#6349](https://github.com/diegosouzapw/OmniRoute/pull/6349)) — a `tinyfish-fetch` executor + `/v1/web/fetch` route + MCP web-fetch tool, registered as a specialty-media provider with request-validation and a search-provider catalog entry. Regression guards: `tests/unit/executor-tinyfish-fetch.test.ts`, `web-fetch-handler.test.ts`, `mcp-web-fetch-tool.test.ts`, `provider-validation-tinyfish.test.ts`. (thanks @dtybnrj)
+- **fix(cli):** `omniroute launch-codex` now spawns `codex.cmd` through a shell on Windows (the npm `.cmd` shim is unresolvable by bare `spawn` → ENOENT), mirroring the qodercli Windows fix (#6263) ([#6312](https://github.com/diegosouzapw/OmniRoute/pull/6312)). Regression guard: `tests/unit/launch-codex-windows-spawn-6312.test.ts`. (thanks @swingtempo)
+- **fix(codex):** isolate the **Spark** quota from the shared Codex quota and stabilize the quota UI ordering / hydration so per-scope limits render consistently ([#6336](https://github.com/diegosouzapw/OmniRoute/pull/6336)). Regression guards: `tests/unit/codex-quota-selection-hydration.test.ts`, `provider-limits-ui.test.ts` + 3 more. (thanks @xz-dev)
+- **feat(api):** add a `hidePaidModels` setting that filters paid-only models out of the `/v1/models` catalog. Regression guard: `tests/unit/models-catalog-hide-paid.test.ts`. (thanks @chirag127)
+- **fix(api-manager):** the fallback model picker now preserves combos instead of dropping them when a primary model is unavailable ([#6443](https://github.com/diegosouzapw/OmniRoute/pull/6443)). Regression guard: `tests/unit/api-manager-page-static.test.ts`. (thanks @jmengit)
+- **fix(providers):** recoverable Antigravity / Cloud-Code (Gemini Code Assist) `403` responses ([#6452](https://github.com/diegosouzapw/OmniRoute/pull/6452)) are now classified as a retryable project-config error instead of a terminal account ban, so a fixable project/API-disabled 403 no longer forces a ~1-year cooldown / full OAuth reconnect. Regression guard: `tests/unit/errorclassifier-antigravity-403.test.ts`. (thanks @developerjillur)
+- **fix(mitm):** `sanitizeHeaders` now redacts `Set-Cookie` response headers so upstream session cookies never leak into logs / diagnostics ([#6451](https://github.com/diegosouzapw/OmniRoute/pull/6451)). Regression guard: `tests/unit/mitm-sanitize-headers.test.ts`. (thanks @developerjillur)
+- **fix(api):** `/api/compression/preview` now accepts `mode: "caveman"` and correctly handles stacked / zero-compression previews ([#6425](https://github.com/diegosouzapw/OmniRoute/issues/6425)). Regression guard: `tests/unit/api/compression-preview-caveman-and-stacked-6425.test.ts`. (thanks @chirag127)
+- **feat(providers):** add **Zed** hosted LLM aggregator as a native-app provider ([#6118](https://github.com/diegosouzapw/OmniRoute/pull/6118)) — OAuth sign-in via the Zed hosted flow, registered through the shared provider registry + executor. Regression guards: `tests/unit/zed-oauth-provider.test.ts`, `zed-import-utils.test.ts`, `zed-docker-detect.test.ts`, `mitm-handler-zed.test.ts`. VPS-validated via live operator login (Hard Rule #18).
+- **fix(oauth):** the Kiro SSO-cache auto-import now **preserves the IDC region** — cross-region Amazon Q / Kiro profiles imported from the SSO cache are no longer collapsed to the default region ([#6113](https://github.com/diegosouzapw/OmniRoute/pull/6113)). Regression guard: `tests/unit/kiro-auto-import-idc-2059.test.ts`. VPS-validated via live operator login (Hard Rule #18).
+- **fix(dashboard):** passthrough model aliases no longer collide when two namespaced model ids share a last segment (port from 9router#1850, [#6431](https://github.com/diegosouzapw/OmniRoute/pull/6431)). `enx/gpt-5.5` and `enx/codebuddy/gpt-5.5` both auto-generated the alias `gpt-5.5`, so the second model could never be added (the UI just alerted "alias already exists"). Aliases are now disambiguated deterministically — bare last segment when free, then parent-qualified (`codebuddy-gpt-5.5`), then a numeric suffix — while re-adding the exact same model id is still blocked. Regression guard: `tests/unit/passthrough-alias-1850.test.ts`. (thanks @arpicato)
+- **fix(translator):** preserve a Gemini `functionResponse` co-located with other parts (another `functionCall`, or trailing `text`) in the same content when translating **Gemini → OpenAI** ([#6376](https://github.com/diegosouzapw/OmniRoute/pull/6376)). `convertGeminiContent()` early-returned the tool message on the first `functionResponse` part, dropping any co-located parts; such contents are now pre-split (one tool message per `functionResponse`, emitted first, plus one message for the remaining parts). Regression guard: `tests/unit/gemini-to-openai-function-response.test.ts`. (thanks @warelik)
+- **fix(headroom):** detect a python interpreter managed by **mise / pyenv / asdf / conda** (port from 9router#2353, [#6382](https://github.com/diegosouzapw/OmniRoute/pull/6382)). Headroom's python probe (`src/lib/headroom/detect.ts`) searched a hardcoded `PATH`, but version managers expose their interpreters via shim dirs that only join `PATH` through interactive-shell activation — which the non-interactive server never runs, so a managed python (≥3.10) was invisible and Headroom reported it missing. The search path now prepends the well-known shim/bin dirs (`~/.local/share/mise/shims`, `~/.pyenv/shims`, `~/.asdf/shims`, `$CONDA_PREFIX/bin`, `~/.local/bin`, respecting `MISE_DATA_DIR`/`PYENV_ROOT`/`ASDF_DATA_DIR` when set), and a new `HEADROOM_PYTHON` env override lets operators point straight at their interpreter (mirroring `HEADROOM_URL`). Still shell-free (`execFileSync`). Regression guard: `tests/unit/headroom-detect.test.ts` (5). (thanks @loopyd)
+- **fix(executors):** strip the OpenAI-Codex/Claude-CLI `client_metadata` passthrough field for **NVIDIA** requests (port from 9router#1887, [#6411](https://github.com/diegosouzapw/OmniRoute/pull/6411)). NVIDIA's OpenAI-compatible wrapper rejects it with `400 Unsupported parameter`, the same class already handled for `cerebras`/`mistral`; `nvidia` (executor `default`) was missing from the strip allowlist so Codex/Claude-Code passthrough requests 400'd. Regression guard: `tests/unit/executor-default-strip-client-metadata.test.ts` (+nvidia case). (thanks @phidinhmanh)
+- **fix(translator):** strip the Claude-style `thinking` field for **NVIDIA `z-ai/glm-5.2`** (port from 9router#2023, [#6413](https://github.com/diegosouzapw/OmniRoute/pull/6413)). NVIDIA's OpenAI-compatible wrapper 400s on `thinking` (a Claude-format client routed here leaves a `thinking:{type:"adaptive"}`); the existing strip rule only dropped `reasoning`. Same class already handled for `minimax-m2.7`. Regression guard: `tests/unit/nvidia-minimax-thinking-strip.test.ts` (+glm-5.2 case). (thanks @phidinhmanh)
+- **fix(translator):** suppress the streamed `</think>` close marker for the **Antigravity IDE** client (port from 9router#1061, [#6415](https://github.com/diegosouzapw/OmniRoute/pull/6415)). On thinking-only turns Antigravity rendered a bare `</think>` as the sole visible content, tripping its loop-detection and wasting requests. Antigravity's UA (`vscode/<v> (Antigravity/<v>)`) is added to the marker-suppress allowlist (alongside OpenCode); Claude Code / Cursor still get the marker, and `x-omniroute-thinking-marker: on` force-restores it. Regression guard: `tests/unit/think-close-marker-suppress-5245.test.ts`. (thanks @abdofallah)
+- **fix(executors):** strip nested `reasoning_content` from messages for **Mistral** (port from 9router#1649, [#6417](https://github.com/diegosouzapw/OmniRoute/pull/6417)). Mistral's API returns `422 extra_forbidden` when an assistant message carries `reasoning_content` (replayed thinking from a prior turn, e.g. via the Codex `/responses` path); the generic top-level 400 field-downgrade retry never covered the nested per-message field. `DefaultExecutor` now strips it for provider `mistral` only, so DeepSeek (which requires replayed `reasoning_content`) is unaffected. Regression guard: `tests/unit/mistral-strip-reasoning-content-1649.test.ts`. (thanks @xxy9468615)
+- **fix(executors):** strip the `client_metadata` passthrough field on the **OpenCode** path (port from 9router#1442, [#6418](https://github.com/diegosouzapw/OmniRoute/pull/6418)). OpenCode upstreams (e.g. `kimi-k2.6` via opencode-go) reject it with `400 "Extra inputs are not permitted, field: 'client_metadata'"`; the DefaultExecutor strip only covered cerebras/mistral and `OpencodeExecutor` extends `BaseExecutor` directly, so nothing removed it there. Regression guard: `tests/unit/opencode-strip-client-metadata-1442.test.ts`. (thanks @yanpaing007)
+- **fix(executors):** inject the `reasoning_content` echo for the native **Moonshot Kimi** provider (port from 9router#1480, [#6419](https://github.com/diegosouzapw/OmniRoute/pull/6419)). Kimi (executor `default`) is a thinking-mode upstream that 400s with "reasoning_content must be passed back" when a prior assistant turn lacks it; the placeholder injection was only wired into the OpenCode meta-provider, so direct multi-turn Kimi conversations failed. Scoped to `kimi` (gateway-served models matching the thinking-model name pattern are unaffected). Regression guard: `tests/unit/kimi-native-reasoning-injected-1480.test.ts`. (thanks @2220258345)
+- **fix(executors):** recover from a strict gateway's `context_management: Extra inputs are not permitted` 400 (port from 9router#1468, [#6420](https://github.com/diegosouzapw/OmniRoute/pull/6420)). **Claude Code** always sends a top-level `context_management` field; strict anthropic-compatible gateways reject it. The dedicated context-editing 400-fallback only fired when OmniRoute's own `contextEditing` feature was enabled (default off), so a client-sent field passed through untouched and 400'd. `context_management` is now in the generic reactive field-strip list, so it's stripped-and-retried once regardless of the feature flag (with correct request re-signing for claude-compatible relays). Regression guard: `tests/unit/provider-field-strips.test.ts`. (thanks @ohahe52-dot)
+- **fix(network):** enable **RFC 8305 Happy Eyeballs** (`autoSelectFamily`) on the direct-egress undici dispatcher (port from 9router#1237, [#6423](https://github.com/diegosouzapw/OmniRoute/pull/6423)). When DNS returns both IPv6 (AAAA) and IPv4 (A) and the IPv6 route is broken (e.g. a NAT64 `64:ff9b::` prefix without routing), undici tried IPv6 first and hung until `ETIMEDOUT` (then a 502 + account lockout), even though `curl` reached the same host. The direct dispatcher now races both families and uses whichever connects first. Proxy paths pin family via `proxyTls` and are unaffected. Regression guard: `tests/unit/direct-dispatcher-pipelining-4580.test.ts`. (thanks @adentdk)
+- **fix(combo):** round-robin now advances the rotation pointer past the model that **actually served**, not the eagerly-scheduled one (port from 9router#948, [#6428](https://github.com/diegosouzapw/OmniRoute/pull/6428)). With `stickyLimit: 1` (true round-robin), when the scheduled model failed and a _different_ model served via fallback, the counter had already advanced +1 from the scheduled index — so the next request reused the fallback-served model, degrading round-robin into hot-spotting on whichever model was healthy. The pointer now advances to the served index + 1 (mirroring the sticky-limit>1 path). Session-stickiness (#3825) and distribution are preserved. Regression guard: `tests/unit/combo-rr-fallback-advance-948.test.ts`. (thanks @binsarjr)
+- **fix(sse):** a non-string `model` field is now rejected with a `400` before the resolver, instead of crashing downstream `.toLowerCase()`/`.split()` calls into an empty-body `500` that escapes the error sanitizer ([#6407](https://github.com/diegosouzapw/OmniRoute/issues/6407)). Regression guard: `tests/unit/chat-non-string-model-6407.test.ts`. (thanks @chirag127)
+- **fix(api):** unknown `/api/*` routes now return a JSON `404` (instead of the dashboard HTML shell) and scalar chat params (`model`/`temperature`/etc.) are validated **before** the provider lookup so malformed requests fail fast with a clear `400` ([#6424](https://github.com/diegosouzapw/OmniRoute/issues/6424), [#6412](https://github.com/diegosouzapw/OmniRoute/issues/6412)). Regression guards: `tests/unit/api/api-catchall-json-404.test.ts`, `tests/unit/chat-early-schema-validation-6412.test.ts`. (thanks @chirag127)
+- **fix(api):** `/v1/chat/completions` now rejects a non-JSON `Content-Type` with a `400` before parsing the body ([#6414](https://github.com/diegosouzapw/OmniRoute/issues/6414)). Regression guard: `tests/unit/v1-chat-completions-content-type-6414.test.ts`. (thanks @chirag127)
+- **fix(api):** the `X-OmniRoute-Compression` response header is now echoed on `/v1/chat/completions` and `/v1/completions` ([#6422](https://github.com/diegosouzapw/OmniRoute/issues/6422)). Regression guard: `tests/unit/compression-header-echo-6422.test.ts`. (thanks @chirag127)
+- **fix(api):** concurrent `GET /v1/models` requests are coalesced into a single catalog build ([#6408](https://github.com/diegosouzapw/OmniRoute/issues/6408)). Regression guard: `tests/unit/v1-models-concurrent-6408.test.ts`. (thanks @chirag127)
+- **fix(api):** `/v1/completions` now echoes the requested `body.model` in its JSON + streamed responses ([#6429](https://github.com/diegosouzapw/OmniRoute/pull/6429)). Regression guard: `tests/unit/completions-body-model-echo.test.ts`. (thanks @chirag127)
+- **fix(api):** env-var master keys now see the full `/v1/models` catalog ([#6406](https://github.com/diegosouzapw/OmniRoute/issues/6406)). Regression guard: `tests/unit/models-catalog-envkey-6406.test.ts`. (thanks @chirag127)
+- **fix(api):** non-streaming `/v1/completions` responses now echo `body.model` aligned with the `X-OmniRoute-Model` header ([#6426](https://github.com/diegosouzapw/OmniRoute/issues/6426)). Regression guard: `tests/unit/v1-completions-model-header-match-6426.test.ts`. (thanks @chirag127)
+- **fix(api):** unknown `/v1/*` routes now return a JSON `404 not_found` instead of the Next.js dashboard HTML shell ([#6405](https://github.com/diegosouzapw/OmniRoute/issues/6405)). Regression guard: `tests/unit/api/v1-catchall-json-404.test.ts`. (thanks @chirag127)
+- **fix(api):** the per-connection provider models route now degrades to the shipped catalog when a provider's `/models` endpoint answers with a redirect ([#6267](https://github.com/diegosouzapw/OmniRoute/issues/6267)) — a `qwen-web` import failed with a raw `Redirect blocked … (307)` 503. `safeOutboundFetch` throws `REDIRECT_BLOCKED` on the 307, `getSafeOutboundFetchErrorStatus` maps it to 503, and `buildDiscoveryErrorFallbackResponse` treated every 503 as a hard error — so the non-empty `getModelsByProviderId("qwen-web")` catalog was never surfaced. A models-endpoint redirect is not a fixable-config error (unlike `URL_GUARD_BLOCKED`/`INVALID_URL`, which stay hard errors), so it now falls back to the local/cached catalog before the 503 short-circuit. General fix — covers any config-driven provider that 307s. Regression guard: `tests/unit/provider-models-qwen-web-redirect-6267.test.ts`. (thanks @chirag127)
+- **fix(api):** the per-connection provider models route (MCP `list_models_catalog` + the dashboard import view) now merges USER-ADDED custom models into its response ([#6247](https://github.com/diegosouzapw/OmniRoute/issues/6247)) — custom models live in the `key_value` namespace `customModels`, which the live REST `/api/v1/models` already merges, but `src/app/api/providers/[id]/models/route.ts` never read `getCustomModels`, so custom models were dropped on both the discovery-success and local_catalog paths. They are now folded into the returned model list (deduped by id, stamped `owned_by: provider`), fixing MCP + the dashboard import view in one place. Regression guard: `tests/unit/provider-models-custom-merge-6247.test.ts`. (thanks @RCrushMe)
+
+- **fix(providers):** GitLab Duo tool-calling follow-up turns no longer fail upstream with `422 {"detail":"Validation error"}` (tokens 0/0, rejected pre-inference). The #6234 tool-result-feedback fix serialized the **entire** multi-turn conversation into GitLab's single-file `code_suggestions` (`small_file`) generation endpoint — folded history that turn-N sent as an oversized `current_file.content_above_cursor` **and** duplicated verbatim into `user_instruction`, tripping the AI-Gateway's `small_file` validation guard. The executor now **bounds** that prompt: it keeps system + latest user message + the most-recent tool round (dropping older turns), caps oversized tool results, and stops duplicating the full prompt into `user_instruction` (which now carries only the short latest user message) — while still feeding the most-recent tool result back so the agent continues (`open-sse/executors/gitlab.ts`, [#6220](https://github.com/diegosouzapw/OmniRoute/issues/6220)). The unit test covers the bounding logic; the upstream 422→200 clearing is VPS-only (Hard Rule #18). Regression guard: `tests/unit/gitlab-tool-exchange-bounded-6220.test.ts`.
+
+- **fix(i18n):** the provider-detail (`/dashboard/providers/[id]`) connection-status filter labels no longer render as `__MISSING__:All` / `__MISSING__:Active` / `__MISSING__:Error` / `__MISSING__:Banned` / `__MISSING__:CreditsExhausted` in non-English locales (notably pt-BR) ([#6290](https://github.com/diegosouzapw/OmniRoute/issues/6290)). Root cause was **not** the namespace mismatch the issue guessed — the `providers.filter*` keys resolve correctly in `en.json`; the debt lived in the locale mirrors (`src/i18n/messages/*.json`), where these five keys carried the `__MISSING__:` sync sentinel in ~15 locales and were absent entirely in ~26 others, so next-intl found the key and echoed the sentinel verbatim. All 40 non-English/-Chinese mirrors now ship real translations for the five `providers.filter*` labels. Regression guard: `tests/unit/i18n-provider-filter-keys-6290.test.ts`. (thanks @diegosouzapw)
+
+- **fix(providers):** the `copilot-m365-web` streaming executor now emits `debug`-level WebSocket diagnostics ([#6210](https://github.com/diegosouzapw/OmniRoute/issues/6210)) — the outbound WS URL (with the `access_token` **redacted** via `redactWsUrl()`), handshake success/failure, and each received SignalR frame's `type`/`target`. Previously the streaming path logged nothing, so an empty `content:null` response (the M365 Education / Starter tier symptom fixed in #6234) was undiagnosable even at `APP_LOG_LEVEL=debug`. The change is debug-level and side-effect-free — it does not alter streaming behavior or the frame parser, and the token never reaches the logs. Regression guard: `tests/unit/copilot-m365-web-logging-6210.test.ts` (thanks @qpeyba)
+
+- **fix(resilience):** a round-robin combo no longer returns `503 all upstream accounts are unavailable` when a compatibility-rejected target is actually healthy ([#6238](https://github.com/diegosouzapw/OmniRoute/issues/6238)). `filterTargetsByRequestCompatibility` drops request-incompatible targets (tool/vision/structured-output unsupported, or below the required context window) **before** any availability check runs, and its `compatible.length === 0` safety net only fired when _all_ targets were filtered — not when the kept targets later all turned out runtime-unavailable (circuit-open / cooldown / no credentials). So a combo could 503 while a compat-rejected-but-healthy provider sat unused. `handleRoundRobinCombo` now keeps the compat-rejected set and, when every compat-kept target was skipped without a single real attempt, probes those rejected targets as a **last-resort fallback tier** (via the new pure `open-sse/services/combo/comboCompatFallback.ts`) before crystallizing the 503. Regression guard: `tests/unit/combo-roundrobin-compat-fallback-6238.test.ts`. (thanks @ThongAccount)
+
+- **fix(startup):** best-effort self-heal for a corrupted Turbopack dev cache on Windows ([#6289](https://github.com/diegosouzapw/OmniRoute/issues/6289)). On Windows, `pnpm dev` can fail at startup when Turbopack `mmap`s a persistent-cache SST file and the OS refuses the mapping (`os error 1455` — "paging file too small"), which Turbopack surfaces as a misleading `Module not found: Can't resolve '@/shared/utils/machine'`. This is a **known upstream Turbopack cache-corruption bug — not our code**. The dev launcher (`scripts/dev/run-next.mjs`) now wraps `nextApp.prepare()` and, when it rejects with that signature (`isTurbopackCacheCorruption` in the new `scripts/dev/turbopackCacheHeal.mjs`), purges `.build/next/**/cache/turbopack` and retries **once** with a clear log. **Caveat — best-effort only:** the corruption often surfaces as a runtime overlay rather than a `prepare()` rejection, so this cannot always intercept it; the reliable remedy remains manually deleting the Turbopack cache dir. Regression guard: `tests/unit/turbopack-cache-heal-6289.test.ts`. (thanks @chirag127)
+
+- **fix(providers):** qodercli PAT auth no longer fails with `spawn qodercli ENOENT` on Windows ([#6263](https://github.com/diegosouzapw/OmniRoute/issues/6263)) — `spawnQoderCli` spawned the bare `qodercli` name with `shell:false` and an unenriched env, so the npm `.cmd` wrapper under `%APPDATA%\npm` (a user-PATH directory) was never resolved. It now resolves the absolute `.cmd`/`.exe` path through the existing `getCliRuntimeStatus("qoder")` resolver in `src/shared/services/cliRuntime.ts` (memoized), spawns with `shell` when the target is a `.cmd`/`.bat`, and uses the cliRuntime-enriched env (PATH + PATHEXT + APPDATA); the ENOENT error now lists the searched paths plus the `CLI_QODER_BIN` override. End-to-end spawn on a real Windows host is host-only (Hard Rule #18); the path-resolution logic is unit-tested. Regression guard: `tests/unit/qodercli-windows-resolve-6263.test.ts`. (thanks @chirag127)
+
+- **fix(sse):** the reasoning-token buffer no longer inflates **probe-sized `max_tokens`** ([#6274](https://github.com/diegosouzapw/OmniRoute/issues/6274)) — Claude Code's `/model` capability check sends `max_tokens: 1`, but for a thinking-capable model with a large output cap (e.g. `glm-5.2`) the #3587 headroom heuristic (`max(current + 1000, ceil(current * 1.5))`) rewrote it to `1001` and forwarded that upstream, wasting tokens on a request that was never a genuine reasoning budget. `resolveReasoningBufferedMaxTokens()` (`open-sse/services/reasoningTokenBuffer.ts`) now short-circuits and returns the caller's value verbatim when it is below the new `REASONING_BUFFER_MIN_TRIGGER` (256) threshold — a tiny explicit limit is a probe, not a reasoning request. Real budgets still receive the #3587 headroom unchanged, and the guard runs after the existing capability checks so unknown / non-reasoning models keep returning `null`. Regression guard: `tests/unit/reasoning-token-buffer-6274.test.ts`. (thanks @brightfiscalband)
+
+- **fix(cli):** `omniroute reset-password` now works as a real subcommand, and password resets over piped (non-TTY) stdin actually apply ([#6261](https://github.com/diegosouzapw/OmniRoute/issues/6261), [#6258](https://github.com/diegosouzapw/OmniRoute/issues/6258)). Two coupled defects: (1) **#6261** — `bin/omniroute.mjs` routed everything through Commander with only two pre-Commander bypasses (`--mcp`, `reset-encrypted-columns`), so `omniroute reset-password` was rejected as an unknown command; only the separate `omniroute-reset-password` bin worked, while the docs falsely advertised the subcommand (incl. a bogus "legacy alias still works"). A pre-Commander bypass mirroring `reset-encrypted-columns` now dynamically imports `bin/reset-password.mjs` (which self-executes) before Commander parses; the three doc lines were corrected. (2) **#6258** — `bin/reset-password.mjs` issued two sequential `rl.question` prompts; under piped stdin the second read never settled at EOF, so `main()` never reached `resetManagementPassword` and the reset was a silent no-op (both prompts printed, no success, password unchanged). The CLI now detects non-TTY stdin and reads it once (first line = password, second line = confirm if present, else reused), adds a `--password-stdin` flag (entire stdin is the password, no confirmation), and exits `0` explicitly so the success line always flushes; interactive TTY behavior is unchanged. Regression guard: `tests/unit/reset-password-cli-6261-6258.test.ts` (3). (thanks @chirag127)
+
+- **fix(db):** the mass-migration **safety abort** now tells the operator how to bypass it and stops flooding the log ([#6260](https://github.com/diegosouzapw/OmniRoute/issues/6260)) — after restoring a backup that wiped the migration tracking table, `runMigrations()` threw the abort on every downstream `ensureDbInitialized()`, re-logging the full banner 11+ times, and the message never mentioned the existing `OMNIROUTE_MAX_PENDING_MIGRATIONS` escape hatch. The abort text now appends a bypass hint (set `OMNIROUTE_MAX_PENDING_MIGRATIONS=0` in `server.env` / `DATA_DIR/.env`), and a new `MigrationSafetyAbortError` is memoized so repeated calls in the same process throw the same instance and emit a single concise line instead of the full cascade. Regression guard: `tests/unit/migration-safety-abort-6260.test.ts`. (thanks @chirag127)
+
+- **fix(auth):** importing a **distinct** Codex/ChatGPT OAuth `auth.json` is no longer falsely rejected as "already exists" when it belongs to a different user in the same workspace ([#6301](https://github.com/diegosouzapw/OmniRoute/issues/6301)). `findExistingCodexConnection` (in `src/lib/oauth/utils/codexAuthImport.ts`) deduped **only** on `providerSpecificData.workspaceId === accountId`, where `accountId` is the shared `chatgpt_account_id`/`tokens.account_id` — so two members of the same ChatGPT Team collapsed onto a single connection (409 `duplicate_account`). The id_token's `https://api.openai.com/auth` claim carries a per-user `chatgpt_user_id` alongside the workspace id (the device-flow path already persisted it as `chatgptUserId`, but the import path did not). Now `parseAndValidateCodexAuth` extracts `userId` (`chatgpt_user_id` → `user_id` → JWT `sub`) into `ParsedCodexAuth`, the create/update paths persist `chatgptUserId` in `providerSpecificData` (mirroring `codex.ts`), and dedup keys on `workspaceId` **AND** `chatgptUserId` — with a backward-compat fallback to legacy accountId-only matching when no stored connection for that workspace records a `chatgptUserId`, so genuinely-same accounts still dedup. Regression guard: `tests/unit/codex-auth-import-userid-dedup-6301.test.ts` (4). (thanks @anungma)
+
+- **fix(providers):** importing models for the **venice-web** provider no longer fails with a red "Provider venice-web does not support models listing" ([#6269](https://github.com/diegosouzapw/OmniRoute/issues/6269)). `venice-web` is a web-cookie provider with an executor but no upstream `/v1/models` endpoint and no registry `models`, so the models route fell through to the tail `400`. Mirroring the `jules`/`linkup-search`/`ollama-search` fix (#5569), it now ships a static local catalog entry in `src/lib/providers/staticModels.ts` — seeding the current Venice lineup (`venice-uncensored`, `llama-3.3-70b`, `qwen3-235b`, `qwen3-4b`, `deepseek-r1-671b`; Venice rotates its catalog, see docs.venice.ai/models/overview) — so the route returns `200` with `source:"local_catalog"`, `intentional:true`. Regression guard: `tests/unit/static-models-venice-web-6269.test.ts`. (thanks @chirag127)
+
+- **fix(api):** the specialty model catalogs (`/v1/embeddings`, `/v1/images`, `/v1/music`, `/v1/videos` model lists) are now derived from the **unified catalog filtered by a predicate** (`getSpecialtyModelsResponse`) instead of ad-hoc per-route logic, so they consistently respect active-credential visibility and stay in sync with the main catalog ([#6303](https://github.com/diegosouzapw/OmniRoute/pull/6303)). Regression guard: `tests/unit/specialty-model-catalog-routes.test.ts`. (thanks @makcimbx)
+- **fix(api):** the agent-bridge server route now resolves the MITM manager via a **dynamic `import("@/mitm/manager.runtime")`** so Turbopack does not statically pull the stub (or over-bundle the manager), and the agent-skills generator anchors its output base path with `path.join(process.cwd(), …)` so Turbopack's static analyzer stops tracing the whole project root ([#6329](https://github.com/diegosouzapw/OmniRoute/issues/6329), [#6366](https://github.com/diegosouzapw/OmniRoute/pull/6366)). Regression guard: `tests/unit/agent-bridge-server-route-dynamic-import.test.ts`. (thanks @Iammilansoni)
+- **fix(api):** internal probes (combo-test, cloud-sync verify) now pick a **management-scoped / allow-all API key** instead of naively grabbing `getApiKeys()[0]` — a restricted `self:usage` first row made the probe fail with "Model X is not allowed for this API key" even when the combo path was healthy (`pickApiKeyForInternalUse` in `src/lib/db/apiKeys.ts`). The API-manager model editor also falls back to `/api/models?all=true` when `/v1/models` is catalog-protected ([#6372](https://github.com/diegosouzapw/OmniRoute/pull/6372)). Regression guard: `tests/unit/pick-internal-api-key-6372.test.ts`. (thanks @jmengit)
+- **fix(live-ws):** the Live Dashboard WebSocket server now **rejects on bind failure** (e.g. `EADDRINUSE` when the API bridge already holds the port) instead of letting the error surface as an unhandled `error` event that crash-loops the process — the `error` listener is attached to `wss` (not `server`) and releases the EventBus subscription on a failed start ([#6324](https://github.com/diegosouzapw/OmniRoute/issues/6324)). Regression guard: `tests/unit/live-ws-eaddrinuse-6324.test.ts`. (thanks @vinayakkulkarni)
+- **fix(dashboard):** the Home provider-topology widget now trusts the live provider-metrics snapshot — it uses `topology.errorProvider` and live `activeRequests` directly instead of re-deriving state from a stale `lastErrorAt` or applying a frontend timeout filter, so the topology reflects real-time provider health ([#6322](https://github.com/diegosouzapw/OmniRoute/pull/6322)). Regression guard: `tests/unit/home-provider-topology-live-state.test.ts`. (thanks @xz-dev)
+- **fix(sse):** strip zero-width markers from streamed **tool-call arguments** — a follow-up to [#5857](https://github.com/diegosouzapw/OmniRoute/pull/5857). That PR removed injected zero-width joiners (U+200D) from streamed assistant text/reasoning but deliberately left tool-call argument JSON byte-exact. The request-side obfuscation (`open-sse/services/claudeCodeObfuscation.ts`) injects ZWJ into agent words — including the temp path inside the Bash tool description — and Claude models copy that verbatim into generated commands, which are delivered as tool-call arguments rather than assistant text. As a result the ZWJ survived and corrupted code blocks (e.g. a temp path rendered with an invisible joiner). Now `open-sse/handlers/responseSanitizer.ts` strips zero-width code points from tool-call argument strings at every emit site (OpenAI non-stream/stream chat `tool_calls` + legacy `function_call`, native Responses `function_call` items, the OpenAI→Responses conversion, and the native Responses streaming `response.function_call_arguments.delta/.done` events). Only zero-width code points are removed; JSON structure and all other bytes stay identical (no parse/restringify), so normal arguments remain byte-exact. Regression guard: 6 new cases in `tests/unit/response-sanitizer.test.ts` (suite 50/50).
+
+- **fix(nodejs):** the default app log path now resolves under `DATA_DIR` (`~/.omniroute/logs/application/app.log`) instead of `process.cwd()` ([#6197](https://github.com/diegosouzapw/OmniRoute/issues/6197)) — the globally-installed CLI runs from an arbitrary working directory, so anchoring the default to cwd made file logging silently write to (or no-op under) an unrelated directory, contradicting the documented `.env.example` default. `getAppLogFilePath()` now computes the default lazily via the pure `resolveDataDir()` resolver (honours a per-process `DATA_DIR`, no directory-creation side effect); an explicit `APP_LOG_FILE_PATH` still wins. Regression guard: `tests/unit/logenv-datadir-path-6197.test.ts` (3).
+
+- **fix(docker):** AgentBridge/`startMitm` no longer aborts in containers/headless when the Antigravity-default DNS step can't write `/etc/hosts` ([#6127](https://github.com/diegosouzapw/OmniRoute/issues/6127)), and the privileged command's stderr now reaches `app.log` instead of only a bare exit code hitting the toast ([#6198](https://github.com/diegosouzapw/OmniRoute/issues/6198)). The default DNS step (`addDNSEntry`) was called unguarded while cert install and the two sibling DNS steps were each best-effort — in the runtime Docker image (`USER node`, no `sudo`, read-only `/etc/hosts`) it threw `Command failed with code 1` out of `startMitmInternal` and killed the whole start, discarding the stderr. The three DNS steps are extracted into a best-effort `provisionDnsEntries()` where each failure is logged with the full `err` (stderr included, folded in by `systemCommands.ts`) and never aborts the start. Regression guard: `tests/unit/mitm-dns-graceful-degrade-6127.test.ts` (4).
+
+- **fix(providers):** copilot-m365-web now supports the M365 Education "Starter / OfficeWebIncludedCopilot" tier and no longer returns an empty `content:null` stream ([#6210](https://github.com/diegosouzapw/OmniRoute/issues/6210)). Two gaps: (1) `buildWsUrl()` hardcoded the individual-consumer scenario (`OfficeWebPaidConsumerCopilot`, `isEdu=false`) — the EDU tier is now opt-in via `providerSpecificData.tier="edu"`, emitting `scenario=OfficeWebIncludedCopilot`/`isEdu=true` (the individual path is unchanged); (2) the EDU/GPT-5.5 path streams deltas via `arguments[0].writeAtCursor` (incremental) instead of only `messages[].text` (accumulated snapshots), which the parser dropped — a new `accumulateBotContent()` folds both formats, with `type:2 item.result.message` as a last-resort fallback. Regression guard: `tests/unit/copilot-m365-edu-writeatcursor-6210.test.ts` (10). (thanks @qpeyba)
+
+- **fix(providers):** GitLab Duo executor now feeds tool results back into the prompt instead of looping ([#6220](https://github.com/diegosouzapw/OmniRoute/issues/6220)) — `buildPrompt()` branched only on `system`/`user` and took `userParts.at(-1)`, silently dropping the `assistant{tool_calls}` + `tool{result}` turns the client appended, so the reconstructed prompt was byte-identical to turn 1 and the model re-emitted the same `<tool>` call forever. When a tool exchange is present the full conversation is now serialized, folding each tool result back keyed by its `tool_call_id`; simple conversations keep the legacy shape. Complements the tool_call emission from [#6051](https://github.com/diegosouzapw/OmniRoute/issues/6051) (the `kilo-duplicate` label was a false positive — different, sequential defect). Regression guard: `tests/unit/gitlab-tool-result-feedback-6220.test.ts` (4).
+
+- **fix(providers):** opencode-go/opencode-zen can now synthesize the OpenCode CLI identity headers Cloudflare requires on VPS egress ([#5997](https://github.com/diegosouzapw/OmniRoute/issues/5997)) — on a datacenter VPS, `opencode.ai/zen/go/v1/chat/completions` 403s (HTML challenge) requests lacking CLI identity, while the reporter's control curl proved that `User-Agent: opencode-cli/1.0.0` + `x-opencode-client: cli` + `x-opencode-project: default` + fresh request/session UUIDs succeed. Opt-in via `OPENCODE_SYNTHESIZE_CLI_HEADERS=true` (values overridable via `OPENCODE_GO_USER_AGENT`/`OPENCODE_USER_AGENT`/`OPENCODE_CLIENT`/`OPENCODE_PROJECT`); it fills only headers the client did not already send. Kept **off by default** — the forward-only path is deliberate (fabricating a wrong value risks upstream rejection; a prior dedup regressed with `opencode/local`), so this replaces the fragile local header-injection shim without changing default behavior. Regression guard: `tests/unit/opencode-cli-headers-synthesis-5997.test.ts` (6). (thanks @aleksesipenko)
+- fix(resilience): sticky session affinity now evicts and fails over to another account when the pinned account is exhausted/unavailable (#6219)
+
+- fix(sse): Responses API passthrough now drops internal commentary-phase output before forwarding to clients (gated by RESPONSES_PASSTHROUGH_DROP_COMMENTARY, default on) (#6199)
+
+- **fix(sse):** tool-call function schemas with a root `type: null` are now coerced to `type: "object"` before dispatch ([#6359](https://github.com/diegosouzapw/OmniRoute/issues/6359)) — clients like the Codex app emit `parameters: { type: null, ... }` for some tools, which OpenAI-compatible upstreams reject with `400 Invalid schema for function '...': schema must be a JSON Schema of 'type: "object"', got 'type: null'`, failing the whole request. `toolSchemaSanitizer` already stripped the null; it now re-adds the mandatory root `"object"` type (and empty `properties`/open `additionalProperties` when absent). Combinator roots (`anyOf`/`oneOf`/`allOf`) and explicit root types are left untouched. Regression guard: 5 new cases in `tests/unit/tool-schema-sanitizer.test.mjs`.
+
+- **fix(docker):** AgentBridge no longer fails to start on npm/Electron/VPS installs with "MITM manager stub reached at runtime" ([#6344](https://github.com/diegosouzapw/OmniRoute/issues/6344)) — v3.8.45 flipped the production bundler default to Turbopack, but `next.config.mjs` aliased `@/mitm/manager` to its Docker-only degraded stub **unconditionally**. That was harmless while Docker (which sets the alias intentionally for #3390 graceful degradation) was the sole Turbopack consumer, but once every artifact built with Turbopack the stub shipped to all non-Docker users and `startMitm` threw on the first Agent-Bridge start. The alias is now opt-in via `OMNIROUTE_MITM_STUB=1` (set only by the Dockerfile) through the shared `scripts/build/mitm-stub-flag.mjs` helper; default builds bundle the real manager. Regression guard: `tests/unit/mitm-stub-alias-6344.test.mjs` (4).
+
+- **fix(proxy):** stop the v3.8.44 proxy regression that leaked the real IP and disabled healthy proxies ([#6246](https://github.com/diegosouzapw/OmniRoute/issues/6246)). Two coupled defects from the new health scheduler: (1) **IP leak** — when a proxy assigned to a connection was marked `inactive`, resolution fell through to a **direct** egress instead of blocking, exposing the operator's real IP; (2) **over-deactivation** — the sweep flipped a proxy to `inactive` on the **first** failed probe and counted our own 5s timeout / a probe-target `5xx` as the proxy's fault, so healthy paid proxies vanished from egress selection ("my proxies are not being used anymore"). Fix: the sweep decision is extracted into a pure, network-free `decideProxyHealthAction` (`src/lib/proxyHealth/decision.ts`) — by default the health check now **only counts/logs and never downgrades status** (a proxy is downgraded/removed only with `PROXY_AUTO_REMOVE=true`, after `PROXY_AUTO_REMOVE_AFTER` **consecutive** conclusive failures); probes are classified tri-state so an inconclusive result (our timeout, or a `5xx` from the probe target) never penalizes the proxy, and the probe timeout is raised 5s→15s. Separately, `safeResolveProxy` now **fails closed** via the existing policy: a connection whose assigned proxy is dead is blocked instead of leaking direct (`hasBlockingProxyAssignment`), honoring the explicit `proxy off` toggles and the `PROXY_FAIL_OPEN=true` opt-out. Existing proxies stuck `inactive` by the old behavior need a one-time manual re-activate (the operator owns proxy status). Regression guards: `tests/unit/proxy-health-decide-action-6246.test.ts`, `tests/unit/proxy-assigned-unavailable-6246.test.ts`.
+
+- **fix(proxy):** make "Test All" read-only and add bulk enable/disable ([#6246](https://github.com/diegosouzapw/OmniRoute/issues/6246)). Complements the core fail-closed / scheduler fix (#6296) with the two remaining reporter asks. (1) The **"Test All" button** (`POST /api/settings/proxies/auto-test`) used to flip a proxy to `inactive` on a failed reachability probe; since the egress selector excludes `inactive` proxies, a flaky probe (an unreachable `httpbin.org`, a proxy that blocks `HEAD`, or a slow paid proxy) silently disabled every proxy that failed — "Test All" is now **read-only by default** (only the operator sets a proxy active/inactive; opt back into the legacy test-and-set with `PROXY_HEALTH_AUTO_DEACTIVATE=true`). (2) Adds a **bulk enable/disable** proxies endpoint + toolbar action (`POST /api/settings/proxies/batch-activate`) so an operator can re-activate proxies in one click. Regression guard: `tests/unit/proxy-health-6246.test.ts`. (thanks @tenshiak)
+
+- **chatcore (tools): stop the default 128-tool cap from silently dropping opencode's `task`/MCP tools.** opencode (used as an MCP/agent host) sends a large tool list; when it exceeds the speculative `MAX_TOOLS_LIMIT` (128) default, `truncateToolList` did a blind `tools.slice(0, 128)`, dropping every tool past index 128 — including opencode's built-in `task` tool (subagent launch) and many MCP tools, so models routed through OmniRoute could no longer spawn subagents or reach part of their tools. The cap exists to avoid upstream `400`s for providers with real hard limits (e.g. grok-cli 200), so it is kept for those: detection of the opencode client (`isOpencodeClient` — any `x-opencode-*` header, or `opencode` in the user-agent) now only bypasses the **speculative 128 default**, never a known provider ceiling. Precedence is explicit — a proactive/detected provider limit always truncates (even for opencode); otherwise opencode forwards its full tool list; otherwise the unchanged 128 default applies to every other client. Refactors `getEffectiveToolLimit` into `getKnownToolLimit(provider) ?? DEFAULT_LIMIT` (byte-identical for existing callers) and fixes a cosmetic debug-log that reported the truncated count instead of the original. Regression guard: `tests/unit/tool-limit-detector.test.ts`.
+
+- **fix(mitm):** the macOS MITM-cert install check now matches the system keychain again. `security find-certificate -a -Z` prints the SHA-1 as a colon-less hex string, but the installed-check compared it against `getCertFingerprint()`'s colon-separated form, so the substring match never hit — the cert was reported as not-installed and re-prompted for the sudo install on every run. Fingerprints are now normalized (colons stripped, upper-cased) on both sides via the extracted `macCertOutputHasFingerprint` helper. Regression guard: `tests/unit/mitm-cert-mac-fingerprint.test.ts`. ([#6204](https://github.com/diegosouzapw/OmniRoute/pull/6204), closes [#6134](https://github.com/diegosouzapw/OmniRoute/issues/6134) — thanks @rianonehub)
+
+- **fix(api):** `/v1/messages/count_tokens` now counts `tool_use`, `tool_result` and `thinking` content blocks (and array-form `system` prompts) in the local-estimation path, instead of only `text`. Real agentic conversations keep ~95% of their tokens inside tool results; the previous estimate returned near-zero for them, which silently broke Claude Code's auto-compaction (context grew past the window with no compaction until the upstream API rejected the request). The real provider-side count path is unchanged. Regression guard: `tests/unit/messages-count-tokens-route.test.ts`. ([#6221](https://github.com/diegosouzapw/OmniRoute/pull/6221) — thanks @luweiCN)
+
+- **fix(antigravity):** strip a trailing assistant prefill turn for Vertex Claude models to avoid upstream 400s ([#6114](https://github.com/diegosouzapw/OmniRoute/pull/6114)). Regression guard: `tests/unit/antigravity-claude-prefill-strip.test.ts`. (thanks @anki1kr)
+
+- **fix(security):** the mutable cloud-agent routes (`/api/cloud/credentials/update`, `/api/cloud/models/alias`) now require management auth instead of being treated as public. They were classified as public API routes, so a request without management credentials could update stored cloud-agent credentials and model aliases. They are removed from the public-route set, classified as management routes in the authz pipeline, and gated by `requireManagementAuth`; cloud **read**/auth routes stay public. Regression guards: `tests/unit/cloud-write-auth.test.ts`, `tests/unit/authz/classify.test.ts`, `tests/unit/public-api-routes.test.ts`. ([#6233](https://github.com/diegosouzapw/OmniRoute/pull/6233) — thanks @vittoroliveira-dev)
+
+- **refactor(dashboard):** extract the onboarding-wizard "Open provider details" link target into a pure, unit-tested `buildProviderDetailsHref(connection)` helper. The wizard already routes by `connection.id` (the node UUID) rather than the provider category slug (#6144/#6145); this hardens that behavior behind a tested helper that guards a missing id/connection. Regression guard: `tests/unit/provider-onboarding-href.test.ts`. ([#6166](https://github.com/diegosouzapw/OmniRoute/pull/6166) — thanks @KooshaPari)
+
+- **fix(security):** the doubao synthetic device-id generator now derives its digits via an unbiased crypto-random draw (rejection sampling over `crypto.randomBytes()`) instead of a `% 10` reduction, closing a CodeQL `js/biased-cryptographic-random` finding.
+- **fix(agentSkills):** the GitHub-skills generator now resolves `outputDir` to an absolute path before writing, fixing a regression introduced by #6366 (relative-to-cwd base path) that could write generated skill files to the wrong directory.
+- **fix(security):** `/api/keys/{id}/devices` now checks the HTTP method before auth/validation, returning a `405` for non-GET/DELETE verbs instead of a misleading `401`/`500` (closes a `dast-smoke` QUERY-method finding).
+- **fix(quality):** clear the last 2 heavy quality-gate reds on the release tip (cycle pre-flight).
+- **fix(mitm):** the test suite and CI can never mutate the OS trust store — `OMNIROUTE_SKIP_SYSTEM_TRUST=1` is set globally for tests/CI so `installCert`/`uninstallCert`/`installTproxyCa` skip the privileged OS dispatch ([#6310](https://github.com/diegosouzapw/OmniRoute/pull/6310); full detail is under the [3.8.45] section below — this branch received it via the parallel-cycle sync-back).
+- **fix(api):** `POST /api/github-skills` now Zod-validates its request body; documented the new quality-gate env vars and pinned the merge-integrity GitHub Actions to a commit SHA.
+- **fix(skills):** generate the missing `omni-github-skills` registry entry and align the agent-skills catalog-count tests (follow-up to #6186).
+- **fix(quality):** clear the cycle's 11 net-new ESLint errors and make `validate-release-green` suppressions-aware.
+
+### 📝 Maintenance
+
+- **i18n(it):** add 118 missing Italian (`it`) translations (net-additive — no existing keys dropped, valid JSON), improving Italian UI coverage. ([#6212](https://github.com/diegosouzapw/OmniRoute/pull/6212) — thanks @serverless83)
+- **chore(providers):** remove deprecated **MiMo V2** model entries from the catalogs (xiaomi-mimo, opencode-go, zenmux-free, audio TTS) — the upstream V2 line is superseded by MiMo V2.5; drops `mimo-v2-tts`, `mimo-v2-pro`, `mimo-v2-omni`, `mimo-v2-flash`, `mimo-v2-flash-free` and realigns the provider-catalog tests. ([#6248](https://github.com/diegosouzapw/OmniRoute/pull/6248) — thanks @backryun)
+- **chore(release):** ~50 commits on this branch are v3.8.45 pre-flight/hardening fixes and CI-perf work that landed here via the parallel-cycle sync-back (`sync-next-cycle.mjs`, Hard Rule #21) **after** the `v3.8.45` git tag was cut, and are already fully documented under the **[3.8.45]** section below — listed here only so the per-cycle commit-coverage check (`npm run release:uncovered`) doesn't flag them as gaps. Provider/catalog/UX/backend fixes: #6041, #6078, #6108, #6135, #6148, #6149, #6154, #6158, #6161, #6162, #6163, #6164, #6165, #6170, #6177, #6178, #6181, #6186, #6187, #6191, #6193, #6194, #6195, #6200, #6205, #6208, #6209, #6211, #6213, #6223, #6224, #6225, #6226, #6227, #6228, #6229, #6230, #6235, #6291, #6292. CI/release-pipeline work: #6167, #6203, #6214, #6215, #6218, #6273, #6275, #6283, #6284, #6285, #6300, #6305.
+- **chore(release):** additional zero-ref release-cycle plumbing on this branch, kept out of `release:uncovered` on purpose (no `#N` in the commit subject to cite): opening the v3.8.46 cycle, opening/closing the v3.8.45 cycle, the finalized [3.8.45] CHANGELOG i18n sync-back to 42 mirrors, the v3.8.45 cognitive/cyclomatic and file-size drift rebaselines, ESLint stale-suppression pruning (4,273 → 4,233), and clearing test-masking/docs-all pre-flight reds for v3.8.45.
+
+### ⚡ Performance & Infrastructure
+
+- **perf(release-green):** the pre-flight validator (`scripts/quality/validate-release-green.mjs`) now runs its 4 slow suites (unit / vitest / integration / pack-artifact) **concurrently** via `Promise.all` — pre-flight wall time drops from ~the sum of the suites to ~the slowest one (~30min saved per round; Phase 0 was the nº1 bottleneck of the v3.8.45 release benchmark, 2h54 of 6h34 e2e). Guard: `tests/unit/validate-release-green.test.ts` ("runs the slow suites CONCURRENTLY"). ([#6319](https://github.com/diegosouzapw/OmniRoute/pull/6319))
+- **fix(ci):** `scripts/release/sync-next-cycle.mjs` — two defects found live in its first production run (v3.8.45 Phase 5): (1) the `git()` helper's default 1 MiB `maxBuffer` crashed with `ENOBUFS` on `git show origin/main:CHANGELOG.md` (the CHANGELOG alone is >1 MiB) — widened to 64 MiB; (2) the i18n resync only propagated the `[NEXT]` (TBD) section, leaving the just-shipped finalized section as "— TBD" in all 42 mirrors — it now also syncs `[prevVersion]` bounded by the heading below it (new exported pure helper `versionAfter`). Guards: +5 tests in `tests/unit/sync-next-cycle.test.ts` (8/8). ([#6327](https://github.com/diegosouzapw/OmniRoute/pull/6327))
+- **test(ci):** concurrency-sensitive flaky tests are **quarantined into a serial pass** (`tests/unit/serial/`, `--test-concurrency=1`, appended to every unit runner incl. sharded variants — the serial pass is sharded too so concurrent shard jobs never self-collide). Initial set: `glm-coding-plan-monthly-3580`, `quota-division-blocks`, `provider-health-autopilot`, `combo-health-autopilot` — the class behind the ~28min CI wedges/re-runs (two live 1h+ wedges cancelled during this PR's own validation). Discovery + TIA gates track the new glob; systemic root cause (async logger writing after teardown) tracked in [#6360](https://github.com/diegosouzapw/OmniRoute/issues/6360). Guard: `tests/unit/test-serial-quarantine.test.ts` (4). ([#6347](https://github.com/diegosouzapw/OmniRoute/pull/6347))
+
+### 🙌 Contributors
+
+Thanks to everyone whose work landed in v3.8.46:
+
+| Contributor                                                    | PRs / Issues           |
+| -------------------------------------------------------------- | ---------------------- |
+| [@2220258345](https://github.com/2220258345)                   | direct commit / report |
+| [@abdofallah](https://github.com/abdofallah)                   | direct commit / report |
+| [@adentdk](https://github.com/adentdk)                         | direct commit / report |
+| [@aleksesipenko](https://github.com/aleksesipenko)             | direct commit / report |
+| [@anki1kr](https://github.com/anki1kr)                         | direct commit / report |
+| [@anungma](https://github.com/anungma)                         | direct commit / report |
+| [@arpicato](https://github.com/arpicato)                       | direct commit / report |
+| [@backryun](https://github.com/backryun)                       | #6248                  |
+| [@binsarjr](https://github.com/binsarjr)                       | direct commit / report |
+| [@brightfiscalband](https://github.com/brightfiscalband)       | direct commit / report |
+| [@chirag127](https://github.com/chirag127)                     | #6501, #6506           |
+| [@developerjillur](https://github.com/developerjillur)         | direct commit / report |
+| [@dilneiss](https://github.com/dilneiss)                       | #6499                  |
+| [@dtybnrj](https://github.com/dtybnrj)                         | direct commit / report |
+| [@Forcerecon](https://github.com/Forcerecon)                   | direct commit / report |
+| [@hao3039032](https://github.com/hao3039032)                   | direct commit / report |
+| [@Iammilansoni](https://github.com/Iammilansoni)               | #6150, #6245           |
+| [@jmengit](https://github.com/jmengit)                         | direct commit / report |
+| [@jordansilly77-stack](https://github.com/jordansilly77-stack) | direct commit / report |
+| [@JxnLexn](https://github.com/JxnLexn)                         | direct commit / report |
+| [@KooshaPari](https://github.com/KooshaPari)                   | #6166                  |
+| [@loopyd](https://github.com/loopyd)                           | direct commit / report |
+| [@luweiCN](https://github.com/luweiCN)                         | #6221                  |
+| [@makcimbx](https://github.com/makcimbx)                       | direct commit / report |
+| [@muflifadla38](https://github.com/muflifadla38)               | direct commit / report |
+| [@newnol](https://github.com/newnol)                           | direct commit / report |
+| [@ofekbetzalel](https://github.com/ofekbetzalel)               | direct commit / report |
+| [@ohahe52-dot](https://github.com/ohahe52-dot)                 | direct commit / report |
+| [@phidinhmanh](https://github.com/phidinhmanh)                 | direct commit / report |
+| [@powellnorma](https://github.com/powellnorma)                 | direct commit / report |
+| [@qpeyba](https://github.com/qpeyba)                           | direct commit / report |
+| [@RaviTharuma](https://github.com/RaviTharuma)                 | direct commit / report |
+| [@RCrushMe](https://github.com/RCrushMe)                       | direct commit / report |
+| [@rianonehub](https://github.com/rianonehub)                   | #6134, #6204           |
+| [@serverless83](https://github.com/serverless83)               | #6212                  |
+| [@swingtempo](https://github.com/swingtempo)                   | direct commit / report |
+| [@tenshiak](https://github.com/tenshiak)                       | direct commit / report |
+| [@ThongAccount](https://github.com/ThongAccount)               | direct commit / report |
+| [@UnrealAryan](https://github.com/UnrealAryan)                 | direct commit / report |
+| [@vinayakkulkarni](https://github.com/vinayakkulkarni)         | direct commit / report |
+| [@vittoroliveira-dev](https://github.com/vittoroliveira-dev)   | #6233                  |
+| [@warelik](https://github.com/warelik)                         | direct commit / report |
+| [@xxy9468615](https://github.com/xxy9468615)                   | direct commit / report |
+| [@xz-dev](https://github.com/xz-dev)                           | direct commit / report |
+| [@yanpaing007](https://github.com/yanpaing007)                 | direct commit / report |
+| [@diegosouzapw](https://github.com/diegosouzapw)               | maintainer             |
+
+---
+
 ## [3.8.45] — 2026-07-06
 
 ### ✨ New Features

docs/openapi.yaml

@@ -1,7 +1,7 @@
 openapi: 3.1.0
 info:
   title: OmniRoute API
-  version: 3.8.45
+  version: 3.8.46
   description: |
     OmniRoute is a local-first AI API proxy router. It provides an OpenAI-compatible
     endpoint that routes requests to multiple AI providers with load balancing,

docs/reference/CLI-TOOLS.md

@@ -648,7 +648,7 @@ omniroute providers validate                        # Local-only structural vali
 ### Recovery & Reset
 
 ```bash
-omniroute reset-password                # Reset the admin password (legacy alias still works)
+omniroute reset-password                # Reset the admin password (also: omniroute-reset-password)
 omniroute reset-encrypted-columns       # Show warning + dry-run for encrypted credential reset
 omniroute reset-encrypted-columns --force  # Actually null out encrypted credentials in SQLite
 ```

docs/routing/AUTO-COMBO.md

@@ -155,10 +155,10 @@ combo's stored config. These apply only to the `auto` strategy and only for the
 that carries them; the combo's saved `modePack`/`budgetCap` are used when the header is
 absent.
 
-| Header               | Accepts                                                                                              | Effect                                                                                                            |
-| :------------------- | :-------------------------------------------------------------------------------------------------- | :-------------------------------------------------------------------------------------------------------------- |
-| `X-OmniRoute-Mode`   | a preset alias (`fast`, `balanced`, `quality`, `cheap`, `reliable`, `offline`) or a raw pack name (`ship-fast`, `cost-saver`, `quality-first`, `offline-friendly`, `reliability-first`) | Overrides the scoring weights for this request. `balanced`/`default` force the default weights (no pack). Unknown values are ignored (config preserved). |
-| `X-OmniRoute-Budget` | a positive number (max USD per request)                                                             | Hard cost ceiling: candidates whose estimated cost exceeds it are filtered before selection, falling back to the cheapest healthy candidate if all exceed. Non-positive/garbage values are ignored. |
+| Header               | Accepts                                                                                                                                                                                 | Effect                                                                                                                                                                                              |
+| :------------------- | :-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | :-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| `X-OmniRoute-Mode`   | a preset alias (`fast`, `balanced`, `quality`, `cheap`, `reliable`, `offline`) or a raw pack name (`ship-fast`, `cost-saver`, `quality-first`, `offline-friendly`, `reliability-first`) | Overrides the scoring weights for this request. `balanced`/`default` force the default weights (no pack). Unknown values are ignored (config preserved).                                            |
+| `X-OmniRoute-Budget` | a positive number (max USD per request)                                                                                                                                                 | Hard cost ceiling: candidates whose estimated cost exceeds it are filtered before selection, falling back to the cheapest healthy candidate if all exceed. Non-positive/garbage values are ignored. |
 
 ```bash
 # Force the fastest profile and cap this request at $0.05
@@ -174,27 +174,28 @@ resolved values feed the engine's existing `config.modePack` / `config.budgetCap
 
 ## All Routing Strategies
 
-OmniRoute's combo engine supports **17 routing strategies** (declared in `src/shared/constants/routingStrategies.ts` → `ROUTING_STRATEGY_VALUES`). The Auto Combo engine itself is exposed under the `auto` strategy; the others are available for persisted combos.
+OmniRoute's combo engine supports **18 routing strategies** (declared in `src/shared/constants/routingStrategies.ts` → `ROUTING_STRATEGY_VALUES`). The Auto Combo engine itself is exposed under the `auto` strategy; the others are available for persisted combos.
 
-| Strategy            | Description                                                                                  |
-| :------------------ | :------------------------------------------------------------------------------------------- |
-| `priority`          | First-target ordered list with explicit priority                                             |
-| `weighted`          | Weighted random by per-target weight                                                         |
-| `round-robin`       | Cycle through targets in order                                                               |
-| `context-relay`     | Hand off context across targets (long conversations)                                         |
-| `fill-first`        | Fill each target's quota before moving to next                                               |
-| `p2c`               | Power-of-2-choices random load balancing                                                     |
-| `random`            | Uniform random selection                                                                     |
-| `least-used`        | Pick target with lowest current load                                                         |
-| `cost-optimized`    | Minimize $ per request given catalog pricing                                                 |
-| `reset-aware` ⭐    | Prioritize by quota reset time — short reset windows ranked higher                           |
-| `reset-window`      | Prefer targets whose quota window resets soonest                                             |
-| `headroom`          | Pick the target with the most remaining quota headroom                                       |
-| `strict-random`     | Random without deduplication of repeats                                                      |
-| `auto`              | Use Auto Combo scoring (9-factor) — **recommended**                                          |
-| `lkgp`              | Last-Known-Good Path (sticky route to last successful target)                                |
-| `context-optimized` | Pick target with best fit for current context size                                           |
-| `fusion` 🧬         | Fan out to a panel of models in parallel, then synthesize one answer via a judge (see below) |
+| Strategy            | Description                                                                                                                  |
+| :------------------ | :--------------------------------------------------------------------------------------------------------------------------- |
+| `priority`          | First-target ordered list with explicit priority                                                                             |
+| `weighted`          | Weighted random by per-target weight                                                                                         |
+| `round-robin`       | Cycle through targets in order                                                                                               |
+| `context-relay`     | Hand off context across targets (long conversations)                                                                         |
+| `fill-first`        | Fill each target's quota before moving to next                                                                               |
+| `p2c`               | Power-of-2-choices random load balancing                                                                                     |
+| `random`            | Uniform random selection                                                                                                     |
+| `least-used`        | Pick target with lowest current load                                                                                         |
+| `cost-optimized`    | Minimize $ per request given catalog pricing                                                                                 |
+| `reset-aware` ⭐    | Prioritize by quota reset time — short reset windows ranked higher                                                           |
+| `reset-window`      | Prefer targets whose quota window resets soonest                                                                             |
+| `headroom`          | Pick the target with the most remaining quota headroom                                                                       |
+| `strict-random`     | Random without deduplication of repeats                                                                                      |
+| `auto`              | Use Auto Combo scoring (9-factor) — **recommended**                                                                          |
+| `lkgp`              | Last-Known-Good Path (sticky route to last successful target)                                                                |
+| `context-optimized` | Pick target with best fit for current context size                                                                           |
+| `fusion` 🧬         | Fan out to a panel of models in parallel, then synthesize one answer via a judge (see below)                                 |
+| `pipeline`          | Run targets sequentially, threading each step's output into the next step's input; only the final answer is returned (#6396) |
 
 ⭐ = New in v3.8.0 · 🧬 = New in v3.8.36
 
@@ -605,7 +606,7 @@ See `docs/marketing/TIERS.md` for tier definitions and provider classification.
 public strategies end-to-end through the real combo pipeline with a mocked upstream.
 Coverage includes:
 
-- All 17 `ROUTING_STRATEGY_VALUES` strategies (ordered, weighted, cost, context, fusion, …).
+- All 18 `ROUTING_STRATEGY_VALUES` strategies (ordered, weighted, cost, context, fusion, …).
 - `quota-share` (internal) end-to-end: DRR fairness + saturation deprioritization via the
   real `selectQuotaShareTarget` seam (`registerQuotaFetcher` / `setLKGP` /
   `__setHeadroomSaturationFetcherForTests`).
@@ -639,5 +640,5 @@ intentionally excluded from CI because they require live credentials and VPS acc
 | `open-sse/services/autoCombo/autoPrefix.ts`               | `auto/` prefix parser + 6 variants                                         |
 | `open-sse/services/autoCombo/virtualFactory.ts`           | Builds in-memory `AutoComboConfig` from live connections                   |
 | `open-sse/services/autoCombo/providerRegistryAccessor.ts` | Test hook for mocking provider registry                                    |
-| `src/shared/constants/routingStrategies.ts`               | `ROUTING_STRATEGY_VALUES` (17 strategies)                                  |
+| `src/shared/constants/routingStrategies.ts`               | `ROUTING_STRATEGY_VALUES` (18 strategies)                                  |
 | `src/sse/handlers/chat.ts`                                | Integration: auto-prefix short-circuit                                     |

electron/package-lock.json

@@ -1,12 +1,12 @@
 {
   "name": "omniroute-desktop",
-  "version": "3.8.45",
+  "version": "3.8.46",
   "lockfileVersion": 3,
   "requires": true,
   "packages": {
     "": {
       "name": "omniroute-desktop",
-      "version": "3.8.45",
+      "version": "3.8.46",
       "license": "MIT",
       "dependencies": {
         "electron-updater": "^6.8.9"

electron/package.json

@@ -1,6 +1,6 @@
 {
   "name": "omniroute-desktop",
-  "version": "3.8.45",
+  "version": "3.8.46",
   "description": "OmniRoute Desktop Application",
   "main": "main.js",
   "author": {

next.config.mjs

@@ -2,6 +2,7 @@ import createNextIntlPlugin from "next-intl/plugin";
 import { createMDX } from "fumadocs-mdx/next";
 import { dirname } from "node:path";
 import { fileURLToPath } from "node:url";
+import { mitmManagerAliasFor } from "./scripts/build/mitm-stub-flag.mjs";
 
 const withNextIntl = createNextIntlPlugin("./src/i18n/request.ts");
 const distDir = process.env.NEXT_DIST_DIR || ".build/next";
@@ -106,8 +107,13 @@ const nextConfig = {
   turbopack: {
     root: projectRoot,
     resolveAlias: {
-      // Point mitm/manager to a stub during build (native child_process/fs can't be bundled)
-      "@/mitm/manager": "./src/mitm/manager.stub.ts",
+      // @/mitm/manager → stub ONLY where the runtime can't run the MITM stack
+      // (Docker sets OMNIROUTE_MITM_STUB=1 — #3390 graceful degradation). The
+      // alias used to be unconditional, which was fine while Docker was the
+      // only Turbopack consumer — but the v3.8.45 bundler-default flip shipped
+      // the stub to every npm/Electron/VPS artifact and broke Agent Bridge
+      // start for all non-Docker users (#6344). See scripts/build/mitm-stub-flag.mjs.
+      ...mitmManagerAliasFor(process.env),
       ...minimalBuildAliases,
     },
   },

open-sse/config/glmProvider.ts

@@ -209,6 +209,79 @@ export function getGlmQuotaUrl(providerSpecificData: unknown): string {
   return GLM_QUOTA_URLS[getGlmApiRegion(providerSpecificData)];
 }
 
+function getProviderSpecificString(data: JsonRecord, keys: readonly string[]): string | null {
+  for (const key of keys) {
+    const value = asString(data[key]);
+    if (value) return value;
+  }
+  return null;
+}
+
+export const GLM_TEAM_QUOTA_ORGANIZATION_KEYS = [
+  "glmOrganizationId",
+  "bigmodelOrganization",
+  "glmOrganization",
+] as const;
+
+export const GLM_TEAM_QUOTA_PROJECT_KEYS = [
+  "glmProjectId",
+  "bigmodelProject",
+  "glmProject",
+] as const;
+
+export const GLM_TEAM_QUOTA_ALIAS_KEYS = [
+  "bigmodelOrganization",
+  "glmOrganization",
+  "bigmodelProject",
+  "glmProject",
+] as const;
+
+export type GlmTeamQuotaConfig =
+  | { state: "none" }
+  | { state: "configured"; organizationId: string; projectId: string }
+  | { state: "incomplete"; missing: "glmOrganizationId" | "glmProjectId" };
+
+export function getGlmTeamQuotaConfig(providerSpecificData: unknown): GlmTeamQuotaConfig {
+  const data = asRecord(providerSpecificData);
+  const organizationId = getProviderSpecificString(data, GLM_TEAM_QUOTA_ORGANIZATION_KEYS);
+  const projectId = getProviderSpecificString(data, GLM_TEAM_QUOTA_PROJECT_KEYS);
+
+  if (!organizationId && !projectId) return { state: "none" };
+  if (organizationId && projectId) {
+    return { state: "configured", organizationId, projectId };
+  }
+  return {
+    state: "incomplete",
+    missing: organizationId ? "glmProjectId" : "glmOrganizationId",
+  };
+}
+
+export function buildGlmQuotaFetch(
+  apiKey: string,
+  providerSpecificData?: unknown
+): { url: string; headers: Record<string, string> } {
+  const teamConfig = getGlmTeamQuotaConfig(providerSpecificData);
+  const baseUrl = getGlmQuotaUrl(providerSpecificData);
+  const url =
+    teamConfig.state === "configured"
+      ? baseUrl.includes("?")
+        ? `${baseUrl}&type=2`
+        : `${baseUrl}?type=2`
+      : baseUrl;
+
+  const headers: Record<string, string> = {
+    Authorization: `Bearer ${apiKey}`,
+    Accept: "application/json",
+  };
+
+  if (teamConfig.state === "configured") {
+    headers["bigmodel-organization"] = teamConfig.organizationId;
+    headers["bigmodel-project"] = teamConfig.projectId;
+  }
+
+  return { url, headers };
+}
+
 function stripKnownGlmEndpointSuffix(baseUrl: string): { base: string; suffix: string } {
   const parts = splitUrlQueryAndHash(baseUrl);
   let base = parts.base;

open-sse/config/providerFieldStrips.ts

@@ -1,10 +1,16 @@
 // Fields that, when literally named in an upstream 400 body, are safe to strip and
 // retry once (FCC NIM-style recovery). Mirrors the existing context_management 400
 // fallback in base.ts, generalized to these OpenAI-compat / NIM reasoning fields.
+// `context_management` (9router#1468): Claude Code sends it top-level; strict
+// anthropic-compatible gateways 400 with "context_management: Extra inputs are not
+// permitted". The dedicated base.ts fallback only fires when OmniRoute's own
+// contextEditing feature is enabled, so a client-sent field passed through
+// untouched when the feature is off — this generic strip covers that case.
 export const KNOWN_OFFENDING_FIELDS: readonly string[] = [
   "reasoning_budget",
   "chat_template",
   "reasoning_content",
+  "context_management",
 ];
 
 /** Return the first known-offending field literally named in a 400 body, or null. */

open-sse/config/providers/index.ts

@@ -128,6 +128,7 @@ import { gitlab_duoProvider } from "./registry/gitlab-duo/index.ts";
 import { command_codeProvider } from "./registry/command-code/index.ts";
 import { novitaProvider } from "./registry/novita/index.ts";
 import { windsurfProvider } from "./registry/windsurf/index.ts";
+import { zed_hostedProvider } from "./registry/zed-hosted/index.ts";
 import { nanogptProvider } from "./registry/nanogpt/index.ts";
 import { scalewayProvider } from "./registry/scaleway/index.ts";
 import { agentrouterProvider } from "./registry/agentrouter/index.ts";
@@ -183,6 +184,8 @@ import { sumopodProvider } from "./registry/sumopod/index.ts";
 import { x5labProvider } from "./registry/x5lab/index.ts";
 import { kenariProvider } from "./registry/kenari/index.ts";
 import { requestyProvider } from "./registry/requesty/index.ts";
+import { digitaloceanProvider } from "./registry/digitalocean/index.ts";
+import { hcnsecProvider } from "./registry/hcnsec/index.ts";
 
 export const REGISTRY: Record<string, RegistryEntry> = {
   aimlapi: aimlapiProvider,
@@ -311,6 +314,7 @@ export const REGISTRY: Record<string, RegistryEntry> = {
   "command-code": command_codeProvider,
   novita: novitaProvider,
   windsurf: windsurfProvider,
+  "zed-hosted": zed_hostedProvider,
   nanogpt: nanogptProvider,
   scaleway: scalewayProvider,
   agentrouter: agentrouterProvider,
@@ -368,4 +372,6 @@ export const REGISTRY: Record<string, RegistryEntry> = {
   x5lab: x5labProvider,
   kenari: kenariProvider,
   requesty: requestyProvider,
+  digitalocean: digitaloceanProvider,
+  hcnsec: hcnsecProvider,
 };

open-sse/config/providers/registry/cerebras/index.ts

@@ -10,6 +10,7 @@ export const cerebrasProvider: RegistryEntry = {
   authHeader: "bearer",
   models: [
     { id: "zai-glm-4.7", name: "GLM 4.7" },
+    { id: "gemma-4-31b", name: "Gemma 4 31B" },
     { id: "gpt-oss-120b", name: "GPT OSS 120B" },
   ],
 };

open-sse/config/providers/registry/digitalocean/index.ts

@@ -0,0 +1,11 @@
+import type { RegistryEntry } from "../../shared.ts";
+import { buildOpenAiCompatibleRegistryEntry } from "../../shared.ts";
+
+export const digitaloceanProvider: RegistryEntry = buildOpenAiCompatibleRegistryEntry({
+  id: "digitalocean",
+  alias: "digitalocean",
+  baseUrl: "https://inference.do-ai.run/v1/chat/completions",
+  modelsUrl: "https://inference.do-ai.run/v1/models",
+  models: [],
+  passthroughModels: true,
+});

open-sse/config/providers/registry/hcnsec/index.ts

@@ -0,0 +1,11 @@
+import type { RegistryEntry } from "../../shared.ts";
+import { buildOpenAiCompatibleRegistryEntry } from "../../shared.ts";
+
+export const hcnsecProvider: RegistryEntry = buildOpenAiCompatibleRegistryEntry({
+  id: "hcnsec",
+  alias: "hcnsec",
+  baseUrl: "https://api.hcnsec.cn/v1/chat/completions",
+  modelsUrl: "https://api.hcnsec.cn/v1/models",
+  models: [],
+  passthroughModels: true,
+});

open-sse/config/providers/registry/zed-hosted/index.ts

@@ -0,0 +1,35 @@
+import type { RegistryEntry } from "../../shared.ts";
+
+/**
+ * Zed Hosted Models — cloud.zed.dev, Zed's own aggregator (fronts
+ * Anthropic/OpenAI/Google/xAI depending on the requested model).
+ *
+ * Distinct from the pre-existing `zed` provider id (Zed IDE local
+ * credential-import surface — src/lib/zed-oauth/ + src/mitm/detection/zed.ts).
+ * This entry is the new cloud-proxy capability; `zed-hosted` avoids an id clash.
+ *
+ * `models: []` + `modelsUrl` intentionally — Zed's hosted catalog changes
+ * frequently and is fetched live per-connection by the executor
+ * (open-sse/shared/zedAuth.ts::resolveZedModels), never hardcoded here.
+ * `modelsUrl` also feeds the dashboard's generic models-discovery fallback
+ * (src/app/api/providers/[id]/models/discoveryConfig.ts).
+ *
+ * No `oauth.clientIdEnv/clientSecretEnv` — Zed's native-app sign-in uses a
+ * self-generated RSA keypair per login attempt, not a registered OAuth
+ * client_id/secret (see open-sse/shared/zedAuth.ts header comment).
+ */
+export const zed_hostedProvider: RegistryEntry = {
+  id: "zed-hosted",
+  format: "openai",
+  executor: "zed-hosted",
+  baseUrl: "https://cloud.zed.dev/completions",
+  authType: "oauth",
+  authHeader: "bearer",
+  timeoutMs: 120000,
+  forceStream: true,
+  models: [],
+  modelsUrl: "https://cloud.zed.dev/models",
+  passthroughModels: true,
+};
+
+export default zed_hostedProvider;

open-sse/executors/copilot-m365-connection.ts

@@ -35,6 +35,19 @@ export const M365_EDU_OVERRIDES = {
   licenseType: "Starter",
 } as const;
 
+/**
+ * Enterprise / "work" (Microsoft 365 Copilot for work) tier overrides (#6334). Enterprise
+ * tenants ride the `agent="work"` BizChat surface with the `officeweb` scenario and a
+ * Premium license. Opt-in via `providerSpecificData.tier="enterprise"` (alias `"work"`) so
+ * the individual and EDU paths are unchanged. A raw `providerSpecificData.agent` override is
+ * also honored for tenants that need a different agent value.
+ */
+export const M365_ENTERPRISE_OVERRIDES = {
+  agent: "work",
+  scenario: "officeweb",
+  licenseType: "Premium",
+} as const;
+
 export const M365_DEFAULT_VARIANTS = [
   "EnableMcpServerWidgets",
   "feature.EnableMcpServerWidgets",
@@ -94,6 +107,7 @@ export interface M365ConnectionParams {
   scenario?: string;
   isEdu?: string;
   licenseType?: string;
+  agent?: string;
 }
 
 /** A new 32-hex chat session id (== XRoutingParameterSessionKey == clientrequestid). */
@@ -174,16 +188,18 @@ export function resolveConnectionParams(
 }
 
 /**
- * Resolve tier overrides (opt-in). `tier="edu"|"included"` applies the EDU overrides;
- * individual fields (`scenario`/`isEdu`/`licenseType`) can also be overridden directly
- * via providerSpecificData. Unset fields fall back to the individual defaults in
- * buildWsUrl. (#6210)
+ * Resolve tier overrides (opt-in). `tier="edu"|"included"` applies the EDU overrides and
+ * `tier="enterprise"|"work"` applies the enterprise/work overrides; individual fields
+ * (`scenario`/`isEdu`/`licenseType`/`agent`) can also be overridden directly via
+ * providerSpecificData. Unset fields fall back to the individual defaults in buildWsUrl.
+ * (#6210, #6334)
  */
 function resolveTierOverrides(
   psd: JsonRecord
-): Pick<M365ConnectionParams, "scenario" | "isEdu" | "licenseType"> {
+): Pick<M365ConnectionParams, "scenario" | "isEdu" | "licenseType" | "agent"> {
   const tier = typeof psd.tier === "string" ? psd.tier.toLowerCase() : "";
   const isEduTier = tier === "edu" || tier === "included";
+  const isEnterpriseTier = tier === "enterprise" || tier === "work";
   const psdIsEdu =
     (typeof psd.isEdu === "string" && psd.isEdu) ||
     (typeof psd.isEdu === "boolean" && String(psd.isEdu)) ||
@@ -191,11 +207,16 @@ function resolveTierOverrides(
   return {
     scenario:
       (typeof psd.scenario === "string" && psd.scenario) ||
-      (isEduTier ? M365_EDU_OVERRIDES.scenario : undefined),
+      (isEduTier ? M365_EDU_OVERRIDES.scenario : undefined) ||
+      (isEnterpriseTier ? M365_ENTERPRISE_OVERRIDES.scenario : undefined),
     isEdu: psdIsEdu || (isEduTier ? M365_EDU_OVERRIDES.isEdu : undefined),
     licenseType:
       (typeof psd.licenseType === "string" && psd.licenseType) ||
-      (isEduTier ? M365_EDU_OVERRIDES.licenseType : undefined),
+      (isEduTier ? M365_EDU_OVERRIDES.licenseType : undefined) ||
+      (isEnterpriseTier ? M365_ENTERPRISE_OVERRIDES.licenseType : undefined),
+    agent:
+      (typeof psd.agent === "string" && psd.agent) ||
+      (isEnterpriseTier ? M365_ENTERPRISE_OVERRIDES.agent : undefined),
   };
 }
 
@@ -219,7 +240,7 @@ export function buildWsUrl(params: M365ConnectionParams): string {
     agentHost: M365_INDIVIDUAL_DEFAULTS.agentHost,
     licenseType: params.licenseType ?? M365_INDIVIDUAL_DEFAULTS.licenseType,
     isEdu: params.isEdu ?? "false",
-    agent: M365_INDIVIDUAL_DEFAULTS.agent,
+    agent: params.agent ?? M365_INDIVIDUAL_DEFAULTS.agent,
     scenario: params.scenario ?? M365_INDIVIDUAL_DEFAULTS.scenario,
   });
   return `wss://${params.host}/m365Copilot/Chathub/${params.chathubPath}?${query.toString()}`;

open-sse/executors/copilot-m365-web.ts

@@ -1,7 +1,7 @@
 import WebSocket from "ws";
 import { FETCH_TIMEOUT_MS } from "../config/constants.ts";
 import { sanitizeErrorMessage } from "../utils/error.ts";
-import { BaseExecutor, type ExecuteInput } from "./base.ts";
+import { BaseExecutor, type ExecuteInput, type ExecutorLog } from "./base.ts";
 import {
   buildPrompt,
   buildWsUrl,
@@ -59,7 +59,11 @@ export class CopilotM365WebExecutor extends BaseExecutor {
     prompt: string;
     model: string;
     signal?: AbortSignal;
+    log?: ExecutorLog | null;
   }): Promise<ReadableStream<Uint8Array>> {
+    // #6210 — observability for the empty-response class. The access_token rides
+    // in the WS query string, so every URL logged here goes through redactWsUrl().
+    const log = input.log ?? null;
     return new ReadableStream<Uint8Array>(
       {
         start: async (controller) => {
@@ -119,6 +123,8 @@ export class CopilotM365WebExecutor extends BaseExecutor {
             const traceId = wsUrlParts.searchParams.get("clientrequestid") ?? crypto.randomUUID().replace(/-/g, "");
             const sessionId = wsUrlParts.searchParams.get("X-SessionId") ?? crypto.randomUUID();
 
+            log?.debug?.("M365_WS", `connecting → ${redactWsUrl(input.wsUrl)}`);
+
             ws = new WebSocketCtor(input.wsUrl, {
               headers: {
                 Origin: "https://m365.cloud.microsoft",
@@ -142,6 +148,7 @@ export class CopilotM365WebExecutor extends BaseExecutor {
             };
 
             ws.on("open", () => {
+              log?.debug?.("M365_WS", "socket open — sending handshake");
               ws?.send(handshakeFrame());
             });
 
@@ -153,14 +160,20 @@ export class CopilotM365WebExecutor extends BaseExecutor {
 
               for (const rawFrame of split.frames) {
                 const frame = parseFrame(rawFrame);
+                log?.debug?.(
+                  "M365_WS",
+                  `frame type=${String(frame?.type)} target=${String(frame?.target)}`
+                );
                 if (!handshakeComplete) {
                   const err = handshakeError(frame);
                   if (err) {
                     clearTimeout(timeout);
+                    log?.debug?.("M365_WS", `handshake failed: ${err}`);
                     abort(`Microsoft 365 Copilot handshake failed: ${err}`);
                     return;
                   }
                   handshakeComplete = true;
+                  log?.debug?.("M365_WS", "handshake complete — sending chat invocation");
                   sendChat();
                   continue;
                 }
@@ -186,6 +199,10 @@ export class CopilotM365WebExecutor extends BaseExecutor {
 
             ws.on("error", (err) => {
               clearTimeout(timeout);
+              log?.debug?.(
+                "M365_WS",
+                `socket error: ${err instanceof Error ? err.message : String(err)}`
+              );
               abort(
                 sanitizeErrorMessage(
                   err instanceof Error ? err.message : "Microsoft 365 Copilot WebSocket error"
@@ -244,7 +261,13 @@ export class CopilotM365WebExecutor extends BaseExecutor {
     const wsUrl = buildWsUrl(connectionParams);
 
     try {
-      const wsStream = await this.wsChat({ wsUrl, prompt, model, signal: input.signal ?? undefined });
+      const wsStream = await this.wsChat({
+        wsUrl,
+        prompt,
+        model,
+        signal: input.signal ?? undefined,
+        log: input.log,
+      });
 
       if (stream) {
         return {

open-sse/executors/default.ts

@@ -17,6 +17,10 @@ import {
 import { isOfficialAnthropicBaseUrl } from "../utils/anthropicHost.ts";
 import { applyProviderRequestDefaults } from "../services/providerRequestDefaults.ts";
 import { stripUnsupportedParams } from "../translator/paramSupport.ts";
+import {
+  injectReasoningContentForThinkingModel,
+  isThinkingMessageModel,
+} from "../utils/reasoningContentInjector.ts";
 import {
   detectFormat,
   getOpenAICompatibleType,
@@ -545,8 +549,9 @@ export class DefaultExecutor extends BaseExecutor {
     withDefaults = this.defaultResponsesTextFormat(withDefaults);
 
     // Port of decolua/9router commit d652300e:
-    // Cerebras returns 400 (wrong_api_format) and Mistral returns 422
-    // (extra_forbidden) when the forwarded body carries `client_metadata`
+    // Cerebras returns 400 (wrong_api_format), Mistral returns 422
+    // (extra_forbidden), and NVIDIA's OpenAI-compatible wrapper returns 400
+    // (Unsupported parameter) when the forwarded body carries `client_metadata`
     // (an OpenAI Codex / Claude CLI passthrough field with no equivalent on
     // these upstreams). Strip it before sending downstream. Other providers
     // (notably `openai` / `codex`) intentionally keep it.
@@ -554,7 +559,9 @@ export class DefaultExecutor extends BaseExecutor {
       withDefaults &&
       typeof withDefaults === "object" &&
       !Array.isArray(withDefaults) &&
-      (this.provider === "cerebras" || this.provider === "mistral") &&
+      (this.provider === "cerebras" ||
+        this.provider === "mistral" ||
+        this.provider === "nvidia") &&
       Object.prototype.hasOwnProperty.call(withDefaults, "client_metadata")
     ) {
       const withoutClientMetadata = { ...(withDefaults as Record<string, unknown>) };
@@ -562,6 +569,41 @@ export class DefaultExecutor extends BaseExecutor {
       withDefaults = withoutClientMetadata;
     }
 
+    // 9router#1649: Mistral's API returns 422 (extra_forbidden) when an
+    // assistant message carries a `reasoning_content` field (replayed thinking
+    // from a prior turn, e.g. via the Codex /responses path). The field is
+    // nested per-message, so the generic top-level 400/field-downgrade retry
+    // doesn't cover it. Strip it from every message before sending — scoped to
+    // Mistral so DeepSeek (which *requires* replayed reasoning_content) is
+    // unaffected.
+    if (
+      this.provider === "mistral" &&
+      withDefaults &&
+      typeof withDefaults === "object" &&
+      !Array.isArray(withDefaults) &&
+      Array.isArray((withDefaults as Record<string, unknown>).messages)
+    ) {
+      const record = withDefaults as Record<string, unknown>;
+      const messages = record.messages as unknown[];
+      let mutated = false;
+      const cleaned = messages.map((msg) => {
+        if (
+          msg &&
+          typeof msg === "object" &&
+          !Array.isArray(msg) &&
+          Object.prototype.hasOwnProperty.call(msg, "reasoning_content")
+        ) {
+          mutated = true;
+          const { reasoning_content: _dropped, ...rest } = msg as Record<string, unknown>;
+          return rest;
+        }
+        return msg;
+      });
+      if (mutated) {
+        withDefaults = { ...record, messages: cleaned };
+      }
+    }
+
     const targetFormat = getTargetFormat(this.provider, credentials?.providerSpecificData);
     const requestFormat =
       withDefaults && typeof withDefaults === "object" && !Array.isArray(withDefaults)
@@ -695,6 +737,23 @@ export class DefaultExecutor extends BaseExecutor {
       this.ensureThinkingBudget(withDefaults as Record<string, unknown>, model);
     }
 
+    // 9router#1480: the native Moonshot `kimi` provider (executor "default")
+    // is a thinking-mode upstream that 400s with "reasoning_content must be
+    // passed back" when a prior assistant turn lacks it. OpencodeExecutor
+    // already injects a placeholder for OpenCode-routed thinking models; the
+    // direct kimi connection hit neither injection path. Scope to `kimi` so
+    // gateway-served models that merely match the thinking-model name pattern
+    // (and may reject an extra field) are unaffected.
+    if (this.provider === "kimi") {
+      const outboundModel =
+        typeof (withDefaults as Record<string, unknown>)?.model === "string"
+          ? ((withDefaults as Record<string, unknown>).model as string)
+          : model;
+      if (isThinkingMessageModel(outboundModel)) {
+        withDefaults = injectReasoningContentForThinkingModel(withDefaults);
+      }
+    }
+
     return withDefaults;
   }
 

open-sse/executors/doubao-web.ts

@@ -48,11 +48,23 @@ function parseJsonRecord(raw: string): JsonRecord | null {
 }
 
 function randomNumericId(length = 19): string {
-  // crypto-backed (CodeQL js/insecure-randomness): a synthetic device/web id,
-  // not a secret — but crypto.getRandomValues costs the same and closes the alert.
-  const bytes = globalThis.crypto.getRandomValues(new Uint8Array(length));
-  let id = String((bytes[0] % 9) + 1);
-  for (let i = 1; i < length; i += 1) id += String(bytes[i] % 10);
+  // crypto-backed, UNBIASED digits (CodeQL js/insecure-randomness +
+  // js/biased-cryptographic-random): a synthetic device/web id, not a secret.
+  // A raw `cryptoByte % 10` is biased (256 is not a multiple of 10), so each
+  // digit is drawn by rejection sampling — discard bytes in the biased tail so
+  // the remaining range divides evenly, then reduce.
+  const digit = (max: number): number => {
+    const limit = 256 - (256 % max); // largest multiple of `max` that fits in a byte
+    const buf = new Uint8Array(1);
+    let b: number;
+    do {
+      globalThis.crypto.getRandomValues(buf);
+      b = buf[0];
+    } while (b >= limit);
+    return b % max;
+  };
+  let id = String(digit(9) + 1);
+  for (let i = 1; i < length; i += 1) id += String(digit(10));
   return id;
 }
 

open-sse/executors/gitlab.ts

@@ -43,6 +43,22 @@ type OpenAIMessage = {
 
 type JsonRecord = Record<string, unknown>;
 
+/**
+ * Bounds for GitLab's `code_suggestions` `small_file` generation contract. Its AI-Gateway
+ * validation guard rejects an oversized/over-structured prompt with
+ * `422 {"detail":"Validation error"}` (tokens 0/0, pre-inference). Turn-1 is small and
+ * passes; a long folded tool-exchange history trips it — so the serialized prompt (and the
+ * `user_instruction` field) must stay bounded (#6220).
+ */
+const MAX_TOOL_EXCHANGE_CHARS = 24_000;
+const MAX_TOOL_RESULT_CHARS = 8_000;
+const MAX_USER_INSTRUCTION_CHARS = 4_000;
+
+function capText(text: string, max: number): string {
+  if (text.length <= max) return text;
+  return `${text.slice(0, max)}\n…[truncated ${text.length - max} chars]`;
+}
+
 type GitLabRequestTarget = {
   mode: "monolith" | "direct";
   url: string;
@@ -145,26 +161,76 @@ function renderConversationTurn(message: OpenAIMessage, role: string, text: stri
   return null;
 }
 
-/** Serialize the full turn history so the model sees the tool result (#6220). */
+/**
+ * Keep only what the `small_file` generation contract can carry: every system message,
+ * the latest user message, and the most-recent tool round (the last assistant `tool_calls`
+ * turn onward). Older turns are dropped so `content_above_cursor` stays bounded (#6220).
+ */
+function selectBoundedMessages(messages: OpenAIMessage[]): OpenAIMessage[] {
+  let lastUserIdx = -1;
+  let lastToolRoundIdx = -1;
+  messages.forEach((message, idx) => {
+    const role = String(message?.role || "user").toLowerCase();
+    if (role === "user") lastUserIdx = idx;
+    if (role === "assistant" && Array.isArray(message?.tool_calls) && message.tool_calls.length) {
+      lastToolRoundIdx = idx;
+    }
+  });
+  const tailStart = lastToolRoundIdx >= 0 ? lastToolRoundIdx : lastUserIdx;
+  const kept: OpenAIMessage[] = [];
+  messages.forEach((message, idx) => {
+    const role = String(message?.role || "user").toLowerCase();
+    if (role === "system" || role === "developer") {
+      kept.push(message);
+    } else if (idx === lastUserIdx || idx >= tailStart) {
+      kept.push(message);
+    }
+  });
+  return kept;
+}
+
+/**
+ * Serialize the most-recent turn history so the model sees the tool result (#6220), but
+ * BOUNDED: only the last tool round is kept, each tool result is capped, and the whole
+ * prompt is capped — otherwise the folded history trips GitLab's `small_file` 422 guard.
+ */
 function buildToolExchangePrompt(messages: OpenAIMessage[]): string {
   const systemParts: string[] = [];
   const convo: string[] = [];
-  for (const message of messages) {
+  for (const message of selectBoundedMessages(messages)) {
     const role = String(message?.role || "user").toLowerCase();
-    const text = extractTextContent(message?.content);
+    let text = extractTextContent(message?.content);
     if (role === "system" || role === "developer") {
       if (text) systemParts.push(text);
       continue;
     }
+    if (role === "tool") text = capText(text, MAX_TOOL_RESULT_CHARS);
     const line = renderConversationTurn(message, role, text);
     if (line) convo.push(line);
   }
   const header = systemParts.length
     ? `System instructions:\n${systemParts.join("\n\n")}\n\n`
     : "";
-  return `${header}${convo.join(
+  const body = `${header}${convo.join(
     "\n\n"
   )}\n\nContinue the response using the tool result above; do not repeat the tool call.`.trim();
+  return capText(body, MAX_TOOL_EXCHANGE_CHARS);
+}
+
+/**
+ * The user's actual instruction — the latest user message, capped. Kept separate from the
+ * folded history so it is NOT duplicated into an oversized `user_instruction`, the likely
+ * offending 422 field on tool-calling follow-up turns (#6220).
+ */
+function buildLatestUserInstruction(messages: OpenAIMessage[]): string {
+  for (let i = messages.length - 1; i >= 0; i--) {
+    const role = String(messages[i]?.role || "user").toLowerCase();
+    if (role === "user") {
+      const text = extractTextContent(messages[i]?.content);
+      if (text) return capText(text, MAX_USER_INSTRUCTION_CHARS);
+    }
+  }
+  return "";
 }
 
 /**
@@ -329,7 +395,16 @@ export class GitlabExecutor extends BaseExecutor {
     _stream: boolean,
     credentials: ExecuteInput["credentials"]
   ): Record<string, unknown> {
-    const prompt = buildPrompt(body.messages as OpenAIMessage[] | undefined);
+    const messages = body.messages as OpenAIMessage[] | undefined;
+    const prompt = buildPrompt(messages);
+    // On a tool-exchange follow-up, `content_above_cursor` already carries the (bounded)
+    // folded history — sending the same long text again as `user_instruction` is the
+    // likely field that trips GitLab's `small_file` 422 guard. Send only the short latest
+    // user message there instead (#6220). Simple turns keep the legacy behavior.
+    const isToolExchange = Array.isArray(messages) && hasToolExchange(messages);
+    const userInstruction = isToolExchange
+      ? buildLatestUserInstruction(messages as OpenAIMessage[])
+      : prompt;
     const providerData =
       credentials?.providerSpecificData && typeof credentials.providerSpecificData === "object"
         ? credentials.providerSpecificData
@@ -354,7 +429,7 @@ export class GitlabExecutor extends BaseExecutor {
       generation_type: "small_file",
       stream: false,
       ...(projectPath ? { project_path: projectPath } : {}),
-      ...(prompt ? { user_instruction: prompt } : {}),
+      ...(userInstruction ? { user_instruction: userInstruction } : {}),
     };
   }
 

open-sse/executors/index.ts

@@ -27,6 +27,7 @@ import { CommandCodeExecutor } from "./commandCode.ts";
 import { GitlabExecutor } from "./gitlab.ts";
 import { NlpCloudExecutor } from "./nlpcloud.ts";
 import { WindsurfExecutor } from "./windsurf.ts";
+import { ZedHostedExecutor } from "./zed-hosted.ts";
 import { DevinCliExecutor } from "./devin-cli.ts";
 import { AuggieExecutor } from "./auggie.ts";
 import { DeepSeekWebExecutor } from "./deepseek-web.ts";
@@ -111,6 +112,7 @@ const executors = {
   "ms-web": new MuseSparkWebExecutor(), // Alias
   windsurf: new WindsurfExecutor(),
   ws: new WindsurfExecutor(), // Alias
+  "zed-hosted": new ZedHostedExecutor(),
   "devin-cli": new DevinCliExecutor(),
   devin: new DevinCliExecutor(), // Alias
   "deepseek-web": new DeepSeekWebWithAutoRefreshExecutor(),
@@ -205,6 +207,7 @@ export { CommandCodeExecutor } from "./commandCode.ts";
 export { GitlabExecutor } from "./gitlab.ts";
 export { NlpCloudExecutor } from "./nlpcloud.ts";
 export { WindsurfExecutor } from "./windsurf.ts";
+export { ZedHostedExecutor } from "./zed-hosted.ts";
 export { DevinCliExecutor } from "./devin-cli.ts";
 export { AuggieExecutor } from "./auggie.ts";
 export { CopilotWebExecutor } from "./copilot-web.ts";

open-sse/executors/opencode.ts

@@ -292,6 +292,19 @@ export class OpencodeExecutor extends BaseExecutor {
     credentials: ProviderCredentials
   ): any {
     let modifiedBody = super.transformRequest(model, body, stream, credentials);
+    // 9router#1442: OpenCode upstreams (e.g. kimi-k2.6 via opencode-go) return
+    // 400 "Extra inputs are not permitted, field: 'client_metadata'" — an
+    // OpenAI-Codex/Claude-CLI passthrough field with no equivalent here. The
+    // DefaultExecutor strip only covers cerebras/mistral, and OpencodeExecutor
+    // extends BaseExecutor directly, so nothing removed it on this path.
+    if (
+      modifiedBody &&
+      typeof modifiedBody === "object" &&
+      !Array.isArray(modifiedBody) &&
+      Object.prototype.hasOwnProperty.call(modifiedBody, "client_metadata")
+    ) {
+      delete (modifiedBody as Record<string, unknown>).client_metadata;
+    }
     if (
       modifiedBody &&
       typeof modifiedBody === "object" &&

open-sse/executors/tinyfish-fetch.ts

@@ -0,0 +1,131 @@
+/**
+ * TinyFish Fetch Executor
+ *
+ * Fetches content from a URL using the TinyFish Fetch API.
+ * POST https://api.fetch.tinyfish.ai
+ *
+ * "Fetch does not use credits" per TinyFish docs — no explicit pricing tier.
+ * Docs: https://docs.tinyfish.ai/fetch-api
+ *
+ * Unlike Firecrawl, TinyFish has no "links" or "screenshot" output modes —
+ * only markdown, html, and json. Requests for those formats fall back to
+ * markdown and return an empty links array / null screenshot, mirroring how
+ * jina-reader-fetch.ts and tavily-fetch.ts handle formats they don't support.
+ */
+
+import { sanitizeErrorMessage, buildErrorBody } from "../utils/error.ts";
+import type { WebFetchResult, WebFetchFormat, WebFetchCredentials } from "../handlers/webFetch.ts";
+
+const TINYFISH_FETCH_URL = "https://api.fetch.tinyfish.ai";
+const TINYFISH_TIMEOUT_MS = 30_000;
+
+function mapFormat(format: WebFetchFormat): "markdown" | "html" {
+  return format === "html" ? "html" : "markdown";
+}
+
+interface TinyFishFetchOptions {
+  url: string;
+  format: WebFetchFormat;
+  includeMetadata: boolean;
+  credentials: WebFetchCredentials;
+}
+
+interface TinyFishResultEntry {
+  url?: string;
+  final_url?: string;
+  title?: string;
+  description?: string;
+  text?: string;
+}
+
+interface TinyFishErrorEntry {
+  url?: string;
+  message?: string;
+  error?: string;
+}
+
+/**
+ * Execute a TinyFish Fetch API request.
+ */
+export async function tinyfishFetch(opts: TinyFishFetchOptions): Promise<WebFetchResult> {
+  const { url, format, includeMetadata, credentials } = opts;
+
+  if (!credentials.apiKey) {
+    const body = buildErrorBody(401, "TinyFish API key required");
+    return { success: false, status: 401, error: body.error.message };
+  }
+
+  const requestBody = {
+    urls: [url],
+    format: mapFormat(format),
+    ttl: 0,
+  };
+
+  const controller = new AbortController();
+  const timeoutId = setTimeout(() => controller.abort(), TINYFISH_TIMEOUT_MS);
+
+  try {
+    const response = await fetch(TINYFISH_FETCH_URL, {
+      method: "POST",
+      headers: {
+        "Content-Type": "application/json",
+        "X-API-Key": credentials.apiKey,
+      },
+      body: JSON.stringify(requestBody),
+      signal: controller.signal,
+    });
+
+    if (!response.ok) {
+      const rawError = await response.text().catch(() => `HTTP ${response.status}`);
+      const msg = sanitizeErrorMessage(`TinyFish error ${response.status}: ${rawError}`);
+      const body = buildErrorBody(response.status, msg);
+      return { success: false, status: response.status, error: body.error.message };
+    }
+
+    const data = (await response.json()) as {
+      results?: TinyFishResultEntry[];
+      errors?: TinyFishErrorEntry[];
+    };
+
+    const result = data.results?.[0];
+
+    if (!result) {
+      const errorEntry = data.errors?.[0];
+      const msg = sanitizeErrorMessage(
+        errorEntry?.message ?? errorEntry?.error ?? "TinyFish could not fetch the requested URL"
+      );
+      const body = buildErrorBody(502, msg);
+      return { success: false, status: 502, error: body.error.message };
+    }
+
+    const metadata = includeMetadata
+      ? {
+          title: result.title != null ? String(result.title) : null,
+          description: result.description != null ? String(result.description) : null,
+        }
+      : null;
+
+    return {
+      success: true,
+      data: {
+        provider: "tinyfish",
+        url,
+        content: String(result.text ?? ""),
+        links: [],
+        metadata,
+        screenshot_url: null,
+      },
+    };
+  } catch (err: unknown) {
+    if (err instanceof Error && err.name === "AbortError") {
+      const body = buildErrorBody(504, "TinyFish request timed out");
+      return { success: false, status: 504, error: body.error.message };
+    }
+    const msg =
+      err instanceof Error ? sanitizeErrorMessage(err.message) : sanitizeErrorMessage(String(err));
+    const body = buildErrorBody(502, msg);
+    return { success: false, status: 502, error: body.error.message };
+  } finally {
+    clearTimeout(timeoutId);
+  }
+}

open-sse/executors/zed-hosted.ts

@@ -0,0 +1,364 @@
+/**
+ * ZedHostedExecutor — routes requests to Zed's hosted LLM aggregator
+ * (cloud.zed.dev/completions), a multi-format proxy that fronts
+ * Anthropic/OpenAI/Google/xAI depending on the requested model.
+ *
+ * Distinct from the pre-existing `zed` provider id, which is a Zed IDE
+ * credential-import surface (src/lib/zed-oauth/ + src/mitm/detection/zed.ts) —
+ * that surface only detects/imports local Zed IDE keychain credentials, it does
+ * not proxy chat completions. This executor is the NEW cloud-proxy
+ * capability; registry id `zed-hosted` avoids colliding with the IDE id.
+ *
+ * Wire protocol: POST /completions with an NDJSON/SSE-ish body-per-line
+ * response stream (`{"event": <provider-shaped-chunk>}` /
+ * `{"status": ...}` / `[DONE]`), authenticated with a short-lived LLM
+ * bearer token (see open-sse/shared/zedAuth.ts). The provider-shaped
+ * chunk is Claude/Gemini/OpenAI-Responses/xAI(OpenAI-shaped) depending on
+ * which upstream Zed is fronting for the requested model — translated back
+ * to OpenAI Chat Completions chunks by reusing OmniRoute's own translators
+ * (the same ones used for the native claude/gemini/codex executors), never
+ * a bespoke per-provider parser.
+ *
+ * Ported from decolua/9router PR #2328 (open-sse/executors/zed.js),
+ * adapted to TypeScript + OmniRoute's BaseExecutor/translator conventions.
+ * Like WindsurfExecutor, this overrides execute() entirely rather than
+ * using BaseExecutor's default Claude-Code-oriented pipeline, because the
+ * Zed wire request/response shape (thread envelope, LLM-token exchange,
+ * NDJSON status frames) doesn't fit the generic transformRequest/buildUrl
+ * contract that pipeline assumes.
+ */
+
+import { BaseExecutor, type ExecuteInput, type ProviderCredentials } from "./base.ts";
+import { PROVIDERS } from "../config/constants.ts";
+import { FORMATS } from "../translator/formats.ts";
+import { initState } from "../translator/index.ts";
+import { openaiToClaudeRequest } from "../translator/request/openai-to-claude.ts";
+import { openaiToGeminiRequest } from "../translator/request/openai-to-gemini.ts";
+import { openaiToOpenAIResponsesRequest } from "../translator/request/openai-responses/toResponses.ts";
+import { claudeToOpenAIResponse } from "../translator/response/claude-to-openai.ts";
+import { geminiToOpenAIResponse } from "../translator/response/gemini-to-openai.ts";
+import { openaiResponsesToOpenAIResponse } from "../translator/response/openai-responses.ts";
+import { ZED_HEADERS, resolveZedModels, zedLlmFetch, type ZedCredentials } from "../shared/zedAuth.ts";
+
+const ZED_PROVIDER = {
+  anthropic: "Anthropic",
+  openai: "OpenAi",
+  google: "Google",
+  xai: "XAi",
+} as const;
+
+type ZedProviderName = (typeof ZED_PROVIDER)[keyof typeof ZED_PROVIDER];
+
+function normalizeZedProvider(value: unknown, model: unknown): ZedProviderName {
+  const raw = String(value || "").toLowerCase();
+  if (raw === "anthropic") return ZED_PROVIDER.anthropic;
+  if (raw === "openai" || raw === "open_ai") return ZED_PROVIDER.openai;
+  if (raw === "google" || raw === "gemini") return ZED_PROVIDER.google;
+  if (raw === "xai" || raw === "x_ai" || raw === "x-ai") return ZED_PROVIDER.xai;
+
+  const m = String(model || "").toLowerCase();
+  if (m.includes("claude")) return ZED_PROVIDER.anthropic;
+  if (m.includes("gemini")) return ZED_PROVIDER.google;
+  if (m.includes("grok") || m.includes("xai")) return ZED_PROVIDER.xai;
+  return ZED_PROVIDER.openai;
+}
+
+function buildProviderRequest(
+  provider: ZedProviderName,
+  model: string,
+  body: unknown,
+  stream: boolean,
+  credentials: ProviderCredentials
+): unknown {
+  if (provider === ZED_PROVIDER.anthropic) {
+    return openaiToClaudeRequest(model, body, true);
+  }
+  if (provider === ZED_PROVIDER.google) {
+    return openaiToGeminiRequest(model, body as Record<string, unknown>, true, credentials);
+  }
+  if (provider === ZED_PROVIDER.openai) {
+    return openaiToOpenAIResponsesRequest(model, body, true, credentials);
+  }
+  return {
+    ...(body as Record<string, unknown>),
+    model,
+    stream: stream !== false,
+  };
+}
+
+function initProviderState(provider: ZedProviderName, model: string): Record<string, unknown> {
+  if (provider === ZED_PROVIDER.anthropic) return initState(FORMATS.CLAUDE);
+  if (provider === ZED_PROVIDER.google) return initState(FORMATS.GEMINI);
+  if (provider === ZED_PROVIDER.openai) return initState(FORMATS.OPENAI_RESPONSES);
+  const state = initState(FORMATS.OPENAI);
+  state.model = model;
+  return state;
+}
+
+function convertProviderEvent(
+  provider: ZedProviderName,
+  event: unknown,
+  state: Record<string, unknown>
+): unknown {
+  if (provider === ZED_PROVIDER.anthropic) return claudeToOpenAIResponse(event, state);
+  if (provider === ZED_PROVIDER.google) return geminiToOpenAIResponse(event, state);
+  if (provider === ZED_PROVIDER.openai) return openaiResponsesToOpenAIResponse(event, state);
+  return event;
+}
+
+function createErrorChunk(model: string, message: string): Record<string, unknown> {
+  return {
+    id: `chatcmpl-zed-error-${Date.now()}`,
+    object: "chat.completion.chunk",
+    created: Math.floor(Date.now() / 1000),
+    model,
+    choices: [{ index: 0, delta: { content: `[Zed error] ${message}` }, finish_reason: "stop" }],
+  };
+}
+
+function enqueueSseObject(
+  controller: ReadableStreamDefaultController<Uint8Array>,
+  encoder: TextEncoder,
+  chunk: unknown
+): void {
+  if (!chunk) return;
+  const items = Array.isArray(chunk) ? chunk : [chunk];
+  for (const item of items) {
+    if (!item) continue;
+    controller.enqueue(encoder.encode(`data: ${JSON.stringify(item)}\n\n`));
+  }
+}
+
+type ZedLine = { done?: true; status?: unknown; event?: unknown } | null;
+
+function unwrapZedLine(line: string): ZedLine {
+  let text = line.replace(/\r$/, "").trim();
+  if (!text) return null;
+  if (text.startsWith("data:")) text = text.slice(5).trimStart();
+  if (text === "[DONE]") return { done: true };
+  try {
+    const parsed = JSON.parse(text);
+    if (parsed && Object.prototype.hasOwnProperty.call(parsed, "event")) {
+      return { event: parsed.event };
+    }
+    if (parsed && Object.prototype.hasOwnProperty.call(parsed, "status")) {
+      return { status: parsed.status };
+    }
+    return { event: parsed };
+  } catch {
+    return null;
+  }
+}
+
+function normalizeStatus(status: unknown): Record<string, unknown> | null {
+  if (!status) return null;
+  if (typeof status === "string") return { type: status };
+  if (typeof status === "object") {
+    const rec = status as Record<string, unknown>;
+    const key = Object.keys(rec)[0];
+    if (key && typeof rec[key] === "object") return { type: key, ...(rec[key] as object) };
+    return rec;
+  }
+  return null;
+}
+
+function wrapZedCompletionStream(
+  response: Response,
+  provider: ZedProviderName,
+  model: string
+): Response {
+  if (!response.ok || !response.body) return response;
+
+  const decoder = new TextDecoder();
+  const encoder = new TextEncoder();
+  const state = initProviderState(provider, model);
+  let buffer = "";
+  let done = false;
+
+  const finish = (controller: ReadableStreamDefaultController<Uint8Array>) => {
+    if (done) return;
+    const finalChunk = convertProviderEvent(provider, null, state);
+    enqueueSseObject(controller, encoder, finalChunk);
+    controller.enqueue(encoder.encode("data: [DONE]\n\n"));
+    done = true;
+  };
+
+  const processLine = (line: string, controller: ReadableStreamDefaultController<Uint8Array>) => {
+    if (done) return;
+    const payload = unwrapZedLine(line);
+    if (!payload) return;
+    if (payload.done) {
+      finish(controller);
+      return;
+    }
+    if (payload.status) {
+      const status = normalizeStatus(payload.status);
+      if (status?.type === "failed" || status?.failed) {
+        const failed = (status.failed as Record<string, unknown>) || status;
+        const message = String(failed.message || failed.error || failed.code || "request failed");
+        enqueueSseObject(controller, encoder, createErrorChunk(model, message));
+        finish(controller);
+      } else if (status?.type === "stream_ended" || status === ("stream_ended" as unknown)) {
+        finish(controller);
+      }
+      return;
+    }
+    const converted = convertProviderEvent(provider, payload.event, state);
+    enqueueSseObject(controller, encoder, converted);
+  };
+
+  const transformed = response.body.pipeThrough(
+    new TransformStream<Uint8Array, Uint8Array>({
+      transform(chunk, controller) {
+        buffer += decoder.decode(chunk, { stream: true });
+        let nl: number;
+        while ((nl = buffer.indexOf("\n")) !== -1) {
+          const line = buffer.slice(0, nl);
+          buffer = buffer.slice(nl + 1);
+          processLine(line, controller);
+        }
+      },
+      flush(controller) {
+        buffer += decoder.decode();
+        if (buffer) {
+          processLine(buffer, controller);
+          buffer = "";
+        }
+        finish(controller);
+      },
+    })
+  );
+
+  return new Response(transformed, {
+    status: response.status,
+    statusText: response.statusText,
+    headers: {
+      "Content-Type": "text/event-stream",
+      "Cache-Control": "no-cache",
+    },
+  });
+}
+
+export class ZedHostedExecutor extends BaseExecutor {
+  constructor() {
+    super("zed-hosted", PROVIDERS["zed-hosted"] || {});
+  }
+
+  async resolveModel(
+    model: string,
+    credentials: ZedCredentials,
+    signal: AbortSignal | null | undefined,
+    log: ExecuteInput["log"]
+  ): Promise<{ raw: Record<string, unknown> | null; provider: ZedProviderName }> {
+    try {
+      const catalog = await resolveZedModels(credentials, { config: this.config, signal });
+      let raw = catalog?.rawById?.get(model) ?? null;
+      if (!raw) {
+        const refreshed = await resolveZedModels(credentials, {
+          config: this.config,
+          signal,
+          forceRefresh: true,
+        });
+        raw = refreshed?.rawById?.get(model) ?? null;
+      }
+      return {
+        raw,
+        provider: normalizeZedProvider(raw?.provider, model),
+      };
+    } catch (error) {
+      const message = error instanceof Error ? error.message : String(error);
+      log?.warn?.("ZED", `model catalog unavailable, inferring provider for ${model}: ${message}`);
+      return { raw: null, provider: normalizeZedProvider(null, model) };
+    }
+  }
+
+  async execute({ model, body, stream, credentials, signal, log }: ExecuteInput): Promise<{
+    response: Response;
+    url: string;
+    headers: Record<string, string>;
+    transformedBody: unknown;
+  }> {
+    const zedCredentials = credentials as ZedCredentials;
+    const { provider } = await this.resolveModel(model, zedCredentials, signal, log);
+    const providerRequest = buildProviderRequest(provider, model, body, stream, credentials);
+    const bodyRecord = (body ?? {}) as Record<string, unknown>;
+    const payload = {
+      thread_id: bodyRecord.thread_id || (credentials as Record<string, unknown>)?._clientSessionId,
+      prompt_id: bodyRecord.prompt_id,
+      provider,
+      model,
+      provider_request: providerRequest,
+    };
+
+    const response = await zedLlmFetch(zedCredentials, "/completions", {
+      config: this.config,
+      signal,
+      fetchOptions: {
+        method: "POST",
+        headers: {
+          "Content-Type": "application/json",
+          Accept: "application/x-ndjson, text/event-stream, */*",
+          "User-Agent": `OmniRoute/zed-hosted`,
+          "x-zed-version": (this.config as Record<string, unknown>)?.appVersion?.toString() || "0.200.0",
+          [ZED_HEADERS.clientSupportsStatus]: "true",
+          [ZED_HEADERS.clientSupportsStreamEnded]: "true",
+        },
+        body: JSON.stringify(payload),
+      },
+    });
+
+    const wrapped = response.ok ? wrapZedCompletionStream(response, provider, model) : response;
+    return {
+      response: wrapped,
+      url: `${(this.config as Record<string, unknown>)?.llmBaseUrl || "https://cloud.zed.dev"}/completions`,
+      headers: { "Content-Type": "application/json", Authorization: "Bearer <zed-llm-token>" },
+      transformedBody: payload,
+    };
+  }
+
+  parseError(response: Response, bodyText: string): { status: number; message: string } {
+    let parsed: Record<string, unknown> | null = null;
+    try {
+      parsed = JSON.parse(bodyText || "{}");
+    } catch {
+      parsed = null;
+    }
+
+    const errorObj = (parsed?.error as Record<string, unknown>) || undefined;
+    const code = (parsed?.code as string) || (errorObj?.code as string) || "";
+    const rawMessage =
+      (parsed?.message as string) || (errorObj?.message as string) || bodyText || response.statusText;
+    if (code === "trial_blocked") {
+      return {
+        status: response.status,
+        message: `Zed trial access is blocked upstream. The account can list hosted models, but Zed is refusing completions until trial/billing access is enabled or unblocked. Zed says: ${rawMessage}`,
+      };
+    }
+    if (code) {
+      return {
+        status: response.status,
+        message: `Zed ${code}: ${rawMessage}`,
+      };
+    }
+    return {
+      status: response.status,
+      message: rawMessage || `Zed upstream error: ${response.status}`,
+    };
+  }
+
+  async refreshCredentials(): Promise<Partial<ProviderCredentials> | null> {
+    return null;
+  }
+
+  needsRefresh(): boolean {
+    return false;
+  }
+}
+
+export default ZedHostedExecutor;
+
+export const __test__ = {
+  normalizeZedProvider,
+  unwrapZedLine,
+  wrapZedCompletionStream,
+};

open-sse/handlers/chatCore.ts

@@ -2107,7 +2107,15 @@ export async function handleChatCore({
   // mode="cliproxyapi": returns the CLIProxyAPI executor instead.
   // mode="fallback": returns a wrapper that tries native first, falls back to CLIProxyAPI on 5xx/network errors.
 
-  const resolveExecutorWithProxy = (prov: string) => resolveExecutorWithProxyFor(prov, log);
+  // #6339: pass the resolved connection's providerSpecificData so a per-connection
+  // cliproxyapiMode="claude-native" override can deep-route this single connection
+  // through CLIProxyAPI regardless of the provider-level upstream_proxy_config mode.
+  const resolveExecutorWithProxy = (prov: string) =>
+    resolveExecutorWithProxyFor(
+      prov,
+      log,
+      (credentials?.providerSpecificData as Record<string, unknown> | null | undefined) ?? null
+    );
 
   // === Quota Share enforcement PRE-hook (B/F7) ===
   // Runs after provider/model/credentials/apiKeyInfo are fully resolved,
@@ -4056,6 +4064,13 @@ export async function handleChatCore({
       requestId: skillRequestId,
       compressionResponseMeta,
     });
+    // #6426: align response body `model` with the `X-OmniRoute-Model` header
+    // (both must be the resolved backend model). Some upstreams (notably legacy
+    // /v1/completions text-completion path) return a body `model` field that
+    // differs from the resolved backend id we advertised in the header, leaving
+    // strict clients unable to reconcile the two. Rewrite body.model to `model`
+    // FIRST, then let #1311 echo override it when the opt-in setting is on.
+    if (typeof model === "string" && model) echoModelInObject(translatedResponse, model);
     // #1311: echo the requested alias/combo name in the non-streaming response model.
     if (echoModel) echoModelInObject(translatedResponse, echoModel);
     return {

open-sse/handlers/chatCore/executorProxy.ts

@@ -10,6 +10,7 @@
  */
 
 import { getExecutor } from "../../executors/index.ts";
+import { isCliproxyapiDeepModeEnabled } from "../../executors/cliproxyapi.ts";
 import { getCachedSettings } from "@/lib/db/readCache";
 import { getUpstreamProxyConfigCached } from "./comboContextCache.ts";
 
@@ -22,7 +23,25 @@ type LoggerLike =
   | null
   | undefined;
 
-export async function resolveExecutorWithProxy(prov: string, log?: LoggerLike) {
+export async function resolveExecutorWithProxy(
+  prov: string,
+  log?: LoggerLike,
+  providerSpecificData?: Record<string, unknown> | null
+) {
+  // Per-connection routing override (#6339): the resolved connection can opt itself
+  // into the CLIProxyAPI passthrough executor via providerSpecificData.cliproxyapiMode
+  // === "claude-native" (UI toggle). This takes precedence over the provider-level
+  // upstream_proxy_config mode — one connection can deep-route while the provider's
+  // default (and its other connections) stay native. Backward-compatible: connections
+  // without the flag fall through to the existing per-provider behaviour untouched.
+  if (isCliproxyapiDeepModeEnabled(providerSpecificData)) {
+    log?.info?.(
+      "UPSTREAM_PROXY",
+      `${prov} routed through CLIProxyAPI (per-connection claude-native override)`
+    );
+    return getExecutor("cliproxyapi");
+  }
+
   const cfg = await getUpstreamProxyConfigCached(prov);
   if (!cfg.enabled || cfg.mode === "native") return getExecutor(prov);
 

open-sse/handlers/webFetch.ts

@@ -2,12 +2,12 @@
  * Web Fetch Handler
  *
  * Handles POST /v1/web/fetch requests.
- * Dispatches to a web-fetch provider executor (Firecrawl, Jina Reader, or Tavily).
+ * Dispatches to a web-fetch provider executor (Firecrawl, Jina Reader, Tavily, or TinyFish).
  *
  * Request format:
  * {
  *   "url": "https://example.com",
- *   "provider": "firecrawl" | "jina-reader" | "tavily-search",  // optional
+ *   "provider": "firecrawl" | "jina-reader" | "tavily-search" | "tinyfish",  // optional
  *   "format": "markdown" | "html" | "links" | "screenshot",
  *   "depth": 0 | 1 | 2,
  *   "wait_for_selector": "main",
@@ -19,12 +19,13 @@ import { buildErrorBody, sanitizeErrorMessage } from "../utils/error.ts";
 import { firecrawlFetch } from "../executors/firecrawl-fetch.ts";
 import { jinaReaderFetch } from "../executors/jina-reader-fetch.ts";
 import { tavilyFetch } from "../executors/tavily-fetch.ts";
+import { tinyfishFetch } from "../executors/tinyfish-fetch.ts";
 
 export type WebFetchFormat = "markdown" | "html" | "links" | "screenshot";
 
 export interface WebFetchRequest {
   url: string;
-  provider?: "firecrawl" | "jina-reader" | "tavily-search";
+  provider?: "firecrawl" | "jina-reader" | "tavily-search" | "tinyfish";
   format?: WebFetchFormat;
   depth?: 0 | 1 | 2;
   wait_for_selector?: string;
@@ -51,7 +52,7 @@ export interface WebFetchCredentials {
   apiKey?: string;
 }
 
-const WEB_FETCH_PROVIDERS = ["firecrawl", "jina-reader", "tavily-search"] as const;
+const WEB_FETCH_PROVIDERS = ["firecrawl", "jina-reader", "tavily-search", "tinyfish"] as const;
 type WebFetchProviderId = (typeof WEB_FETCH_PROVIDERS)[number];
 
 /**
@@ -99,6 +100,14 @@ export async function handleWebFetch(
           credentials,
         });
 
+      case "tinyfish":
+        return await tinyfishFetch({
+          url: req.url,
+          format,
+          includeMetadata,
+          credentials,
+        });
+
       default: {
         const _exhaustive: never = provider;
         return {

open-sse/mcp-server/schemas/tools.ts

@@ -453,7 +453,7 @@ export const webFetchInput = z.object({
     .min(1, "URL is required")
     .describe("The URL to fetch content from"),
   provider: z
-    .enum(["firecrawl", "jina-reader", "tavily-search"])
+    .enum(["firecrawl", "jina-reader", "tavily-search", "tinyfish"])
     .optional()
     .describe("Specific fetch provider to use (default: first available)"),
   format: z
@@ -496,7 +496,7 @@ export const webFetchOutput = z.object({
 export const webFetchTool: McpToolDefinition<typeof webFetchInput, typeof webFetchOutput> = {
   name: "omniroute_web_fetch",
   description:
-    "Fetches and extracts content from a URL using OmniRoute's web fetch gateway. Supports multiple providers (Firecrawl, Jina Reader, Tavily) with automatic failover. Returns the page content as markdown, HTML, links, or screenshot, along with metadata.",
+    "Fetches and extracts content from a URL using OmniRoute's web fetch gateway. Supports multiple providers (Firecrawl, Jina Reader, Tavily, TinyFish) with automatic failover. Returns the page content as markdown, HTML, links, or screenshot, along with metadata.",
   inputSchema: webFetchInput,
   outputSchema: webFetchOutput,
   scopes: ["execute:search"],

open-sse/mcp-server/server.ts

@@ -578,7 +578,7 @@ async function handleWebSearch(args: {
 
 async function handleWebFetch(args: {
   url: string;
-  provider?: "firecrawl" | "jina-reader" | "tavily-search";
+  provider?: "firecrawl" | "jina-reader" | "tavily-search" | "tinyfish";
   format?: "markdown" | "html" | "links" | "screenshot";
   include_metadata?: boolean;
   depth?: number;
@@ -866,7 +866,16 @@ export function createMcpServer(): McpServer {
     )
   );
 
-  server.registerTool("omniroute_pick_fastest_model", { description: "Picks the fastest reliable provider-model pair from live telemetry.", inputSchema: pickFastestModelInput }, withScopeEnforcement("omniroute_pick_fastest_model", (args) => handlePickFastestModel(pickFastestModelInput.parse(args))));
+  server.registerTool(
+    "omniroute_pick_fastest_model",
+    {
+      description: "Picks the fastest reliable provider-model pair from live telemetry.",
+      inputSchema: pickFastestModelInput,
+    },
+    withScopeEnforcement("omniroute_pick_fastest_model", (args) =>
+      handlePickFastestModel(pickFastestModelInput.parse(args))
+    )
+  );
 
   server.registerTool(
     "omniroute_get_session_snapshot",
@@ -1073,17 +1082,21 @@ export function createMcpServer(): McpServer {
         // @ts-ignore: dynamic zod access
         inputSchema: toolDef.inputSchema,
       },
-      withScopeEnforcement(toolDef.name, async (args) => {
-        try {
-          const parsedArgs = toolDef.inputSchema.parse(args ?? {});
-          // @ts-expect-error - handler type lost through dynamic Object.values() access
-          const result = await toolDef.handler(parsedArgs);
-          return { content: [{ type: "text" as const, text: JSON.stringify(result, null, 2) }] };
-        } catch (err) {
-          const msg = err instanceof Error ? err.message : String(err);
-          return { content: [{ type: "text" as const, text: `Error: ${msg}` }], isError: true };
-        }
-      }, toolDef.scopes)
+      withScopeEnforcement(
+        toolDef.name,
+        async (args) => {
+          try {
+            const parsedArgs = toolDef.inputSchema.parse(args ?? {});
+            // @ts-expect-error - handler type lost through dynamic Object.values() access
+            const result = await toolDef.handler(parsedArgs);
+            return { content: [{ type: "text" as const, text: JSON.stringify(result, null, 2) }] };
+          } catch (err) {
+            const msg = err instanceof Error ? err.message : String(err);
+            return { content: [{ type: "text" as const, text: `Error: ${msg}` }], isError: true };
+          }
+        },
+        toolDef.scopes
+      )
     );
   });
 

open-sse/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@omniroute/open-sse",
-  "version": "3.8.45",
+  "version": "3.8.46",
   "description": "Express SSE sidecar for OmniRoute — handles streaming, protocol translation, and provider orchestration",
   "type": "module",
   "main": "index.js",

open-sse/services/autoCombo/builtinCatalog.ts

@@ -1,6 +1,9 @@
 import type { AutoVariant } from "./autoPrefix";
 import { VALID_VARIANTS } from "./autoPrefix";
 import { parseAutoSuffix } from "./suffixComposition";
+import { isValidModelFamily, AUTO_FAMILY_IDS } from "./modelFamily";
+
+export { AUTO_FAMILY_IDS };
 
 /**
  * Built-in `auto/*` catalog → AutoVariant resolution.
@@ -78,7 +81,11 @@ export function resolveAutoVariant(modelStr: string, suffix: string): ResolvedAu
  * decide whether an `auto/` model is a valid built-in before materializing it.
  */
 export function isRecognizedBuiltinAuto(modelStr: string, suffix: string): boolean {
-  return resolveAutoVariant(modelStr, suffix).recognized || parseAutoSuffix(suffix).valid;
+  return (
+    resolveAutoVariant(modelStr, suffix).recognized ||
+    parseAutoSuffix(suffix).valid ||
+    isValidModelFamily(suffix)
+  );
 }
 
 export async function createBuiltinAutoCombo(modelStr: string, suffix: string) {
@@ -105,5 +112,15 @@ export async function createBuiltinAutoCombo(modelStr: string, suffix: string) {
     return virtualCombo;
   }
 
+  // #6453: `auto/<family>` (e.g. auto/glm, auto/minimax, auto/zai, auto/mimo,
+  // auto/gemma, auto/llama, auto/gemini) — spans whatever installed backends
+  // currently expose that model family, degrading gracefully as backends rotate.
+  if (isValidModelFamily(suffix)) {
+    const virtualCombo = await createVirtualAutoCombo(undefined, { family: suffix });
+    virtualCombo.name = modelStr;
+    virtualCombo.id = modelStr;
+    return virtualCombo;
+  }
+
   throw new Error(`Unknown built-in auto combo: ${modelStr}`);
 }

open-sse/services/autoCombo/modelFamily.ts

@@ -0,0 +1,95 @@
+/**
+ * #6453 — Provider-family auto combos.
+ *
+ * `auto/<family>` (e.g. `auto/glm`, `auto/minimax`) is a NEW axis alongside the
+ * existing task-mode (`auto/best-coding`) and category:tier (`auto/coding:fast`)
+ * combos: instead of grouping by routing intent, it groups by underlying MODEL
+ * FAMILY and materializes an on-demand virtual combo spanning whatever installed
+ * backends currently expose that family, degrading gracefully as backends
+ * rotate — same on-demand mechanism (`createVirtualAutoCombo`), a different
+ * candidate filter.
+ *
+ * Kept as a pure, dependency-free module so `detectModelFamily` is unit-testable
+ * in isolation without touching the DB/registry-backed virtual factory.
+ */
+
+export type ModelFamily = "glm" | "minimax" | "mimo" | "zai" | "gemma" | "llama" | "gemini";
+
+export const MODEL_FAMILIES: readonly ModelFamily[] = [
+  "glm",
+  "minimax",
+  "mimo",
+  "zai",
+  "gemma",
+  "llama",
+  "gemini",
+];
+
+const MODEL_FAMILY_SET: ReadonlySet<string> = new Set(MODEL_FAMILIES);
+
+/** Model-id prefix → family. Matched against the bare model id (provider prefix, if
+ * any, stripped) so both `glm-5.2` and `zai/glm-5.2` resolve the same way. */
+const FAMILY_ID_PATTERNS: ReadonlyArray<{ family: ModelFamily; pattern: RegExp }> = [
+  { family: "glm", pattern: /^glm-/i },
+  { family: "minimax", pattern: /^minimax-/i },
+  { family: "mimo", pattern: /^mimo-/i },
+  { family: "gemma", pattern: /^gemma-/i },
+  { family: "llama", pattern: /^llama-/i },
+  { family: "gemini", pattern: /^gemini-/i },
+];
+
+/**
+ * `zai` is deliberately NOT a model-id prefix rule: Zhipu's z.ai hosted API serves
+ * the same `glm-*` model ids as every other GLM backend (`glm` provider, custom
+ * OpenAI/Anthropic-compatible connections, etc — see
+ * `open-sse/config/providers/registry/zai/index.ts`). Aliasing `auto/zai` to a
+ * model-name prefix would make it identical to `auto/glm`, so instead it is
+ * resolved by PROVIDER id: "route to my z.ai backend specifically", distinct from
+ * `auto/glm` ("route to any connected provider currently serving a GLM model,
+ * z.ai included"). Documented here rather than deferred because the distinction
+ * is meaningful and the rule is a one-line lookup.
+ */
+export const FAMILY_PROVIDER_OVERRIDE: Readonly<Partial<Record<ModelFamily, string>>> = {
+  zai: "zai",
+};
+
+export function isValidModelFamily(value: string | null | undefined): value is ModelFamily {
+  return typeof value === "string" && MODEL_FAMILY_SET.has(value);
+}
+
+/**
+ * Detect the model family from a bare or provider-prefixed model id.
+ * Returns `null` when the id doesn't match any known family prefix — including
+ * `zai`, which is never detected from a model id (see `FAMILY_PROVIDER_OVERRIDE`).
+ */
+export function detectModelFamily(modelId: string | null | undefined): ModelFamily | null {
+  if (typeof modelId !== "string" || modelId.trim().length === 0) return null;
+  const bare = modelId.includes("/") ? modelId.slice(modelId.lastIndexOf("/") + 1) : modelId;
+  for (const { family, pattern } of FAMILY_ID_PATTERNS) {
+    if (pattern.test(bare)) return family;
+  }
+  return null;
+}
+
+interface FamilyPoolCandidate {
+  provider: string;
+  model: string;
+}
+
+/**
+ * Build the candidate filter for `auto/<family>`. Provider-override families
+ * (currently only `zai`) filter by connection provider id; every other family
+ * filters by `detectModelFamily(model) === family`.
+ */
+export function buildFamilyCandidateFilter(
+  family: ModelFamily
+): (candidate: FamilyPoolCandidate) => boolean {
+  const providerOverride = FAMILY_PROVIDER_OVERRIDE[family];
+  if (providerOverride) {
+    return (candidate) => candidate.provider === providerOverride;
+  }
+  return (candidate) => detectModelFamily(candidate.model) === family;
+}
+
+/** Advertised `auto/<family>` catalog ids (#6453), e.g. `auto/glm`, `auto/minimax`. */
+export const AUTO_FAMILY_IDS: readonly string[] = MODEL_FAMILIES.map((family) => `auto/${family}`);

open-sse/services/autoCombo/paidModelFilter.ts

@@ -0,0 +1,44 @@
+/**
+ * #6512 (follow-up to #6328 / #6495) — exclude paid-only backends from `auto/*`
+ * candidate pools when the operator opts into the `hidePaidModels` setting.
+ *
+ * PR #6495 added `hidePaidModels` to hide paid models from the `GET /v1/models`
+ * listing, but `auto/*` combos could still pick a paid-only backend into their
+ * candidate pool → a 402/403 at request time, exactly what #6328 wanted to avoid.
+ * This applies the SAME free-model predicate #6495 uses in `catalog.ts` to every
+ * virtual auto-combo candidate pool.
+ *
+ * Kept as a pure, dependency-light function so the filter is unit-testable in
+ * isolation without seeding the DB-backed virtual factory.
+ */
+import { isFreeModel, providerHasFreeModels } from "@/shared/utils/freeModels";
+
+interface PaidFilterCandidate {
+  provider: string;
+  model: string;
+}
+
+/** A candidate is kept only when its provider has documented free models AND the
+ * selected model itself qualifies as free — mirrors `shouldHidePaid` in
+ * `src/app/api/v1/models/catalog.ts`. */
+function isFreeCandidate(candidate: PaidFilterCandidate): boolean {
+  return (
+    providerHasFreeModels(candidate.provider) &&
+    isFreeModel(candidate.provider, { id: candidate.model })
+  );
+}
+
+/**
+ * Return the candidate pool filtered to free-only backends when
+ * `hidePaidModels` is on; otherwise return the pool unchanged (identity — the
+ * default, opt-in-off path). If every candidate is paid the result is empty, and
+ * the caller's existing graceful empty-pool path handles it (consistent with the
+ * opt-in intent — the operator asked not to route to paid models).
+ */
+export function filterPaidOnlyCandidates<T extends PaidFilterCandidate>(
+  pool: T[],
+  hidePaidModels: boolean
+): T[] {
+  if (!hidePaidModels) return pool;
+  return pool.filter(isFreeCandidate);
+}

open-sse/services/autoCombo/virtualFactory.ts

@@ -17,12 +17,17 @@ import {
   type AutoCategory,
   type AutoTier,
 } from "./suffixComposition";
+import { buildFamilyCandidateFilter, type ModelFamily } from "./modelFamily";
 import { getHiddenModelsByProvider } from "@/models";
+import { filterPaidOnlyCandidates } from "./paidModelFilter";
 
-/** #4235 Phase B: optional category/tier overlay for `auto/<category>:<tier>` combos. */
+/** #4235 Phase B: optional category/tier overlay for `auto/<category>:<tier>` combos.
+ * #6453: optional `family` overlay for `auto/<family>` combos (e.g. `auto/glm`) —
+ * mutually exclusive with category/tier, applied instead of them when present. */
 export interface AutoComboSpec {
   category?: AutoCategory;
   tier?: AutoTier;
+  family?: ModelFamily;
 }
 
 /** Minimal connection shape needed for virtual auto-combo factory */
@@ -270,6 +275,17 @@ export async function createVirtualAutoCombo(
     ...getNoAuthCandidates(new Set(validConnections.map((conn) => conn.provider)), blockedProviders)
   );
 
+  // #6512 (follow-up to #6328/#6495): when the operator opts into `hidePaidModels`,
+  // exclude paid-only backends from EVERY `auto/*` candidate pool — not just the
+  // `/v1/models` listing — so auto-routing never picks a model that will 402/403.
+  // If this empties the pool the existing graceful empty-pool path below handles it
+  // (consistent with the opt-in intent). Default OFF → pool unchanged.
+  const paidFilteredPool = filterPaidOnlyCandidates(candidatePool, settings.hidePaidModels === true);
+  if (paidFilteredPool !== candidatePool) {
+    candidatePool.length = 0;
+    candidatePool.push(...paidFilteredPool);
+  }
+
   if (candidatePool.length === 0) {
     log.warn("AUTO", "No connected providers with valid credentials for virtual auto-combo");
     const emptyPool: string[] = [];
@@ -306,26 +322,40 @@ export async function createVirtualAutoCombo(
   // connected. Operators who want the old "never break routing, lose the bias"
   // behavior can opt back in via the env var below.
   let effectivePool = candidatePool;
-  const candidateFilter = spec ? buildAutoCandidateFilter(spec.category, spec.tier) : null;
+  // #6453: `auto/<family>` narrows by model family instead of category/tier. The
+  // two overlays are mutually exclusive on the spec (family takes precedence when
+  // both are somehow present, which callers never do in practice).
+  const candidateFilter = spec?.family
+    ? buildFamilyCandidateFilter(spec.family)
+    : spec
+      ? buildAutoCandidateFilter(spec.category, spec.tier)
+      : null;
   if (candidateFilter) {
     const narrowed = candidatePool.filter((c) =>
       candidateFilter({ provider: c.provider, model: c.model })
     );
+    const label = spec?.family
+      ? `auto/${spec.family}`
+      : `auto/${spec?.category ?? ""}${spec?.tier ? `:${spec.tier}` : ""}`;
     if (narrowed.length > 0) {
       effectivePool = narrowed;
     } else if (
-      process.env.OMNIROUTE_AUTO_FREE_FALLBACK_TO_FULL_POOL === "true" ||
-      process.env.OMNIROUTE_AUTO_FREE_FALLBACK_TO_FULL_POOL === "1"
+      !spec?.family &&
+      (process.env.OMNIROUTE_AUTO_FREE_FALLBACK_TO_FULL_POOL === "true" ||
+        process.env.OMNIROUTE_AUTO_FREE_FALLBACK_TO_FULL_POOL === "1")
     ) {
-      // Opt-in legacy behavior: warn loudly, then keep the full pool.
+      // Opt-in legacy behavior (category/tier only): warn loudly, then keep the full pool.
       log.warn(
         "AUTO",
-        `auto/${spec?.category ?? ""}${spec?.tier ? `:${spec.tier}` : ""} matched no connected models; falling back to the full pool (OMNIROUTE_AUTO_FREE_FALLBACK_TO_FULL_POOL=true)`
+        `${label} matched no connected models; falling back to the full pool (OMNIROUTE_AUTO_FREE_FALLBACK_TO_FULL_POOL=true)`
       );
     } else {
+      // Family combos always degrade to an empty pool when unavailable — a family
+      // is a hard identity constraint, not a soft optimization bias, so there is
+      // no sensible "fall back to the full pool" behavior for it.
       log.warn(
         "AUTO",
-        `auto/${spec?.category ?? ""}${spec?.tier ? `:${spec.tier}` : ""} matched no connected models; returning an empty pool. Set OMNIROUTE_AUTO_FREE_FALLBACK_TO_FULL_POOL=true to restore the legacy "use full pool" behavior.`
+        `${label} matched no connected models; returning an empty pool.${spec?.family ? "" : ' Set OMNIROUTE_AUTO_FREE_FALLBACK_TO_FULL_POOL=true to restore the legacy "use full pool" behavior.'}`
       );
       effectivePool = [];
     }

open-sse/services/combo.ts

@@ -119,6 +119,7 @@ import {
   waitForCooldownAwareRetry,
 } from "../../src/sse/services/cooldownAwareRetry.ts";
 import { handleFusionChat, type FusionTuning } from "./fusion.ts";
+import { handlePipelineChat, type PipelineStep } from "./pipeline.ts";
 import {
   TRANSIENT_FOR_SEMAPHORE,
   MAX_FALLBACK_WAIT_MS,
@@ -142,6 +143,7 @@ import {
   expandProviderWildcardsInCollection,
 } from "./combo/providerWildcard.ts";
 import { resolveShadowTargets, scheduleShadowRouting } from "./combo/shadowRouting.ts";
+import { attemptCompatRejectedFallback } from "./combo/comboCompatFallback.ts";
 import {
   filterTargetsByRequestCompatibility,
   resolveComboRuntimeUnits,
@@ -796,6 +798,37 @@ export async function handleComboChat({
     });
   }
 
+  // Pipeline strategy: sequential chain — each step's output feeds the next step's
+  // input, only the final step's response is returned. Handled in a separate module
+  // because it neither iterates targets as fallbacks nor needs the failover/retry
+  // machinery below — it runs targets in order, threading output → input. The step
+  // list is `combo.models` (in order); an optional per-step `prompt` is read off the
+  // target object (comboModelStepInputSchema.prompt).
+  if (strategy === "pipeline") {
+    const pipelineSteps = (combo.models || [])
+      .map((m): PipelineStep | null => {
+        if (typeof m === "string") return { model: m };
+        if (m && typeof m === "object") {
+          const obj = m as Record<string, unknown>;
+          if (typeof obj.model === "string") {
+            return {
+              model: obj.model,
+              prompt: typeof obj.prompt === "string" ? obj.prompt : undefined,
+            };
+          }
+        }
+        return null;
+      })
+      .filter((s): s is PipelineStep => Boolean(s));
+    return handlePipelineChat({
+      body,
+      steps: pipelineSteps,
+      handleSingleModel: handleSingleModelWithTimeout,
+      log,
+      comboName: combo.name,
+    });
+  }
+
   const nestingContext = nesting || {
     depth: 0,
     maxDepth: clampComboDepth(config.maxComboDepth),
@@ -2371,6 +2404,15 @@ async function handleRoundRobinCombo({
     log,
     "Context-aware round-robin fallback"
   );
+  // #6238: keep the targets the compat pre-filter rejected so they can serve as a
+  // last-resort fallback tier. The pre-filter drops request-incompatible targets
+  // BEFORE availability is known; if every compat-kept target then turns out to be
+  // runtime-unavailable, we must reconsider these before returning 503, instead of
+  // permanently dropping a compat-rejected-but-healthy provider.
+  const compatKeptSet = new Set(filteredTargets);
+  const compatRejectedTargets = evalRankedTargets.filter(
+    (target) => !compatKeptSet.has(target)
+  );
   const modelCount = filteredTargets.length;
   if (modelCount === 0) {
     return comboModelNotFoundResponse("Round-robin combo has no executable targets");
@@ -2712,6 +2754,15 @@ async function handleRoundRobinCombo({
 
           if (stickyRoundRobinEnabled) {
             recordStickyRoundRobinSuccess(combo.name, target, stickyLimit, filteredTargets);
+          } else {
+            // #948: true round-robin (stickyLimit <= 1). The counter was advanced
+            // eagerly (+1 from the scheduled start index) before this loop ran, so
+            // when the scheduled model failed and a *different* model served via
+            // fallback, the next request reused the fallback-served model. Advance
+            // the pointer past the model that ACTUALLY served (modelIndex) instead,
+            // mirroring recordStickyRoundRobinSuccess's served-index logic. Read
+            // side applies `% modelCount`, so storing modelIndex + 1 is correct.
+            rrCounters.set(combo.name, modelIndex + 1);
           }
 
           // #3825: (re)record the sticky binding so the next turn re-pins (prompt-cache).
@@ -2964,6 +3015,37 @@ async function handleRoundRobinCombo({
 
   // All models exhausted
   const latencyMs = Date.now() - startTime;
+
+  // #6238: every compat-kept target was skipped as unavailable and NONE was ever
+  // attempted (recordedAttempts === 0). Before crystallizing 503, probe the targets
+  // the compat pre-filter rejected — a compat-rejected-but-healthy provider is a
+  // valid last-resort fallback tier, not a permanently dropped target.
+  if (recordedAttempts === 0 && compatRejectedTargets.length > 0) {
+    const compatFallbackResult = await attemptCompatRejectedFallback(compatRejectedTargets, body, {
+      handleSingleModel,
+      isModelAvailable,
+      isProviderInCooldown: (target) =>
+        resilienceSettings.providerCooldown.enabled &&
+        Boolean(target.provider && target.provider !== "unknown") &&
+        isProviderInCooldown(
+          target.provider as string,
+          target.connectionId as string | undefined,
+          resilienceSettings
+        ),
+      log,
+      strategy: "round-robin",
+    });
+    if (compatFallbackResult) {
+      recordComboRequest(combo.name, null, {
+        success: true,
+        latencyMs: Date.now() - startTime,
+        fallbackCount,
+        strategy: "round-robin",
+      });
+      return compatFallbackResult;
+    }
+  }
+
   if (recordedAttempts === 0) {
     recordComboRequest(combo.name, null, {
       success: false,

open-sse/services/combo/comboCompatFallback.ts

@@ -0,0 +1,82 @@
+import type { ComboLogger, HandleSingleModel, IsModelAvailable, ResolvedComboTarget } from "./types";
+
+/**
+ * Last-resort fallback tier for combo routing (#6238).
+ *
+ * `filterTargetsByRequestCompatibility` drops targets that look request-incompatible
+ * (tool/vision/structured-output unsupported, or below the required context window)
+ * BEFORE any runtime availability check runs. Its only safety net triggers when ALL
+ * targets are filtered — not when the kept targets are later all runtime-unavailable
+ * (circuit-open / cooldown / no credentials). In that case a combo would return
+ * `503 ALL_ACCOUNTS_INACTIVE` without ever reconsidering a compat-rejected-but-healthy
+ * target.
+ *
+ * This helper makes those compat-rejected targets a genuine fallback tier: only after
+ * the primary (compat-kept) targets were all skipped without a single real attempt do
+ * we probe the rejected set. A compat-rejected target is used only if it is actually
+ * available and returns a successful upstream response; otherwise the caller keeps its
+ * original error surface unchanged.
+ */
+export interface CompatFallbackContext {
+  handleSingleModel: HandleSingleModel;
+  isModelAvailable?: IsModelAvailable;
+  /** Returns true when the target's provider connection is in resilience cooldown. */
+  isProviderInCooldown?: (target: ResolvedComboTarget) => boolean;
+  log: ComboLogger;
+  /** Effective combo strategy, threaded into the single-model dispatch for telemetry. */
+  strategy: string;
+}
+
+/**
+ * Attempt each compatibility-rejected target in order, returning the first successful
+ * upstream `Response`. Targets that are unavailable, in cooldown, or return a non-ok
+ * response are skipped. Returns `null` when no rejected target yields a success, so the
+ * caller can fall through to its existing error/503 path with an unchanged error surface.
+ */
+export async function attemptCompatRejectedFallback(
+  rejectedTargets: ResolvedComboTarget[],
+  body: Record<string, unknown>,
+  ctx: CompatFallbackContext
+): Promise<Response | null> {
+  if (rejectedTargets.length === 0) return null;
+
+  for (const target of rejectedTargets) {
+    if (ctx.isModelAvailable) {
+      const available = await ctx.isModelAvailable(target.modelStr, target);
+      if (!available) {
+        ctx.log.debug(
+          "COMBO",
+          `Last-resort compat fallback: ${target.modelStr} still unavailable — skipping`
+        );
+        continue;
+      }
+    }
+
+    if (ctx.isProviderInCooldown?.(target)) {
+      ctx.log.debug(
+        "COMBO",
+        `Last-resort compat fallback: ${target.modelStr} provider in cooldown — skipping`
+      );
+      continue;
+    }
+
+    ctx.log.info(
+      "COMBO",
+      `Last-resort compat fallback → ${target.modelStr} (all compat-kept targets were unavailable)`
+    );
+    const result = await ctx.handleSingleModel(body, target.modelStr, {
+      ...target,
+      effectiveComboStrategy: ctx.strategy,
+    });
+    if (result.ok) {
+      ctx.log.info("COMBO", `Last-resort compat fallback succeeded via ${target.modelStr}`);
+      return result;
+    }
+    ctx.log.debug(
+      "COMBO",
+      `Last-resort compat fallback: ${target.modelStr} failed (${result.status}) — trying next`
+    );
+  }
+
+  return null;
+}

open-sse/services/compression/diffHelper.ts

@@ -28,6 +28,7 @@ export interface CompressionPreviewDiff {
   validationWarnings: string[];
   validationErrors: string[];
   fallbackApplied: boolean;
+  fallbackReason?: string;
   heatmap?: CompressionHeatmap;
 }
 
@@ -187,6 +188,15 @@ export function buildCompressionPreviewDiff(
     ? [{ type: "same", text: "[diff omitted: input too large]" }]
     : buildCompressionDiff(original, compressed);
 
+  let fallbackReason: string | undefined;
+  if (validation.fallbackApplied) {
+    fallbackReason = validation.errors.length > 0
+      ? `validation-failed: ${validation.errors[0]}`
+      : "validation-failed";
+  } else if (stats?.fallbackApplied) {
+    fallbackReason = "compression-fallback";
+  }
+
   const result: CompressionPreviewDiff = {
     segments,
     preservedBlocks: preserved,
@@ -198,6 +208,7 @@ export function buildCompressionPreviewDiff(
     ],
     validationErrors: [...(stats?.validationErrors ?? []), ...validation.errors],
     fallbackApplied: Boolean(stats?.fallbackApplied || validation.fallbackApplied),
+    ...(fallbackReason && { fallbackReason }),
   };
 
   if (heatmapMode) {

open-sse/services/compression/engines/cavemanAdapter.ts

@@ -289,9 +289,22 @@ export const cavemanEngine: CompressionEngine = {
   },
   apply(body, options) {
     const adapter = adaptBodyForCompression(body);
+    // Mirror rtkAdapter's default-enabled behavior (see rtk/index.ts:530-535). When this engine
+    // is invoked as a stacked step without explicit `enabled` on either the cavemanConfig or the
+    // stepConfig, default `enabled: true` so the rules actually run. Without this,
+    // DEFAULT_CAVEMAN_CONFIG.enabled=false made cavemanCompress() a silent no-op, and the
+    // preview route's default [rtk, caveman] pipeline reported 0% savings even on trigger prose.
+    // (Issue #6425.)
+    const explicitCavemanConfig = options?.config?.cavemanConfig;
+    const explicitStepConfig = options?.stepConfig;
+    const explicitEnabled =
+      (explicitCavemanConfig && "enabled" in explicitCavemanConfig) ||
+      (explicitStepConfig && "enabled" in explicitStepConfig);
+    const enabledDefault = explicitEnabled ? {} : { enabled: true };
     const cavemanConfig = {
-      ...(options?.config?.cavemanConfig ?? {}),
-      ...(options?.stepConfig ?? {}),
+      ...enabledDefault,
+      ...(explicitCavemanConfig ?? {}),
+      ...(explicitStepConfig ?? {}),
       ...(options?.config?.languageConfig?.enabled
         ? {
             language: options.config.languageConfig.defaultLanguage,

open-sse/services/compression/engines/session-dedup/index.ts

@@ -88,6 +88,53 @@ function findSuffixBlocks(
 
 // ─── two-pass dedup on message texts ─────────────────────────────────────────
 
+/**
+ * Deduplicates repeated lines within a single message (intra-message dedup).
+ * Replaces repeated suffix blocks with markers.
+ */
+function dedupeWithinMessage(
+  text: string,
+  minBlockChars: number
+): { deduped: string; changed: boolean } {
+  const lines = text.split("\n");
+  const blocks = findSuffixBlocks(lines, minBlockChars);
+
+  if (blocks.length < 2) return { deduped: text, changed: false };
+
+  // Find the most common block (likely candidate for intra-message dedup).
+  const blockFreq = new Map<string, number>();
+  for (const { block } of blocks) {
+    blockFreq.set(block, (blockFreq.get(block) || 0) + 1);
+  }
+
+  // Sort by frequency descending, then by length descending (prefer replacing more common, longer blocks first).
+  const sortedBlocks = [...blocks].sort((a, b) => {
+    const freqDiff = (blockFreq.get(b.block) || 0) - (blockFreq.get(a.block) || 0);
+    return freqDiff !== 0 ? freqDiff : b.block.length - a.block.length;
+  });
+
+  let result = text;
+  let changed = false;
+
+  for (const { block } of sortedBlocks) {
+    // Only dedup blocks that appear 2+ times in the text.
+    const occurrences = (result.match(new RegExp(block.replace(/[.*+?^${}()|[\]\\]/g, "\\$&"), "g")) || []).length;
+    if (occurrences < 2) continue;
+
+    const sha = hashBlock(block);
+    const marker = `[dedup:ref sha=${sha}]`;
+    // Replace ALL occurrences except the first (keep the original once).
+    let count = 0;
+    result = result.replace(new RegExp(block.replace(/[.*+?^${}()|[\]\\]/g, "\\$&"), "g"), () => {
+      count++;
+      return count === 1 ? block : marker;
+    });
+    changed = true;
+  }
+
+  return { deduped: result, changed };
+}
+
 /**
  * Runs two-pass dedup over an ordered list of (msgIdx, text) pairs.
  * Returns the replaced texts for duplicate messages, a reverse map, and a count.
@@ -99,6 +146,21 @@ function dedupMessageTexts(
   deduped: Map<number, string>;
   dedupCount: number;
 } {
+  const deduped = new Map<number, string>();
+  let dedupCount = 0;
+
+  // Single-message case: apply intra-message dedup.
+  if (msgTexts.length === 1) {
+    const { text, msgIdx } = msgTexts[0];
+    const { deduped: dedupedText, changed } = dedupeWithinMessage(text, minBlockChars);
+    if (changed) {
+      deduped.set(msgIdx, dedupedText);
+      dedupCount++;
+    }
+    return { deduped, dedupCount };
+  }
+
+  // Multi-message case: apply cross-turn dedup.
   // Pass 1: for each message, extract suffix blocks and record first ownership.
   // `firstSeen`: sha → { ownerMsgIdx, block }
   const firstSeen = new Map<string, { ownerMsgIdx: number; block: string }>();
@@ -115,9 +177,6 @@ function dedupMessageTexts(
   }
 
   // Pass 2: for each message, find blocks that were FIRST seen in an earlier message.
-  const deduped = new Map<number, string>();
-  let dedupCount = 0;
-
   for (const { msgIdx, text } of msgTexts) {
     const lines = text.split("\n");
     const blocks = findSuffixBlocks(lines, minBlockChars);
@@ -202,7 +261,7 @@ function processMessages(
     }
   }
 
-  if (msgTexts.length < 2) {
+  if (msgTexts.length === 0) {
     return { messages, dedupCount: 0 };
   }
 

open-sse/services/compression/strategySelector.ts

@@ -839,7 +839,10 @@ function runStackedCompression(
 
   for (const step of steps) {
     const engine = getCompressionEngine(step.engine);
-    if (!engine) continue;
+    if (!engine) {
+      acc.validationErrors.add(`Unknown compression engine: "${step.engine}"`);
+      continue;
+    }
     // Respect the registry enabled flag: a step naming a disabled engine is skipped, so an
     // operator can turn an engine off (setEngineEnabled) without editing every pipeline.
     if (getEngineEntry(step.engine)?.enabled === false) continue;
@@ -939,7 +942,10 @@ async function runStackedCompressionAsync(
 
   for (const step of steps) {
     const engine = getCompressionEngine(step.engine);
-    if (!engine) continue;
+    if (!engine) {
+      acc.validationErrors.add(`Unknown compression engine: "${step.engine}"`);
+      continue;
+    }
     // Respect the registry enabled flag (same as the sync loop) — keep both in lockstep.
     if (getEngineEntry(step.engine)?.enabled === false) continue;
     // T02: skip an engine whose breaker is OPEN (verbatim body kept — fail-open). Lockstep w/ sync.

open-sse/services/errorClassifier.ts

@@ -158,7 +158,29 @@ export function classifyProviderError(
     return PROVIDER_ERROR_TYPES.ACCOUNT_DEACTIVATED;
   }
   if (statusCode === 403) {
-    if (bodyStr.includes("has not been used in project")) {
+    // Cloud Code / Antigravity (Gemini Code Assist) 403s are almost always a
+    // RECOVERABLE project-config issue — the Cloud AI Companion API not enabled
+    // on the project ("has not been used in project …", SERVICE_DISABLED,
+    // accessNotConfigured), a stale/mismatched project, or PERMISSION_DENIED on
+    // the project — NOT an account ban. Real account bans are already caught by
+    // isAccountDeactivated above (→ ACCOUNT_DEACTIVATED). Classifying these as
+    // PROJECT_ROUTE_ERROR keeps the account active and recoverable once the
+    // project/API is fixed, instead of permanently disabling it on a single
+    // fixable 403 (which previously required a full OAuth reconnect). (antigravity-403)
+    const p = (provider || "").toLowerCase();
+    const isCloudCodeProvider =
+      p === "antigravity" ||
+      p === "gemini-cli" ||
+      p.includes("cloudcode") ||
+      p.includes("cloud-code");
+    const recoverableProject403 =
+      bodyStr.includes("has not been used in project") ||
+      bodyStr.includes("SERVICE_DISABLED") ||
+      bodyStr.includes("accessNotConfigured") ||
+      bodyStr.includes("PERMISSION_DENIED") ||
+      /\bit is disabled\b/i.test(bodyStr) ||
+      isCloudCodeProvider;
+    if (recoverableProject403) {
       return PROVIDER_ERROR_TYPES.PROJECT_ROUTE_ERROR;
     }
     if (provider && getProviderCategory(provider) === "apikey") {

open-sse/services/fusion.ts

@@ -268,6 +268,7 @@ export async function handleFusionChat({
 
   // 2. Collect successful answers.
   const answers: Array<{ model: string; text: string }> = [];
+  const rateLimited: string[] = [];
   for (let i = 0; i < settled.length; i++) {
     const res = settled[i];
     const model = panel[i];
@@ -288,7 +289,12 @@ export async function handleFusionChat({
     }
     const resp = res as Response;
     if (!resp.ok) {
-      log.warn("FUSION", `Panel ${model} failed`, { status: resp.status });
+      if (resp.status === 429) {
+        rateLimited.push(model);
+        log.warn("FUSION", `Panel ${model} rate-limited`, { status: resp.status });
+      } else {
+        log.warn("FUSION", `Panel ${model} failed`, { status: resp.status });
+      }
       continue;
     }
     try {
@@ -309,8 +315,12 @@ export async function handleFusionChat({
 
   // 3. Degrade gracefully when the panel is too thin to fuse.
   if (answers.length === 0) {
-    log.warn("FUSION", "All panel models failed");
-    return errorResponse(503, "All fusion panel models failed");
+    const detail =
+      rateLimited.length > 0
+        ? `${rateLimited.length} models rate-limited, ${panel.length - rateLimited.length} failed`
+        : `all ${panel.length} models failed`;
+    log.warn("FUSION", `No live models: ${detail}`);
+    return errorResponse(503, `All fusion panel models failed (${detail})`);
   }
   if (answers.length === 1) {
     log.info(

open-sse/services/pipeline.ts

@@ -0,0 +1,189 @@
+/**
+ * Pipeline combo strategy — sequential chain.
+ *
+ * A pipeline combo runs its targets IN ORDER: step N's output is fed into step
+ * N+1 as input, each step carries its own optional `prompt` (instruction), and
+ * only the FINAL step's response is returned to the client. This is the sequential
+ * counterpart to `fusion` (parallel fan-out + judge synthesis).
+ *
+ * ── Per-step config shape ─────────────────────────────────────────────────────
+ * The ordered step list IS `combo.models` — we reuse the existing target order
+ * rather than introducing a parallel `pipelineSteps` array that could drift out of
+ * sync with the models. Each step's optional instruction is read from a `prompt`
+ * field on the target object (`comboModelStepInputSchema.prompt`); a plain-string
+ * model entry is simply a step with no prompt. The field is optional and ignored by
+ * every other strategy, so this is fully backward-compatible.
+ *
+ * ── Prompt injection ──────────────────────────────────────────────────────────
+ * The engine passes prompts through the request's message array (OpenAI `messages`,
+ * Responses `input`, or Gemini `contents`). Each step's `prompt` is injected as a
+ * leading system instruction in whichever format the request uses:
+ *   - step 1 keeps the client's original conversation and (if set) prepends its
+ *     prompt as an extra system turn, so the first model sees the real user request;
+ *   - steps 2..N are transforms — the conversation is replaced with the previous
+ *     step's output as the user turn, plus this step's prompt as the system turn.
+ *
+ * Intermediate steps are forced non-streaming with tools stripped (we need the
+ * complete text to thread forward). The FINAL step keeps the client's original
+ * `stream` flag + tools, so streaming and downstream tool use still work.
+ *
+ * A step failure fails the whole pipeline EXPLICITLY (never silently swallowed):
+ * a non-OK intermediate response, an unparseable body, or an intermediate step that
+ * yields no text short-circuits with a sanitized error response.
+ */
+import { errorResponse } from "../utils/error.ts";
+import type { ComboLogger, HandleSingleModel } from "./combo/types.ts";
+// extractPanelText is a generic assistant-text extractor (OpenAI chat / Claude /
+// Gemini / Responses) — reused here to read each step's output, not fusion-specific.
+import { extractPanelText } from "./fusion.ts";
+
+type Body = Record<string, unknown>;
+
+export type PipelineStep = { model: string; prompt?: string | null };
+
+/**
+ * Prepend a system instruction to the client's original conversation (format-aware),
+ * so step 1 sees the real user request plus its own step prompt. No-op when the
+ * step has no prompt.
+ */
+export function prependSystemInstruction(body: Body, prompt: string | null | undefined): Body {
+  const sys = typeof prompt === "string" && prompt.trim() ? prompt.trim() : null;
+  const next: Body = { ...body };
+  if (!sys) return next;
+  if (Array.isArray(body.input)) {
+    next.input = [{ role: "system", content: sys }, ...(body.input as unknown[])];
+  } else if (Array.isArray(body.contents)) {
+    // Gemini contents have no system role — a leading user turn is the closest analog.
+    next.contents = [{ role: "user", parts: [{ text: sys }] }, ...(body.contents as unknown[])];
+  } else if (Array.isArray(body.messages)) {
+    next.messages = [{ role: "system", content: sys }, ...(body.messages as unknown[])];
+  } else {
+    next.messages = [{ role: "system", content: sys }];
+  }
+  return next;
+}
+
+/**
+ * Replace the request's conversation with a fresh transform turn set (the previous
+ * step's output as the user turn + this step's prompt as the system turn),
+ * preserving whichever message-array shape the request format uses. Non-message
+ * fields are carried over; the caller overrides the model per step.
+ */
+export function buildTransformBody(
+  body: Body,
+  prompt: string | null | undefined,
+  input: string
+): Body {
+  const next: Body = { ...body };
+  const sys = typeof prompt === "string" && prompt.trim() ? prompt.trim() : null;
+  if (Array.isArray(body.input)) {
+    const turns: unknown[] = [];
+    if (sys) turns.push({ role: "system", content: sys });
+    turns.push({ role: "user", content: input });
+    next.input = turns;
+    delete next.messages;
+    delete next.contents;
+  } else if (Array.isArray(body.contents)) {
+    // Gemini contents have no system role — fold the instruction into the user turn.
+    const text = sys ? `${sys}\n\n${input}` : input;
+    next.contents = [{ role: "user", parts: [{ text }] }];
+    delete next.messages;
+    delete next.input;
+  } else {
+    const turns: unknown[] = [];
+    if (sys) turns.push({ role: "system", content: sys });
+    turns.push({ role: "user", content: input });
+    next.messages = turns;
+  }
+  return next;
+}
+
+/** Force non-streaming and strip tools so an intermediate step yields complete prose. */
+function stripStreaming(body: Body): Body {
+  const { tools: _tools, tool_choice: _tc, ...rest } = body;
+  void _tools;
+  void _tc;
+  return { ...rest, stream: false };
+}
+
+export type HandlePipelineChatOptions = {
+  body: Body;
+  steps: PipelineStep[];
+  handleSingleModel: HandleSingleModel;
+  log: ComboLogger;
+  comboName?: string;
+};
+
+/**
+ * Handle a pipeline combo: run the steps in order, threading each step's output
+ * into the next step's input, and return only the final step's response.
+ */
+export async function handlePipelineChat({
+  body,
+  steps,
+  handleSingleModel,
+  log,
+  comboName,
+}: HandlePipelineChatOptions): Promise<Response> {
+  const chain = (Array.isArray(steps) ? steps : []).filter((s) => s && s.model);
+  if (chain.length === 0) {
+    return errorResponse(400, "Pipeline combo has no models");
+  }
+  log.info(
+    "PIPELINE",
+    `Combo "${comboName ?? ""}" | steps=${chain.length} [${chain.map((s) => s.model).join(" -> ")}]`
+  );
+
+  // Single-step pipeline: nothing to chain — run it directly (streams to client).
+  if (chain.length === 1) {
+    return handleSingleModel(prependSystemInstruction(body, chain[0].prompt), chain[0].model);
+  }
+
+  let prevOutput = "";
+  for (let i = 0; i < chain.length; i++) {
+    const step = chain[i];
+    const isFinal = i === chain.length - 1;
+    const isFirst = i === 0;
+
+    let stepBody: Body = isFirst
+      ? prependSystemInstruction(body, step.prompt)
+      : buildTransformBody(body, step.prompt, prevOutput);
+    // Intermediate steps: complete prose only (no stream, no tools). The final step
+    // keeps the client's original stream flag + tools.
+    if (!isFinal) stepBody = stripStreaming(stepBody);
+
+    const t0 = Date.now();
+    const res = await handleSingleModel(stepBody, step.model);
+
+    if (isFinal) {
+      log.info("PIPELINE", `Final step ${step.model} responded (${Date.now() - t0}ms)`);
+      return res;
+    }
+
+    // An intermediate step must succeed with usable text — otherwise fail the whole
+    // pipeline (never silently swallow; the client gets a clear, sanitized error).
+    if (!res.ok) {
+      log.warn("PIPELINE", `Step ${i + 1} (${step.model}) failed`, { status: res.status });
+      const status = res.status >= 400 && res.status <= 599 ? res.status : 502;
+      return errorResponse(status, `Pipeline step ${i + 1} (${step.model}) failed`);
+    }
+    try {
+      const json = await res.clone().json();
+      prevOutput = extractPanelText(json);
+    } catch {
+      log.warn("PIPELINE", `Step ${i + 1} (${step.model}) returned an unparseable body`);
+      return errorResponse(502, `Pipeline step ${i + 1} (${step.model}) returned an unparseable body`);
+    }
+    if (!prevOutput.trim()) {
+      log.warn("PIPELINE", `Step ${i + 1} (${step.model}) returned empty output`);
+      return errorResponse(502, `Pipeline step ${i + 1} (${step.model}) returned empty output`);
+    }
+    log.info(
+      "PIPELINE",
+      `Step ${i + 1} ${step.model} ok (${prevOutput.length} chars, ${Date.now() - t0}ms)`
+    );
+  }
+
+  // Unreachable — the final step returns inside the loop.
+  return errorResponse(500, "Pipeline produced no final response");
+}

open-sse/services/qoderCli.ts

@@ -3,6 +3,9 @@ import crypto from "crypto";
 import fs from "fs";
 import os from "os";
 import path from "path";
+import { getLookupEnv } from "@/shared/services/cliRuntime";
+import { buildQoderCliNotFoundHint, resolveQoderCliInvocation } from "./qoderCliResolve";
+export { getQoderCliCommand } from "./qoderCliResolve"; // #6263 public entry point
 
 const DEFAULT_TIMEOUT_MS = 45_000;
 const DEFAULT_MODELS_TIMEOUT_MS = 20_000;
@@ -63,11 +66,6 @@ function getString(value: unknown): string {
   return typeof value === "string" ? value : "";
 }
 
-export function getQoderCliCommand(): string {
-  const explicit = String(process.env.CLI_QODER_BIN || "").trim();
-  return explicit || "qodercli";
-}
-
 export function getQoderCliWorkspace(): string {
   const explicit = String(
     process.env.QODER_CLI_WORKSPACE || process.env.OMNIROUTE_QODER_WORKSPACE || ""
@@ -124,10 +122,13 @@ type SpawnQoderCliOptions = {
  * honors for headless PAT auth — and the prompt is piped through stdin so no
  * untrusted value is ever interpolated into a shell command (Hard Rule #13).
  */
-function spawnQoderCli(options: SpawnQoderCliOptions): Promise<QoderCliRunResult> {
-  const command = String(options.command || "").trim() || getQoderCliCommand();
+async function spawnQoderCli(options: SpawnQoderCliOptions): Promise<QoderCliRunResult> {
   const timeoutMs = options.timeoutMs ?? DEFAULT_TIMEOUT_MS;
-  const env: NodeJS.ProcessEnv = { ...process.env };
+  // #6263: resolve the real qodercli command (absolute .cmd/.exe on Windows) and
+  // whether it needs a shell, then spawn with the cliRuntime-enriched env (PATH +
+  // PATHEXT + APPDATA) so the npm `.cmd` wrapper under %APPDATA%\npm is found.
+  const { command, useShell } = await resolveQoderCliInvocation(options.command);
+  const env: NodeJS.ProcessEnv = { ...getLookupEnv() };
   const token = String(options.token || "").trim();
   if (token) env.QODER_PERSONAL_ACCESS_TOKEN = token;
 
@@ -143,6 +144,7 @@ function spawnQoderCli(options: SpawnQoderCliOptions): Promise<QoderCliRunResult
         env,
         cwd: options.cwd || undefined,
         stdio: ["pipe", "pipe", "pipe"],
+        ...(useShell ? { shell: true } : {}),
       });
     } catch (err) {
       resolve({
@@ -944,10 +946,7 @@ export async function validateQoderCliPat({
   if (run.error && /enoent|not found|no such file|spawn/i.test(run.error)) {
     return {
       valid: false,
-      error:
-        `Qoder CLI (qodercli) was not found on the OmniRoute host (${run.error}). ` +
-        "Install it from https://qoder.com or point CLI_QODER_BIN at the binary. " +
-        "PAT auth is driven through the local qodercli binary.",
+      error: buildQoderCliNotFoundHint(run.error),
       unsupported: false,
     };
   }

open-sse/services/qoderCliResolve.ts

@@ -0,0 +1,111 @@
+/**
+ * qodercli command resolution (#6263).
+ *
+ * Extracted from `qoderCli.ts` (frozen at the file-size baseline) so the
+ * Windows-aware resolution logic can grow without bloating the transport module.
+ *
+ * The bare `"qodercli"` name does not resolve on Windows, where npm installs the
+ * CLI as a `qodercli.cmd` wrapper under `%APPDATA%\npm` (a user-PATH directory)
+ * that `spawn` cannot find with `shell:false` and an unenriched env. OmniRoute
+ * already has a Windows-aware resolver for this exact tool in `cliRuntime.ts`, so
+ * we reuse it: `getCliRuntimeStatus("qoder")` returns an absolute `.cmd`/`.exe`
+ * `commandPath`, and `shouldUseShellForCommand()` tells us whether it needs cmd.exe.
+ */
+import {
+  getCliRuntimeStatus,
+  getKnownToolPaths,
+  shouldUseShellForCommand,
+} from "@/shared/services/cliRuntime";
+
+export function getQoderCliCommand(): string {
+  const explicit = String(process.env.CLI_QODER_BIN || "").trim();
+  return explicit || "qodercli";
+}
+
+export type QoderCliInvocation = { command: string; useShell: boolean };
+
+// Resolving through cliRuntime does synchronous fs walks plus a `--version`
+// healthcheck spawn; memoize the result so we don't repeat that on every chat /
+// quota request. The install location is effectively static for a running host.
+const QODER_RESOLVE_TTL_MS = 5 * 60 * 1000;
+// Keyed on the fallback command (which folds in CLI_QODER_BIN) so changing the
+// override — or a test pointing at a fresh stub — invalidates a stale entry
+// instead of spawning a since-deleted binary.
+let qoderInvocationCache: (QoderCliInvocation & { key: string; expiresAt: number }) | null = null;
+
+/** Test-only: drop the memoized qodercli command resolution. */
+export function __clearQoderCliInvocationCache(): void {
+  qoderInvocationCache = null;
+}
+
+/**
+ * Resolve the exact command + shell flag to spawn qodercli. `deps` is injectable
+ * for unit tests; production uses the real cliRuntime exports.
+ */
+export async function resolveQoderCliInvocation(
+  explicitCommand?: string | null,
+  deps: {
+    getStatus?: typeof getCliRuntimeStatus;
+    shouldUseShell?: typeof shouldUseShellForCommand;
+  } = {}
+): Promise<QoderCliInvocation> {
+  const explicit = String(explicitCommand || "").trim();
+  const getStatus = deps.getStatus || getCliRuntimeStatus;
+  const shouldUseShell = deps.shouldUseShell || shouldUseShellForCommand;
+  // Only the default path is cached; an explicit per-call command or an injected
+  // resolver (tests) always resolves fresh and never touches the shared cache.
+  const cacheable = !explicit && !deps.getStatus && !deps.shouldUseShell;
+  const fallback = explicit || getQoderCliCommand();
+
+  if (
+    cacheable &&
+    qoderInvocationCache &&
+    qoderInvocationCache.key === fallback &&
+    qoderInvocationCache.expiresAt > Date.now()
+  ) {
+    return { command: qoderInvocationCache.command, useShell: qoderInvocationCache.useShell };
+  }
+
+  let command = fallback;
+  try {
+    const status = await getStatus("qoder");
+    if (status && status.installed && status.commandPath) {
+      command = status.commandPath;
+    }
+  } catch {
+    /* fall back to the bare/explicit command — spawn will surface a real ENOENT */
+  }
+
+  const invocation: QoderCliInvocation = { command, useShell: shouldUseShell(command) };
+  if (cacheable) {
+    qoderInvocationCache = {
+      ...invocation,
+      key: fallback,
+      expiresAt: Date.now() + QODER_RESOLVE_TTL_MS,
+    };
+  }
+  return invocation;
+}
+
+/**
+ * Build the operator-facing "qodercli not found" error, listing the paths the
+ * resolver searched plus the `CLI_QODER_BIN` override hint (#6263).
+ */
+export function buildQoderCliNotFoundHint(runError: string): string {
+  let searchedHint = "";
+  try {
+    const candidates = getKnownToolPaths("qoder");
+    if (candidates.length > 0) {
+      searchedHint = ` Searched: ${candidates.slice(0, 6).join(", ")}.`;
+    }
+  } catch {
+    /* best-effort — the path list is only advisory for the error message */
+  }
+  return (
+    `Qoder CLI (qodercli) was not found on the OmniRoute host (${runError}).` +
+    searchedHint +
+    " Install it from https://qoder.com, or set CLI_QODER_BIN to the absolute path " +
+    "of the qodercli binary (e.g. %APPDATA%\\npm\\qodercli.cmd on Windows). " +
+    "PAT auth is driven through the local qodercli binary."
+  );
+}