fix(responses-api): tool call after a text message collided on the same output_index

Live incident (2026-08-08): an OpenClaw agent sent a short preamble line
("Kör nu, på riktigt — apply_patch på vibe-scriptet:") followed by an
apply_patch tool call in the same turn. The client only spoke the preamble
and never executed the patch, even though OmniRoute's own recorded
responseBody had a complete, valid tool_calls entry.

Root cause: emitToolCall/closeToolCall computed a tool call's output_index
as `reasoningIndex + 1 + tcIdx`, assuming reasoningIndex + 1 was free for
the first tool call (tcIdx=0). But a text message emitted in the same turn
ALSO claims reasoningIndex + 1 (or index 0 with no reasoning) — so a
turn with reasoning + text content + a tool call collided the tool call's
added/delta/done events onto the same output_index as the just-closed
message. A client that tracks response items by output_index (as expected
for the Responses API) sees the tool call events land on an index it
already marked complete and can silently drop them.

Fix: track whether a message item was actually emitted at that index
(state.msgItemAdded) and, if so, tool calls start one slot after it.
Extracted a shared toolCallOutputIndexBase() helper so emitToolCall and
closeToolCall can no longer compute this independently and drift apart.

Confirmed via the live call log artifact (id 1786223153235-770a1c):
response.output_item.done for the text message and response.output_item.added
for the tool call both carried output_index=1 in the raw SSE stream, 1.84s
apart, exactly matching the reported symptom.

.github/workflows/quality.yml

@@ -151,63 +151,6 @@ jobs:
           key: eslint-${{ runner.os }}-${{ hashFiles('eslint.config.mjs', 'eslint.complexity-ratchets.config.mjs', 'config/quality/eslint-suppressions.json', 'package-lock.json') }}
           restore-keys: |
             eslint-${{ runner.os }}-
-      - run: npm run check:provider-consistency
-      - run: npm run check:fetch-targets
-      # docs-all / openapi-routes / docs-symbols live in docs-gates (path-filtered).
-      - run: npm run check:deps
-      # #8522: --base-ref mode for PR events — compare against max(frozen, base) so
-      # inherited drift (base already over frozen cap) doesn't red an innocent PR.
-      # workflow_dispatch (no PR base) falls back to absolute comparison.
-      - name: File-size ratchet (base-relative on PR)
-        env:
-          PR_BASE_SHA: ${{ github.event.pull_request.base.sha }}
-        run: |
-          if [ -n "$PR_BASE_SHA" ]; then
-            npm run check:file-size -- --base-ref "$PR_BASE_SHA"
-          else
-            npm run check:file-size
-          fi
-      - run: npm run check:error-helper
-      - run: npm run check:migration-numbering
-      - run: npm run check:public-creds
-      - run: npm run check:db-rules
-      - run: npm run check:known-symbols
-      - run: npm run check:route-guard-membership
-      - run: npm run check:test-discovery
-      - run: npm run check:test-runner-api
-      # Guards tap.testFiles drift: a covering unit test absent from stryker.conf.json
-      # tap.testFiles makes its module's mutants survive on a cold nightly-mutation run,
-      # false-failing the blocking mutationScore ratchet. See check-mutation-test-coverage.mjs.
-      - run: npm run check:mutation-test-coverage
-      - run: npm run check:any-budget:t11
-      # Build-scope guard: fails if worktrees/cruft leak into the tsconfig include
-      # scope (would OOM `next build`). Instant. See incident 2026-06-25 / #5031.
-      - run: npm run check:build-scope
-      # Pack-policy (unexpected-files allowlist) WITHOUT a build — catches a stray file
-      # leaking into the npm tarball (v3.8.36: 6 ops bin/*.sh) per-PR instead of only on
-      # the release PR's heavy Package Artifact job.
-      - run: npm run check:pack-policy
-      # Complexity + cognitive-complexity: ONE ESLint walk (both baselines still
-      # enforced separately by ruleId). Avoids two cold tree walks on fast-path.
-      - run: npm run check:complexity-ratchets
-      # ── G0 (trilho .50): gates do trilho A que faltavam no trilho B ──────────────
-      # The god-file refactor happens in PRs→release/**; without these, the release
-      # rail never sees a new import cycle, dead code, duplication or a security
-      # regression until the release PR to main. Deliberately NOT brought here:
-      # bundle-size (self-skips without a build — this rail's build job is advisory
-      # and uploads nothing, so it would be dead configuration) and the coverage
-      # run (fast-unit already runs the full suite; the coverage ratchet stays on
-      # the main rail via --allow-missing in lint-guard).
-      - run: npm run check:cycles
-      - run: npm run check:lockfile
-      - name: Duplication ratchet
-        run: npm run check:duplication
-      - name: Dead-code ratchet (knip)
-        run: npm run check:dead-code
-      - name: Type coverage ratchet
-        run: npm run check:type-coverage
-      - name: Compression budget ratchet
-        run: npm run check:compression-budget
       # Security scanners — same hardened install as ci.yml quality-extended
       # (gh release download = authenticated, 5000 req/hr; curl to api.github.com
       # is rate-limited to 60/hr and silently no-ops when throttled). The blocking
@@ -251,30 +194,63 @@ jobs:
           "$HOME/.local/bin/osv-scanner" --version || true
           "$HOME/.local/bin/oasdiff" --version || true
           zizmor --version || true
-      - name: Secret scan (gitleaks, ratchet, blocking)
-        run: npm run check:secrets -- --ratchet
-      - name: Vulnerability ratchet (osv-scanner, ratchet, blocking)
-        run: npm run check:vuln-ratchet -- --ratchet
-      - name: Workflow lint (actionlint+zizmor, ratchet, blocking)
-        run: npm run check:workflows -- --ratchet
-      # BASE_REF is read by the script from the env (never interpolated into a
-      # shell body) — workflow-injection-safe. actions/checkout fetches remote
-      # refs, not a local branch named github.base_ref, so prefix origin/ or this
-      # gate self-skips every PR with reason=base-unresolved.
-      - name: OpenAPI breaking-change (oasdiff, ratchet, blocking)
+      # Quality gates (all, non-fail-fast) — #8542: replaces 17 bare check:* steps,
+      # 6 G0 gates, 4 ratchet gates, and 3 typecheck steps with a single aggregation
+      # step. Each gate runs in a loop with ::group::; failures are collected and
+      # reported at the end. set -uo pipefail (NOT set -e) so one failing gate does
+      # not abort the job and mask every later gate. Release-added gates are folded
+      # in: open-sse typecheck (#8781) and file-size base-relative mode (#8522).
+      - name: Quality gates (all, non-fail-fast)
         env:
+          # #8522: base-relative file-size mode on PR events — inherited drift (base
+          # already over frozen cap) must not red an innocent PR. Unset on
+          # workflow_dispatch (no PR base) → absolute comparison.
+          PR_BASE_SHA: ${{ github.event.pull_request.base.sha }}
           BASE_REF: ${{ github.base_ref && format('origin/{0}', github.base_ref) || '' }}
-        run: npm run check:openapi-breaking -- --ratchet
-      - name: Typecheck (core)
-        run: npm run typecheck:core
-      # #7033: dashboard-scoped typecheck gate — src/app/(dashboard) TSX is not
-      # covered by typecheck:core's curated allowlist. See check-dashboard-typecheck.mjs.
-      - name: Typecheck (dashboard)
-        run: npm run check:dashboard-typecheck
-      # #8781: open-sse workspace typecheck gate — the workspace imports @/ which
-      # escapes to src/ via undeclared path aliases. See check-open-sse-typecheck.mjs.
-      - name: Typecheck (open-sse)
-        run: npm run check:open-sse-typecheck
+        run: |
+          set -uo pipefail
+          gates=(
+            provider-consistency fetch-targets deps file-size error-helper
+            migration-numbering public-creds db-rules known-symbols
+            route-guard-membership test-discovery test-runner-api
+            mutation-test-coverage any-budget:t11 build-scope pack-policy
+            complexity-ratchets
+            cycles lockfile duplication dead-code type-coverage compression-budget
+            # #8781: open-sse workspace typecheck gate — the workspace imports @/ which
+            # escapes to src/ via undeclared path aliases. See check-open-sse-typecheck.mjs.
+            open-sse-typecheck
+          )
+          ratchet_gates=(
+            secrets vuln-ratchet workflows openapi-breaking
+          )
+          failed=()
+          for g in "${gates[@]}"; do
+            echo "::group::check:$g"
+            # #8522: file-size is base-relative on PR events (compare against
+            # max(frozen, base)) so inherited drift doesn't red an innocent PR;
+            # workflow_dispatch (no PR base) falls back to absolute comparison.
+            if [ "$g" = "file-size" ] && [ -n "${PR_BASE_SHA:-}" ]; then
+              npm run "check:$g" -- --base-ref "$PR_BASE_SHA" || failed+=("$g")
+            else
+              npm run "check:$g" || failed+=("$g")
+            fi
+            echo "::endgroup::"
+          done
+          for g in "${ratchet_gates[@]}"; do
+            echo "::group::check:$g (ratchet)"
+            npm run "check:$g" -- --ratchet || failed+=("$g")
+            echo "::endgroup::"
+          done
+          echo "::group::typecheck:core"
+          npm run typecheck:core || failed+=("typecheck:core")
+          echo "::endgroup::"
+          echo "::group::check:dashboard-typecheck"
+          npm run check:dashboard-typecheck || failed+=("check:dashboard-typecheck")
+          echo "::endgroup::"
+          if (( ${#failed[@]} )); then
+            printf '::error::%d gate(s) failed: %s\n' "${#failed[@]}" "${failed[*]}"
+            exit 1
+          fi
       # WS4.2 (v3.8.49 plan): TypeScript 7 native-compiler SHADOW — advisory only.
       # TS7 went GA 2026-07-08 with 8-12x type-check speedups; its Compiler API only
       # arrives in 7.1, so typescript-eslint / type-coverage / Stryker stay on 6.x

.gitignore

@@ -265,3 +265,7 @@ docker-compose.yml.bak
 # output/**/.playwright-cli/ (covered above), but anything written directly to
 # output/ would otherwise show up as untracked.
 /output/
+
+# _tasks e um repo git SEPARADO (ver AGENTS.md). _tasks/ (com barra) NAO ignora um
+# SYMLINK _tasks; /_tasks (ancorado) cobre symlink/dir na raiz (incidente 2026-08-08).
+/_tasks

@omniroute/opencode-plugin/src/index.ts

@@ -4403,11 +4403,11 @@ export function buildStaticProviderEntry(
       entry.release_date = raw.release_date;
     }
 
-    // OC's static-catalog reader parses each key on `/` and rejects the
-    // entire provider block if ANY key resolves to a parsed providerID that
-    // has no corresponding provider block. So bare keys (no `/`) MUST be
-    // prefixed with the resolved providerId. Already-prefixed keys
-    // (e.g. `cc/claude-opus-4-7`) are left as-is to avoid double-prefixing.
+    // #9175: OC's `getModel` looks the model up by BARE id — the part after
+    // the first `/` in the user's request — so a dict key with an embedded
+    // provider prefix (`<providerId>/<raw-id>`) is unreachable. Keys are the
+    // raw id verbatim; ids that already contain `/` (e.g. `cc/claude-opus-4-7`)
+    // keep it because the slash is part of the upstream model id itself.
     models[raw.id] = entry;
   }
 

@omniroute/opencode-plugin/tests/config-shim.test.ts

@@ -227,7 +227,7 @@ test("config: with valid auth.json + apiKey + baseURL → mutates input.provider
   // Stripped per-model shape: name + cap flags + modalities + (optional)
   // cost. OC's SDK static schema accepts only `limit.{context,output}` —
   // `limit.input` is NOT in the SDK shape and gets dropped silently.
-  const claude = entry.models["opencode-omniroute/claude-sonnet-4-6"];
+  const claude = entry.models["claude-sonnet-4-6"];
   assert.ok(claude, "claude model surfaced");
   assert.equal(claude.name, "claude-sonnet-4-6");
   assert.equal(claude.attachment, true);
@@ -248,7 +248,7 @@ test("config: with valid auth.json + apiKey + baseURL → mutates input.provider
 
   // Combo surfaces under bare key + LCD'd
   // (gemini's reasoning=false → combo reasoning=false).
-  const combo = entry.models["omniroute/claude-tier"];
+  const combo = entry.models["claude-tier"];
   assert.ok(combo, "combo surfaced under bare key");
   assert.equal(combo.name, "Claude Tier");
   assert.equal(combo.reasoning, false, "LCD: any member reasoning=false → combo reasoning=false");
@@ -471,10 +471,10 @@ test("config: combos fetcher throws → emit models-only catalog (no combos in m
   assert.ok(entry);
   const ids = Object.keys(entry.models).sort();
   assert.deepEqual(ids, [
-    "opencode-omniroute/claude-sonnet-4-6",
-    "opencode-omniroute/gemini-3-flash",
+    "claude-sonnet-4-6",
+    "gemini-3-flash",
   ]);
-  assert.equal(entry.models["omniroute/claude-tier"], undefined, "no combo entry");
+  assert.equal(entry.models["claude-tier"], undefined, "no combo entry");
   assert.ok(
     logger.entries.some((e) => String(e[0]).includes("/api/combos fetch failed")),
     "combos-fetch breadcrumb emitted"
@@ -723,7 +723,7 @@ test("buildStaticProviderEntry: stripped per-model shape matches sibling @omniro
   }
 
   // Sanity: claude entry has all expected stripped fields.
-  const claude = block.models["opencode-omniroute/claude-sonnet-4-6"];
+  const claude = block.models["claude-sonnet-4-6"];
   assert.equal(typeof claude.name, "string");
   assert.equal(typeof claude.attachment, "boolean");
   assert.equal(typeof claude.reasoning, "boolean");
@@ -748,8 +748,8 @@ test("buildStaticProviderEntry: hidden combos are excluded", () => {
     "https://or.example/v1",
     "sk-test"
   );
-  assert.equal(block.models["omniroute/claude-tier"], undefined);
-  assert.ok(block.models["opencode-omniroute/claude-sonnet-4-6"]);
+  assert.equal(block.models["claude-tier"], undefined);
+  assert.ok(block.models["claude-sonnet-4-6"]);
 });
 
 // ────────────────────────────────────────────────────────────────────────────
@@ -765,7 +765,7 @@ test("buildStaticProviderEntry: emits modalities.input from raw.input_modalities
     "https://or.example/v1",
     "sk-test"
   );
-  const claude = block.models["opencode-omniroute/claude-sonnet-4-6"];
+  const claude = block.models["claude-sonnet-4-6"];
   assert.deepEqual(claude.modalities?.input, ["text", "image"]);
   assert.deepEqual(claude.modalities?.output, ["text"]);
 });
@@ -779,7 +779,7 @@ test("buildStaticProviderEntry: never emits limit.input (OC SDK rejects it)", ()
     "https://or.example/v1",
     "sk-test"
   );
-  const claude = block.models["opencode-omniroute/claude-sonnet-4-6"];
+  const claude = block.models["claude-sonnet-4-6"];
   assert.equal((claude.limit as Record<string, unknown>).input, undefined);
   assert.equal(typeof claude.limit?.context, "number");
   assert.equal(typeof claude.limit?.output, "number");
@@ -807,7 +807,7 @@ test("buildStaticProviderEntry: emits cost when enrichment carries pricing", ()
     "sk-test",
     enrichment
   );
-  const claude = block.models["opencode-omniroute/claude-sonnet-4-6"];
+  const claude = block.models["claude-sonnet-4-6"];
   assert.equal(claude.cost?.input, 3);
   assert.equal(claude.cost?.output, 15);
   assert.equal(claude.cost?.cache_read, 0.3);
@@ -828,8 +828,8 @@ test("buildStaticProviderEntry: emits release_date when raw carries it; omits wh
     "https://or.example/v1",
     "sk-test"
   );
-  assert.equal(block.models["opencode-omniroute/claude-with-date"].release_date, "2026-02-19");
-  assert.equal(block.models["opencode-omniroute/gemini-3-flash"].release_date, undefined);
+  assert.equal(block.models["claude-with-date"].release_date, "2026-02-19");
+  assert.equal(block.models["gemini-3-flash"].release_date, undefined);
 });
 
 test("buildStaticProviderEntry: combo modalities = intersection of members (LCD)", () => {
@@ -858,7 +858,7 @@ test("buildStaticProviderEntry: combo modalities = intersection of members (LCD)
     "https://or.example/v1",
     "sk-test"
   );
-  const combo = block.models["omniroute/mixed-tier"];
+  const combo = block.models["mixed-tier"];
   assert.ok(combo, "combo emitted under slug key");
   // claude has text+image, text-only has text → intersection drops image.
   assert.deepEqual(combo.modalities?.input, ["text"]);
@@ -967,10 +967,10 @@ test("config: enrichment fetched + name overlaid on raw-model entries", async ()
     "opencode-omniroute"
   ];
   assert.ok(entry);
-  assert.equal(entry.models["opencode-omniroute/claude-sonnet-4-6"].name, "Claude Sonnet 4.6");
-  assert.equal(entry.models["opencode-omniroute/gemini-3-flash"].name, "Gemini 3 Flash");
+  assert.equal(entry.models["claude-sonnet-4-6"].name, "Claude Sonnet 4.6");
+  assert.equal(entry.models["gemini-3-flash"].name, "Gemini 3 Flash");
   // Combo names still come from /api/combos — enrichment overlay does NOT touch combos.
-  assert.equal(entry.models["omniroute/claude-tier"].name, "Claude Tier");
+  assert.equal(entry.models["claude-tier"].name, "Claude Tier");
   assert.equal(enrichmentFetcher.callCount(), 1);
 });
 
@@ -1000,7 +1000,7 @@ test("config: features.enrichment=false skips enrichment fetch + keeps raw-id na
   assert.ok(entry);
   assert.equal(enrichmentFetcher.callCount(), 0, "enrichment fetch suppressed by feature flag");
   assert.equal(
-    entry.models["opencode-omniroute/claude-sonnet-4-6"].name,
+    entry.models["claude-sonnet-4-6"].name,
     "claude-sonnet-4-6",
     "raw id retained"
   );
@@ -1027,7 +1027,7 @@ test("config: enrichment fetcher throws → soft-fail (warn + raw-id static cata
   ];
   assert.ok(entry, "static block still published on enrichment failure");
   assert.equal(
-    entry.models["opencode-omniroute/claude-sonnet-4-6"].name,
+    entry.models["claude-sonnet-4-6"].name,
     "claude-sonnet-4-6",
     "raw id retained"
   );
@@ -1229,11 +1229,11 @@ test("config: diskCache hydrates stale snapshot when /v1/models throws", async (
     "opencode-omniroute"
   ];
   assert.ok(
-    entry.models["opencode-omniroute/claude-sonnet-4-6"],
+    entry.models["claude-sonnet-4-6"],
     "stale snapshot hydrated into static block"
   );
   assert.equal(
-    entry.models["opencode-omniroute/claude-sonnet-4-6"].name,
+    entry.models["claude-sonnet-4-6"].name,
     "Claude Sonnet 4.6 (cached)",
     "stale enrichment also reused"
   );
@@ -1281,7 +1281,7 @@ test("config: cached rawEnrichment from earlier provider hook is reused (no refe
   const entry = (input as { provider: Record<string, OmniRouteStaticProviderEntry> }).provider[
     "opencode-omniroute"
   ];
-  assert.equal(entry.models["opencode-omniroute/claude-sonnet-4-6"].name, "Claude Sonnet 4.6");
+  assert.equal(entry.models["claude-sonnet-4-6"].name, "Claude Sonnet 4.6");
 });
 
 // ─────────────────────────────────────────────────────────────────────
@@ -1332,12 +1332,12 @@ test("config: providerTag (default-on) prepends '<provider> - ' to enriched raw-
   ];
   assert.ok(entry);
   assert.equal(
-    entry.models["opencode-omniroute/claude-sonnet-4-6"].name,
+    entry.models["claude-sonnet-4-6"].name,
     "Claude - Claude Sonnet 4.6"
   );
-  assert.equal(entry.models["opencode-omniroute/gemini-3-flash"].name, "Gemini - Gemini 3 Flash");
+  assert.equal(entry.models["gemini-3-flash"].name, "Gemini - Gemini 3 Flash");
   // Combos stay untouched — `Combo: ` prefix already conveys multi-upstream.
-  assert.equal(entry.models["omniroute/claude-tier"].name, "Claude Tier");
+  assert.equal(entry.models["claude-tier"].name, "Claude Tier");
 });
 
 test("config: providerTag=false suppresses the suffix", async () => {
@@ -1364,7 +1364,7 @@ test("config: providerTag=false suppresses the suffix", async () => {
     "opencode-omniroute"
   ];
   assert.equal(
-    entry.models["opencode-omniroute/claude-sonnet-4-6"].name,
+    entry.models["claude-sonnet-4-6"].name,
     "Claude Sonnet 4.6",
     "enriched name kept, provider tag suppressed"
   );
@@ -1396,7 +1396,7 @@ test("config: providerTag falls back to UPPER(alias) when providerDisplayName mi
   const entry = (input as { provider: Record<string, OmniRouteStaticProviderEntry> }).provider[
     "opencode-omniroute"
   ];
-  assert.equal(entry.models["opencode-omniroute/claude-sonnet-4-6"].name, "CC - Claude Sonnet 4.6");
+  assert.equal(entry.models["claude-sonnet-4-6"].name, "CC - Claude Sonnet 4.6");
 });
 
 test("config: providerTag skipped entirely when neither providerDisplayName nor providerAlias set", async () => {
@@ -1423,7 +1423,7 @@ test("config: providerTag skipped entirely when neither providerDisplayName nor
   const entry = (input as { provider: Record<string, OmniRouteStaticProviderEntry> }).provider[
     "opencode-omniroute"
   ];
-  assert.equal(entry.models["opencode-omniroute/claude-sonnet-4-6"].name, "Claude Sonnet 4.6");
+  assert.equal(entry.models["claude-sonnet-4-6"].name, "Claude Sonnet 4.6");
 });
 
 test("config: providerTag is idempotent — second hook call doesn't double-suffix", async () => {
@@ -1451,7 +1451,7 @@ test("config: providerTag is idempotent — second hook call doesn't double-suff
     "opencode-omniroute"
   ];
   assert.equal(
-    entryA.models["opencode-omniroute/claude-sonnet-4-6"].name,
+    entryA.models["claude-sonnet-4-6"].name,
     "Claude - Claude Sonnet 4.6"
   );
 
@@ -1462,7 +1462,7 @@ test("config: providerTag is idempotent — second hook call doesn't double-suff
     "opencode-omniroute"
   ];
   assert.equal(
-    entryB.models["opencode-omniroute/claude-sonnet-4-6"].name,
+    entryB.models["claude-sonnet-4-6"].name,
     "Claude - Claude Sonnet 4.6"
   );
 });
@@ -1516,7 +1516,7 @@ test("buildStaticProviderEntry: nested combo-ref context is the bottleneck acros
   );
   // Pre-fix: Parent would advertise 200_000 (only raw-big counted).
   // Post-fix: Parent should advertise 8_000 (TinyCombo bottleneck).
-  const parent = block.models["omniroute/parent"];
+  const parent = block.models["parent"];
   assert.ok(parent, "Parent combo must be in the static catalog");
   assert.equal(parent.limit?.context, 8_000);
 });

@omniroute/opencode-plugin/tests/provider-id-routing.test.ts

@@ -111,7 +111,9 @@ test("#6859: createOmniRouteProviderHook end-to-end — catalog keys/providerID
 // `opencode-omniroute`. Confirmed against the issue's own curl repro
 // (`model: "opencode-omniroute/hermes-smart-stack"` → "No active
 // credentials for provider: opencode-omniroute").
-test("#7976: buildStaticProviderEntry keys bare-slug combo ids with the unprefixed omnirouteProviderId (no double OC-gate prefix)", () => {
+// #9175 tightened this further: OC's `getModel` looks models up by BARE id,
+// so combo dict keys now carry NO prefix at all (not even `omniroute/`).
+test("#7976/#9175: buildStaticProviderEntry keys combos by bare slug (no prefix at all — never the OC-gate providerId)", () => {
   const resolved = resolveOmniRoutePluginOptions({ providerId: "omniroute" });
   assert.equal(resolved.providerId, "opencode-omniroute");
   assert.equal(resolved.omnirouteProviderId, "omniroute");
@@ -131,7 +133,7 @@ test("#7976: buildStaticProviderEntry keys bare-slug combo ids with the unprefix
     "sk-test"
   );
 
-  assert.deepEqual(Object.keys(block.models), ["omniroute/hermes-smart-stack"]);
+  assert.deepEqual(Object.keys(block.models), ["hermes-smart-stack"]);
   assert.equal(
     block.models["opencode-omniroute/hermes-smart-stack"],
     undefined,

_tasks

@@ -1 +0,0 @@
-/home/diegosouzapw/dev/proxys/OmniRoute/_tasks

bin/cli/runtime/processSupervisor.mjs

@@ -52,8 +52,11 @@ export class ServerSupervisor {
     // silently, so a boot that never becomes ready looked like a dead hang with zero
     // output even at APP_LOG_LEVEL=debug. Pipe stdout too and buffer it alongside
     // stderr so a readiness timeout can surface what the child actually printed.
+    // #9156: macOS launchd cannot resolve bare "node" because its PATH is
+    // minimal. Always use process.execPath (the absolute path to the running
+    // Node.js binary) so the supervisor never depends on PATH resolution.
     this.child = spawn(
-      process.versions.bun ? process.execPath : "node",
+      process.execPath,
       process.versions.bun
         ? [this.serverPath]
         : buildNodeRuntimeArgs(process.env, this.memoryLimit, this.serverPath),

changelog.d/fixes/7754-best-free-fallback.md

@@ -0,0 +1 @@
+- test(combo): guard auto/best-free never leaks the combo name as a model (#7754)

changelog.d/fixes/8542-fix.plan.md

@@ -0,0 +1 @@
+- fix(ci): aggregate all fast-gates into non-fail-fast loop so one red gate no longer masks later gates (#8542)

changelog.d/fixes/8847-bun-prebuilds.md

@@ -0,0 +1 @@
+- fix(build): include better-sqlite3 prebuilds in standalone bun bundle

changelog.d/fixes/8906-quota-pool-combo-cleanup.md

@@ -0,0 +1 @@
+- **fix(quota):** Deleting a quota pool now removes its scoped managed combos without racing in-flight pool mutations ([#8906](https://github.com/diegosouzapw/OmniRoute/pull/8906)) — thanks @xiaoyaner0201

changelog.d/fixes/8951-fix.plan.md

@@ -0,0 +1 @@
+- fix(github): add targetFormat to GPT-5.6 Sol/Terra/Luna models (#8951)

changelog.d/fixes/8997-gpt56-max-reasoning.plan.md

@@ -0,0 +1 @@
+- **fix(translator):** the Responses-to-Chat promotion path called `normalizeResponsesReasoningEffort` without the model argument, so GPT-5.6 Sol/Terra/Luna requests with `reasoning.effort: "max""` were downgraded to `"xhigh"`. The model is now threaded through, preserving `max` for GPT-5.6 while keeping the legacy downgrade for older models ([#8997](https://github.com/diegosouzapw/OmniRoute/pull/8997))

changelog.d/fixes/9034-fix.plan.md

@@ -0,0 +1 @@
+- fix(api): use configured prefix instead of raw node UUID for alias-backed model id in /v1/models (#9034)

changelog.d/fixes/9057-fix.plan.md

@@ -0,0 +1 @@
+- fix(api): auto/* routing aliases bypass API-key allowedConnections/disableNonPublicModels (#9057)

changelog.d/fixes/9096-fix-audio-speech-transcriptions-translations-provider-nodes.plan.md

@@ -0,0 +1 @@
+- fix(providers): admit audio-speech/audio-transcriptions apiType in audio route provider-node filters (#9096)

changelog.d/fixes/9134-fix.plan.md

@@ -0,0 +1 @@
+- fix(providers): resolve combo names in audio transcriptions route so /v1/models stays honest (#9134)

changelog.d/fixes/9140-fix.plan.md

@@ -0,0 +1 @@
+- fix(vscode): allow built-in auto-routing models in VS Code model filter (#9140)

changelog.d/fixes/9142-fix.plan.md

@@ -0,0 +1 @@
+- fix(background): detect Anthropic top-level system prompts for background task detection (#9142)

changelog.d/fixes/9156-macos-autostart-execpath.md

@@ -0,0 +1 @@
+- fix(cli): use process.execPath for macOS launchd autostart

changelog.d/fixes/9160-fix.plan.md

@@ -0,0 +1 @@
+- fix(model-discovery): ingest capabilities.effort_tiers for synced models (#9160)

changelog.d/fixes/9168-streamed-responses-tool-null.plan.md

@@ -0,0 +1 @@
+- fix(translator): buffer and normalize upstream tool-call argument deltas so optional null values are stripped before reaching the client (#9168)

changelog.d/fixes/9177-fix.plan.md

@@ -0,0 +1 @@
+- fix(translator): avoid double-normalizing tool names in Gemini-to-Claude response path (#9177)

changelog.d/fixes/9269-fix.plan.md

@@ -0,0 +1 @@
+- **fix(classify429):** add missing `have exhausted their quota` pattern so the synthetic 429 from auth.ts is recognized as quota exhaustion, preventing the combo loop from burning retries against the same provider instead of falling back to a healthy one ([#9269](https://github.com/diegosouzapw/OmniRoute/issues/9269))

changelog.d/fixes/9305-fix.plan.md

@@ -0,0 +1 @@
+- fix(sse): broaden OMNIROUTE_SSE_COMMENTS to accept 'false','0','no' and gate metadata comment emission (#9305)

changelog.d/fixes/9306-fix.plan.md

@@ -0,0 +1 @@
+- fix(lmarena): encode SSE stream chunks as Uint8Array to prevent TextDecoder TypeError (#9306)

changelog.d/fixes/9486-claude-400-quota.md

@@ -0,0 +1 @@
+- fix(providers): classify 400 out of extra usage as quota_exhausted for Anthropic OAuth

changelog.d/fixes/9615-docker-colocate-partial-trace.md

@@ -0,0 +1 @@
+- **fix(docker):** standalone co-location now completes packages Next's file tracing materialized partially (package.json without its `main` payload) — unblocks the Docker Hub publish that failed on every v3.8.50 push with `Cannot find module '@atjsh/llmlingua-2/dist/index.js'` ([#9615](https://github.com/diegosouzapw/OmniRoute/pull/9615))

changelog.d/fixes/9623-connection-test-recovery.md

@@ -0,0 +1 @@
+- fix(resilience): failed connection test now sets a short cooldown so connections recover after transient outages (#9623)

changelog.d/fixes/9624-telemetry-cleanup-wiring.md

@@ -0,0 +1 @@
+- fix(db): wire telemetry cleanup scheduler in Next.js startup path (#9624)

changelog.d/fixes/9625-domain-cost-ms.md

@@ -0,0 +1 @@
+- fix(db): align domain_cost_history cleanup cutoff with millisecond column (#9625)

changelog.d/fixes/9626-playground-errors.md

@@ -0,0 +1 @@
+- fix(playground): surface provider model loading errors and offer retry (#9626)

changelog.d/fixes/9633-npm-build-files.md

@@ -0,0 +1 @@
+- fix(build): add build-next-isolated.mjs sibling imports to package.json files array

changelog.d/fixes/9737-memory-id-route-backend.md

@@ -0,0 +1 @@
+- Fixed `GET`/`PUT`/`DELETE /api/memory/[id]` always failing with a 500 (`Primary backend "sqlite" not registered`) when the route was reached before any other memory endpoint in the same process.

changelog.d/fixes/9737-route-body-validation.md

@@ -0,0 +1 @@
+- Replaced hand-rolled body type checks with Zod validation in the plugins marketplace install route and the three Dario admin routes, restoring the `t06:route-validation` gate (Hard Rule #7).

changelog.d/fixes/9737-vietnamese-locale-parity.md

@@ -0,0 +1 @@
+- Restore Vietnamese locale parity after the entity-normalization sync dropped Radar, provider, and mini-playground messages.

changelog.d/fixes/9826-command-code-responses-usage.md

@@ -0,0 +1 @@
+- **fix(executors):** preserve Command Code usage in `/v1/responses` streams so Codex clients receive real input, output, cache, and reasoning token counts ([#9826](https://github.com/diegosouzapw/OmniRoute/pull/9826)) — thanks @MrShitFox

changelog.d/fixes/9828-codex-redundant-oneof-enum.md

@@ -0,0 +1 @@
+- **fix(executors):** prevent intermittent Codex `upstream_empty_response` errors for tool schemas that combine `oneOf` const branches with a matching sibling `enum` by removing only the semantically redundant `oneOf`; bare, narrowing, non-matching, and type-discriminated `oneOf` schemas remain unchanged. ([#9828](https://github.com/diegosouzapw/OmniRoute/pull/9828))

changelog.d/fixes/9834-cursor-selected-image-blobid.md

@@ -0,0 +1 @@
+- **fix(cursor):** SelectedImage uses `blobIdWithData` + session blobStore, with JPEG soft-cap prep via sharp ([#9834](https://github.com/diegosouzapw/OmniRoute/pull/9834)) — thanks @yansigit

changelog.d/maintenance/9614-opencode-plugin-bare-key-suite.md

@@ -0,0 +1 @@
+- **test(cli):** OpenCode plugin suite realigned to the bare-key static-catalog contract from #9178/#9175 (21 tests were red on every opencode-plugin CI run; 287/287 after) ([#9614](https://github.com/diegosouzapw/OmniRoute/pull/9614))

changelog.d/maintenance/9738-deadcode-radar-referrals.md

@@ -0,0 +1 @@
+- Removed the unused `RadarReferrals` type export left by the radar referral-links feature (#9697), returning the dead-code ratchet to its 227 baseline.

changelog.d/maintenance/9839-v3850-quality-ratchets.md

@@ -0,0 +1 @@
+- Reconcile the final v3.8.50 bundle-size and file-size ratchets against the measured release tip, preserving exact direction-down ceilings and their source attribution.

config/quality/dependency-allowlist.json

@@ -96,6 +96,7 @@
     "node-machine-id",
     "omniglyph",
     "open",
+    "opencode-ai",
     "ora",
     "parse5",
     "pino",

config/quality/eslint-suppressions.json

@@ -81,7 +81,7 @@
   },
   "open-sse/handlers/search.ts": {
     "@typescript-eslint/no-explicit-any": {
-      "count": 34
+      "count": 33
     }
   },
   "open-sse/handlers/sseParser.ts": {

config/quality/file-size-baseline.json

@@ -1,5 +1,5 @@
 {
-  "_rebaseline_2026_08_02_9259_rolling_rpm": "PR #9259 (issue #8733) own growth: open-sse/services/rateLimitManager.ts baseline 1060->1167 (+107; final source 1153). The existing withRateLimit chokepoint now composes process-local rolling RPM leases with Bottleneck admission, releases pre-dispatch leases on queue timeout/abort/connection disable, preserves caller abort reasons, and wires 429/header state into the extracted rollingRpmGate.ts. The remaining growth is irreducible lifecycle wiring at the dispatch boundary plus the real watchdog test hooks needed to verify queued-wedge recovery; moving it further would obscure lease ownership and Bottleneck cleanup. Covered by the focused rate-limit manager/sliding-window suite (33/33); distributed multi-instance coordination remains explicitly out of scope.",
+  "_rebaseline_2026_08_09_9296_adobe_media_capabilities": "PR #9296 (artickc, fix/adobe-firefly-model-capabilities) own growth: src/app/api/v1/models/catalog.ts 1590->1597 (+7). The image and video catalog serializers now expose the already-normalized Adobe Firefly discovery capability data (media_capabilities, plus the existing video modality/size fields) at their only response-emission chokepoints. The discovery parser and capability normalization remain in open-sse/services/adobeFireflyModels.ts; extracting these seven serialization fields would obscure the catalog contract. Covered by tests/unit/adobe-firefly.test.ts and tests/unit/image-upscale.test.ts.",
   "_rebaseline_2026_07_24_8470_hyperagent_sticky_thread": "PR #8470 (artickc, fix/hyperagent-tool-loop-thread-sticky) own growth: open-sse/executors/hyperagent.ts 936->1025 (wc -l; check-file-size.mjs counts via split(\"\\n\").length so the gate sees 937->1026, +89, crosses the 1000 cap). Fixes a real bug where a reverse-conversion proxy (text-Intent/JSON to Claude Code native tool_calls) rewrites assistant messages between agentic tool-loop turns, breaking HyperAgent’s conversation-prefix fingerprint and cold-starting the thread mid tool-loop. Adds Anthropic tool_use/tool_result flattening to extractMessageText() plus a new rootUserFingerprint()/root-key lookup tier in resolveHyperAgentThreadBinding()/storeHyperAgentThreadAfterTurn() so the thread stays sticky across the tool loop. Cohesive additions inside the existing single-file executor; not extractable without splitting the executor mid-request-flow. Covered by tests/unit/executor-hyperagent.test.ts (19/19, +5 new cases for tool_result/tool_use flattening + root-key stickiness). Pre-merge review flagged a cross-conversation root-key collision risk (tracked in the PR’s own mandatory pre-merge checklist, not yet addressed) — unrelated to this file-size ratchet, tracked separately by /fix-prs.",
   "_rebaseline_2026_07_25_8494_capability_filter_fail_closed": "PR #8494 (fix/capability-filters-fail-closed, #8488) own growth: open-sse/services/combo.ts 3640->3693 (+53) adds a fail-closed guard after filterTargetsByRequestCompatibility() — when every eligible target is excluded by request-capability filtering (vision/tools/etc) instead of quota/health, the combo now returns an explicit `capability_mismatch` 400 (describeCapabilityFilterExhaustion, imported from combo/comboStructure.ts) rather than silently falling through to a generic no-targets error, plus a `compatFilterFailOpen` escape hatch (combo config OR settings) mirrored at both the main/auto and round-robin call sites for symmetry. combo/comboStructure.ts (previously under cap, un-frozen) grows 794->918 (+124) — new home for describeCapabilityFilterExhaustion + providerSupportsEmulatedToolCalling (#5240 emulated tool-calling exemption so fail-closed does not regress prompt-emulation-only combos like all-chatgpt-web). Irreducible orchestration wiring at the existing filter chokepoint (same precedent as #7301's universal-cooldown-retry generalization). Companion test tests/unit/combo-routing-engine.test.ts 3409->3449 (+40, fail-closed/fail-open coverage across both call sites) also rebaselined. Covered by tests/unit/8488-capability-filter-fail-closed.test.ts (new) + 95/95 passing across both files. Structural shrink of combo.ts tracked in #3501.",
   "_rebaseline_2026_07_25_8499_ts7_result_union_predicates": "PR #8499 (backryun, chore/ts7-types-executor-scattered) own growth: muse-spark-web.ts 1396->1405 (+9, irreducible). Under this workspace's `strictNullChecks: false`, the boolean-literal discriminant on `GraphqlResult` (`{ ok: true } | { ok: false; error: string }`) narrows the positive `.ok===true` branch but leaves `!result.ok` at the full union under TS7, making `.error` unreachable to the checker at the two call sites (warmup, mode-switch). Fixed by adding a single `isGraphqlFailure()` type-predicate helper (doc comment + 3-line body) reused at both call sites instead of duplicating the predicate inline — not extractable to a shared module without splitting a single-file executor's local narrowing helper out of its own file. Covered by the existing muse-spark-web executor test suite (no behavior change, pure narrowing fix).",
@@ -14,7 +14,6 @@
   "_rebaseline_2026_07_19_7546_ghe_copilot_route": "PR #7546 (GHE Copilot OAuth provider) own growth: oauth/[provider]/[action]/route.ts 960->963 (gate units, +3 = ghe-copilot device-code wiring at the existing multi-provider device-code branch — reading + HTTPS-validating the gheUrl search param (isValidGheUrl guards at both raw entry points, security-review hardening, 963->970), adding ghe-copilot to the no-PKCE provider set, and building the provider config override / threading gheUrl through poll->postExchange extraData). Mirrors the existing kiro/amazon-q startUrl override pattern right above it in the same branch; cohesive with the existing device-code dispatch chokepoint, not separately extractable without splitting a single provider-switch mid-branch. Frozen so can only shrink; structural shrink tracked in #3501.",
   "_rebaseline_2026_07_19_6846_nvidia_concurrency_gate": "Issue #6846 Phase 1 (nvidia NIM local RPM budget + per-model lockout + per-connection concurrency cap) own growth: open-sse/executors/default.ts 877->890 (+13 = the irreducible call-site wiring at DefaultExecutor.execute(), the only place nvidia requests dispatch through — the existing session-pool body was extracted verbatim into a new private executeWithSessionPool() so the outer execute() can wrap it in the nvidia concurrency-gate acquire/finally-release). All actual gating logic (semaphore key + cap resolution) lives in the new leaf open-sse/executors/default/nvidiaConcurrencyGate.ts (not frozen, well under cap). Covered by tests/unit/nvidia-quota-phase1.test.ts.",
   "_rebaseline_2026_07_18_v3849_provider_detail_wiring": "Merge campaign R2/R3 (2026-07-18): three authorized PRs each add irreducible call-site wiring to ProviderDetailPageClient.tsx — #7360 +5 (ProviderQuotaVisibilityToggle render, component extracted), #7419 +4 (NoAuthProviderControls wiring), #7062 +3 (Dahl provider hook) = 786->798. All three follow the extracted-component pattern (AgentrouterConsoleFields precedent); the frozen file only takes the wiring. Structural shrink tracked in #3501.",
-  "_rebaseline_2026_07_25_dario_upstream_proxy_selector": "PR #8523 (Dario embedded service): upstream-proxy mode selector replaces the binary CLIProxyAPI toggle with Native/CLIProxyAPI/Dario/Fallback + a fallback-backend picker. ProviderDetailPageClient.tsx 798->804 (+6, new hook fields threaded through to ConnectionsListPanel), ConnectionRow.tsx 942->958 (+16, the mode <select> + conditional fallback-backend <select> replacing a single pill button), useProviderConnections.ts 954->986 (+32, upstreamProxyMode/upstreamProxyFallbackBackend state + handleSetUpstreamProxyMode, handleToggleCliproxyapiMode kept as a thin backward-compat wrapper for the existing hook-shape test). All additive UI/state for the new modes — no unrelated refactor.",
   "_rebaseline_2026_07_18_pr7653_chat_tracker_import": "PR #7653 merge-interaction growth: release moved chat.ts to its 1796 cap while this PR adds the single side-effect import 'quotaTrackersBatch.ts' (line 130) — chat.ts IS the canonical quota-fetcher registration point (codex/bailian/deepseek/openrouter/opencode/generic all import+register there), so the +1 is irreducible call-site wiring. 1796->1797. Covered by tests/unit/{agentrouter,v0,freemodel}-quota-fetcher.test.ts.",
   "_rebaseline_2026_07_17_pr7653_agentrouter_console_fields": "PR #7653 own growth (missing acceptance criterion: the AgentRouter quota tracker (#6850) read providerSpecificData.consoleApiKey/newApiUserId but neither field had dashboard UI for provider agentrouter — consoleApiKey was gated to bailian-coding-plan only and newApiUserId had zero UI). AddApiKeyModal.tsx 961->967 (+6) and EditConnectionModal.tsx 1278->1286 (+8) = import + a single <AgentrouterConsoleFields .../> render call plus the newApiUserId formData init field. The actual Input rendering (both consoleApiKey reuse + the new newApiUserId field) was EXTRACTED into a new leaf src/app/(dashboard)/dashboard/providers/[id]/components/modals/AgentrouterConsoleFields.tsx (48 LOC, <cap), mirroring the QuotaScrapingFields.tsx / GlmTeamQuotaFields.tsx precedent (#6351) so the frozen modals only carry the irreducible call-site wiring. Persist logic lives in connectionProviderSpecificData.ts (not frozen). Covered by tests/unit/agentrouter-connection-modal-fields.test.ts.",
   "_rebaseline_2026_07_17_v3849_6842_free_window_wiring": "PR #7651 (openrouter :free-window quota tracking) follow-up: the counter shipped built but never wired into the request pipeline, so combos kept spending guaranteed-429 requests on exhausted free-tier targets. Own growth: src/sse/services/auth.ts 2461->2462 (+1, irreducible at the existing model-aware preflight chokepoint — the `provider === \"codex\"` check that forwards requestedModel into the connection arg is extended to also cover `openrouter`, one added boolean + a doc comment, offset to a single net line by dropping the now-redundant inline condition). Enforcement itself lives in open-sse/services/openrouterQuotaFetcher.ts (not frozen) and the dispatch-time record/correct hooks live in open-sse/executors/base.ts (not frozen). Covered by tests/unit/openrouter-free-window-wiring-6842.test.ts.",
@@ -159,133 +158,8 @@
   "_rebaseline_2026_06_20_1409_1294_models": "Re-baseline src/lib/db/models.ts 1184->1221: combined growth of sibling fixes #1409 (cascade-delete orphaned model aliases when a provider is removed) + #1294 (persist max_input_tokens/max_output_tokens on custom models), both adding CRUD at the existing models domain module. Cohesive db module; not extractable.",
   "_rebaseline_2026_06_20_4389_thinking_toolchoice": "Re-baseline base.ts 1387->1399 (#4389): tool_choice-forced thinking guard at the existing Claude wire-image injection chokepoint (effThinking gate avoids the Anthropic 400 when tool_choice forces a tool). Cohesive guard; structural shrink tracked in #3501.",
   "_rebaseline_2026_07_18_6979_codex_test": "PR #6979 own growth: executor-codex.test.ts 1340->1347 (+7 = generalized ensureThinkingBudget assertion added to the existing codex thinking-budget cases). antigravity-test bump 942->977 REVERTED here: #7408's test split dropped that file to 888, so this PR's +35 fits under the original 942 frozen cap.",
+  "_rebaseline_2026_07_24_8354_logs_timeline_sidebar": "PR #8354 (hartmark, feature/scrolling-log) own growth: src/shared/constants/sidebarVisibility/sections.ts 812->820 (+8, the single new logs-timeline SidebarItemDefinition entry added to LOGS_GROUP.items for the new /dashboard/logs/timeline scrolling request-timeline page). Irreducible data-literal wiring at the existing sidebar-sections chokepoint, same shape as every other item in the file; not extractable without an ad-hoc single-item exception to the file's otherwise-uniform multi-line item style.",
   "cap": 1000,
-  "frozen": {
-    "_rebaseline_2026_06_22_4644_deepseek_web_tools": "PR #4644 (BugsBag/robust deepseek-web tool-call parsing): open-sse/executors/deepseek-web.ts 1117->1125 (+8). The new agentic tool-call path emits surrounding text + reasoning before tool_calls and swaps to the dedicated deepseekWebTools.ts parser; the +8 lines are cohesive wiring at the existing transformSSE chokepoint (the parser itself lives in the new deepseekWebTools.ts file, already under cap). The PR's own fast-gate (PR->release) does not run check:file-size, so this surfaced only at release reconcile. Covered by tests/unit/deepseek-web-tools-variants.test.ts + deepseek-web-tools-execute.test.ts.",
-    "_rebaseline_2026_06_23_4712_deepseek_web_tool_results": "PR for #4712 (deepseek-web drops role:tool): open-sse/executors/deepseek-web.ts 1125->1148 (+23). messagesToPrompt() now folds role:\"tool\" results into the single-prompt transcript (recovering the tool name from the preceding assistant tool_calls by tool_call_id) instead of silently dropping them; the lines are cohesive wiring inside the existing function. Covered by tests/unit/deepseek-web-tool-result-prompt-4712.test.ts.",
-    "_rebaseline_2026_06_24_headroom_strategy": "Headroom-aware connection selection (dario technique): combo.ts 3168->3180 (+12 = a new `else if (strategy === \"headroom\")` dispatch branch in handleComboChat that delegates to orderTargetsByHeadroom + its log line, plus the import). The actual logic lives OUT of the god-file: the pure ranker rankByHeadroom/computeHeadroom is the new leaf open-sse/services/combo/headroomRanking.ts (91 LOC, <cap) and the async orderer orderTargetsByHeadroom is appended to the existing open-sse/services/combo/quotaStrategies.ts (<cap) next to its sibling reset-aware/reset-window orderers (reuses their connection-expansion machinery). headroom = 1 - max(util_5h, util_7d) from getSaturation (src/lib/quota/saturationSignals.ts), prefers the connection with the most free capacity. Only the dispatch wiring is irreducible at the existing combo strategy chokepoint (mirrors the reset-aware/reset-window/context-optimized branches); not extractable without hiding the call site. fill-first stays default; all existing strategies untouched. Covered by tests/unit/combo-headroom-ranking.test.ts (pure helper) + tests/unit/combo-headroom-strategy.test.ts (orderer, saturation injected). Structural shrink of combo.ts tracked in #3501.",
-    "_rebaseline_2026_06_24_quota_share_strategy": "Dedicated quota-share strategy (Phase 3 #9): combo.ts 3180->3190 (+10 = one new `else if (strategy === \"quota-share\")` dispatch branch in handleComboChat that delegates 100% to selectQuotaShareTarget + its log line, plus the import). All the new logic lives OUT of the god-file in two new leaves under open-sse/services/combo/: quotaShareInflight.ts (in-flight counter with TTL/lease, ~150 LOC <cap) and quotaShareStrategy.ts (per-model bucket gating via isBucketSaturated + DRR proportional to weight + P2C over in-flight, ~240 LOC <cap). Only the dispatch wiring is irreducible at the existing combo strategy chokepoint (mirrors the headroom/reset-aware/reset-window/context-optimized branches); not extractable without hiding the call site. ZERO existing strategy cases were modified — only this branch was added, and the qtSd/ combos switched from fill-first to quota-share in src/lib/quota/quotaCombos.ts. Covered by tests/unit/quota-share-strategy.test.ts (gating, DRR fairness, P2C in-flight, fail-open, activation). Structural shrink of combo.ts tracked in #3501.",
-    "_rebaseline_2026_06_24_task_aware_routing": "Task-aware routing strategy (port PR #2045, OmniRoute #4945): combo.ts 3190->3225 (+35) = one new `else if (strategy === \"task-aware\")` dispatch branch delegating 100% to selectTaskAwareTarget + its imports/log lines. All scoring/classification logic lives OUT of the god-file in the new leaf open-sse/services/taskAwareRouting.ts (553 LOC <cap). Only the dispatch wiring is irreducible at the existing combo strategy chokepoint (mirrors quota-share/headroom/reset-aware branches). ZERO existing strategy cases modified. Covered by tests/unit/combo-task-aware.test.ts (35 tests). Structural shrink of combo.ts tracked in #3501.",
-    "_rebaseline_2026_06_26_fidelity_gate_extraction": "Milestone-B fidelity-gate wiring residual: bodyToText+gateAdvance extracted to fidelityGateStep.ts (889->854, -35), but the StackOptions.fidelityGate field, the `const fidelityGate` reads at the two stacked-loop dispatch chokepoints, and the import of FidelityGateConfig are irreducible wiring that cannot leave strategySelector without an architectural refactor of the pre-existing stacked pipeline. Net: 889->854 (+6 vs the pre-Milestone-B frozen 848). Covered by tests/unit/compression/*.test.ts (940 pass).",
-    "_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.",
-    "_rebaseline_2026_06_27_5193_antigravity_basered": "Base-red (pre-existing release drift, fast-gate PR->release skips check:file-size): accountFallback.ts 1773->1777 and src/app/api/providers/[id]/test/route.ts 924->940 were already over their frozen caps on release/v3.8.39 independent of any antigravity change. Owner chose to rebaseline (keep the documented issue-reference comments #1846/#1449/#347 etc.) rather than accept the contributor comment-stripping in #5200/#5198. Reverted #5200 to restore the comments; bumped these two frozen caps to the actual base sizes. No logic change.",
-    "_rebaseline_2026_06_28_5237_impersonation_ua_refresh": "PR #5237 (refresh impersonation UAs): grok-web.ts 1871->1873 (+2), muse-spark-web.ts 1284->1302 (+18), perplexity-web.ts 1013->1032 (+19). Net semantic change in each file is a single User-Agent constant (Chrome 147->149 for grok/muse; perplexity kept at Firefox 148 to stay matched with the firefox_148 TLS profile — the contributor's 152 bump was reverted to avoid a UA-vs-JA3 mismatch, #2459). The growth is Prettier reflow that lint-staged unavoidably applies to these grandfathered long-line files the moment they are touched; not extractable. src/sse/services/auth.ts 2336->2401 in the same reconcile is #5222's antigravity-LRU-retry growth that merged via --admin without a baseline bump.",
-    "_rebaseline_2026_06_28_5243_risk_gate_prepass": "PR #5243 (compression risk-gate pre-pass) own growth: open-sse/services/compression/strategySelector.ts 854->899 (+45). The three exported entry points (applyCompression/applyStackedCompression/applyStackedCompressionAsync) become thin wrappers over pure-extracted private bodies (runCompression/runStackedCompression/runStackedCompressionAsync) so the risk-gate mask->run->restore wrapper sits strictly OUTSIDE the per-step loop — a single universal integration point. The wrapper logic itself (resolveRiskGate/withRiskGate) lives in the new riskGate/strategyWrap.ts (<cap); the residual growth is the duplicated thin-wrapper signatures + the extracted bodies' dispatch boundary, guarded by a byte-identical parity test (riskGateIntegration). Default off (DEFAULT_COMPRESSION_CONFIG unchanged). Not extractable without hiding the dispatch boundary, mirroring prior compression rebaselines. Structural shrink tracked in #3501.",
-    "_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.",
-    "_rebaseline_2026_06_29_4038_cas_guard": "PR (#4038) own growth: tokenRefresh.ts 2103->2181 (+78 = the compare-and-swap guard on the refresh persist — runWithCasGuard/getActiveCasGuard AsyncLocalStorage pair mirroring runWithOnPersist, casGuardShouldSkipPersist that rereads the row right before persisting and skips the write when a concurrent writer already rotated the refresh_token past the one presented, plus getCasGuardStats counters). Fixes the sibling-rotation-revert → token-family-revocation storm. Gated behind an active guard (opt-in; no guard => byte-identical). Wiring lives at the two persist chokepoints inside getAccessToken; the comparison reuses wasRefreshTokenRotated from refreshSerializer. Not extractable without splitting the refresh hot path.",
-    "_rebaseline_2026_06_29_5286_memoization": "PR #5286 own growth: strategySelector.ts 899->960 (+61 = the opt-in result-memoization branches in applyCompression/applyCompressionAsync — principal+determinism gate, makeMemoKey lookup/store with model+supportsVision folded into the key, recompute-with-memo-off). Default off (memoizeCompressionResults), so zero behavior change. The memo helpers live in the leaf resultMemo.ts (<cap); the chokepoint wiring here is not extractable. Structural shrink of this hot-path file tracked in #3501.",
-    "_rebaseline_2026_07_01_v3843_release_5609": "Rebaseline v3.8.43 (PR #5609 release reconciliation). DRIFT dos 109 commits do ciclo: 8 god-files existentes cresceram (ApiManagerPageClient 2983->3017, combos/page 4594->4608, AddApiKeyModal 868->869, providerPageHelpers 974->996, chat.ts 1635->1647, auth.ts 2401->2403, batchProcessor 828->915, combo.ts 3368->3387) + 2 novos acima do cap (huggingchat.ts 813, tests web-cookie-providers-new 827) + 4 test files cresceram. Modularizacao deferida (blast-radius mid-release); congelado no estado atual p/ o proximo ciclo ratchetar daqui.",
-    "_rebaseline_2026_07_02_5816_qoder": "PR #5816 (@AgentKiller45, qoder PAT via qodercli): qoderCli.ts 666->989, new-above-cap frozen (owner-approved baseline freeze). The growth is the legitimate PAT job-token exchange + quota parsing CLI transport (the pure-JS Cosy path 500'd on every PAT request); extracting the spawn/parse helpers now would just add indirection to a contributor PR mid-merge. Test frozen also raised for this PR's coverage growth: providers-page-utils.test.ts 1052->1092. Additionally clears an inherited base-red from the already-merged #5933 (codex json_schema->text.format): translator-openai-responses-req.test.ts 1097->1172 (+75 regression tests, no offending branch left). All remain frozen (cannot grow further); release captain's rebaseline-at-release supersedes.",
-    "_rebaseline_2026_07_09_6126_clinepass_dual_auth": "PR #6126 (@hajilok, dual-auth ClinePass) own growth: tokenRefresh.ts 2181->2182 (+1 = a single `case \"clinepass\":` fallthrough label added to the existing `case \"cline\":` in _getAccessTokenInternal's provider switch, so clinepass token refresh dispatches to the already-shared refreshClineToken() instead of silently falling through to the generic OAuth refresh). Irreducible 1-line switch-case wiring at the existing chokepoint; the header-building logic for the same feature was extracted to a new leaf src/shared/utils/clineAuth.ts::buildClinepassHeaders() (well under cap) to avoid growing open-sse/executors/default.ts. Covered by tests/unit/clinepass-provider.test.ts.",
-    "_rebaseline_2026_07_09_6363_kiro_external_idp": "PR #6363 (@artickc, Kiro external IdP) own growth: tokenRefresh.ts 2182->2249 (+67 = the external_idp refresh branch inside refreshKiroToken — standard public-client OAuth2 refresh_token grant against the org IdP tokenEndpoint via buildExternalIdpRefreshParams/isExternalIdpAuthMethod from the new leaf open-sse/services/kiroExternalIdp.ts, with invalid_grant/invalid_client -> unrecoverable_refresh_error mapping). Cohesive addition at the existing refreshKiroToken chokepoint. Covered by tests/unit/kiro-external-idp.test.ts.",
-    "_rebaseline_2026_07_09_6587_kiro_api_key_auth": "PR #6587 (@strangersp) own growth for Kiro long-lived API-key auth, merged onto v3.8.47 tip: openai-to-kiro.ts 890->912 (+22, auth-header selection for API-key-vs-OAuth-token connections), providerLimits.ts 998->1000 (+2, API-key auth-type branch), translator-openai-to-kiro.test.ts 1234->1257 (+23), providers-page-utils.test.ts 1109->1107 (net -2 after merging with parallel release drift; connectionMatchesProviderCard api_key coverage added), provider-validation-specialty.test.ts 2856->2980 (+124 net after merge with parallel release drift; this PR also removed the file's `@typescript-eslint/no-explicit-any` eslint-suppression entry by fixing all `any` usages, adding typed replacements). Cohesive additive feature growth, well tested; not extractable without splitting the existing chokepoints mid-merge.",
-    "_rebaseline_2026_07_09_6678_routing_strategy_9router": "#6678 (SeaXen) — 9router-parity Routing Strategy settings card + per-provider/combo sticky-round-robin override. Own growth: ProviderDetailPageClient.tsx 784->786 (single ProviderAccountRoutingCard mount + import), auth.ts 2448->2458 (providerStrategies override resolution: fallbackStrategy/stickyRoundRobinLimit per-provider cascade in getProviderCredentials). Both additive, zero unrelated refactor; new UI/logic lives in new files (ProviderAccountRoutingCard.tsx, RoutingStrategyCard.tsx, rrState.ts::resolveComboStickyRoundRobinLimit). chat.ts value below reflects the current release tip (grown by other concurrent PRs, e.g. #6640), not this PR own change.",
-    "_rebaseline_2026_07_10_6318_omp_letta": "PR #6318 (@hamsa0x7, omp+letta CLI integrations) own growth: cliTools.ts (+53 = 2 registry entries incl. omp docsUrl) and cliRuntime.ts (+18 = runtime-detection wiring for the 2 new tools). Cohesive registry/wiring growth at the existing chokepoints; scope reduced from the original 5 tools (pi/codewhale/jcode shipped separately).",
-    "_rebaseline_2026_07_10_gcf_v3_2_decode": "PR #6838 own growth: new vendored file open-sse/services/compression/engines/headroom/gcf/decode_generic.ts frozen at 880 (> 800 cap). It is the vendored GCF generic-profile decoder (spec v3.2 nested flattening plus the prototype-pollution / hasOwnProperty hardening added in this PR's Gemini review). Kept as one file faithful to upstream gcf-typescript so re-vendoring stays a clean copy rather than a re-split each cycle (sibling generic.ts/scalar.ts stay < cap; extraction would also fragment the file's frozen eslint no-explicit-any suppressions). Round-trip + prototype-pollution regression coverage in tests/unit/compression/headroom-smartcrusher.test.ts. Frozen: only shrinks from here.",
-    "_rebaseline_2026_07_12_v3847_mergeprs_tail": "v3.8.47 /merge-prs tail (owner-approved): src/lib/localDb.ts NEW>800 (799->805, +6 re-exports countFreeProxies + recordFreeProxySyncErrors/clearFreeProxySyncErrors/getFreeProxySyncErrors + FreeProxySyncErrors type for #6909 free-pool relay-repair; re-export-only per Hard Rule #2, not extractable).",
-    "_rebaseline_2026_07_15_7070_combos_memo": "PR #7070 (perf/p1-memo) own growth: src/app/(dashboard)/dashboard/combos/page.tsx 4655->4656 (+1 = React.memo wrapping of ComboCard). Covered by tests/unit/ui/combos-page-smoke.test.tsx.",
-    "_rebaseline_2026_07_18_7399_xai_oauth_modal": "PR #7399 (xAI OAuth PKCE) own growth: OAuthModal.tsx 993->998 (+5 = provider entry + PKCE flow branch wiring at the existing provider-switch chokepoint; the provider logic itself lives in src/lib/oauth/providers/xai-oauth.ts, new leaf). Third irreducible wiring bump on this modal (969->989->993->998); structural shrink tracked in #3501.",
-    "_rebaseline_2026_07_19_6636_codex_session_json": "#6636 own growth: OAuthModal.tsx 998->1030 (gate units, split(\"\\n\").length incl. trailing newline; +32 = session-JSON paste branch for handleManualSubmit plus a shared submitCodexAccessToken() helper extracted from the pre-existing bare-JWT branch, mirroring the #5203 oauthBlobSubmit.ts extraction precedent; the normalizer logic itself lives in the new src/lib/oauth/utils/codexSessionImport.ts leaf module, not here). Fourth irreducible wiring bump on this modal (969->989->993->998->1030); structural shrink tracked in #3501.",
-    "_rebaseline_2026_07_19_7546_ghe_copilot_modal": "PR #7546 (GHE Copilot OAuth provider) own growth: OAuthModal.tsx 1030->1056 (gate units). Adds a gheUrl input state, routes ghe-copilot through the existing device-code branch, and threads gheUrl into the device-code request/poll extraData at the existing provider-switch chokepoints (+~24 lines, cohesive with the same pattern as #7399/#6636). The standalone GHE enterprise-URL config step JSX (originally +31 lines inline) was extracted to the new src/shared/components/oauthModal/GheConfigStep.tsx leaf component to minimize the bump; what remains is the irreducible provider-branch wiring. Fifth bump on this modal (969->989->993->998->1030->1056); structural shrink tracked in #3501.",
-    "_rebaseline_2026_07_19_7787_ic2_localdb_reexports": "PR #7787 (IC2 raw connections cache + lazy-decrypt) own growth: localDb.ts 805->807 (gate units, +2). localDb.ts is the re-export-only layer (hard rule #2 — no logic); the PR adds 4 new db/readCache re-exports (touchConnectionLastUsed, getCachedRawProviderConnections, getCachedProviderConnectionById, getCachedProviderNodes) required by existing barrel importers. Irreducible for a re-export list; frozen so it can only shrink.",
-    "_rebaseline_2026_07_20_7779_routingcombo_thread": "PR #7779 own growth: chatHelpers.ts 876->877 (+1, thread routingComboId into executeChatWithBreaker for compression-combo assignment). Frozen so it can only shrink.",
-    "_rebaseline_2026_07_20_7819_autocandidateoverrides_reexport": "PR for #7819 (Level 1+2: read-only auto/* candidate transparency + per-API-key exclusions) own growth: localDb.ts 807->808 (+1). Adds a single `export * from \"./db/autoCandidateOverrides\"` barrel re-export (hard rule #2 — no logic) for the new DB module backing per-apiKey candidate exclusions. Irreducible for a re-export list; frozen so it can only shrink.",
-    "_rebaseline_2026_07_21_8027_grok_cli_auth_json_paste": "PR #8027 (RaviTharuma, fix(grok-cli) #7610) own growth: OAuthModal.tsx 1080->1100 (gate units). Requires the full ~/.grok/auth.json (with refresh_token) on the paste-import path instead of a bare JWT, at the existing paste-token chokepoint (renamed tab label, updated instructions/placeholder, textarea for the auth.json blob, inline error surface). The validation logic itself (parseGrokCliPasteToken, previously an inline ~75-line function) was extracted to the new src/lib/oauth/utils/grokCliAuthJson.ts leaf module — mirroring the #6636/#7546 extraction precedent — so only the irreducible UI wiring remains here. Sixth bump on this modal (969->989->993->998->1030->1056->1100); structural shrink tracked in #3501.",
-    "_rebaseline_2026_07_21_8034_compression_exclusions_sidebar": "#8034 (compression exclusions dashboard tab) own growth: sections.ts 796->806 (+10, one new COMPRESSION_CONTEXT_GROUP sidebar item linking /dashboard/compression/exclusions). The file was already 796/800 before this PR (organic growth from prior sidebar entries), so a single new nav item pushed it 6 lines over cap. Freezing at 806 (cannot grow further); the sidebar item array is data, not extractable logic.",
-    "_rebaseline_2026_07_22_7936_namespace_roundtrip": "#7936 (@RCrushMe, Responses-Chat namespace round-trip identity seam) own growth: open-sse/translator/response/openai-responses.ts 1092->1125 (+33) and open-sse/utils/stream.ts 2814->2869 (+55) — threading the namespace-identity seam through the Responses↔Chat translation + stream paths so tool-call namespaces survive the round-trip. Cohesive translation/stream wiring at existing chokepoints, frozen at new size.",
-    "_rebaseline_2026_07_22_8010_codex_responses_engine": "PR #8010 (@JxnLexn) own growth: open-sse/mcp-server/schemas/tools.ts 1497->1505 (+8 = threading the new \"codex-responses\" literal into the compressionConfigureInput strategy/autoTriggerMode Zod enums and setCompressionEngineInput engine enum, mirroring the existing rtk/omniglyph enum entries; no new tool). open-sse/services/compression/strategySelector.ts 1043->1054 (+11 = one new `if (mode === \"codex-responses\")` dispatch branch in runCompression that delegates 100% to the new codexResponsesEngine.apply, mirroring the existing rtk single-mode dispatch, plus threading config.codexResponsesConfig.preserveToolNames into the shared adaptBodyForCompression call at the 3 existing call sites). src/lib/db/compression.ts (untracked, new-file cap 800) 794->845 (+51 = normalizeCodexResponsesConfig, mirroring the existing normalizeRtkConfig normalizer, plus registering \"codex-responses\" in the COMPRESSION_MODES/STACKED_PIPELINE_ENGINE_IDS/SINGLE_MODE_ENGINE sets and the getCompressionSettings load/save switch) — added to the baseline at its current size. All three are cohesive dispatch/normalizer wiring at existing chokepoints (mirroring the prior compression-mode rebaselines #6534/#6556), not extractable without hiding the mode-dispatch boundary. Covered by tests/unit/compression/codex-responses.test.ts (6) + omniglyph-registries.test.ts/types.test.ts (22, updated for the new mode).",
-    "_rebaseline_2026_07_22_8034_compression_exclusions_persistence": "#8034 (compression exclusions) own growth: src/lib/db/compression.ts 845->850 (+5 = threading the new compressionExclusions field through the existing getCompressionSettings/saveCompressionSettings load/save switch over the shared key_value compression namespace — no new table, no raw SQL). Mirrors the prior compression-field rebaselines (#8010 codex-responses normalizer at the same chokepoint); the load/save switch is a single dispatch boundary, not extractable without hiding it. Covered by the PR's 8 node:test + 3 vitest cases.",
-    "_rebaseline_2026_07_22_8050_model_lockout_exact_family": "#8050 (@AndrianBalanescu) own growth: accountFallback.ts 1864->1892 (+28) — exact-vs-family model-lockout scoping (getModelLockKey/isModelLocked/clearModelLock/getModelLockoutInfo) so an Antigravity 404 for one bare model no longer hijacks the whole family cooldown. Cohesive lockout logic; frozen at new size.",
-    "_rebaseline_2026_07_22_8056_headroom_minrows": "#8056 (@RaviTharuma, persist Headroom minRows) own growth: src/lib/db/compression.ts 850->866 (+16 HeadroomConfig+DEFAULT_HEADROOM_CONFIG+normalize/store in get/updateCompressionSettings) and open-sse/services/compression/strategySelector.ts 1054->1060 (+6 merge settings.headroom into stacked stepConfig). Cohesive settings-persistence + stacked-merge wiring at existing chokepoints, frozen at new size.",
-    "_rebaseline_2026_07_22_8081_reasoning_placeholder_guard": "#8081 (@Dingding-leo) own growth: openai-responses.ts 1125->1137 (+12) restructuring the reasoning-placeholder guard so it skips only the empty content block and still emits finish_reason/tool_calls in the same chunk. Cohesive translator wiring; frozen at new size.",
-    "_rebaseline_2026_07_22_8210_openrouter_midstream_error": "PR #8210 (hartmark, fix/openrouter-midstream-error-surfacing) own growth: open-sse/translator/response/openai-responses.ts 1137->1163 (+26) measured on the merged tip (release 1137 + this PR own growth). Adds a single new branch inside openaiToOpenAIResponsesResponse() that detects an OpenRouter-style mid-stream aggregator error (HTTP 200 SSE chunk with empty choices + a top-level error object) and surfaces it as state.upstreamError instead of silently falling through to the no-op/awaitingTrailingUsage path, which previously masked the failure as a false empty-success completion and skipped combo fallback. Irreducible call-site addition at the existing chunk-dispatch chokepoint (mirrors the Gemini-to-OpenAI translator's #4177 precedent for the same class of upstream error surfacing). Note: this baseline entry does NOT cover the separate pre-existing +11 drift already on the release tip from #8081/#8162 (1125->1136, unrelated reasoning-placeholder-stripping fix merged after this PR branched) — that drift belongs to the maintainer's rebaseline, not this PR.",
-    "_rebaseline_2026_07_22_8211_gemini_malformed_tool_choice": "PR #8211 (hartmark, fix/gemini-malformed-function-call-tool-choice) own growth: open-sse/translator/response/gemini-to-openai.ts 771->821 (+50, entirely this PR's diff — no other commit touched this file between the PR's merge-base and the release tip). Adds MALFORMED_FUNCTION_CALL/UNEXPECTED_TOOL_CALL handling inside geminiToOpenAIResponse(): synthesizes a `malformed_tool_call` tool_calls entry so finish_reason normalizes to the standard \"tool_calls\" instead of an unrecognized raw enum value that OpenAI-compatible clients (e.g. OpenClaw) silently ignore, and always synthesizes (rather than skipping when a real tool call already exists) so a malformed attempt alongside a real one in the same turn is not silently discarded. Irreducible cohesive addition at the existing candidate/finishReason translation chokepoint (mirrors the 9router#2462 raw-finish-reason precedent immediately below it in the same function). Covered by the PR's own tests/unit test additions for both the malformed-only and malformed-plus-real-call cases.",
-    "_rebaseline_2026_07_22_8213_chat_abandoned_target_abort": "PR #8213 (hartmark, fix/gemini-tpm-quota-cooldown-wait) own growth: src/sse/handlers/chat.ts 1794->1860 (+66, measured against the PR's own merge-base — the release tip separately carries an unrelated -5 net shrink from #8013's antigravity callable-catalog alignment, which this PR's branch does not include and this entry does not cover). Adds resolveDispatchClientRawRequest(): merges a per-target modelAbortSignal into clientRawRequest.signal (via mergeAbortSignals) so a combo target abandoned by comboTargetTimeoutMs actually observes its own abort and reaches its cleanup path, instead of hanging forever inside withRateLimit/acquireAccountSemaphore and leaking a permanent 'pending' dashboard entry (live incident, log id 1784418258231-14961a). Also wires combo-exhausted rejection logging to capture request body + attempted models via the new rejectedRequestUsage helper. Irreducible additions at the existing chat dispatch chokepoint. Covered by the PR's own combo-config + integration test additions.",
-    "_rebaseline_2026_07_22_8213_combo_cooldown_wait_recording": "PR #8213 (hartmark, fix/gemini-tpm-quota-cooldown-wait) own growth: open-sse/services/combo.ts 3548->3604 (+56, entirely this PR's diff — no other commit touched this file between the PR's merge-base and the release tip). Fixes combo cooldown-wait state recording so a bogus 503 is no longer crystallized when the cooldown-wait vars reset every setTry, adds an OpenAI-format SSE error frame path for combo-exhausted rejections (capturing request body + attempted models), and gives an abandoned per-target dispatch its own timeout instead of leaking a permanent 'pending' dashboard entry. Irreducible additions at the existing handleComboChat dispatch/retry chokepoint (mirrors the prior quota-share/headroom/task-aware strategy-branch precedents already frozen in this file). Covered by the PR's own combo-config + Gemini TPM-ceiling benchmark test additions.",
-    "_rebaseline_2026_07_22_8213_gemini_tpm_quota_cooldown_wait": "PR #8213 (hartmark, fix/gemini-tpm-quota-cooldown-wait) own growth: open-sse/services/accountFallback.ts 1857->1932 on the merged tip (release 1892 incl #8050 +35, plus this PR own growth +40); measured against the PR own merge-base was 1857->1898 (+41 — the release tip separately carries an unrelated +34 from #8050's antigravity 404 model-not-found lockout scoping, which this PR's branch does not include and this entry does not cover). Own growth is the Gemini TPM-ceiling classification + cooldown-wait wiring feeding into the combo cooldown-wait state machine (rate-limit wedge recovery) introduced by this PR's commit series. Irreducible additions at the existing account-fallback/model-lockout chokepoint. Covered by the PR's own gemini-rate-limit-tracker and TPM-ceiling benchmark test additions.",
-    "_rebaseline_2026_07_22_8213_health_unblock_model_cooldowns": "PR #8213 (hartmark, fix/gemini-tpm-quota-cooldown-wait) own growth: src/app/(dashboard)/dashboard/health/page.tsx 1094->1165 (+71, entirely this PR's diff — no other commit touched this file between the PR's merge-base and the release tip). Adds handleUnblockAll/handleUnblockOne dashboard actions (DELETE /api/resilience/model-cooldowns) so an operator can manually clear a Gemini TPM-wedge model lockout surfaced by this PR's cooldown-wait fixes, instead of waiting out the ceiling. Irreducible UI wiring at the existing health-page action chokepoint. Covered by the PR's own dashboard/resilience test additions.",
-    "_rebaseline_2026_07_22_8213_requestloggerdetail_unblock_ui": "PR #8213 (hartmark, fix/gemini-tpm-quota-cooldown-wait) own growth: src/shared/components/RequestLoggerDetail.tsx 799->941 (+142, entirely this PR's diff — no other commit touched this file between the PR's merge-base and the release tip; crosses the general 800-line new-file cap so is frozen here for the first time). Adds a collapsible section header (open/expand-less toggle) plus per-log-entry unblock (`unblocking`/`cleared` state, isCombo503 detection) so the request-logger detail panel surfaces the same Gemini TPM cooldown-wait / model-lockout unblock action introduced by this PR at the individual-request level (mirrors the health-page bulk unblock action added in the same PR). Covered by the PR's own dashboard/resilience test additions.",
-    "_rebaseline_2026_07_22_fusion_8013_8098_antigravity": "Fusion of #8013 (backryun, catalog/IDE-CLI-split rewrite) + #8098 (nguyenha935, protocol-fidelity/fail-closed/credits/tool-cloaking): open-sse/services/usage/antigravity.ts NEW 802 (>cap 800, +2 — #8098 credits/tier usage service on #8013's profile-aware headers). Test growth (models-catalog-route 1605->1608, provider-models-route 1752->1757 from #8013 Gemini 3.6 catalog) tracked in testFrozen.",
-    "_rebaseline_2026_07_23_8127_grok_weekly_quota": "#8127 (@apoapostolov) own growth: src/sse/handlers/chat.ts 1861->1865 (+4) — weekly quota tracking for grok-web wires a quota-fetch hook at the existing dispatch chokepoint. Thin wiring mirroring adjacent provider-quota branches; not extractable. Covered by tests/unit/grok-quota-fetcher.test.ts.",
-    "_rebaseline_2026_07_23_8143_empty_catch_logging": "#8143 (@chirag127) own growth: open-sse/utils/stream.ts 2869->2887 (+18) — replacing empty catch blocks in the SSE stream subsystem with console.debug logging (Rule #6 silent-swallow fix, issues #8138-#8142). Cohesive logging additions at the existing catch chokepoints, not extractable; frozen at new size. Covered by tests/unit/stream-handler-catch-logging-8143.test.ts.",
-    "_rebaseline_2026_07_23_8219_cache_ttl_settings_sidebar": "#8219 (@oyi77) own growth: sections.ts 806->813 (+7) — configurable model-catalog cache-TTL settings adds a new sidebar nav entry + its visibility wiring. Sidebar item array is data, not extractable logic; frozen at new size.",
-    "_rebaseline_2026_07_23_8247_8248_model_unhealthy": "#8247+#8248 own growth: accountFallback.ts 1940->1941 (+1, irreducible import statement only — the substantive #8248 DEGRADED-pattern classifier was extracted into open-sse/config/errorConfig.ts, which has ample headroom, instead of growing this frozen file; #8247's fix is a single existing-line condition change, net zero lines). Scoping the credits-exhausted 403/429 branch to isCompatibleProvider() (per-model-quota openai/anthropic-compatible-* nicknames) so it stays model-scoped instead of terminalling the whole connection, and classifying NVIDIA NIM 'Function ... DEGRADED' 400 bodies as model-access-denied instead of a raw passthrough 400. Covered by tests/unit/8247-accountfallback-model-unhealthy.test.ts and tests/unit/8248-accountfallback-nvidia-degraded.test.ts.",
-    "_rebaseline_2026_07_23_8252_combo_400_advance": "#8252 (@RaviTharuma) own growth: accountFallback.ts 1932->1940 (+8) + combo.ts 3604->3630 (+26) — advance combo on model-scoped 400s wrapped as invalid/Bad-Request. Irreducible wiring at existing account-fallback + combo dispatch chokepoints. Covered by combo-model-scoped-400-advance.test.ts.",
-    "_rebaseline_2026_07_23_8266_alibaba_media": "#8266 (@backryun) own growth: imageRegistry.ts 821->979 (+158) — Alibaba-family media models (Qwen image/video, Bailian, Wan) added to the image/video registry. Registry model data, not extractable logic; frozen at new size.",
-    "_rebaseline_2026_07_24_8388_compression_detail_persist": "#8388 (compression engine DETAIL settings — Headroom/session-dedup/CCR — dropped on save) own growth: src/lib/db/compression.ts 866->872 (+6 = irreducible call-site wiring at the existing getCompressionSettings/updateCompressionSettings chokepoint: one import line, one `...buildDetailConfigDefaults()` spread in the seed config, and one `case \"sessionDedup\": case \"ccr\": applyDetailConfigUpdate(config, key, parsed); break;` load-switch case, mirroring the existing headroom/#8056 case immediately above it). The actual normalizer logic (normalizeSessionDedupConfig/normalizeCcrConfig, matching SESSION_DEDUP_SCHEMA/CCR_SCHEMA bounds) was EXTRACTED into a new leaf src/lib/db/compressionDetailNormalizers.ts (well under cap) so this frozen file only carries the minimal dispatch wiring. Covered by tests/unit/8388-compression-detail-persist.test.ts (schema-accept + full DB save->reload round-trip for both new sub-objects, plus a no-regression assertion on the existing headroom round-trip).",
-    "_rebaseline_2026_07_24_responses_toolcalls_log_summary": "hartmark, fix/responses-tool-calls-log-summary own growth: open-sse/translator/response/openai-responses.ts 1163->1174 (+11). closeToolCall() now also writes the completed tool call into the shared state.toolCalls Map (already populated by the openai-to-claude / claude-to-openai / gemini-to-openai response translators) so stream.ts's completion-log summary builder (which reads state.toolCalls, not this translator's own funcCallIds/funcNames/funcArgsBuf bookkeeping) reports finish_reason \"tool_calls\" and message.tool_calls for openai->openai-responses translated streams instead of always logging \"stop\" with no tool_calls — the actual client-facing SSE events were already correct; only the persisted call-log summary was wrong. Irreducible call-site addition at the existing tool-call-close chokepoint. Covered by the new regression test in tests/unit/translator-resp-openai-responses.test.ts.",
-    "_rebaseline_2026_07_25_8476_combo_input_bound_homogeneous_scope": "PR #8476 (herjarsa, fix/8375-8459-combo-image-fixes, #8375) own growth: open-sse/services/combo.ts 3642->3679 (+37 net: +29 the PR's own isInputBoundFailure short-circuit for deterministic context_length_exceeded/context_window_exceeded failures, +8 a /green-prs pre-merge fix scoping that short-circuit to homogeneous remainders only — the shipped code fired unconditionally on ANY target, regressing the intentional heterogeneous-combo fallback #6637/isContextOverflow400 protects, exactly as flagged by this PR's own review evidence but never actually implemented in the branch). The fix compares orderedTargets[i+1..] modelStr against the failing target's modelStr at the existing executeTarget dispatch chokepoint (mirrors the sameProviderNext precedent a few lines below) — irreducible call-site wiring, not extractable without hiding the dispatch boundary. Covered by tests/unit/combo-input-bound-failure-8375.test.ts (homogeneous pool still short-circuits) and the new tests/unit/combo-input-bound-heterogeneous-8375.test.ts (heterogeneous combo now correctly falls through to the larger-context target).",
-    "_rebaseline_2026_07_25_adobe_firefly_reference_images": "Follow-up to #8006: storage upload + referenceBlobs for image/video and /v1/images/edits dispatch. adobeFireflyClient.ts 1958->2317 (+upload helpers, extract sources, resolve blob ids). Note: 2317 not 2316 — check-file-size.mjs counts LOC via split(\"\\n\").length (counts the trailing-newline empty element), which is 1 higher than `wc -l` on a file ending in \\n; the PR's original entry (2316) was measured with wc -l and undercounted by 1 against the actual gate.",
-    "_rebaseline_pr1043_minimax_tts": "Upstream port decolua/9router#1043 (toanalien) own growth: audioSpeech.ts 965->1061 (+96). Adds MiniMax T2A v2 TTS dispatch (handleMinimaxSpeech + hexToBytes helper) — provider entry was already in audioRegistry (format: minimax-tts) but no handler existed, falling through to the OpenAI-compatible default that fails (T2A has custom shape + hex-encoded audio + base_resp envelope). New branch sits next to the other inline provider branches (xiaomi-mimo, coqui, tortoise, aws-polly) — extracting would just create indirection. Covered by tests/unit/minimax-tts-1043.test.ts (3 tests, GREEN: success, base_resp error, invalid-hex).",
-    "_rebaseline_pr4592_exclude_exhausted_auto": "Reconcile #4592 already-merged growth: combo.ts 2991->3036 (+45, terminal-status quota-cutoff exclusion in buildAutoCandidates + opt-in gate). Fast-gate PR->release does not run check:file-size.",
-    "open-sse/executors/antigravity.ts": 1528,
-    "open-sse/executors/base.ts": 1640,
-    "open-sse/executors/chatgpt-web.ts": 3241,
-    "open-sse/executors/codex.ts": 1562,
-    "open-sse/executors/cursor.ts": 1563,
-    "open-sse/executors/deepseek-web.ts": 1148,
-    "open-sse/executors/grok-web.ts": 1044,
-    "open-sse/executors/muse-spark-web.ts": 1405,
-    "open-sse/handlers/chatCore.ts": 5034,
-    "open-sse/handlers/imageGeneration.ts": 3101,
-    "open-sse/handlers/responseSanitizer.ts": 1128,
-    "open-sse/handlers/search.ts": 1536,
-    "open-sse/handlers/videoGeneration.ts": 1063,
-    "open-sse/mcp-server/schemas/tools.ts": 1505,
-    "open-sse/mcp-server/server.ts": 1411,
-    "open-sse/mcp-server/tools/advancedTools.ts": 1120,
-    "open-sse/services/accountFallback.ts": 1972,
-    "open-sse/services/adobeFireflyClient.ts": 2385,
-    "open-sse/services/claudeCodeCompatible.ts": 1202,
-    "open-sse/services/combo.ts": 3648,
-    "open-sse/services/compression/strategySelector.ts": 1060,
-    "open-sse/services/rateLimitManager.ts": 1167,
-    "open-sse/translator/response/openai-responses.ts": 1204,
-    "open-sse/utils/cursorAgentProtobuf.ts": 1505,
-    "open-sse/utils/stream.ts": 2889,
-    "src/app/(dashboard)/dashboard/HomePageClient.tsx": 1388,
-    "src/app/(dashboard)/dashboard/analytics/ComboHealthTab.tsx": 1031,
-    "src/app/(dashboard)/dashboard/api-manager/ApiManagerPageClient.tsx": 3117,
-    "src/app/(dashboard)/dashboard/cache/media/MediaPageClient.tsx": 1067,
-    "src/app/(dashboard)/dashboard/combos/page.tsx": 4647,
-    "src/app/(dashboard)/dashboard/costs/CostOverviewTab.tsx": 1283,
-    "src/app/(dashboard)/dashboard/costs/quota-share/components/PoolWizard.tsx": 1022,
-    "src/app/(dashboard)/dashboard/endpoint/EndpointPageClient.tsx": 2615,
-    "src/app/(dashboard)/dashboard/health/page.tsx": 1165,
-    "src/app/(dashboard)/dashboard/providers/[id]/components/modals/EditConnectionModal.tsx": 1316,
-    "src/app/(dashboard)/dashboard/providers/page.tsx": 1944,
-    "src/app/(dashboard)/dashboard/runtime/RuntimePageClient.tsx": 1201,
-    "src/app/(dashboard)/dashboard/settings/components/PricingTab.tsx": 1019,
-    "src/app/(dashboard)/dashboard/settings/components/ProxyRegistryManager.tsx": 1464,
-    "src/app/(dashboard)/dashboard/settings/components/ResilienceTab.tsx": 1123,
-    "src/app/(dashboard)/dashboard/settings/components/RoutingTab.tsx": 1629,
-    "src/app/(dashboard)/dashboard/settings/components/SystemStorageTab.tsx": 1573,
-    "src/app/(dashboard)/dashboard/usage/components/BudgetTab.tsx": 1028,
-    "src/app/(dashboard)/dashboard/usage/components/EvalsTab.tsx": 2148,
-    "src/app/(dashboard)/dashboard/usage/components/ProviderLimits/index.tsx": 1119,
-    "src/app/api/providers/[id]/models/route.ts": 2250,
-    "src/app/api/v1/models/catalog.ts": 1549,
-    "src/lib/tokenHealthCheck.ts": 1021,
-    "src/lib/db/apiKeys.ts": 1529,
-    "src/lib/db/core.ts": 1637,
-    "src/lib/db/migrationRunner.ts": 1084,
-    "src/lib/db/models.ts": 1097,
-    "src/lib/db/providers.ts": 1034,
-    "src/lib/memory/retrieval.ts": 1073,
-    "src/lib/tailscaleTunnel.ts": 1202,
-    "src/lib/usage/providerLimits.ts": 1013,
-    "src/shared/components/OAuthModal.tsx": 1134,
-    "src/shared/components/RequestLoggerV2.tsx": 1629,
-    "src/shared/components/analytics/charts.tsx": 1035,
-    "src/shared/services/cliRuntime.ts": 1122,
-    "src/sse/handlers/chat.ts": 1877,
-    "src/sse/services/auth.ts": 2508,
-    "tests/unit/account-fallback-service.test.ts": 1572,
-    "tests/unit/provider-validation-specialty.test.ts": 2985,
-    "open-sse/executors/hyperagent.ts": 1026,
-    "open-sse/executors/default.ts": 1042,
-    "open-sse/executors/kiro.ts": 1069
-  },
   "testCap": 1000,
   "testFrozen": {
     "_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).",
@@ -410,7 +284,134 @@
   "_rebaseline_2026_07_27_3850_relax_filesize_cap_v2_20pct": "OWNER-APPROVED TEMPORARY relax for v3.8.50-3.8.54 PREPARE phase (docs/ROADMAP.md). v1 was cap 800->900 / testCap 800->900 on 2026-07-27; v2 = v1 +20% buffer = cap 900->1000 (+100), testCap 900->1000 (+100). Justification: same as complexity v2 — the v3.8.50 release cut coincides with high-merge activity; owner accepted enlarging the headroom to cover the entire PREPARE phase (5 minor cycles .50-.54) without per-PR rebaseline noise. Targets: decompose-existing-frozen unchanged (frozen still only-shrink — see frozen[] entries and the 105 files >900 that still need structural decomposition regardless of cap); this only relaxes the cap for NEW files in the decompose/extract-while-PREPARE phase (.51='executor registry in-place' and .52='combo.ts decomposition' create new leaf modules above 800). RE-TIGHTENING MANDATORY in v3.8.51: cap target 850 = 850 once decomposition wave stabilizes (gives 150 units of post-tighten headroom vs the new 1000 ceiling). Tracked via same roadmap issue as complexity v2. Window: v3.8.50 (release cut) → v3.8.54 close (RE-TIGHTEN at v3.8.51 prep merge per ROADMAP.md). Last entry unless measured regression. v1 entry retained below for audit trail.",
   "_rebaseline_2026_07_27_3850_relax_filesize_cap": "OWNER-APPROVED TEMPORARY relax for v3.8.50-3.8.54 PREPARE phase (docs/ROADMAP.md). cap 800->900 (+100), testCap 800->900 (+100). Targets: decompose-existing-frozen unchanged (frozen still only-shrink); this only relaxes the cap for NEW files in the decompose/extract-while-PREPARE phase (.51='executor registry in-place' and .52='combo.ts decomposition' create new leaf modules above 800). RE-TIGHTENING MANDATORY in v3.8.51: cap target 850 = 850 once decomposition wave stabilizes. SUPERSEDED by _rebaseline_2026_07_27_3850_relax_filesize_cap_v2_20pct (v1 +20% buffer) — retained for audit. Tracked via same roadmap issue.",
   "_rebaseline_2026_07_27_v3849_train1h": "Merge-train 1H (31 PRs) — owner-approved 2026-07-27. Two distinct causes, kept separate on purpose: (1) GENUINE irreducible growth at existing chokepoints — providerLimits/auth (#8632 Kimi quota-reset recovery), rateLimitManager (#8616 idle wedged limiters), models-catalog-route.test (#8610 OpenCode Go effort aliases); (2) COLLISION with #8585, which banked shrinks measured on the pre-train release tip while 30 sibling PRs in the SAME train grew those files again — chat/accountFallback (#8628), chatCore (#8613), videoGeneration (#8581), imageGeneration. The zero-headroom frozen entries cannot absorb either. Ceilings re-pinned to the post-merge tip; #8612 (also in this train) automates shrink-banking so this self-inflicted drift stops recurring. Detail: src/lib/usage/providerLimits.ts 1006->1013 (#8632); src/sse/services/auth.ts 2492->2508 (#8632); open-sse/services/rateLimitManager.ts 1014->1060 (#8616); src/sse/handlers/chat.ts 1842->1845 (#8628); open-sse/handlers/chatCore.ts 4939->4955 (#8613); open-sse/handlers/imageGeneration.ts 3100->3101 ((sem PR — teto do #8585)); open-sse/handlers/videoGeneration.ts 1038->1063 (#8581); open-sse/services/accountFallback.ts 1965->1966 (#8628); tests/unit/models-catalog-route.test.ts 1608->1636 (#8610)",
-  "_rebaseline_2026_08_02_9242_token_health_transient": "PR #9242 (fix/refresh-circuit-transient): src/lib/tokenHealthCheck.ts 1021 (new file, above cap 1000). The file consolidates token-refresh health checking logic that was previously scattered across auth.ts and tokenRefresh.ts. Cohesive single-responsibility module for refresh circuit state management; not extractable without splitting the refresh state machine. Covered by tests/unit/tokenHealthCheck-transient.test.ts.",
+  "frozen": {
+    "_rebaseline_2026_06_22_4644_deepseek_web_tools": "PR #4644 (BugsBag/robust deepseek-web tool-call parsing): open-sse/executors/deepseek-web.ts 1117->1125 (+8). The new agentic tool-call path emits surrounding text + reasoning before tool_calls and swaps to the dedicated deepseekWebTools.ts parser; the +8 lines are cohesive wiring at the existing transformSSE chokepoint (the parser itself lives in the new deepseekWebTools.ts file, already under cap). The PR's own fast-gate (PR->release) does not run check:file-size, so this surfaced only at release reconcile. Covered by tests/unit/deepseek-web-tools-variants.test.ts + deepseek-web-tools-execute.test.ts.",
+    "_rebaseline_2026_06_23_4712_deepseek_web_tool_results": "PR for #4712 (deepseek-web drops role:tool): open-sse/executors/deepseek-web.ts 1125->1148 (+23). messagesToPrompt() now folds role:\"tool\" results into the single-prompt transcript (recovering the tool name from the preceding assistant tool_calls by tool_call_id) instead of silently dropping them; the lines are cohesive wiring inside the existing function. Covered by tests/unit/deepseek-web-tool-result-prompt-4712.test.ts.",
+    "_rebaseline_2026_06_24_headroom_strategy": "Headroom-aware connection selection (dario technique): combo.ts 3168->3180 (+12 = a new `else if (strategy === \"headroom\")` dispatch branch in handleComboChat that delegates to orderTargetsByHeadroom + its log line, plus the import). The actual logic lives OUT of the god-file: the pure ranker rankByHeadroom/computeHeadroom is the new leaf open-sse/services/combo/headroomRanking.ts (91 LOC, <cap) and the async orderer orderTargetsByHeadroom is appended to the existing open-sse/services/combo/quotaStrategies.ts (<cap) next to its sibling reset-aware/reset-window orderers (reuses their connection-expansion machinery). headroom = 1 - max(util_5h, util_7d) from getSaturation (src/lib/quota/saturationSignals.ts), prefers the connection with the most free capacity. Only the dispatch wiring is irreducible at the existing combo strategy chokepoint (mirrors the reset-aware/reset-window/context-optimized branches); not extractable without hiding the call site. fill-first stays default; all existing strategies untouched. Covered by tests/unit/combo-headroom-ranking.test.ts (pure helper) + tests/unit/combo-headroom-strategy.test.ts (orderer, saturation injected). Structural shrink of combo.ts tracked in #3501.",
+    "_rebaseline_2026_06_24_quota_share_strategy": "Dedicated quota-share strategy (Phase 3 #9): combo.ts 3180->3190 (+10 = one new `else if (strategy === \"quota-share\")` dispatch branch in handleComboChat that delegates 100% to selectQuotaShareTarget + its log line, plus the import). All the new logic lives OUT of the god-file in two new leaves under open-sse/services/combo/: quotaShareInflight.ts (in-flight counter with TTL/lease, ~150 LOC <cap) and quotaShareStrategy.ts (per-model bucket gating via isBucketSaturated + DRR proportional to weight + P2C over in-flight, ~240 LOC <cap). Only the dispatch wiring is irreducible at the existing combo strategy chokepoint (mirrors the headroom/reset-aware/reset-window/context-optimized branches); not extractable without hiding the call site. ZERO existing strategy cases were modified — only this branch was added, and the qtSd/ combos switched from fill-first to quota-share in src/lib/quota/quotaCombos.ts. Covered by tests/unit/quota-share-strategy.test.ts (gating, DRR fairness, P2C in-flight, fail-open, activation). Structural shrink of combo.ts tracked in #3501.",
+    "_rebaseline_2026_06_24_task_aware_routing": "Task-aware routing strategy (port PR #2045, OmniRoute #4945): combo.ts 3190->3225 (+35) = one new `else if (strategy === \"task-aware\")` dispatch branch delegating 100% to selectTaskAwareTarget + its imports/log lines. All scoring/classification logic lives OUT of the god-file in the new leaf open-sse/services/taskAwareRouting.ts (553 LOC <cap). Only the dispatch wiring is irreducible at the existing combo strategy chokepoint (mirrors quota-share/headroom/reset-aware branches). ZERO existing strategy cases modified. Covered by tests/unit/combo-task-aware.test.ts (35 tests). Structural shrink of combo.ts tracked in #3501.",
+    "_rebaseline_2026_06_26_fidelity_gate_extraction": "Milestone-B fidelity-gate wiring residual: bodyToText+gateAdvance extracted to fidelityGateStep.ts (889->854, -35), but the StackOptions.fidelityGate field, the `const fidelityGate` reads at the two stacked-loop dispatch chokepoints, and the import of FidelityGateConfig are irreducible wiring that cannot leave strategySelector without an architectural refactor of the pre-existing stacked pipeline. Net: 889->854 (+6 vs the pre-Milestone-B frozen 848). Covered by tests/unit/compression/*.test.ts (940 pass).",
+    "_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.",
+    "_rebaseline_2026_06_27_5193_antigravity_basered": "Base-red (pre-existing release drift, fast-gate PR->release skips check:file-size): accountFallback.ts 1773->1777 and src/app/api/providers/[id]/test/route.ts 924->940 were already over their frozen caps on release/v3.8.39 independent of any antigravity change. Owner chose to rebaseline (keep the documented issue-reference comments #1846/#1449/#347 etc.) rather than accept the contributor comment-stripping in #5200/#5198. Reverted #5200 to restore the comments; bumped these two frozen caps to the actual base sizes. No logic change.",
+    "_rebaseline_2026_06_28_5237_impersonation_ua_refresh": "PR #5237 (refresh impersonation UAs): grok-web.ts 1871->1873 (+2), muse-spark-web.ts 1284->1302 (+18), perplexity-web.ts 1013->1032 (+19). Net semantic change in each file is a single User-Agent constant (Chrome 147->149 for grok/muse; perplexity kept at Firefox 148 to stay matched with the firefox_148 TLS profile — the contributor's 152 bump was reverted to avoid a UA-vs-JA3 mismatch, #2459). The growth is Prettier reflow that lint-staged unavoidably applies to these grandfathered long-line files the moment they are touched; not extractable. src/sse/services/auth.ts 2336->2401 in the same reconcile is #5222's antigravity-LRU-retry growth that merged via --admin without a baseline bump.",
+    "_rebaseline_2026_06_28_5243_risk_gate_prepass": "PR #5243 (compression risk-gate pre-pass) own growth: open-sse/services/compression/strategySelector.ts 854->899 (+45). The three exported entry points (applyCompression/applyStackedCompression/applyStackedCompressionAsync) become thin wrappers over pure-extracted private bodies (runCompression/runStackedCompression/runStackedCompressionAsync) so the risk-gate mask->run->restore wrapper sits strictly OUTSIDE the per-step loop — a single universal integration point. The wrapper logic itself (resolveRiskGate/withRiskGate) lives in the new riskGate/strategyWrap.ts (<cap); the residual growth is the duplicated thin-wrapper signatures + the extracted bodies' dispatch boundary, guarded by a byte-identical parity test (riskGateIntegration). Default off (DEFAULT_COMPRESSION_CONFIG unchanged). Not extractable without hiding the dispatch boundary, mirroring prior compression rebaselines. Structural shrink tracked in #3501.",
+    "_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.",
+    "_rebaseline_2026_06_29_4038_cas_guard": "PR (#4038) own growth: tokenRefresh.ts 2103->2181 (+78 = the compare-and-swap guard on the refresh persist — runWithCasGuard/getActiveCasGuard AsyncLocalStorage pair mirroring runWithOnPersist, casGuardShouldSkipPersist that rereads the row right before persisting and skips the write when a concurrent writer already rotated the refresh_token past the one presented, plus getCasGuardStats counters). Fixes the sibling-rotation-revert → token-family-revocation storm. Gated behind an active guard (opt-in; no guard => byte-identical). Wiring lives at the two persist chokepoints inside getAccessToken; the comparison reuses wasRefreshTokenRotated from refreshSerializer. Not extractable without splitting the refresh hot path.",
+    "_rebaseline_2026_06_29_5286_memoization": "PR #5286 own growth: strategySelector.ts 899->960 (+61 = the opt-in result-memoization branches in applyCompression/applyCompressionAsync — principal+determinism gate, makeMemoKey lookup/store with model+supportsVision folded into the key, recompute-with-memo-off). Default off (memoizeCompressionResults), so zero behavior change. The memo helpers live in the leaf resultMemo.ts (<cap); the chokepoint wiring here is not extractable. Structural shrink of this hot-path file tracked in #3501.",
+    "_rebaseline_2026_07_01_v3843_release_5609": "Rebaseline v3.8.43 (PR #5609 release reconciliation). DRIFT dos 109 commits do ciclo: 8 god-files existentes cresceram (ApiManagerPageClient 2983->3017, combos/page 4594->4608, AddApiKeyModal 868->869, providerPageHelpers 974->996, chat.ts 1635->1647, auth.ts 2401->2403, batchProcessor 828->915, combo.ts 3368->3387) + 2 novos acima do cap (huggingchat.ts 813, tests web-cookie-providers-new 827) + 4 test files cresceram. Modularizacao deferida (blast-radius mid-release); congelado no estado atual p/ o proximo ciclo ratchetar daqui.",
+    "_rebaseline_2026_07_02_5816_qoder": "PR #5816 (@AgentKiller45, qoder PAT via qodercli): qoderCli.ts 666->989, new-above-cap frozen (owner-approved baseline freeze). The growth is the legitimate PAT job-token exchange + quota parsing CLI transport (the pure-JS Cosy path 500'd on every PAT request); extracting the spawn/parse helpers now would just add indirection to a contributor PR mid-merge. Test frozen also raised for this PR's coverage growth: providers-page-utils.test.ts 1052->1092. Additionally clears an inherited base-red from the already-merged #5933 (codex json_schema->text.format): translator-openai-responses-req.test.ts 1097->1172 (+75 regression tests, no offending branch left). All remain frozen (cannot grow further); release captain's rebaseline-at-release supersedes.",
+    "_rebaseline_2026_07_09_6126_clinepass_dual_auth": "PR #6126 (@hajilok, dual-auth ClinePass) own growth: tokenRefresh.ts 2181->2182 (+1 = a single `case \"clinepass\":` fallthrough label added to the existing `case \"cline\":` in _getAccessTokenInternal's provider switch, so clinepass token refresh dispatches to the already-shared refreshClineToken() instead of silently falling through to the generic OAuth refresh). Irreducible 1-line switch-case wiring at the existing chokepoint; the header-building logic for the same feature was extracted to a new leaf src/shared/utils/clineAuth.ts::buildClinepassHeaders() (well under cap) to avoid growing open-sse/executors/default.ts. Covered by tests/unit/clinepass-provider.test.ts.",
+    "_rebaseline_2026_07_09_6363_kiro_external_idp": "PR #6363 (@artickc, Kiro external IdP) own growth: tokenRefresh.ts 2182->2249 (+67 = the external_idp refresh branch inside refreshKiroToken — standard public-client OAuth2 refresh_token grant against the org IdP tokenEndpoint via buildExternalIdpRefreshParams/isExternalIdpAuthMethod from the new leaf open-sse/services/kiroExternalIdp.ts, with invalid_grant/invalid_client -> unrecoverable_refresh_error mapping). Cohesive addition at the existing refreshKiroToken chokepoint. Covered by tests/unit/kiro-external-idp.test.ts.",
+    "_rebaseline_2026_07_09_6587_kiro_api_key_auth": "PR #6587 (@strangersp) own growth for Kiro long-lived API-key auth, merged onto v3.8.47 tip: openai-to-kiro.ts 890->912 (+22, auth-header selection for API-key-vs-OAuth-token connections), providerLimits.ts 998->1000 (+2, API-key auth-type branch), translator-openai-to-kiro.test.ts 1234->1257 (+23), providers-page-utils.test.ts 1109->1107 (net -2 after merging with parallel release drift; connectionMatchesProviderCard api_key coverage added), provider-validation-specialty.test.ts 2856->2980 (+124 net after merge with parallel release drift; this PR also removed the file's `@typescript-eslint/no-explicit-any` eslint-suppression entry by fixing all `any` usages, adding typed replacements). Cohesive additive feature growth, well tested; not extractable without splitting the existing chokepoints mid-merge.",
+    "_rebaseline_2026_07_09_6678_routing_strategy_9router": "#6678 (SeaXen) — 9router-parity Routing Strategy settings card + per-provider/combo sticky-round-robin override. Own growth: ProviderDetailPageClient.tsx 784->786 (single ProviderAccountRoutingCard mount + import), auth.ts 2448->2458 (providerStrategies override resolution: fallbackStrategy/stickyRoundRobinLimit per-provider cascade in getProviderCredentials). Both additive, zero unrelated refactor; new UI/logic lives in new files (ProviderAccountRoutingCard.tsx, RoutingStrategyCard.tsx, rrState.ts::resolveComboStickyRoundRobinLimit). chat.ts value below reflects the current release tip (grown by other concurrent PRs, e.g. #6640), not this PR own change.",
+    "_rebaseline_2026_07_10_6318_omp_letta": "PR #6318 (@hamsa0x7, omp+letta CLI integrations) own growth: cliTools.ts (+53 = 2 registry entries incl. omp docsUrl) and cliRuntime.ts (+18 = runtime-detection wiring for the 2 new tools). Cohesive registry/wiring growth at the existing chokepoints; scope reduced from the original 5 tools (pi/codewhale/jcode shipped separately).",
+    "_rebaseline_2026_07_10_gcf_v3_2_decode": "PR #6838 own growth: new vendored file open-sse/services/compression/engines/headroom/gcf/decode_generic.ts frozen at 880 (> 800 cap). It is the vendored GCF generic-profile decoder (spec v3.2 nested flattening plus the prototype-pollution / hasOwnProperty hardening added in this PR's Gemini review). Kept as one file faithful to upstream gcf-typescript so re-vendoring stays a clean copy rather than a re-split each cycle (sibling generic.ts/scalar.ts stay < cap; extraction would also fragment the file's frozen eslint no-explicit-any suppressions). Round-trip + prototype-pollution regression coverage in tests/unit/compression/headroom-smartcrusher.test.ts. Frozen: only shrinks from here.",
+    "_rebaseline_2026_07_12_v3847_mergeprs_tail": "v3.8.47 /merge-prs tail (owner-approved): src/lib/localDb.ts NEW>800 (799->805, +6 re-exports countFreeProxies + recordFreeProxySyncErrors/clearFreeProxySyncErrors/getFreeProxySyncErrors + FreeProxySyncErrors type for #6909 free-pool relay-repair; re-export-only per Hard Rule #2, not extractable).",
+    "_rebaseline_2026_07_15_7070_combos_memo": "PR #7070 (perf/p1-memo) own growth: src/app/(dashboard)/dashboard/combos/page.tsx 4655->4656 (+1 = React.memo wrapping of ComboCard). Covered by tests/unit/ui/combos-page-smoke.test.tsx.",
+    "_rebaseline_2026_07_18_7399_xai_oauth_modal": "PR #7399 (xAI OAuth PKCE) own growth: OAuthModal.tsx 993->998 (+5 = provider entry + PKCE flow branch wiring at the existing provider-switch chokepoint; the provider logic itself lives in src/lib/oauth/providers/xai-oauth.ts, new leaf). Third irreducible wiring bump on this modal (969->989->993->998); structural shrink tracked in #3501.",
+    "_rebaseline_2026_07_19_6636_codex_session_json": "#6636 own growth: OAuthModal.tsx 998->1030 (gate units, split(\"\\n\").length incl. trailing newline; +32 = session-JSON paste branch for handleManualSubmit plus a shared submitCodexAccessToken() helper extracted from the pre-existing bare-JWT branch, mirroring the #5203 oauthBlobSubmit.ts extraction precedent; the normalizer logic itself lives in the new src/lib/oauth/utils/codexSessionImport.ts leaf module, not here). Fourth irreducible wiring bump on this modal (969->989->993->998->1030); structural shrink tracked in #3501.",
+    "_rebaseline_2026_07_19_7546_ghe_copilot_modal": "PR #7546 (GHE Copilot OAuth provider) own growth: OAuthModal.tsx 1030->1056 (gate units). Adds a gheUrl input state, routes ghe-copilot through the existing device-code branch, and threads gheUrl into the device-code request/poll extraData at the existing provider-switch chokepoints (+~24 lines, cohesive with the same pattern as #7399/#6636). The standalone GHE enterprise-URL config step JSX (originally +31 lines inline) was extracted to the new src/shared/components/oauthModal/GheConfigStep.tsx leaf component to minimize the bump; what remains is the irreducible provider-branch wiring. Fifth bump on this modal (969->989->993->998->1030->1056); structural shrink tracked in #3501.",
+    "_rebaseline_2026_07_19_7787_ic2_localdb_reexports": "PR #7787 (IC2 raw connections cache + lazy-decrypt) own growth: localDb.ts 805->807 (gate units, +2). localDb.ts is the re-export-only layer (hard rule #2 — no logic); the PR adds 4 new db/readCache re-exports (touchConnectionLastUsed, getCachedRawProviderConnections, getCachedProviderConnectionById, getCachedProviderNodes) required by existing barrel importers. Irreducible for a re-export list; frozen so it can only shrink.",
+    "_rebaseline_2026_07_20_7779_routingcombo_thread": "PR #7779 own growth: chatHelpers.ts 876->877 (+1, thread routingComboId into executeChatWithBreaker for compression-combo assignment). Frozen so it can only shrink.",
+    "_rebaseline_2026_07_20_7819_autocandidateoverrides_reexport": "PR for #7819 (Level 1+2: read-only auto/* candidate transparency + per-API-key exclusions) own growth: localDb.ts 807->808 (+1). Adds a single `export * from \"./db/autoCandidateOverrides\"` barrel re-export (hard rule #2 — no logic) for the new DB module backing per-apiKey candidate exclusions. Irreducible for a re-export list; frozen so it can only shrink.",
+    "_rebaseline_2026_07_21_8027_grok_cli_auth_json_paste": "PR #8027 (RaviTharuma, fix(grok-cli) #7610) own growth: OAuthModal.tsx 1080->1100 (gate units). Requires the full ~/.grok/auth.json (with refresh_token) on the paste-import path instead of a bare JWT, at the existing paste-token chokepoint (renamed tab label, updated instructions/placeholder, textarea for the auth.json blob, inline error surface). The validation logic itself (parseGrokCliPasteToken, previously an inline ~75-line function) was extracted to the new src/lib/oauth/utils/grokCliAuthJson.ts leaf module — mirroring the #6636/#7546 extraction precedent — so only the irreducible UI wiring remains here. Sixth bump on this modal (969->989->993->998->1030->1056->1100); structural shrink tracked in #3501.",
+    "_rebaseline_2026_07_21_8034_compression_exclusions_sidebar": "#8034 (compression exclusions dashboard tab) own growth: sections.ts 796->806 (+10, one new COMPRESSION_CONTEXT_GROUP sidebar item linking /dashboard/compression/exclusions). The file was already 796/800 before this PR (organic growth from prior sidebar entries), so a single new nav item pushed it 6 lines over cap. Freezing at 806 (cannot grow further); the sidebar item array is data, not extractable logic.",
+    "_rebaseline_2026_07_22_7936_namespace_roundtrip": "#7936 (@RCrushMe, Responses-Chat namespace round-trip identity seam) own growth: open-sse/translator/response/openai-responses.ts 1092->1125 (+33) and open-sse/utils/stream.ts 2814->2869 (+55) — threading the namespace-identity seam through the Responses↔Chat translation + stream paths so tool-call namespaces survive the round-trip. Cohesive translation/stream wiring at existing chokepoints, frozen at new size.",
+    "_rebaseline_2026_07_22_8010_codex_responses_engine": "PR #8010 (@JxnLexn) own growth: open-sse/mcp-server/schemas/tools.ts 1497->1505 (+8 = threading the new \"codex-responses\" literal into the compressionConfigureInput strategy/autoTriggerMode Zod enums and setCompressionEngineInput engine enum, mirroring the existing rtk/omniglyph enum entries; no new tool). open-sse/services/compression/strategySelector.ts 1043->1054 (+11 = one new `if (mode === \"codex-responses\")` dispatch branch in runCompression that delegates 100% to the new codexResponsesEngine.apply, mirroring the existing rtk single-mode dispatch, plus threading config.codexResponsesConfig.preserveToolNames into the shared adaptBodyForCompression call at the 3 existing call sites). src/lib/db/compression.ts (untracked, new-file cap 800) 794->845 (+51 = normalizeCodexResponsesConfig, mirroring the existing normalizeRtkConfig normalizer, plus registering \"codex-responses\" in the COMPRESSION_MODES/STACKED_PIPELINE_ENGINE_IDS/SINGLE_MODE_ENGINE sets and the getCompressionSettings load/save switch) — added to the baseline at its current size. All three are cohesive dispatch/normalizer wiring at existing chokepoints (mirroring the prior compression-mode rebaselines #6534/#6556), not extractable without hiding the mode-dispatch boundary. Covered by tests/unit/compression/codex-responses.test.ts (6) + omniglyph-registries.test.ts/types.test.ts (22, updated for the new mode).",
+    "_rebaseline_2026_07_22_8034_compression_exclusions_persistence": "#8034 (compression exclusions) own growth: src/lib/db/compression.ts 845->850 (+5 = threading the new compressionExclusions field through the existing getCompressionSettings/saveCompressionSettings load/save switch over the shared key_value compression namespace — no new table, no raw SQL). Mirrors the prior compression-field rebaselines (#8010 codex-responses normalizer at the same chokepoint); the load/save switch is a single dispatch boundary, not extractable without hiding it. Covered by the PR's 8 node:test + 3 vitest cases.",
+    "_rebaseline_2026_07_22_8050_model_lockout_exact_family": "#8050 (@AndrianBalanescu) own growth: accountFallback.ts 1864->1892 (+28) — exact-vs-family model-lockout scoping (getModelLockKey/isModelLocked/clearModelLock/getModelLockoutInfo) so an Antigravity 404 for one bare model no longer hijacks the whole family cooldown. Cohesive lockout logic; frozen at new size.",
+    "_rebaseline_2026_07_22_8056_headroom_minrows": "#8056 (@RaviTharuma, persist Headroom minRows) own growth: src/lib/db/compression.ts 850->866 (+16 HeadroomConfig+DEFAULT_HEADROOM_CONFIG+normalize/store in get/updateCompressionSettings) and open-sse/services/compression/strategySelector.ts 1054->1060 (+6 merge settings.headroom into stacked stepConfig). Cohesive settings-persistence + stacked-merge wiring at existing chokepoints, frozen at new size.",
+    "_rebaseline_2026_07_22_8081_reasoning_placeholder_guard": "#8081 (@Dingding-leo) own growth: openai-responses.ts 1125->1137 (+12) restructuring the reasoning-placeholder guard so it skips only the empty content block and still emits finish_reason/tool_calls in the same chunk. Cohesive translator wiring; frozen at new size.",
+    "_rebaseline_2026_07_22_8210_openrouter_midstream_error": "PR #8210 (hartmark, fix/openrouter-midstream-error-surfacing) own growth: open-sse/translator/response/openai-responses.ts 1137->1163 (+26) measured on the merged tip (release 1137 + this PR own growth). Adds a single new branch inside openaiToOpenAIResponsesResponse() that detects an OpenRouter-style mid-stream aggregator error (HTTP 200 SSE chunk with empty choices + a top-level error object) and surfaces it as state.upstreamError instead of silently falling through to the no-op/awaitingTrailingUsage path, which previously masked the failure as a false empty-success completion and skipped combo fallback. Irreducible call-site addition at the existing chunk-dispatch chokepoint (mirrors the Gemini-to-OpenAI translator's #4177 precedent for the same class of upstream error surfacing). Note: this baseline entry does NOT cover the separate pre-existing +11 drift already on the release tip from #8081/#8162 (1125->1136, unrelated reasoning-placeholder-stripping fix merged after this PR branched) — that drift belongs to the maintainer's rebaseline, not this PR.",
+    "_rebaseline_2026_07_22_8211_gemini_malformed_tool_choice": "PR #8211 (hartmark, fix/gemini-malformed-function-call-tool-choice) own growth: open-sse/translator/response/gemini-to-openai.ts 771->821 (+50, entirely this PR's diff — no other commit touched this file between the PR's merge-base and the release tip). Adds MALFORMED_FUNCTION_CALL/UNEXPECTED_TOOL_CALL handling inside geminiToOpenAIResponse(): synthesizes a `malformed_tool_call` tool_calls entry so finish_reason normalizes to the standard \"tool_calls\" instead of an unrecognized raw enum value that OpenAI-compatible clients (e.g. OpenClaw) silently ignore, and always synthesizes (rather than skipping when a real tool call already exists) so a malformed attempt alongside a real one in the same turn is not silently discarded. Irreducible cohesive addition at the existing candidate/finishReason translation chokepoint (mirrors the 9router#2462 raw-finish-reason precedent immediately below it in the same function). Covered by the PR's own tests/unit test additions for both the malformed-only and malformed-plus-real-call cases.",
+    "_rebaseline_2026_07_22_8213_chat_abandoned_target_abort": "PR #8213 (hartmark, fix/gemini-tpm-quota-cooldown-wait) own growth: src/sse/handlers/chat.ts 1794->1860 (+66, measured against the PR's own merge-base — the release tip separately carries an unrelated -5 net shrink from #8013's antigravity callable-catalog alignment, which this PR's branch does not include and this entry does not cover). Adds resolveDispatchClientRawRequest(): merges a per-target modelAbortSignal into clientRawRequest.signal (via mergeAbortSignals) so a combo target abandoned by comboTargetTimeoutMs actually observes its own abort and reaches its cleanup path, instead of hanging forever inside withRateLimit/acquireAccountSemaphore and leaking a permanent 'pending' dashboard entry (live incident, log id 1784418258231-14961a). Also wires combo-exhausted rejection logging to capture request body + attempted models via the new rejectedRequestUsage helper. Irreducible additions at the existing chat dispatch chokepoint. Covered by the PR's own combo-config + integration test additions.",
+    "_rebaseline_2026_07_22_8213_combo_cooldown_wait_recording": "PR #8213 (hartmark, fix/gemini-tpm-quota-cooldown-wait) own growth: open-sse/services/combo.ts 3548->3604 (+56, entirely this PR's diff — no other commit touched this file between the PR's merge-base and the release tip). Fixes combo cooldown-wait state recording so a bogus 503 is no longer crystallized when the cooldown-wait vars reset every setTry, adds an OpenAI-format SSE error frame path for combo-exhausted rejections (capturing request body + attempted models), and gives an abandoned per-target dispatch its own timeout instead of leaking a permanent 'pending' dashboard entry. Irreducible additions at the existing handleComboChat dispatch/retry chokepoint (mirrors the prior quota-share/headroom/task-aware strategy-branch precedents already frozen in this file). Covered by the PR's own combo-config + Gemini TPM-ceiling benchmark test additions.",
+    "_rebaseline_2026_07_22_8213_gemini_tpm_quota_cooldown_wait": "PR #8213 (hartmark, fix/gemini-tpm-quota-cooldown-wait) own growth: open-sse/services/accountFallback.ts 1857->1932 on the merged tip (release 1892 incl #8050 +35, plus this PR own growth +40); measured against the PR own merge-base was 1857->1898 (+41 — the release tip separately carries an unrelated +34 from #8050's antigravity 404 model-not-found lockout scoping, which this PR's branch does not include and this entry does not cover). Own growth is the Gemini TPM-ceiling classification + cooldown-wait wiring feeding into the combo cooldown-wait state machine (rate-limit wedge recovery) introduced by this PR's commit series. Irreducible additions at the existing account-fallback/model-lockout chokepoint. Covered by the PR's own gemini-rate-limit-tracker and TPM-ceiling benchmark test additions.",
+    "_rebaseline_2026_07_22_8213_health_unblock_model_cooldowns": "PR #8213 (hartmark, fix/gemini-tpm-quota-cooldown-wait) own growth: src/app/(dashboard)/dashboard/health/page.tsx 1094->1165 (+71, entirely this PR's diff — no other commit touched this file between the PR's merge-base and the release tip). Adds handleUnblockAll/handleUnblockOne dashboard actions (DELETE /api/resilience/model-cooldowns) so an operator can manually clear a Gemini TPM-wedge model lockout surfaced by this PR's cooldown-wait fixes, instead of waiting out the ceiling. Irreducible UI wiring at the existing health-page action chokepoint. Covered by the PR's own dashboard/resilience test additions.",
+    "_rebaseline_2026_07_22_8213_requestloggerdetail_unblock_ui": "PR #8213 (hartmark, fix/gemini-tpm-quota-cooldown-wait) own growth: src/shared/components/RequestLoggerDetail.tsx 799->941 (+142, entirely this PR's diff — no other commit touched this file between the PR's merge-base and the release tip; crosses the general 800-line new-file cap so is frozen here for the first time). Adds a collapsible section header (open/expand-less toggle) plus per-log-entry unblock (`unblocking`/`cleared` state, isCombo503 detection) so the request-logger detail panel surfaces the same Gemini TPM cooldown-wait / model-lockout unblock action introduced by this PR at the individual-request level (mirrors the health-page bulk unblock action added in the same PR). Covered by the PR's own dashboard/resilience test additions.",
+    "_rebaseline_2026_07_22_fusion_8013_8098_antigravity": "Fusion of #8013 (backryun, catalog/IDE-CLI-split rewrite) + #8098 (nguyenha935, protocol-fidelity/fail-closed/credits/tool-cloaking): open-sse/services/usage/antigravity.ts NEW 802 (>cap 800, +2 — #8098 credits/tier usage service on #8013's profile-aware headers). Test growth (models-catalog-route 1605->1608, provider-models-route 1752->1757 from #8013 Gemini 3.6 catalog) tracked in testFrozen.",
+    "_rebaseline_2026_07_23_8127_grok_weekly_quota": "#8127 (@apoapostolov) own growth: src/sse/handlers/chat.ts 1861->1865 (+4) — weekly quota tracking for grok-web wires a quota-fetch hook at the existing dispatch chokepoint. Thin wiring mirroring adjacent provider-quota branches; not extractable. Covered by tests/unit/grok-quota-fetcher.test.ts.",
+    "_rebaseline_2026_07_23_8143_empty_catch_logging": "#8143 (@chirag127) own growth: open-sse/utils/stream.ts 2869->2887 (+18) — replacing empty catch blocks in the SSE stream subsystem with console.debug logging (Rule #6 silent-swallow fix, issues #8138-#8142). Cohesive logging additions at the existing catch chokepoints, not extractable; frozen at new size. Covered by tests/unit/stream-handler-catch-logging-8143.test.ts.",
+    "_rebaseline_2026_07_23_8219_cache_ttl_settings_sidebar": "#8219 (@oyi77) own growth: sections.ts 806->813 (+7) — configurable model-catalog cache-TTL settings adds a new sidebar nav entry + its visibility wiring. Sidebar item array is data, not extractable logic; frozen at new size.",
+    "_rebaseline_2026_07_23_8247_8248_model_unhealthy": "#8247+#8248 own growth: accountFallback.ts 1940->1941 (+1, irreducible import statement only — the substantive #8248 DEGRADED-pattern classifier was extracted into open-sse/config/errorConfig.ts, which has ample headroom, instead of growing this frozen file; #8247's fix is a single existing-line condition change, net zero lines). Scoping the credits-exhausted 403/429 branch to isCompatibleProvider() (per-model-quota openai/anthropic-compatible-* nicknames) so it stays model-scoped instead of terminalling the whole connection, and classifying NVIDIA NIM 'Function ... DEGRADED' 400 bodies as model-access-denied instead of a raw passthrough 400. Covered by tests/unit/8247-accountfallback-model-unhealthy.test.ts and tests/unit/8248-accountfallback-nvidia-degraded.test.ts.",
+    "_rebaseline_2026_07_23_8252_combo_400_advance": "#8252 (@RaviTharuma) own growth: accountFallback.ts 1932->1940 (+8) + combo.ts 3604->3630 (+26) — advance combo on model-scoped 400s wrapped as invalid/Bad-Request. Irreducible wiring at existing account-fallback + combo dispatch chokepoints. Covered by combo-model-scoped-400-advance.test.ts.",
+    "_rebaseline_2026_07_23_8266_alibaba_media": "#8266 (@backryun) own growth: imageRegistry.ts 821->979 (+158) — Alibaba-family media models (Qwen image/video, Bailian, Wan) added to the image/video registry. Registry model data, not extractable logic; frozen at new size.",
+    "_rebaseline_2026_07_24_8388_compression_detail_persist": "#8388 (compression engine DETAIL settings — Headroom/session-dedup/CCR — dropped on save) own growth: src/lib/db/compression.ts 866->872 (+6 = irreducible call-site wiring at the existing getCompressionSettings/updateCompressionSettings chokepoint: one import line, one `...buildDetailConfigDefaults()` spread in the seed config, and one `case \"sessionDedup\": case \"ccr\": applyDetailConfigUpdate(config, key, parsed); break;` load-switch case, mirroring the existing headroom/#8056 case immediately above it). The actual normalizer logic (normalizeSessionDedupConfig/normalizeCcrConfig, matching SESSION_DEDUP_SCHEMA/CCR_SCHEMA bounds) was EXTRACTED into a new leaf src/lib/db/compressionDetailNormalizers.ts (well under cap) so this frozen file only carries the minimal dispatch wiring. Covered by tests/unit/8388-compression-detail-persist.test.ts (schema-accept + full DB save->reload round-trip for both new sub-objects, plus a no-regression assertion on the existing headroom round-trip).",
+    "_rebaseline_2026_07_24_responses_toolcalls_log_summary": "hartmark, fix/responses-tool-calls-log-summary own growth: open-sse/translator/response/openai-responses.ts 1163->1174 (+11). closeToolCall() now also writes the completed tool call into the shared state.toolCalls Map (already populated by the openai-to-claude / claude-to-openai / gemini-to-openai response translators) so stream.ts's completion-log summary builder (which reads state.toolCalls, not this translator's own funcCallIds/funcNames/funcArgsBuf bookkeeping) reports finish_reason \"tool_calls\" and message.tool_calls for openai->openai-responses translated streams instead of always logging \"stop\" with no tool_calls — the actual client-facing SSE events were already correct; only the persisted call-log summary was wrong. Irreducible call-site addition at the existing tool-call-close chokepoint. Covered by the new regression test in tests/unit/translator-resp-openai-responses.test.ts.",
+    "_rebaseline_2026_07_25_8476_combo_input_bound_homogeneous_scope": "PR #8476 (herjarsa, fix/8375-8459-combo-image-fixes, #8375) own growth: open-sse/services/combo.ts 3642->3679 (+37 net: +29 the PR's own isInputBoundFailure short-circuit for deterministic context_length_exceeded/context_window_exceeded failures, +8 a /green-prs pre-merge fix scoping that short-circuit to homogeneous remainders only — the shipped code fired unconditionally on ANY target, regressing the intentional heterogeneous-combo fallback #6637/isContextOverflow400 protects, exactly as flagged by this PR's own review evidence but never actually implemented in the branch). The fix compares orderedTargets[i+1..] modelStr against the failing target's modelStr at the existing executeTarget dispatch chokepoint (mirrors the sameProviderNext precedent a few lines below) — irreducible call-site wiring, not extractable without hiding the dispatch boundary. Covered by tests/unit/combo-input-bound-failure-8375.test.ts (homogeneous pool still short-circuits) and the new tests/unit/combo-input-bound-heterogeneous-8375.test.ts (heterogeneous combo now correctly falls through to the larger-context target).",
+    "_rebaseline_2026_07_25_adobe_firefly_reference_images": "Follow-up to #8006: storage upload + referenceBlobs for image/video and /v1/images/edits dispatch. adobeFireflyClient.ts 1958->2317 (+upload helpers, extract sources, resolve blob ids). Note: 2317 not 2316 — check-file-size.mjs counts LOC via split(\"\\n\").length (counts the trailing-newline empty element), which is 1 higher than `wc -l` on a file ending in \\n; the PR's original entry (2316) was measured with wc -l and undercounted by 1 against the actual gate.",
+    "_rebaseline_pr1043_minimax_tts": "Upstream port decolua/9router#1043 (toanalien) own growth: audioSpeech.ts 965->1061 (+96). Adds MiniMax T2A v2 TTS dispatch (handleMinimaxSpeech + hexToBytes helper) — provider entry was already in audioRegistry (format: minimax-tts) but no handler existed, falling through to the OpenAI-compatible default that fails (T2A has custom shape + hex-encoded audio + base_resp envelope). New branch sits next to the other inline provider branches (xiaomi-mimo, coqui, tortoise, aws-polly) — extracting would just create indirection. Covered by tests/unit/minimax-tts-1043.test.ts (3 tests, GREEN: success, base_resp error, invalid-hex).",
+    "_rebaseline_pr4592_exclude_exhausted_auto": "Reconcile #4592 already-merged growth: combo.ts 2991->3036 (+45, terminal-status quota-cutoff exclusion in buildAutoCandidates + opt-in gate). Fast-gate PR->release does not run check:file-size.",
+    "open-sse/executors/antigravity.ts": 1528,
+    "open-sse/executors/base.ts": 1640,
+    "open-sse/executors/chatgpt-web.ts": 3241,
+    "open-sse/executors/codex.ts": 1562,
+    "open-sse/executors/cursor.ts": 1563,
+    "open-sse/executors/deepseek-web.ts": 1148,
+    "open-sse/executors/grok-web.ts": 1044,
+    "open-sse/executors/muse-spark-web.ts": 1405,
+    "open-sse/handlers/chatCore.ts": 5034,
+    "open-sse/handlers/imageGeneration.ts": 3101,
+    "open-sse/handlers/responseSanitizer.ts": 1128,
+    "open-sse/handlers/search.ts": 1536,
+    "open-sse/handlers/videoGeneration.ts": 1063,
+    "open-sse/mcp-server/schemas/tools.ts": 1553,
+    "open-sse/mcp-server/server.ts": 1448,
+    "open-sse/mcp-server/tools/advancedTools.ts": 1120,
+    "open-sse/services/accountFallback.ts": 1978,
+    "open-sse/services/adobeFireflyClient.ts": 2385,
+    "open-sse/services/claudeCodeCompatible.ts": 1202,
+    "open-sse/services/combo.ts": 3648,
+    "open-sse/services/compression/strategySelector.ts": 1060,
+    "open-sse/services/rateLimitManager.ts": 1167,
+    "open-sse/translator/response/openai-responses.ts": 1224,
+    "open-sse/utils/cursorAgentProtobuf.ts": 1505,
+    "open-sse/utils/stream.ts": 2889,
+    "src/app/(dashboard)/dashboard/HomePageClient.tsx": 1388,
+    "src/app/(dashboard)/dashboard/analytics/ComboHealthTab.tsx": 1031,
+    "src/app/(dashboard)/dashboard/api-manager/ApiManagerPageClient.tsx": 3117,
+    "src/app/(dashboard)/dashboard/cache/media/MediaPageClient.tsx": 1067,
+    "src/app/(dashboard)/dashboard/combos/page.tsx": 4703,
+    "src/app/(dashboard)/dashboard/costs/CostOverviewTab.tsx": 1283,
+    "src/app/(dashboard)/dashboard/costs/quota-share/components/PoolWizard.tsx": 1022,
+    "src/app/(dashboard)/dashboard/endpoint/EndpointPageClient.tsx": 2615,
+    "src/app/(dashboard)/dashboard/health/page.tsx": 1165,
+    "src/app/(dashboard)/dashboard/providers/[id]/components/modals/EditConnectionModal.tsx": 1324,
+    "src/app/(dashboard)/dashboard/providers/page.tsx": 1944,
+    "src/app/(dashboard)/dashboard/runtime/RuntimePageClient.tsx": 1201,
+    "src/app/(dashboard)/dashboard/settings/components/PricingTab.tsx": 1019,
+    "src/app/(dashboard)/dashboard/settings/components/ProxyRegistryManager.tsx": 1470,
+    "src/app/(dashboard)/dashboard/settings/components/ResilienceTab.tsx": 1123,
+    "src/app/(dashboard)/dashboard/settings/components/RoutingTab.tsx": 1629,
+    "src/app/(dashboard)/dashboard/settings/components/SystemStorageTab.tsx": 1573,
+    "src/app/(dashboard)/dashboard/usage/components/BudgetTab.tsx": 1028,
+    "src/app/(dashboard)/dashboard/usage/components/EvalsTab.tsx": 2148,
+    "src/app/(dashboard)/dashboard/usage/components/ProviderLimits/index.tsx": 1119,
+    "src/app/api/providers/[id]/models/route.ts": 2361,
+    "src/app/api/v1/models/catalog.ts": 1597,
+    "src/lib/db/apiKeys.ts": 1529,
+    "src/lib/db/core.ts": 1639,
+    "src/lib/db/migrationRunner.ts": 1094,
+    "src/lib/db/models.ts": 1097,
+    "src/lib/db/providers.ts": 1034,
+    "src/lib/memory/retrieval.ts": 1073,
+    "src/lib/tailscaleTunnel.ts": 1202,
+    "src/lib/usage/providerLimits.ts": 1013,
+    "src/shared/components/OAuthModal.tsx": 1134,
+    "src/shared/components/RequestLoggerV2.tsx": 1629,
+    "src/shared/components/analytics/charts.tsx": 1035,
+    "src/shared/services/cliRuntime.ts": 1122,
+    "src/sse/handlers/chat.ts": 1904,
+    "src/sse/services/auth.ts": 2520,
+    "tests/unit/account-fallback-service.test.ts": 1572,
+    "tests/unit/provider-validation-specialty.test.ts": 2985,
+    "open-sse/executors/hyperagent.ts": 1026,
+    "src/lib/tokenHealthCheck.ts": 1053,
+    "open-sse/executors/default.ts": 1042,
+    "open-sse/executors/kiro.ts": 1069,
+    "open-sse/translator/request/openai-to-kiro.ts": 1057,
+    "open-sse/utils/sseHeartbeat.ts": 149
+  },
   "_rebaseline_2026_07_27_v3849_train2": "Merge-train 2 (7 PRs) — owner-approved 2026-07-27. Single entry: chatCore.ts 4955->5006 (#8595, Responses multi-turn image compaction before the context hard-reject). Genuine irreducible growth at the existing compaction chokepoint in handleChatCore — the PR adds a last-resort retry against the concrete budget plus the estimateFinalInputTokens helper, both wired at the pre-existing call site rather than a new branch. Covered by tests/unit/8560-responses-image-compaction.test.ts (4 tests).",
   "_rebaseline_2026_07_27_v3849_train3": "Merge-train 3 (13 PRs) — owner-approved 2026-07-27. Both entries are genuine irreducible growth at existing chokepoints, not new branches: src/lib/db/apiKeys.ts 1518->1529 (#8805 cx/* ≡ codex/* API-key model permissions); open-sse/handlers/chatCore.ts 5006->5020 (#8806 real response payload into plugin onResponse hooks). Covered by tests/unit/db-apiKeys-crud.test.ts (4 new cases) and the two plugin-hook test files updated in #8806 respectively.",
   "_rebaseline_2026_07_28_8842_antigravity_projectid_refresh": "PR #8842 (fix/antigravity-projectid-refresh) own growth: open-sse/executors/antigravity.ts 1493->1528 (+35 = projectId discovery in refreshCredentials: import ensureAntigravityProjectAssigned + trim projectId + call ensureAntigravityProjectAssigned with 8s timeout + persistDiscoveredAntigravityProjectId + log success/failure). Irreducible wiring at the existing credential-refresh chokepoint. Covered by tests/unit/executor-antigravity.test.ts (4 new test cases).",
@@ -418,15 +419,148 @@
   "_rebaseline_2026_07_28_8860_tokenrefresh_projectid": "PR #8860 (fix/antigravity-projectid-centralized) own test growth: tests/unit/token-refresh-service.test.ts 1311->1378 (+67 = 4 cases covering projectId discovery on the tokenRefresh.ts path — the Dashboard/health-check refresh route, which #8842 did not reach since that fixed the executor path). Covered by the same file.",
   "_rebaseline_2026_07_28_8861_xiaomi_token_plan": "PR #8861 (feat/xiaomi-token-plan-protocol-selector) own growth: EditConnectionModal.tsx 1283->1316 (+33 = the per-connection API-protocol selector field) and open-sse/executors/base.ts 1540->1562 (+22 = alternate-format resolution at the existing buildUrl/headers chokepoint). Both are irreducible wiring at existing call sites.",
   "_rebaseline_2026_07_28_8863_firefly_detail_level": "PR #8863 (fix/adobe-firefly-gpt-detail-level-max) own growth: adobeFireflyClient.ts 2317->2322 (+5 = gpt-image detailLevel defaulting to maximal at the existing payload-build site). Covered by tests/unit/adobe-firefly.test.ts.",
-  "_rebaseline_2026_07_28_8870_firefly_ref_cap_timeout": "PR #8870 (fix/adobe-firefly-gpt-ref-cap-timeout) own growth: adobeFireflyClient.ts 2322->2385 (+63 = gpt-image subject-ref hard cap at 2 + adaptive poll timeout budget (base 300s + 60s/ref, max 600s) + defensive .slice on referenceBlobs for gpt/nano/generic families). Fixes live 504s on multi-screenshot listing jobs (Featured Promo / Box Art) where 3–4+ subject refs stall colligo until the old 180s poll budget expires. Helpers adobeFireflyMaxImageRefs/adobeFireflyImageTimeoutMs live next to the existing payload/poll chokepoint (not extractable without splitting the wire recipe mid-PR). Covered by tests/unit/adobe-firefly.test.ts (ref-cap + timeout cases). Structural shrink tracked in #3501.",
   "_rebaseline_2026_07_29_8281_home_quickstart_prefetch": "Release v3.8.49 base-red fix (no PR — captain sweep): src/app/(dashboard)/dashboard/HomePageClient.tsx 1377->1381 (+4). #8292 added prefetch={false} to the sidebar but left /home's five quick-start Links prefetching, so first paint still fired 12 speculative RSC requests — caught by navigation.spec.ts only after the e2e helper bug (APP_ROUTE_PATTERN missing /home) was repaired in the same cycle. Growth is the five prefetch attributes; it was offset first by extracting the repeated className literals (INLINE_LINK x4, DOCS_LINK x1), which collapsed five wrapped <Link> blocks back to one line each — a naive fix measured 1391. Guard: tests/unit/sidebar-prefetch-policy-8281.test.ts.",
+  "_rebaseline_2026_08_02_v3850_agentrouter_responses": "Release v3.8.50 AgentRouter/Codex compatibility reconciliation. open-sse/executors/base.ts 1562->1578: #9190 wires AgentRouter's selected Claude/OpenAI/Responses protocol through the existing executor URL, auth, identity-header and fingerprint chokepoints; the reusable alternate resolver remains outside base.ts. open-sse/utils/stream.ts 2887->2889: #9213 evaluates Responses ID and usage normalization independently so response.completed always receives finite usage.total_tokens instead of short-circuiting after an ID rewrite. tests/unit/chatcore-translation-paths.test.ts 2769->2776: #9191 updates the existing Claude-Code bridge assertions for the dynamic AgentRouter wire image. PR #9224 offsets its own chatCore growth by extracting the AgentRouter protocol decisions into chatCore/agentRouterProtocol.ts, leaving chatCore below its frozen ceiling. Covered by agentrouter executor/chatCore protocol tests, chatcore translation-path tests, and responses-commentary-passthrough tests.",
+  "_rebaseline_2026_08_08_v3850_base_drift_batch_9757": "Base drift on release/v3.8.50, not own growth: the 08-06..08-08 merge batches grew 12 already-frozen (or newly-landed) files without carrying their rebaselines — the dedicated rebaseline PR #9616 was closed as 'superseded' but its file-size entries never actually reached the base, and later merges (#8894 combos page, #9539 EditConnectionModal, #8895 models route, #9294/#9293 catalog, #9541 db/core, #8970 tokenHealthCheck, #8925 mcp schemas+server, #8890 accountFallback, #9467 chat.ts, #8931 openai-to-kiro, ProxyRegistryManager) kept growing them. All 12 values re-measured on THIS branch's tree (= pure tip + this PR's 1-line chat.ts fix, which adds zero lines). This PR's own source changes (chat.ts identifier restore, stream.ts format carve-out) do not grow any frozen file past these values.",
+  "_rebaseline_2026_08_08_migration_135_collision": "fix(db): resolve migration version 135 numbering collision — #9449's 135_connection_runtime_state.sql and #8908's 135_migrate_model_capability_max_token.sql both claimed version 135 (#9449 branched before #8908 merged and never got renumbered before landing on release/v3.8.50), which threw 'Migration version collision detected' the moment ANY code touched the database — a fresh install/deploy from this tip cannot even boot. Renumbered the later-landing file to 140 (next free slot) and added the matching isSchemaAlreadyApplied('140') retroactive guard, matching the established pattern already used for the prior 135/136 -> 137/138 renumber in the same file. Own growth: src/lib/db/migrationRunner.ts 1084->1094 (+10, the new case block) — irreducible, matches the existing per-case guard pattern exactly. Covered by tests/unit/migration-135-numbering-collision.test.ts (2/2), confirmed failing (reproducing the exact live crash) against the pre-fix colliding filenames, passing after.",
+  "_rebaseline_2026_08_02_9259_rolling_rpm": "PR #9259 (issue #8733) own growth: open-sse/services/rateLimitManager.ts baseline 1060->1167 (+107; final source 1153). The existing withRateLimit chokepoint now composes process-local rolling RPM leases with Bottleneck admission, releases pre-dispatch leases on queue timeout/abort/connection disable, preserves caller abort reasons, and wires 429/header state into the extracted rollingRpmGate.ts. The remaining growth is irreducible lifecycle wiring at the dispatch boundary plus the real watchdog test hooks needed to verify queued-wedge recovery; moving it further would obscure lease ownership and Bottleneck cleanup. Covered by the focused rate-limit manager/sliding-window suite (33/33); distributed multi-instance coordination remains explicitly out of scope.",
+  "_rebaseline_2026_07_25_dario_upstream_proxy_selector": "PR #8523 (Dario embedded service): upstream-proxy mode selector replaces the binary CLIProxyAPI toggle with Native/CLIProxyAPI/Dario/Fallback + a fallback-backend picker. ProviderDetailPageClient.tsx 798->804 (+6, new hook fields threaded through to ConnectionsListPanel), ConnectionRow.tsx 942->958 (+16, the mode <select> + conditional fallback-backend <select> replacing a single pill button), useProviderConnections.ts 954->986 (+32, upstreamProxyMode/upstreamProxyFallbackBackend state + handleSetUpstreamProxyMode, handleToggleCliproxyapiMode kept as a thin backward-compat wrapper for the existing hook-shape test). All additive UI/state for the new modes — no unrelated refactor.",
+  "_rebaseline_2026_08_02_9242_token_health_transient": "PR #9242 (fix/refresh-circuit-transient): src/lib/tokenHealthCheck.ts 1021 (new file, above cap 1000). The file consolidates token-refresh health checking logic that was previously scattered across auth.ts and tokenRefresh.ts. Cohesive single-responsibility module for refresh circuit state management; not extractable without splitting the refresh state machine. Covered by tests/unit/tokenHealthCheck-transient.test.ts.",
+  "_rebaseline_2026_07_28_8870_firefly_ref_cap_timeout": "PR #8870 (fix/adobe-firefly-gpt-ref-cap-timeout) own growth: adobeFireflyClient.ts 2322->2385 (+63 = gpt-image subject-ref hard cap at 2 + adaptive poll timeout budget (base 300s + 60s/ref, max 600s) + defensive .slice on referenceBlobs for gpt/nano/generic families). Fixes live 504s on multi-screenshot listing jobs (Featured Promo / Box Art) where 3–4+ subject refs stall colligo until the old 180s poll budget expires. Helpers adobeFireflyMaxImageRefs/adobeFireflyImageTimeoutMs live next to the existing payload/poll chokepoint (not extractable without splitting the wire recipe mid-PR). Covered by tests/unit/adobe-firefly.test.ts (ref-cap + timeout cases). Structural shrink tracked in #3501.",
   "_rebaseline_2026_08_01_8964_xai_agent_tools": "PR #8964 own growth: chatCore.ts 5020->5034 at the existing native-passthrough chokepoint. Adds xAI Agent Tools passthrough for /v1/responses (xai/xai-oauth/xao): resolve nativeXaiResponsesPassthrough, force openai-responses targetFormat, stamp body marker, and OR into the existing nativeCodexPassthrough sites (web-search bypass + requestEndpointPath). Leaf logic in passthroughHelpers, responsesEndpoint, targetFormat, xai executor, responseSanitizer, usageTracking. Cohesive wiring at the Codex passthrough boundary.",
   "_rebaseline_2026_08_01_8964_response_sanitizer": "PR #8964 own growth: responseSanitizer.ts 1115->1128. Keep cost_in_usd_ticks / server_side_tool_usage(_details) through sanitizeResponsesApiResponse allowlists so native xAI tool responses retain usage.",
-  "_rebaseline_2026_08_02_v3850_agentrouter_responses": "Release v3.8.50 AgentRouter/Codex compatibility reconciliation. open-sse/executors/base.ts 1562->1578: #9190 wires AgentRouter's selected Claude/OpenAI/Responses protocol through the existing executor URL, auth, identity-header and fingerprint chokepoints; the reusable alternate resolver remains outside base.ts. open-sse/utils/stream.ts 2887->2889: #9213 evaluates Responses ID and usage normalization independently so response.completed always receives finite usage.total_tokens instead of short-circuiting after an ID rewrite. tests/unit/chatcore-translation-paths.test.ts 2769->2776: #9191 updates the existing Claude-Code bridge assertions for the dynamic AgentRouter wire image. PR #9224 offsets its own chatCore growth by extracting the AgentRouter protocol decisions into chatCore/agentRouterProtocol.ts, leaving chatCore below its frozen ceiling. Covered by agentrouter executor/chatCore protocol tests, chatcore translation-path tests, and responses-commentary-passthrough tests.",
   "_rebaseline_2026_08_05_9323_agentrouter_waf_retry": "PR #9323 (fix(agentrouter): retry on 400 content-blocked + burst guard) own growth: open-sse/executors/base.ts 1578->1623 (check-file-size.mjs conta via split(\"\\n\").length; wc -l ve 1622). As +45 linhas sao o WAF_RETRY_CONFIG + o burst guard via gateOutboundRequest() para o WAF do agentrouter.org, com comentarios explicando o porque de cada mitigacao e cobertos por tests/unit/base-executor-waf-retry.test.ts e tests/unit/wafRateLimit.test.ts. Crescimento funcional legitimo, nao inchaco.",
   "_rebaseline_2026_08_05_9529_own_growth": "PR #9529 own growth (base release/v3.8.50 medida EXATAMENTE nos frozen antigos, entao o modo base-relative #8522 nao cobre): open-sse/services/rateLimitManager.ts 1060->1105 (+45: helper applyLimiterSettings() que re-arma o heartbeat do reservoir apos updateSettings — fix do bug Bottleneck 2.19.5 que congelava a fila weighted; TDD em tests/unit/ratelimit-reservoir-refresh.test.ts); tests/integration/chat-pipeline.test.ts 1592->1598 (+6: User-Agent do codex derivado de getCodexClientVersion() em vez de literal pinado — teste-irmao alinhado ao contrato); tests/unit/provider-validation-specialty.test.ts 2980->2985 (+5: cobertura NOVA claude-web 429 -> valid:false, alinhamento #9406); open-sse/translator/response/openai-responses.ts 1174->1204 (+30: buildResponsesReasoningSummaryDelta MOVIDA do leaf pureHelpers.ts para o host — a funcao do #9500 muta stream state e violava o contrato do leaf puro; o LOC total do par host+leaf nao cresceu, o pureHelpers encolheu o mesmo tanto). Crescimento por fix de producao + cobertura adicional + realocacao arquitetural, nao inchaco.",
   "_rebaseline_2026_08_06_v3850_inherited_drift_reconcile": "Reconciliacao 2026-08-06 do drift ACUMULADO da release/v3.8.50 apos o lote de merges de 08-05/06: 13 arquivos acima do frozen no tip puro 8180b49ce1 (medidos pelo proprio gate). O modo PR base-relative (#8522) deixa PRs inocentes passarem, e os rebaselines individuais dos PRs se perderam nas resolucoes sucessivas de conflito deste hot-file — o drift so aparece no modo absoluto (nightly/local). Crescimentos funcionais dos PRs mergeados: #9024 topology click-nav src/app/(dashboard)/dashboard/HomePageClient.tsx; #9324 OpenRouter enrich src/app/(dashboard)/dashboard/providers/page.tsx; #9329 quota card ordering src/app/(dashboard)/dashboard/usage/components/ProviderLimits/index.tsx; #9193 context-window suffixes src/sse/handlers/chat.ts; #9332 nested Claude server tool ids open-sse/executors/base.ts; #9228 strip orphaned tool outputs open-sse/executors/codex.ts; #9236 nvidia tool-name normalize open-sse/executors/default.ts; #9314 nested tool_call validation open-sse/executors/kiro.ts; #9260 caller identity REST hops open-sse/mcp-server/server.ts; #8934 cache breakpoints tests tests/unit/chatcore-translation-paths.test.ts; #9193 suffix tests tests/unit/combo-routing-engine.test.ts; #9196 reasoning-on-tool-finish tests tests/unit/sse-auth.test.ts; #9163 GPT-5.6 Max reasoning tests tests/unit/translator-openai-to-kiro.test.ts. default.ts e kiro.ts entram no frozen (estavam sem entrada, acima do cap 1000). Atualizacao pos-medicao (a base avancou durante o ciclo do PR): src/sse/handlers/chat.ts 1857->1877 (#9184 affinity EOF evict) e open-sse/executors/default.ts 1027->1042 (#9005 Kimi K3 tool-name backfill).",
   "_rebaseline_2026_08_06b_v3850_sweepreds_drift": "Segunda reconciliacao de 2026-08-06 (/sweep-reds sobre o tip puro 2ddbbc61a6): 3 arquivos voltaram a passar do frozen apos os merges do mesmo dia, com atribuicao 1:1 por commit. (1) src/app/(dashboard)/dashboard/providers/page.tsx 1928->1944 e (2) open-sse/executors/base.ts 1635->1640, ambos do #9515 (feat(radar): flag-gated signed free-model catalog overlay, commit e7f6b1d130) — o overlay do Radar entra por wiring nos chokepoints ja existentes (a resolucao/verificacao do catalogo assinado mora fora destes dois arquivos); +16 e +5 linhas liquidas nao sao extraiveis sem inventar um leaf por callsite. (3) open-sse/services/accountFallback.ts 1966->1972 do #8704 (commit c4527f97bd), +6 linhas de dados em CREDITS_EXHAUSTED_SIGNALS ('has been exhausted', fixes #8631). src/sse/handlers/chat.ts 1880>1877 tambem estava violando e NAO entra aqui de proposito: e drenado por encolhimento na PR #9598, sem rebaseline. Crescimento proprio DESTA PR: src/lib/db/migrationRunner.ts 1077->1084 (+7) — o guard retroativo em isSchemaAlreadyApplied para os arquivos renumerados 137/138, exigido pela propria mensagem de erro de colisao do runner (ambas as migracoes sao ALTER TABLE ADD COLUMN puro, nao idempotente). Dois `case` + dois `return hasColumn(...)` + 3 linhas de comentario dentro do switch existente; nao extraivel.",
   "_rebaseline_2026_08_06c_v3850_sweepreds_pr2": "Segunda PR do /sweep-reds (fix/release-v3.8.50-basereds-0806b): tests/unit/provider-models-route.test.ts 1784->1787 (medido pelo gate, que conta split(\"\\n\").length) (+2 apos compressao de comentarios) — alinhamento de contrato forcado por dois merges do dia: #9106 tornou gemini-3.1-pro-high user-callable (a entry do alias entra na lista esperada do teste de discovery-retry, +1 linha de dado + 1 de comentario) e ff012ff420 adicionou onboardUser como bootstrap hop (exclusao no mock, ja comprimida a 1 linha). Nao ha o que encolher sem apagar o comentario que explica o porque.",
-  "_rebaseline_2026_08_07_v3850_sweepreds_pr2_toolnamemap": "tests/unit/translator-openai-to-gemini.test.ts 1616->1619 (+3). O frozen estava EXATAMENTE no tamanho da base, entao qualquer linha nova viola. #9568 (c9a3361e5a) fez buildChangedToolNameMap emitir entradas IDENTIDADE (o Gemini minusculiza nomes de tool nas respostas, entao o tradutor de resposta precisa da chave para mapear de volta), o que passou a incluir `_toolNameMap` no envelope Antigravity de qualquer request com tools. As 3 linhas sao: a chave nova na lista esperada de Object.keys, 1 comentario explicando POR QUE ela aparece (sem ele o proximo leitor tenta remove-la de novo) e 1 assert do CONTEUDO do map — presenca de chave sozinha nao provaria a entrada identidade, que e justamente o comportamento novo. Nao ha o que extrair: e alinhamento de contrato dentro de um teste existente."
+  "_rebaseline_2026_08_07_v3850_sweepreds_pr2_toolnamemap": "tests/unit/translator-openai-to-gemini.test.ts 1616->1619 (+3). O frozen estava EXATAMENTE no tamanho da base, entao qualquer linha nova viola. #9568 (c9a3361e5a) fez buildChangedToolNameMap emitir entradas IDENTIDADE (o Gemini minusculiza nomes de tool nas respostas, entao o tradutor de resposta precisa da chave para mapear de volta), o que passou a incluir `_toolNameMap` no envelope Antigravity de qualquer request com tools. As 3 linhas sao: a chave nova na lista esperada de Object.keys, 1 comentario explicando POR QUE ela aparece (sem ele o proximo leitor tenta remove-la de novo) e 1 assert do CONTEUDO do map — presenca de chave sozinha nao provaria a entrada identidade, que e justamente o comportamento novo. Nao ha o que extrair: e alinhamento de contrato dentro de um teste existente.",
+  "_rebaseline_2026_06_22_4644_deepseek_web_tools": "PR #4644 (BugsBag/robust deepseek-web tool-call parsing): open-sse/executors/deepseek-web.ts 1117->1125 (+8). The new agentic tool-call path emits surrounding text + reasoning before tool_calls and swaps to the dedicated deepseekWebTools.ts parser; the +8 lines are cohesive wiring at the existing transformSSE chokepoint (the parser itself lives in the new deepseekWebTools.ts file, already under cap). The PR's own fast-gate (PR->release) does not run check:file-size, so this surfaced only at release reconcile. Covered by tests/unit/deepseek-web-tools-variants.test.ts + deepseek-web-tools-execute.test.ts.",
+  "_rebaseline_2026_06_23_4712_deepseek_web_tool_results": "PR for #4712 (deepseek-web drops role:tool): open-sse/executors/deepseek-web.ts 1125->1148 (+23). messagesToPrompt() now folds role:\\\"tool\\\" results into the single-prompt transcript (recovering the tool name from the preceding assistant tool_calls by tool_call_id) instead of silently dropping them; the lines are cohesive wiring inside the existing function. Covered by tests/unit/deepseek-web-tool-result-prompt-4712.test.ts.",
+  "_rebaseline_2026_06_24_headroom_strategy": "Headroom-aware connection selection (dario technique): combo.ts 3168->3180 (+12 = a new `else if (strategy === \\\"headroom\\\")` dispatch branch in handleComboChat that delegates to orderTargetsByHeadroom + its log line, plus the import). The actual logic lives OUT of the god-file: the pure ranker rankByHeadroom/computeHeadroom is the new leaf open-sse/services/combo/headroomRanking.ts (91 LOC, <cap) and the async orderer orderTargetsByHeadroom is appended to the existing open-sse/services/combo/quotaStrategies.ts (<cap) next to its sibling reset-aware/reset-window orderers (reuses their connection-expansion machinery). headroom = 1 - max(util_5h, util_7d) from getSaturation (src/lib/quota/saturationSignals.ts), prefers the connection with the most free capacity. Only the dispatch wiring is irreducible at the existing combo strategy chokepoint (mirrors the reset-aware/reset-window/context-optimized branches); not extractable without hiding the call site. fill-first stays default; all existing strategies untouched. Covered by tests/unit/combo-headroom-ranking.test.ts (pure helper) + tests/unit/combo-headroom-strategy.test.ts (orderer, saturation injected). Structural shrink of combo.ts tracked in #3501.",
+  "_rebaseline_2026_06_24_quota_share_strategy": "Dedicated quota-share strategy (Phase 3 #9): combo.ts 3180->3190 (+10 = one new `else if (strategy === \\\"quota-share\\\")` dispatch branch in handleComboChat that delegates 100% to selectQuotaShareTarget + its log line, plus the import). All the new logic lives OUT of the god-file in two new leaves under open-sse/services/combo/: quotaShareInflight.ts (in-flight counter with TTL/lease, ~150 LOC <cap) and quotaShareStrategy.ts (per-model bucket gating via isBucketSaturated + DRR proportional to weight + P2C over in-flight, ~240 LOC <cap). Only the dispatch wiring is irreducible at the existing combo strategy chokepoint (mirrors the headroom/reset-aware/reset-window/context-optimized branches); not extractable without hiding the call site. ZERO existing strategy cases were modified — only this branch was added, and the qtSd/ combos switched from fill-first to quota-share in src/lib/quota/quotaCombos.ts. Covered by tests/unit/quota-share-strategy.test.ts (gating, DRR fairness, P2C in-flight, fail-open, activation). Structural shrink of combo.ts tracked in #3501.",
+  "_rebaseline_2026_06_24_task_aware_routing": "Task-aware routing strategy (port PR #2045, OmniRoute #4945): combo.ts 3190->3225 (+35) = one new `else if (strategy === \\\"task-aware\\\")` dispatch branch delegating 100% to selectTaskAwareTarget + its imports/log lines. All scoring/classification logic lives OUT of the god-file in the new leaf open-sse/services/taskAwareRouting.ts (553 LOC <cap). Only the dispatch wiring is irreducible at the existing combo strategy chokepoint (mirrors quota-share/headroom/reset-aware branches). ZERO existing strategy cases modified. Covered by tests/unit/combo-task-aware.test.ts (35 tests). Structural shrink of combo.ts tracked in #3501.",
+  "_rebaseline_2026_06_26_fidelity_gate_extraction": "Milestone-B fidelity-gate wiring residual: bodyToText+gateAdvance extracted to fidelityGateStep.ts (889->854, -35), but the StackOptions.fidelityGate field, the `const fidelityGate` reads at the two stacked-loop dispatch chokepoints, and the import of FidelityGateConfig are irreducible wiring that cannot leave strategySelector without an architectural refactor of the pre-existing stacked pipeline. Net: 889->854 (+6 vs the pre-Milestone-B frozen 848). Covered by tests/unit/compression/*.test.ts (940 pass).",
+  "_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.",
+  "_rebaseline_2026_06_27_5193_antigravity_basered": "Base-red (pre-existing release drift, fast-gate PR->release skips check:file-size): accountFallback.ts 1773->1777 and src/app/api/providers/[id]/test/route.ts 924->940 were already over their frozen caps on release/v3.8.39 independent of any antigravity change. Owner chose to rebaseline (keep the documented issue-reference comments #1846/#1449/#347 etc.) rather than accept the contributor comment-stripping in #5200/#5198. Reverted #5200 to restore the comments; bumped these two frozen caps to the actual base sizes. No logic change.",
+  "_rebaseline_2026_06_28_5237_impersonation_ua_refresh": "PR #5237 (refresh impersonation UAs): grok-web.ts 1871->1873 (+2), muse-spark-web.ts 1284->1302 (+18), perplexity-web.ts 1013->1032 (+19). Net semantic change in each file is a single User-Agent constant (Chrome 147->149 for grok/muse; perplexity kept at Firefox 148 to stay matched with the firefox_148 TLS profile — the contributor's 152 bump was reverted to avoid a UA-vs-JA3 mismatch, #2459). The growth is Prettier reflow that lint-staged unavoidably applies to these grandfathered long-line files the moment they are touched; not extractable. src/sse/services/auth.ts 2336->2401 in the same reconcile is #5222's antigravity-LRU-retry growth that merged via --admin without a baseline bump.",
+  "_rebaseline_2026_06_28_5243_risk_gate_prepass": "PR #5243 (compression risk-gate pre-pass) own growth: open-sse/services/compression/strategySelector.ts 854->899 (+45). The three exported entry points (applyCompression/applyStackedCompression/applyStackedCompressionAsync) become thin wrappers over pure-extracted private bodies (runCompression/runStackedCompression/runStackedCompressionAsync) so the risk-gate mask->run->restore wrapper sits strictly OUTSIDE the per-step loop — a single universal integration point. The wrapper logic itself (resolveRiskGate/withRiskGate) lives in the new riskGate/strategyWrap.ts (<cap); the residual growth is the duplicated thin-wrapper signatures + the extracted bodies' dispatch boundary, guarded by a byte-identical parity test (riskGateIntegration). Default off (DEFAULT_COMPRESSION_CONFIG unchanged). Not extractable without hiding the dispatch boundary, mirroring prior compression rebaselines. Structural shrink tracked in #3501.",
+  "_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.",
+  "_rebaseline_2026_06_29_4038_cas_guard": "PR (#4038) own growth: tokenRefresh.ts 2103->2181 (+78 = the compare-and-swap guard on the refresh persist — runWithCasGuard/getActiveCasGuard AsyncLocalStorage pair mirroring runWithOnPersist, casGuardShouldSkipPersist that rereads the row right before persisting and skips the write when a concurrent writer already rotated the refresh_token past the one presented, plus getCasGuardStats counters). Fixes the sibling-rotation-revert → token-family-revocation storm. Gated behind an active guard (opt-in; no guard => byte-identical). Wiring lives at the two persist chokepoints inside getAccessToken; the comparison reuses wasRefreshTokenRotated from refreshSerializer. Not extractable without splitting the refresh hot path.",
+  "_rebaseline_2026_06_29_5286_memoization": "PR #5286 own growth: strategySelector.ts 899->960 (+61 = the opt-in result-memoization branches in applyCompression/applyCompressionAsync — principal+determinism gate, makeMemoKey lookup/store with model+supportsVision folded into the key, recompute-with-memo-off). Default off (memoizeCompressionResults), so zero behavior change. The memo helpers live in the leaf resultMemo.ts (<cap); the chokepoint wiring here is not extractable. Structural shrink of this hot-path file tracked in #3501.",
+  "_rebaseline_2026_07_01_v3843_release_5609": "Rebaseline v3.8.43 (PR #5609 release reconciliation). DRIFT dos 109 commits do ciclo: 8 god-files existentes cresceram (ApiManagerPageClient 2983->3017, combos/page 4594->4608, AddApiKeyModal 868->869, providerPageHelpers 974->996, chat.ts 1635->1647, auth.ts 2401->2403, batchProcessor 828->915, combo.ts 3368->3387) + 2 novos acima do cap (huggingchat.ts 813, tests web-cookie-providers-new 827) + 4 test files cresceram. Modularizacao deferida (blast-radius mid-release); congelado no estado atual p/ o proximo ciclo ratchetar daqui.",
+  "_rebaseline_2026_07_02_5816_qoder": "PR #5816 (@AgentKiller45, qoder PAT via qodercli): qoderCli.ts 666->989, new-above-cap frozen (owner-approved baseline freeze). The growth is the legitimate PAT job-token exchange + quota parsing CLI transport (the pure-JS Cosy path 500'd on every PAT request); extracting the spawn/parse helpers now would just add indirection to a contributor PR mid-merge. Test frozen also raised for this PR's coverage growth: providers-page-utils.test.ts 1052->1092. Additionally clears an inherited base-red from the already-merged #5933 (codex json_schema->text.format): translator-openai-responses-req.test.ts 1097->1172 (+75 regression tests, no offending branch left). All remain frozen (cannot grow further); release captain's rebaseline-at-release supersedes.",
+  "_rebaseline_2026_07_09_6126_clinepass_dual_auth": "PR #6126 (@hajilok, dual-auth ClinePass) own growth: tokenRefresh.ts 2181->2182 (+1 = a single `case \\\"clinepass\\\":` fallthrough label added to the existing `case \\\"cline\\\":` in _getAccessTokenInternal's provider switch, so clinepass token refresh dispatches to the already-shared refreshClineToken() instead of silently falling through to the generic OAuth refresh). Irreducible 1-line switch-case wiring at the existing chokepoint; the header-building logic for the same feature was extracted to a new leaf src/shared/utils/clineAuth.ts::buildClinepassHeaders() (well under cap) to avoid growing open-sse/executors/default.ts. Covered by tests/unit/clinepass-provider.test.ts.",
+  "_rebaseline_2026_07_09_6363_kiro_external_idp": "PR #6363 (@artickc, Kiro external IdP) own growth: tokenRefresh.ts 2182->2249 (+67 = the external_idp refresh branch inside refreshKiroToken — standard public-client OAuth2 refresh_token grant against the org IdP tokenEndpoint via buildExternalIdpRefreshParams/isExternalIdpAuthMethod from the new leaf open-sse/services/kiroExternalIdp.ts, with invalid_grant/invalid_client -> unrecoverable_refresh_error mapping). Cohesive addition at the existing refreshKiroToken chokepoint. Covered by tests/unit/kiro-external-idp.test.ts.",
+  "_rebaseline_2026_07_09_6587_kiro_api_key_auth": "PR #6587 (@strangersp) own growth for Kiro long-lived API-key auth, merged onto v3.8.47 tip: openai-to-kiro.ts 890->912 (+22, auth-header selection for API-key-vs-OAuth-token connections), providerLimits.ts 998->1000 (+2, API-key auth-type branch), translator-openai-to-kiro.test.ts 1234->1257 (+23), providers-page-utils.test.ts 1109->1107 (net -2 after merging with parallel release drift; connectionMatchesProviderCard api_key coverage added), provider-validation-specialty.test.ts 2856->2980 (+124 net after merge with parallel release drift; this PR also removed the file's `@typescript-eslint/no-explicit-any` eslint-suppression entry by fixing all `any` usages, adding typed replacements). Cohesive additive feature growth, well tested; not extractable without splitting the existing chokepoints mid-merge.",
+  "_rebaseline_2026_07_09_6678_routing_strategy_9router": "#6678 (SeaXen) — 9router-parity Routing Strategy settings card + per-provider/combo sticky-round-robin override. Own growth: ProviderDetailPageClient.tsx 784->786 (single ProviderAccountRoutingCard mount + import), auth.ts 2448->2458 (providerStrategies override resolution: fallbackStrategy/stickyRoundRobinLimit per-provider cascade in getProviderCredentials). Both additive, zero unrelated refactor; new UI/logic lives in new files (ProviderAccountRoutingCard.tsx, RoutingStrategyCard.tsx, rrState.ts::resolveComboStickyRoundRobinLimit). chat.ts value below reflects the current release tip (grown by other concurrent PRs, e.g. #6640), not this PR own change.",
+  "_rebaseline_2026_07_10_6318_omp_letta": "PR #6318 (@hamsa0x7, omp+letta CLI integrations) own growth: cliTools.ts (+53 = 2 registry entries incl. omp docsUrl) and cliRuntime.ts (+18 = runtime-detection wiring for the 2 new tools). Cohesive registry/wiring growth at the existing chokepoints; scope reduced from the original 5 tools (pi/codewhale/jcode shipped separately).",
+  "_rebaseline_2026_07_10_gcf_v3_2_decode": "PR #6838 own growth: new vendored file open-sse/services/compression/engines/headroom/gcf/decode_generic.ts frozen at 880 (> 800 cap). It is the vendored GCF generic-profile decoder (spec v3.2 nested flattening plus the prototype-pollution / hasOwnProperty hardening added in this PR's Gemini review). Kept as one file faithful to upstream gcf-typescript so re-vendoring stays a clean copy rather than a re-split each cycle (sibling generic.ts/scalar.ts stay < cap; extraction would also fragment the file's frozen eslint no-explicit-any suppressions). Round-trip + prototype-pollution regression coverage in tests/unit/compression/headroom-smartcrusher.test.ts. Frozen: only shrinks from here.",
+  "_rebaseline_2026_07_12_v3847_mergeprs_tail": "v3.8.47 /merge-prs tail (owner-approved): src/lib/localDb.ts NEW>800 (799->805, +6 re-exports countFreeProxies + recordFreeProxySyncErrors/clearFreeProxySyncErrors/getFreeProxySyncErrors + FreeProxySyncErrors type for #6909 free-pool relay-repair; re-export-only per Hard Rule #2, not extractable).",
+  "_rebaseline_2026_07_15_7070_combos_memo": "PR #7070 (perf/p1-memo) own growth: src/app/(dashboard)/dashboard/combos/page.tsx 4655->4656 (+1 = React.memo wrapping of ComboCard). Covered by tests/unit/ui/combos-page-smoke.test.tsx.",
+  "_rebaseline_2026_07_18_7399_xai_oauth_modal": "PR #7399 (xAI OAuth PKCE) own growth: OAuthModal.tsx 993->998 (+5 = provider entry + PKCE flow branch wiring at the existing provider-switch chokepoint; the provider logic itself lives in src/lib/oauth/providers/xai-oauth.ts, new leaf). Third irreducible wiring bump on this modal (969->989->993->998); structural shrink tracked in #3501.",
+  "_rebaseline_2026_07_19_6636_codex_session_json": "#6636 own growth: OAuthModal.tsx 998->1030 (gate units, split(\\\"\\\\n\\\").length incl. trailing newline; +32 = session-JSON paste branch for handleManualSubmit plus a shared submitCodexAccessToken() helper extracted from the pre-existing bare-JWT branch, mirroring the #5203 oauthBlobSubmit.ts extraction precedent; the normalizer logic itself lives in the new src/lib/oauth/utils/codexSessionImport.ts leaf module, not here). Fourth irreducible wiring bump on this modal (969->989->993->998->1030); structural shrink tracked in #3501.",
+  "_rebaseline_2026_07_19_7546_ghe_copilot_modal": "PR #7546 (GHE Copilot OAuth provider) own growth: OAuthModal.tsx 1030->1056 (gate units). Adds a gheUrl input state, routes ghe-copilot through the existing device-code branch, and threads gheUrl into the device-code request/poll extraData at the existing provider-switch chokepoints (+~24 lines, cohesive with the same pattern as #7399/#6636). The standalone GHE enterprise-URL config step JSX (originally +31 lines inline) was extracted to the new src/shared/components/oauthModal/GheConfigStep.tsx leaf component to minimize the bump; what remains is the irreducible provider-branch wiring. Fifth bump on this modal (969->989->993->998->1030->1056); structural shrink tracked in #3501.",
+  "_rebaseline_2026_07_19_7787_ic2_localdb_reexports": "PR #7787 (IC2 raw connections cache + lazy-decrypt) own growth: localDb.ts 805->807 (gate units, +2). localDb.ts is the re-export-only layer (hard rule #2 — no logic); the PR adds 4 new db/readCache re-exports (touchConnectionLastUsed, getCachedRawProviderConnections, getCachedProviderConnectionById, getCachedProviderNodes) required by existing barrel importers. Irreducible for a re-export list; frozen so it can only shrink.",
+  "_rebaseline_2026_07_20_7779_routingcombo_thread": "PR #7779 own growth: chatHelpers.ts 876->877 (+1, thread routingComboId into executeChatWithBreaker for compression-combo assignment). Frozen so it can only shrink.",
+  "_rebaseline_2026_07_20_7819_autocandidateoverrides_reexport": "PR for #7819 (Level 1+2: read-only auto/* candidate transparency + per-API-key exclusions) own growth: localDb.ts 807->808 (+1). Adds a single `export * from \\\"./db/autoCandidateOverrides\\\"` barrel re-export (hard rule #2 — no logic) for the new DB module backing per-apiKey candidate exclusions. Irreducible for a re-export list; frozen so it can only shrink.",
+  "_rebaseline_2026_07_21_8027_grok_cli_auth_json_paste": "PR #8027 (RaviTharuma, fix(grok-cli) #7610) own growth: OAuthModal.tsx 1080->1100 (gate units). Requires the full ~/.grok/auth.json (with refresh_token) on the paste-import path instead of a bare JWT, at the existing paste-token chokepoint (renamed tab label, updated instructions/placeholder, textarea for the auth.json blob, inline error surface). The validation logic itself (parseGrokCliPasteToken, previously an inline ~75-line function) was extracted to the new src/lib/oauth/utils/grokCliAuthJson.ts leaf module — mirroring the #6636/#7546 extraction precedent — so only the irreducible UI wiring remains here. Sixth bump on this modal (969->989->993->998->1030->1056->1100); structural shrink tracked in #3501.",
+  "_rebaseline_2026_07_21_8034_compression_exclusions_sidebar": "#8034 (compression exclusions dashboard tab) own growth: sections.ts 796->806 (+10, one new COMPRESSION_CONTEXT_GROUP sidebar item linking /dashboard/compression/exclusions). The file was already 796/800 before this PR (organic growth from prior sidebar entries), so a single new nav item pushed it 6 lines over cap. Freezing at 806 (cannot grow further); the sidebar item array is data, not extractable logic.",
+  "_rebaseline_2026_07_22_7936_namespace_roundtrip": "#7936 (@RCrushMe, Responses-Chat namespace round-trip identity seam) own growth: open-sse/translator/response/openai-responses.ts 1092->1125 (+33) and open-sse/utils/stream.ts 2814->2869 (+55) — threading the namespace-identity seam through the Responses↔Chat translation + stream paths so tool-call namespaces survive the round-trip. Cohesive translation/stream wiring at existing chokepoints, frozen at new size.",
+  "_rebaseline_2026_07_22_8010_codex_responses_engine": "PR #8010 (@JxnLexn) own growth: open-sse/mcp-server/schemas/tools.ts 1497->1505 (+8 = threading the new \\\"codex-responses\\\" literal into the compressionConfigureInput strategy/autoTriggerMode Zod enums and setCompressionEngineInput engine enum, mirroring the existing rtk/omniglyph enum entries; no new tool). open-sse/services/compression/strategySelector.ts 1043->1054 (+11 = one new `if (mode === \\\"codex-responses\\\")` dispatch branch in runCompression that delegates 100% to the new codexResponsesEngine.apply, mirroring the existing rtk single-mode dispatch, plus threading config.codexResponsesConfig.preserveToolNames into the shared adaptBodyForCompression call at the 3 existing call sites). src/lib/db/compression.ts (untracked, new-file cap 800) 794->845 (+51 = normalizeCodexResponsesConfig, mirroring the existing normalizeRtkConfig normalizer, plus registering \\\"codex-responses\\\" in the COMPRESSION_MODES/STACKED_PIPELINE_ENGINE_IDS/SINGLE_MODE_ENGINE sets and the getCompressionSettings load/save switch) — added to the baseline at its current size. All three are cohesive dispatch/normalizer wiring at existing chokepoints (mirroring the prior compression-mode rebaselines #6534/#6556), not extractable without hiding the mode-dispatch boundary. Covered by tests/unit/compression/codex-responses.test.ts (6) + omniglyph-registries.test.ts/types.test.ts (22, updated for the new mode).",
+  "_rebaseline_2026_07_22_8034_compression_exclusions_persistence": "#8034 (compression exclusions) own growth: src/lib/db/compression.ts 845->850 (+5 = threading the new compressionExclusions field through the existing getCompressionSettings/saveCompressionSettings load/save switch over the shared key_value compression namespace — no new table, no raw SQL). Mirrors the prior compression-field rebaselines (#8010 codex-responses normalizer at the same chokepoint); the load/save switch is a single dispatch boundary, not extractable without hiding it. Covered by the PR's 8 node:test + 3 vitest cases.",
+  "_rebaseline_2026_07_22_8050_model_lockout_exact_family": "#8050 (@AndrianBalanescu) own growth: accountFallback.ts 1864->1892 (+28) — exact-vs-family model-lockout scoping (getModelLockKey/isModelLocked/clearModelLock/getModelLockoutInfo) so an Antigravity 404 for one bare model no longer hijacks the whole family cooldown. Cohesive lockout logic; frozen at new size.",
+  "_rebaseline_2026_07_22_8081_reasoning_placeholder_guard": "#8081 (@Dingding-leo) own growth: openai-responses.ts 1125->1137 (+12) restructuring the reasoning-placeholder guard so it skips only the empty content block and still emits finish_reason/tool_calls in the same chunk. Cohesive translator wiring; frozen at new size.",
+  "_rebaseline_2026_07_22_8210_openrouter_midstream_error": "PR #8210 (hartmark, fix/openrouter-midstream-error-surfacing) own growth: open-sse/translator/response/openai-responses.ts 1137->1163 (+26) measured on the merged tip (release 1137 + this PR own growth). Adds a single new branch inside openaiToOpenAIResponsesResponse() that detects an OpenRouter-style mid-stream aggregator error (HTTP 200 SSE chunk with empty choices + a top-level error object) and surfaces it as state.upstreamError instead of silently falling through to the no-op/awaitingTrailingUsage path, which previously masked the failure as a false empty-success completion and skipped combo fallback. Irreducible call-site addition at the existing chunk-dispatch chokepoint (mirrors the Gemini-to-OpenAI translator's #4177 precedent for the same class of upstream error surfacing). Note: this baseline entry does NOT cover the separate pre-existing +11 drift already on the release tip from #8081/#8162 (1125->1136, unrelated reasoning-placeholder-stripping fix merged after this PR branched) — that drift belongs to the maintainer's rebaseline, not this PR.",
+  "_rebaseline_2026_07_22_8211_gemini_malformed_tool_choice": "PR #8211 (hartmark, fix/gemini-malformed-function-call-tool-choice) own growth: open-sse/translator/response/gemini-to-openai.ts 771->821 (+50, entirely this PR's diff — no other commit touched this file between the PR's merge-base and the release tip). Adds MALFORMED_FUNCTION_CALL/UNEXPECTED_TOOL_CALL handling inside geminiToOpenAIResponse(): synthesizes a `malformed_tool_call` tool_calls entry so finish_reason normalizes to the standard \\\"tool_calls\\\" instead of an unrecognized raw enum value that OpenAI-compatible clients (e.g. OpenClaw) silently ignore, and always synthesizes (rather than skipping when a real tool call already exists) so a malformed attempt alongside a real one in the same turn is not silently discarded. Irreducible cohesive addition at the existing candidate/finishReason translation chokepoint (mirrors the 9router#2462 raw-finish-reason precedent immediately below it in the same function). Covered by the PR's own tests/unit test additions for both the malformed-only and malformed-plus-real-call cases.",
+  "_rebaseline_2026_07_22_8213_chat_abandoned_target_abort": "PR #8213 (hartmark, fix/gemini-tpm-quota-cooldown-wait) own growth: src/sse/handlers/chat.ts 1794->1860 (+66, measured against the PR's own merge-base — the release tip separately carries an unrelated -5 net shrink from #8013's antigravity callable-catalog alignment, which this PR's branch does not include and this entry does not cover). Adds resolveDispatchClientRawRequest(): merges a per-target modelAbortSignal into clientRawRequest.signal (via mergeAbortSignals) so a combo target abandoned by comboTargetTimeoutMs actually observes its own abort and reaches its cleanup path, instead of hanging forever inside withRateLimit/acquireAccountSemaphore and leaking a permanent 'pending' dashboard entry (live incident, log id 1784418258231-14961a). Also wires combo-exhausted rejection logging to capture request body + attempted models via the new rejectedRequestUsage helper. Irreducible additions at the existing chat dispatch chokepoint. Covered by the PR's own combo-config + integration test additions.",
+  "_rebaseline_2026_07_22_8213_combo_cooldown_wait_recording": "PR #8213 (hartmark, fix/gemini-tpm-quota-cooldown-wait) own growth: open-sse/services/combo.ts 3548->3604 (+56, entirely this PR's diff — no other commit touched this file between the PR's merge-base and the release tip). Fixes combo cooldown-wait state recording so a bogus 503 is no longer crystallized when the cooldown-wait vars reset every setTry, adds an OpenAI-format SSE error frame path for combo-exhausted rejections (capturing request body + attempted models), and gives an abandoned per-target dispatch its own timeout instead of leaking a permanent 'pending' dashboard entry. Irreducible additions at the existing handleComboChat dispatch/retry chokepoint (mirrors the prior quota-share/headroom/task-aware strategy-branch precedents already frozen in this file). Covered by the PR's own combo-config + Gemini TPM-ceiling benchmark test additions.",
+  "_rebaseline_2026_07_22_8213_gemini_tpm_quota_cooldown_wait": "PR #8213 (hartmark, fix/gemini-tpm-quota-cooldown-wait) own growth: open-sse/services/accountFallback.ts 1857->1932 on the merged tip (release 1892 incl #8050 +35, plus this PR own growth +40); measured against the PR own merge-base was 1857->1898 (+41 — the release tip separately carries an unrelated +34 from #8050's antigravity 404 model-not-found lockout scoping, which this PR's branch does not include and this entry does not cover). Own growth is the Gemini TPM-ceiling classification + cooldown-wait wiring feeding into the combo cooldown-wait state machine (rate-limit wedge recovery) introduced by this PR's commit series. Irreducible additions at the existing account-fallback/model-lockout chokepoint. Covered by the PR's own gemini-rate-limit-tracker and TPM-ceiling benchmark test additions.",
+  "_rebaseline_2026_07_22_8213_health_unblock_model_cooldowns": "PR #8213 (hartmark, fix/gemini-tpm-quota-cooldown-wait) own growth: src/app/(dashboard)/dashboard/health/page.tsx 1094->1165 (+71, entirely this PR's diff — no other commit touched this file between the PR's merge-base and the release tip). Adds handleUnblockAll/handleUnblockOne dashboard actions (DELETE /api/resilience/model-cooldowns) so an operator can manually clear a Gemini TPM-wedge model lockout surfaced by this PR's cooldown-wait fixes, instead of waiting out the ceiling. Irreducible UI wiring at the existing health-page action chokepoint. Covered by the PR's own dashboard/resilience test additions.",
+  "_rebaseline_2026_07_22_8213_requestloggerdetail_unblock_ui": "PR #8213 (hartmark, fix/gemini-tpm-quota-cooldown-wait) own growth: src/shared/components/RequestLoggerDetail.tsx 799->941 (+142, entirely this PR's diff — no other commit touched this file between the PR's merge-base and the release tip; crosses the general 800-line new-file cap so is frozen here for the first time). Adds a collapsible section header (open/expand-less toggle) plus per-log-entry unblock (`unblocking`/`cleared` state, isCombo503 detection) so the request-logger detail panel surfaces the same Gemini TPM cooldown-wait / model-lockout unblock action introduced by this PR at the individual-request level (mirrors the health-page bulk unblock action added in the same PR). Covered by the PR's own dashboard/resilience test additions.",
+  "_rebaseline_2026_07_22_fusion_8013_8098_antigravity": "Fusion of #8013 (backryun, catalog/IDE-CLI-split rewrite) + #8098 (nguyenha935, protocol-fidelity/fail-closed/credits/tool-cloaking): open-sse/services/usage/antigravity.ts NEW 802 (>cap 800, +2 — #8098 credits/tier usage service on #8013's profile-aware headers). Test growth (models-catalog-route 1605->1608, provider-models-route 1752->1757 from #8013 Gemini 3.6 catalog) tracked in testFrozen.",
+  "_rebaseline_2026_07_23_8127_grok_weekly_quota": "#8127 (@apoapostolov) own growth: src/sse/handlers/chat.ts 1861->1865 (+4) — weekly quota tracking for grok-web wires a quota-fetch hook at the existing dispatch chokepoint. Thin wiring mirroring adjacent provider-quota branches; not extractable. Covered by tests/unit/grok-quota-fetcher.test.ts.",
+  "_rebaseline_2026_07_23_8143_empty_catch_logging": "#8143 (@chirag127) own growth: open-sse/utils/stream.ts 2869->2887 (+18) — replacing empty catch blocks in the SSE stream subsystem with console.debug logging (Rule #6 silent-swallow fix, issues #8138-#8142). Cohesive logging additions at the existing catch chokepoints, not extractable; frozen at new size. Covered by tests/unit/stream-handler-catch-logging-8143.test.ts.",
+  "_rebaseline_2026_07_23_8219_cache_ttl_settings_sidebar": "#8219 (@oyi77) own growth: sections.ts 806->813 (+7) — configurable model-catalog cache-TTL settings adds a new sidebar nav entry + its visibility wiring. Sidebar item array is data, not extractable logic; frozen at new size.",
+  "_rebaseline_2026_07_23_8247_8248_model_unhealthy": "#8247+#8248 own growth: accountFallback.ts 1940->1941 (+1, irreducible import statement only — the substantive #8248 DEGRADED-pattern classifier was extracted into open-sse/config/errorConfig.ts, which has ample headroom, instead of growing this frozen file; #8247's fix is a single existing-line condition change, net zero lines). Scoping the credits-exhausted 403/429 branch to isCompatibleProvider() (per-model-quota openai/anthropic-compatible-* nicknames) so it stays model-scoped instead of terminalling the whole connection, and classifying NVIDIA NIM 'Function ... DEGRADED' 400 bodies as model-access-denied instead of a raw passthrough 400. Covered by tests/unit/8247-accountfallback-model-unhealthy.test.ts and tests/unit/8248-accountfallback-nvidia-degraded.test.ts.",
+  "_rebaseline_2026_07_23_8252_combo_400_advance": "#8252 (@RaviTharuma) own growth: accountFallback.ts 1932->1940 (+8) + combo.ts 3604->3630 (+26) — advance combo on model-scoped 400s wrapped as invalid/Bad-Request. Irreducible wiring at existing account-fallback + combo dispatch chokepoints. Covered by combo-model-scoped-400-advance.test.ts.",
+  "_rebaseline_2026_07_23_8266_alibaba_media": "#8266 (@backryun) own growth: imageRegistry.ts 821->979 (+158) — Alibaba-family media models (Qwen image/video, Bailian, Wan) added to the image/video registry. Registry model data, not extractable logic; frozen at new size.",
+  "_rebaseline_2026_07_24_8388_compression_detail_persist": "#8388 (compression engine DETAIL settings — Headroom/session-dedup/CCR — dropped on save) own growth: src/lib/db/compression.ts 866->872 (+6 = irreducible call-site wiring at the existing getCompressionSettings/updateCompressionSettings chokepoint: one import line, one `...buildDetailConfigDefaults()` spread in the seed config, and one `case \\\"sessionDedup\\\": case \\\"ccr\\\": applyDetailConfigUpdate(config, key, parsed); break;` load-switch case, mirroring the existing headroom/#8056 case immediately above it). The actual normalizer logic (normalizeSessionDedupConfig/normalizeCcrConfig, matching SESSION_DEDUP_SCHEMA/CCR_SCHEMA bounds) was EXTRACTED into a new leaf src/lib/db/compressionDetailNormalizers.ts (well under cap) so this frozen file only carries the minimal dispatch wiring. Covered by tests/unit/8388-compression-detail-persist.test.ts (schema-accept + full DB save->reload round-trip for both new sub-objects, plus a no-regression assertion on the existing headroom round-trip).",
+  "_rebaseline_2026_07_24_responses_toolcalls_log_summary": "hartmark, fix/responses-tool-calls-log-summary own growth: open-sse/translator/response/openai-responses.ts 1163->1174 (+11). closeToolCall() now also writes the completed tool call into the shared state.toolCalls Map (already populated by the openai-to-claude / claude-to-openai / gemini-to-openai response translators) so stream.ts's completion-log summary builder (which reads state.toolCalls, not this translator's own funcCallIds/funcNames/funcArgsBuf bookkeeping) reports finish_reason \\\"tool_calls\\\" and message.tool_calls for openai->openai-responses translated streams instead of always logging \\\"stop\\\" with no tool_calls — the actual client-facing SSE events were already correct; only the persisted call-log summary was wrong. Irreducible call-site addition at the existing tool-call-close chokepoint. Covered by the new regression test in tests/unit/translator-resp-openai-responses.test.ts.",
+  "_rebaseline_2026_07_25_8476_combo_input_bound_homogeneous_scope": "PR #8476 (herjarsa, fix/8375-8459-combo-image-fixes, #8375) own growth: open-sse/services/combo.ts 3642->3679 (+37 net: +29 the PR's own isInputBoundFailure short-circuit for deterministic context_length_exceeded/context_window_exceeded failures, +8 a /green-prs pre-merge fix scoping that short-circuit to homogeneous remainders only — the shipped code fired unconditionally on ANY target, regressing the intentional heterogeneous-combo fallback #6637/isContextOverflow400 protects, exactly as flagged by this PR's own review evidence but never actually implemented in the branch). The fix compares orderedTargets[i+1..] modelStr against the failing target's modelStr at the existing executeTarget dispatch chokepoint (mirrors the sameProviderNext precedent a few lines below) — irreducible call-site wiring, not extractable without hiding the dispatch boundary. Covered by tests/unit/combo-input-bound-failure-8375.test.ts (homogeneous pool still short-circuits) and the new tests/unit/combo-input-bound-heterogeneous-8375.test.ts (heterogeneous combo now correctly falls through to the larger-context target).",
+  "_rebaseline_2026_07_25_adobe_firefly_reference_images": "Follow-up to #8006: storage upload + referenceBlobs for image/video and /v1/images/edits dispatch. adobeFireflyClient.ts 1958->2317 (+upload helpers, extract sources, resolve blob ids). Note: 2317 not 2316 — check-file-size.mjs counts LOC via split(\\\"\\\\n\\\").length (counts the trailing-newline empty element), which is 1 higher than `wc -l` on a file ending in \\\\n; the PR's original entry (2316) was measured with wc -l and undercounted by 1 against the actual gate.",
+  "_rebaseline_pr1043_minimax_tts": "Upstream port decolua/9router#1043 (toanalien) own growth: audioSpeech.ts 965->1061 (+96). Adds MiniMax T2A v2 TTS dispatch (handleMinimaxSpeech + hexToBytes helper) — provider entry was already in audioRegistry (format: minimax-tts) but no handler existed, falling through to the OpenAI-compatible default that fails (T2A has custom shape + hex-encoded audio + base_resp envelope). New branch sits next to the other inline provider branches (xiaomi-mimo, coqui, tortoise, aws-polly) — extracting would just create indirection. Covered by tests/unit/minimax-tts-1043.test.ts (3 tests, GREEN: success, base_resp error, invalid-hex).",
+  "_rebaseline_pr4592_exclude_exhausted_auto": "Reconcile #4592 already-merged growth: combo.ts 2991->3036 (+45, terminal-status quota-cutoff exclusion in buildAutoCandidates + opt-in gate). Fast-gate PR->release does not run check:file-size.",
+  "open-sse/executors/antigravity.ts": "1528",
+  "open-sse/executors/base.ts": "1640",
+  "open-sse/executors/chatgpt-web.ts": "3241",
+  "open-sse/executors/codex.ts": "1562",
+  "open-sse/executors/cursor.ts": "1563",
+  "open-sse/executors/deepseek-web.ts": "1148",
+  "open-sse/executors/grok-web.ts": "1044",
+  "open-sse/executors/muse-spark-web.ts": "1405",
+  "open-sse/handlers/chatCore.ts": "5034",
+  "open-sse/handlers/imageGeneration.ts": "3101",
+  "open-sse/handlers/responseSanitizer.ts": "1128",
+  "open-sse/handlers/search.ts": "1536",
+  "open-sse/handlers/videoGeneration.ts": "1063",
+  "open-sse/mcp-server/schemas/tools.ts": "1553",
+  "open-sse/mcp-server/server.ts": "1448",
+  "open-sse/mcp-server/tools/advancedTools.ts": "1120",
+  "open-sse/services/accountFallback.ts": "1978",
+  "open-sse/services/adobeFireflyClient.ts": "2385",
+  "open-sse/services/claudeCodeCompatible.ts": "1202",
+  "open-sse/services/combo.ts": "3648",
+  "open-sse/services/compression/strategySelector.ts": "1060",
+  "open-sse/services/rateLimitManager.ts": "1167",
+  "open-sse/translator/response/openai-responses.ts": "1204",
+  "open-sse/utils/cursorAgentProtobuf.ts": "1505",
+  "open-sse/utils/stream.ts": "2889",
+  "src/app/(dashboard)/dashboard/HomePageClient.tsx": "1388",
+  "src/app/(dashboard)/dashboard/analytics/ComboHealthTab.tsx": "1031",
+  "src/app/(dashboard)/dashboard/api-manager/ApiManagerPageClient.tsx": "3117",
+  "src/app/(dashboard)/dashboard/cache/media/MediaPageClient.tsx": "1067",
+  "src/app/(dashboard)/dashboard/combos/page.tsx": "4703",
+  "src/app/(dashboard)/dashboard/costs/CostOverviewTab.tsx": "1283",
+  "src/app/(dashboard)/dashboard/costs/quota-share/components/PoolWizard.tsx": "1022",
+  "src/app/(dashboard)/dashboard/endpoint/EndpointPageClient.tsx": "2615",
+  "src/app/(dashboard)/dashboard/health/page.tsx": "1165",
+  "src/app/(dashboard)/dashboard/providers/[id]/components/modals/EditConnectionModal.tsx": "1324",
+  "src/app/(dashboard)/dashboard/providers/page.tsx": "1944",
+  "src/app/(dashboard)/dashboard/runtime/RuntimePageClient.tsx": "1201",
+  "src/app/(dashboard)/dashboard/settings/components/PricingTab.tsx": "1019",
+  "src/app/(dashboard)/dashboard/settings/components/ProxyRegistryManager.tsx": "1470",
+  "src/app/(dashboard)/dashboard/settings/components/ResilienceTab.tsx": "1123",
+  "src/app/(dashboard)/dashboard/settings/components/RoutingTab.tsx": "1629",
+  "src/app/(dashboard)/dashboard/settings/components/SystemStorageTab.tsx": "1573",
+  "src/app/(dashboard)/dashboard/usage/components/BudgetTab.tsx": "1028",
+  "src/app/(dashboard)/dashboard/usage/components/EvalsTab.tsx": "2148",
+  "src/app/(dashboard)/dashboard/usage/components/ProviderLimits/index.tsx": "1119",
+  "src/app/api/providers/[id]/models/route.ts": "2361",
+  "src/app/api/v1/models/catalog.ts": "1597",
+  "src/lib/tokenHealthCheck.ts": "1053",
+  "src/lib/db/apiKeys.ts": "1529",
+  "src/lib/db/core.ts": "1639",
+  "src/lib/db/migrationRunner.ts": "1094",
+  "src/lib/db/models.ts": "1097",
+  "src/lib/db/providers.ts": "1034",
+  "src/lib/memory/retrieval.ts": "1073",
+  "src/lib/tailscaleTunnel.ts": "1202",
+  "src/lib/usage/providerLimits.ts": "1013",
+  "src/shared/components/OAuthModal.tsx": "1134",
+  "src/shared/components/RequestLoggerV2.tsx": "1629",
+  "src/shared/components/analytics/charts.tsx": "1035",
+  "src/shared/services/cliRuntime.ts": "1122",
+  "src/sse/handlers/chat.ts": "1904",
+  "src/sse/services/auth.ts": "2508",
+  "tests/unit/account-fallback-service.test.ts": "1572",
+  "tests/unit/provider-validation-specialty.test.ts": "2985",
+  "open-sse/executors/hyperagent.ts": "1026",
+  "open-sse/executors/default.ts": "1042",
+  "open-sse/executors/kiro.ts": "1069",
+  "open-sse/translator/request/openai-to-kiro.ts": "1057",
+  "open-sse/utils/sseHeartbeat.ts": "142",
+  "_rebaseline_2026_08_04_9305_sse_comments": "#9305 fix: broadened sseCommentsEnabled()",
+  "_rebaseline_2026_08_09_v3850_release_close": "Release v3.8.50 close reconciliation on e0ce95c592: src/sse/handlers/chat.ts 1904->1918 is the irreducible request-pipeline wiring from #9759 that invokes the Modality Bridge guardrail without moving its implementation into the handler; covered by the 17 Vision Bridge canaries plus the PR-1 focused suite. open-sse/translator/response/openai-responses.ts 1204->1215 is #9168's Responses tool-call argument delta buffering/normalization at the existing translator state-machine chokepoint; covered by its dedicated translator regression tests. Both values are measured by check:file-size (split-newline semantics), and the gate remains frozen at the new exact sizes.",
+  "_rebaseline_2026_08_08_toolcall_message_index_collision": "fix(responses-api): tool call after a text message collided on the same output_index. own growth: open-sse/translator/response/openai-responses.ts 1204->1224 (+20, extracted toolCallOutputIndexBase() shared helper so emitToolCall/closeToolCall can no longer compute a tool call's output_index independently and collide with a text message emitted in the same turn). Live incident (2026-08-08, OpenClaw agent): a client that tracks response items by output_index saw the tool call's added/delta/done events land on an index it had already marked complete (the just-closed text message), and silently dropped them — the agent spoke its preamble and never executed the tool call, even though OmniRoute's own recorded responseBody had a complete, valid tool_calls entry. Covered by the new regression test in tests/unit/translator-resp-openai-responses.test.ts reproducing the exact live scenario."
 }

config/quality/quality-baseline.json

@@ -177,12 +177,13 @@
       "dedicatedGate": true
     },
     "bundleSize": {
-      "value": 7666,
+      "value": 8045,
       "direction": "down",
       "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.",
       "_rebaseline_2026_07_19_7808_codeql_alias_resolver_hook": "6534->6762 (+228). PR #7808 (CodeQL js/incomplete-url-substring-sanitization fix): the ESM loader hook source moved out of the inline `HOOK_SOURCE` template literal in bin/aliasResolver.mjs into a real file bin/aliasResolverHook.mjs, loaded via pathToFileURL() instead of a dynamically-built `data:text/javascript,...` URL. The new file is now counted by size-limit as a 5th bin/*.mjs entrypoint. Net +228 = the hook's gzip size (previously hidden inside aliasResolver.mjs because the template literal was compressed away). Security-driven; no shrink opportunity.",
-      "_rebaseline_2026_07_28_v3849_release_preflight": "6762 -> 7666 (+904). Fechamento do ciclo v3.8.49: gzip dos entrypoints bin/*.mjs (size-limit + @size-limit/file) cresceu com o que os merges do ciclo puxam transitivamente para o CLI (novos provedores — 271->290, seletor de protocolo por conexão #8861, catálogos de busca #8814, resiliência). Crescimento legítimo de ciclo, medido localmente com `npm run check:bundle-size` = 7666 (gzip determinístico, bate com o CI). Encolher é dívida separada."
+      "_rebaseline_2026_07_28_v3849_release_preflight": "6762 -> 7666 (+904). Fechamento do ciclo v3.8.49: gzip dos entrypoints bin/*.mjs (size-limit + @size-limit/file) cresceu com o que os merges do ciclo puxam transitivamente para o CLI (novos provedores — 271->290, seletor de protocolo por conexão #8861, catálogos de busca #8814, resiliência). Crescimento legítimo de ciclo, medido localmente com `npm run check:bundle-size` = 7666 (gzip determinístico, bate com o CI). Encolher é dívida separada.",
+      "_rebaseline_2026_08_09_v3850_release_close": "7666 -> 8045 (+379 gzip bytes, +4.9%). Release v3.8.50 close reconciliation measured twice with the real size-limit + @size-limit/file path on tip e0ce95c592. Per-entry measurements remain below their absolute budgets: omniroute.mjs 4380/15000, mcp-server.mjs 1195/5000, nodeRuntimeSupport.mjs 887/8000, reset-password.mjs 1583/6000. The growth accumulated through legitimate CLI/runtime work in this cycle, including global-install ESM alias resolution, Termux cache preparation, and MCP stdio startup hardening; no entrypoint is near its absolute ceiling. The direction:down ratchet stays blocking from this exact measured tip."
     },
     "openapiBreaking": {
       "value": 0,

docs/compression/COMPRESSION_GUIDE.md

@@ -158,7 +158,7 @@ error lines produced **95.93% token savings / 96.26% character savings** — squ
 range. But the same pipeline run against normal, non-redundant tool output (a clean `grep` match list,
 a short file read, ordinary conversational text) correctly produces **near-zero savings**, because
 there is nothing repetitive to remove and `validateCompression()` (`validation.ts`) refuses to ship a
-rewrite that would drop or alter code blocks, URLs, headings, versions, or `CONST_CASE` identifiers.
+rewrite that would drop or alter code blocks, URLs, headings, versions, or ALL-CAPS constant identifiers.
 
 This is expected, safe behavior, not a bug: a coding session that mostly reads/greps clean files will
 see modest total savings even with compression fully enabled, while a session that hits a failing

docs/frameworks/RADAR.md

@@ -107,15 +107,18 @@ client component never reads `process.env` itself.
 | `RADAR_CONTRIBUTOR_CLAIM_URL`   | Overrides the contributor-claim URL (default `https://radar.omniroute.online/auth/github`).   |
 | `RADAR_SUPPORTER_PLANS_URL`     | Overrides the supporter-plans URL (default `https://radar.omniroute.online/planos`).          |
 
-Once a visitor has a key (`omr_` + 40 hex chars), it is set with `POST
-/api/radar/settings` (`{ supporterKey }`) — the same endpoint documented under
-[Data sync](#data-sync-is-a-separate-opt-in--the-privacy-promise) above.
-
-**Known gap:** the dashboard activation screen does not yet have a dedicated
-key-paste input — pasting a key today requires calling `POST /api/radar/settings`
-directly (curl, a script, or a future UI). This release only adds the two claim/plans
-buttons; the API already accepts and masks the key, but no `<input>` for it exists in
-`src/app/(dashboard)/dashboard/radar/page.tsx` yet.
+Once a visitor has a key (`omr_` + 40 hex chars), the activation screen
+(`src/app/(dashboard)/dashboard/radar/page.tsx`) has a paste-key input as the primary
+path: pasting a key and submitting sends `POST /api/radar/settings`
+(`{ optIn: true, supporterKey }`) in one call — pasting a key both sets it and opts in,
+unlocking the screen. The format (`omr_` + 40 hex chars) is checked client-side first
+with the shared `isValidSupporterKeyFormat()` helper (`src/lib/radar/supporterKey.ts`)
+as a UX nicety; the server's Zod schema is the authoritative check either way. Once a
+key is set, the activation screen shows the masked form (`supporterKeyMasked` from
+`GET /api/radar/settings`) instead of an empty input, with a "change key" control to
+paste a new one — the raw key is never redisplayed. The two claim/plans buttons above
+remain the way to *obtain* a key in the first place; this input is where an operator
+who already has one activates it.
 
 ---
 
@@ -296,36 +299,84 @@ auth state — only the masked form and a `hasSupporterKey` boolean.
 
 ## Referral links (free credits)
 
-The server-published feed carries a `referrals` section (server-side D28 work, already
-in production — this section documents the **client** consumption only):
+Referral links are served from a **standalone, always-current** feed —
+`GET /v1/referrals/latest` — separate from the catalog feed. This is deliberate: the
+catalog feed on the community tier is a snapshot that can be up to 30 days old, so a
+referral link extracted from it used to lag the server's real link list by the same
+amount (a newly-added referral wouldn't reach a free/community user for up to a month).
+The referrals feed removes that delay by syncing on its own, much shorter cadence.
 
 ```ts
-referrals: {
-  fixed: RadarReferral[],      // present in EVERY tier, including community
-  campaigns: RadarReferral[],  // only populated on the live (supporter) tier;
-                                // the community artifact always publishes []
+// GET /v1/referrals/latest response body (Ed25519-signed, same pinned key as
+// the catalog feed):
+{
+  feed: "omniroute-radar-referrals",
+  schemaVersion: 1,
+  generatedAt: string,           // ISO — deterministic: max(updatedAt) across referral
+                                  // links, so two identical requests produce the exact
+                                  // same signed bytes/signature
+  referrals: {
+    fixed: RadarReferral[],      // present in EVERY tier, including no-auth/community
+    campaigns: RadarReferral[],  // only populated for a valid live (supporter) Bearer
+                                  // key; no-auth/expired-key requests get []
+  },
 }
 // RadarReferral = { provider, url, kind: "fixo" | "campanha", validUntil,
 //                    requiredAction, isDefault }
 ```
 
-The client never decides which tier it received or which referrals belong in which
-tier — the server already publishes two artifacts (`live`/`community`) with
-`campaigns` gated server-side, same principle as the [tiers](#tiers-community-and-live)
-section above. `RadarFeedSchema` (`src/lib/radar/feedSchema.ts`) validates `referrals`
-as a whole-object `.default({fixed:[],campaigns:[]})`, and `campaigns` defaults
-independently inside it — so a feed cached before this section existed on the server
-still parses cleanly, and `campaigns` alone can also be absent without failing
-validation. Every `RadarReferral.url` must be `https://` — a `http://` url fails
-schema validation.
+Unlike the catalog feed, this body carries no `tier` field at all — the server decides
+what to include per-request based on the `Authorization` key, so the
+`x-omniroute-feed-tier` response header is the ONLY source for the served tier
+(`referralsSync.ts::syncRadarReferrals`); an absent/unrecognized header degrades to
+`"community"`, the least-privileged assumption. `RadarReferralsFeedSchema`
+(`src/lib/radar/referralsFeedSchema.ts`) validates the whole body, reusing the same
+per-referral `RadarReferralSchema` exported from `feedSchema.ts` so both feeds validate
+individual referrals identically. Every `RadarReferral.url` must be `https://` — a
+`http://` url fails schema validation.
+
+The OLD catalog-embedded `referrals` field on `RadarFeedSchema` (`feedSchema.ts`) is
+kept for backward-compat with already-cached catalog feeds, but `getRadarReferrals()`
+no longer reads it — see [Accessor](#accessor) below.
+
+### Sync
+
+`syncRadarReferrals()` (`src/lib/radar/referralsSync.ts`) is the ONLY module that
+touches the network for referrals, mirroring `syncRadar()`'s contract exactly: flag off
+→ `disabled`; opt-in false → `opt_out`; downloads `${RADAR_FEED_URL}/v1/referrals/latest`
+(same `RADAR_FEED_URL`/`RADAR_FEED_PUBKEY` fork overrides as the catalog), verifies the
+Ed25519 signature over the exact response bytes (`verifyFeedBytes`), validates against
+`RadarReferralsFeedSchema`, and caches into the `radar_referrals_cache` table
+(migration `142_radar_referrals_cache.sql`) — a table entirely separate from the
+catalog's `radar_feed_cache`. A 10 MB response cap and a `generatedAt` floor (an
+incoming feed with a `generatedAt` no newer than the cached one is treated as `stale`
+and never overwrites the cache — guards against a replay of an older signed artifact)
+mirror the catalog sync's own `MAX_FEED_BYTES`/version-floor guards. Never throws —
+always returns a status object; errors never carry a stack trace in `reason`.
+
+Two triggers keep the referrals cache warm, both independent of the catalog's own
+24h cadence:
+
+- **Sync-on-read** — `GET /api/radar/referrals` itself calls `syncRadarReferrals()`
+  inline whenever the cache is missing or older than `REFERRALS_STALE_MS` (1h,
+  `shouldSyncReferralsOnRead()`), before serving the response. This is what makes fixed
+  links "always current" for the very next dashboard load, without waiting on any
+  background timer.
+- **Scheduler side-sync** — `radarSchedulerTick()` (`scheduler.ts`) independently
+  evaluates referrals staleness on the same hourly tick used for the catalog, calling
+  `syncRadarReferrals()` when due. This runs regardless of whether the catalog itself
+  was due that tick, and never affects `RadarTickResult`'s shape (best-effort side
+  effect only, swallowed on error).
 
 ### Accessor
 
 `src/lib/radar/index.ts` exports two read-only accessors, both never throwing (same
-defensive contract as `getRadarCatalog()` — flag off, no cache, or a corrupt/old cached
+defensive contract as `getRadarCatalog()` — flag off, no cache, or a corrupt cached
 payload all resolve to the empty shape instead of an error):
 
-- `getRadarReferrals()` → `{ fixed: RadarReferral[], campaigns: RadarReferral[] }`.
+- `getRadarReferrals()` → `{ fixed: RadarReferral[], campaigns: RadarReferral[] }`,
+  reading from `radar_referrals_cache` (via `getRadarReferralsCache()`) and validating
+  through `RadarReferralsFeedSchema` — **not** the catalog cache.
 - `getDefaultReferralFor(provider)` → the `fixed` referral with `isDefault: true` for
   that provider, or `null`. Only looks at `fixed` — a campaign is never used as a
   provider's "default" link.
@@ -340,11 +391,13 @@ server-only; the providers dashboard imports `referrals.ts` directly instead of
 ### `GET /api/radar/referrals`
 
 Follows the exact same gate order as every other Radar route: `RADAR_ENABLED` off →
-`404` (checked first, byte-identical inertia); unauthenticated → `401`; otherwise `200`
-with `{ fixed, campaigns, tier }` — `tier` comes straight from the cache row and is
-purely informative (drives the UI's soft upsell copy below), the route does no
-gating of its own. Never proxies the feed server — same local-cache-only contract as
-`/api/radar/catalog`.
+`404` (checked first, byte-identical inertia); unauthenticated → `401`; otherwise
+triggers a sync-on-read (see above) when stale, then `200` with
+`{ fixed, campaigns, tier }` — `tier` comes straight from the (possibly just-refreshed)
+cache row and is purely informative (drives the UI's soft upsell copy below). Never
+proxies the feed server directly — the route's own source contains no `fetch(` call;
+the network only ever happens inside `syncRadarReferrals()`, same local-cache-only
+principle as `/api/radar/catalog`.
 
 ### Dashboard UI — "Free credits" tab on `/dashboard/radar`
 
@@ -412,6 +465,16 @@ automatically (`getFeedPublicKeys()` in `src/lib/radar/pinnedKeys.ts`), and vers
 comparison, schema validation, and the merge rules apply identically to a self-hosted
 feed.
 
+Referral links (see [Referral links (free credits)](#referral-links-free-credits)
+above) are a separate, optional artifact: a fork that only serves `/v1/catalog/latest`
+still works fully — `syncRadarReferrals()` degrades to `{ status: "error" }` on a `404`
+from `/v1/referrals/latest` and the cache simply stays empty, so
+`GET /api/radar/referrals` keeps returning `{ fixed: [], campaigns: [], tier: null }`
+instead of failing the rest of the page. To also offer referral links, serve
+`GET /v1/referrals/latest` satisfying `RadarReferralsFeedSchema`
+(`src/lib/radar/referralsFeedSchema.ts`) and sign it with the same Ed25519 key pair as
+the catalog feed.
+
 ---
 
 ## Related docs

docs/openapi.yaml

@@ -5294,6 +5294,19 @@ paths:
         "200":
           description: Caches cleared
 
+  /api/modality-bridge/stats:
+    get:
+      tags: [System]
+      summary: Get Modality Bridge telemetry
+      description: In-memory per-modality bridge counters (bridged, cacheHits, failures, lastUsedAt). Counters reset on process restart.
+      security:
+        - ManagementSessionAuth: []
+      responses:
+        "200":
+          description: Per-modality bridge stats (vision, audio)
+        "401":
+          description: Unauthorized
+
   /api/cache/stats:
     get:
       tags: [System]

docs/reference/API_REFERENCE.md

@@ -573,6 +573,7 @@ Response example:
 | `/api/rate-limits`       | GET        | Per-account rate limits                                                                              |
 | `/api/monitoring/health` | GET        | Health check + provider summary (`catalogCount`, `configuredCount`, `activeCount`, `monitoredCount`) |
 | `/api/cache/stats`       | GET/DELETE | Cache stats / clear                                                                                  |
+| `/api/modality-bridge/stats` | GET    | In-memory Modality Bridge telemetry — per-modality `bridged`/`cacheHits`/`failures`/`lastUsedAt` counters (reset on restart; management auth) |
 
 ### Backup & Export/Import
 

docs/security/GUARDRAILS.md

@@ -1,13 +1,13 @@
 ---
 title: "Guardrails"
-version: 3.8.40
-lastUpdated: 2026-06-28
+version: 3.8.50
+lastUpdated: 2026-08-07
 ---
 
 # Guardrails
 
 > **Source of truth:** `src/lib/guardrails/`
-> **Last updated:** 2026-06-28 — v3.8.40 (injection-guard coverage + 16 KB scan bound + red-team)
+> **Last updated:** 2026-08-07 — v3.8.50 (Modality Bridge PR-1: mode selector, task-aware prompt, describe cache, transparency header + stats)
 
 Guardrails enforce safety, policy, and content transformations at the boundary
 between OmniRoute and upstream providers. Each guardrail can inspect (and
@@ -32,31 +32,107 @@ The registry auto-loads four guardrails in priority order on import
 
 Lower priority numbers run **first**.
 
-### Vision Bridge (`visionBridge.ts`)
+### Vision Bridge (`visionBridge.ts`) — Modality Bridge PR-1
 
-Intercepts image-bearing requests aimed at **non-vision models** and replaces
-the image parts with text descriptions produced by a configurable vision model
-before the upstream call. This lets text-only providers transparently handle
+Intercepts image-bearing requests aimed at **non-vision models** and either
+reroutes the whole request to a vision-capable model or replaces the image
+parts with text descriptions produced by a configurable vision model before
+the upstream call. This lets text-only providers transparently handle
 multimodal payloads.
 
 Flow:
 
 1. Skip if the target model already supports vision (unless it appears in the
    forced-bridge list `isVisionBridgeForcedModel`).
-2. Extract image parts via `extractImageParts(messages)`. Skip if none.
-   `extractImageParts` recognizes all three image shapes: OpenAI `image_url`,
-   Anthropic base64 `source.type:"base64"`, and Anthropic URL
-   `source.type:"url"` — so Claude-Code-compatible clients (e.g. Zoo Code)
-   sending `{ type: "image", source: { type: "url", url } }` are described
-   instead of silently dropped.
-3. Load runtime config from `getSettings()` (`visionBridgeEnabled`,
-   `visionBridgeModel`, `visionBridgePrompt`, `visionBridgeTimeout`,
-   `visionBridgeMaxImages`).
-4. Cap images at `maxImages`, call the vision model **in parallel**
-   (`Promise.allSettled`), and inject `[Image N]: <description>` text parts
-   in their place — failed images become `[Image N]: (unavailable)`.
-5. Return `modifiedPayload` + meta (`imagesProcessed`, `processingTimeMs`,
-   `visionModel`).
+2. Extract image parts via `extractImageParts(messages)`
+   (`visionBridgeHelpers.ts`), which delegates to the **unified media
+   detector** `detectMediaParts()` in `open-sse/utils/mediaParts.ts` — the
+   single source of truth shared with the combo compatibility filter.
+   Extraction is allowlisted to top-level parts of the shapes
+   `replaceImageParts` can splice back (the extract↔replace contract): OpenAI
+   `image_url`, Anthropic base64 `source.type:"base64"`, Anthropic URL
+   `source.type:"url"`, and Responses API `input_image`. Nested hits and
+   indicator-only shapes are combo-filter material and are never extracted.
+   Skip if none found.
+3. Resolve runtime config via `resolveVisionBridgeRuntimeSettings()`
+   (`src/shared/constants/modalityBridgeDefaults.ts`): new `modalityBridge*`
+   settings keys win; legacy `visionBridge*` keys remain a **one-cycle
+   fallback** (rollback window). Skip before any media traversal when the
+   bridge is disabled.
+4. Mode selector (`modalityBridgeVisionMode`, see table below) decides
+   reroute vs describe. Reroute returns `modifiedPayload` with only `model`
+   swapped, plus meta `{ rerouted, fromModel, toModel, imagesKept }`.
+5. Describe path: cap images at `maxImages`, compose the task-aware prompt,
+   consult the describe cache, call the vision model **in parallel**
+   (`Promise.allSettled`), and inject `[Image N]: <description>` text parts in
+   their place. A failed describe yields `null` and the original image part is
+   **preserved** (#4012) — except on the combo describe path when every
+   describe failed, where a confirmed non-vision upstream gets an
+   `(unavailable — no vision-capable provider connected)` stub instead (#8430).
+6. Return `modifiedPayload` + meta (`imagesProcessed`, `descriptions`,
+   `processingTimeMs`, `visionModel`).
+
+#### Mode selector (`modalityBridgeVisionMode`)
+
+| Mode       | Default | Behavior                                                                                                                                                                                                                                                |
+| ---------- | ------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| `auto`     | ✔       | Legacy heuristic, untouched (#6640/#7204): non-combo/`auto/` models reroute to the best vision model unless the original model already has usable credentials (then describe); combo targets always describe.                                           |
+| `describe` |         | Always describe — the reroute block is skipped entirely; the user's chosen model always answers.                                                                                                                                                        |
+| `reroute`  |         | Force reroute: the keep-credentialed-model guard is bypassed. The reroute-**target** credential guard still applies — when no usable vision target exists, the request falls through to describe so raw images never reach a text-only backend (#8430). |
+
+Forced modes short-circuit **before** the auto heuristic runs; `auto` behavior
+is byte-identical to the pre-PR-1 guardrail.
+
+#### Task-aware describe prompt (`modalityBridgeVisionTaskAware`)
+
+Default **true**. `composeVisionPrompt()` (`visionBridgeHelpers.ts`) appends
+the text of the **last user message** (truncated to 500 chars) to the base
+describe prompt, steering the description toward what the user actually asked
+(codex-vision-proxy pattern) and asking the vision model to transcribe visible
+text. With the flag off — or no user text — the base prompt is used unchanged.
+
+#### Describe cache (`modalityBridge/bridgeCache.ts`)
+
+In-memory LRU + TTL cache for describe outputs, shared process-wide.
+Key = `sha256(imageRef + composedPrompt + configuredBridgeModel)` with
+length-prefix framing (no field-boundary collisions). The model component is
+the **configured** bridge model, not the model that actually answered —
+`callVisionModel` may fall back internally, and keying per attempt would
+fragment the cache. Failed describes are never cached. Settings:
+
+| Key                             | Default | Range   |
+| ------------------------------- | ------- | ------- |
+| `modalityBridgeCacheEnabled`    | `true`  | —       |
+| `modalityBridgeCacheTtlMinutes` | `60`    | 1–1440  |
+| `modalityBridgeCacheMaxEntries` | `200`   | 10–5000 |
+
+#### Settings schema + migration
+
+The new `modalityBridge*` keys are Zod-validated in `updateSettingsSchema`
+(`src/shared/validation/settingsSchemas.ts`): `modalityBridgeVisionEnabled`,
+`modalityBridgeVisionMode`, `modalityBridgeVisionModel`,
+`modalityBridgeVisionTaskAware`, `modalityBridgeVisionPrompt`,
+`modalityBridgeVisionTimeout`, `modalityBridgeVisionMaxImages`, the
+`modalityBridgeCache*` trio, and the PR-3-reserved `modalityBridgeAudio*`
+group. Migration `141_modality_bridge_settings.sql` copies existing legacy
+`visionBridge*` values to the matching new keys (idempotent, never overwrites
+an operator-set `modalityBridge*` value); the legacy keys stay accepted as a
+read fallback for one release cycle.
+
+#### Transparency header + stats
+
+Describe-transformed responses carry
+`x-omniroute-modality-bridge: image->text;model=<visionModel>;parts=<n>`
+(built by `buildModalityBridgeHeader()` in `modalityBridge/bridgeStats.ts`,
+stamped by `withModalityBridgeHeader()` in `src/sse/handlers/chatHelpers.ts`).
+Rerouted requests get **no** header — the payload was untouched and the model
+swap is already visible in the response body's `model` field.
+
+`GET /api/modality-bridge/stats` (management auth, same tier as
+`GET /api/settings`) returns the in-memory per-modality counters
+`{ bridged, cacheHits, failures, lastUsedAt }` for `vision` (and the
+PR-3-reserved `audio`). Counters reset on process restart by design
+(telemetry, not accounting).
 
 **Self-loop admission bypass:** when the describe call routes through OmniRoute's
 own `/v1` self-loop (non-standard provider model), the sub-request sends
@@ -67,8 +143,10 @@ operator-configured `OMNIROUTE_API_KEY` / `ROUTER_API_KEY` env key (#1350) so
 is only honored for those exact credentials, so external clients cannot use the
 header to skip admission.
 
-Defaults live in `src/shared/constants/visionBridgeDefaults.ts`. The guardrail
-exposes a `deps` constructor option so tests can inject fake `getSettings` and
+Legacy defaults live in `src/shared/constants/visionBridgeDefaults.ts`; the
+new mode/task-aware/cache defaults and the settings resolver live in
+`src/shared/constants/modalityBridgeDefaults.ts`. The guardrail exposes a
+`deps` constructor option so tests can inject fake `getSettings` and
 `callVisionModel` implementations.
 
 ### PII Masker (`piiMasker.ts`)

open-sse/config/audioRegistry.ts

@@ -603,7 +603,7 @@ export interface ProviderNodeRow {
 }
 
 /** Hosts reachable only from the operator's machine/Docker network. */
-function isLoopbackNodeHost(baseUrl: string): boolean {
+export function isLoopbackNodeHost(baseUrl: string): boolean {
   try {
     const hostname = new URL(baseUrl).hostname;
     return (

open-sse/config/errorConfig.ts

@@ -149,6 +149,18 @@ export const ERROR_RULES: ErrorRule[] = [
     backoff: true,
     reason: "quota_exhausted",
   },
+  {
+    id: "out_of_extra_usage",
+    text: "out of extra usage",
+    backoff: true,
+    reason: "quota_exhausted",
+  },
+  {
+    id: "extra_usage_required",
+    text: "extra usage required",
+    backoff: true,
+    reason: "quota_exhausted",
+  },
   { id: "capacity", text: "capacity", backoff: true, reason: "model_capacity" },
   { id: "overloaded", text: "overloaded", backoff: true, reason: "model_capacity" },
   { id: "high_demand", text: "high demand", backoff: true, reason: "model_capacity" },

open-sse/config/freeModelCatalog.data.ts

@@ -311,7 +311,6 @@ export const FREE_MODEL_BUDGETS: FreeModelBudget[] = [
   { provider: "nscale", modelId: "openai/gpt-oss-20b", displayName: "openai/gpt-oss-20b", monthlyTokens: 0, creditTokens: 5000000, freeType: "one-time-initial", poolKey: "nscale", tos: "caution" },
   { provider: "nscale", modelId: "meta-llama/Llama-4-Scout-17B-16E-Instruct", displayName: "meta-llama/Llama-4-Scout-17B-16E-Instruct", monthlyTokens: 0, creditTokens: 5000000, freeType: "one-time-initial", poolKey: "nscale", tos: "caution" },
   { provider: "nscale", modelId: "meta-llama/Llama-3.3-70B-Instruct", displayName: "meta-llama/Llama-3.3-70B-Instruct", monthlyTokens: 0, creditTokens: 5000000, freeType: "one-time-initial", poolKey: "nscale", tos: "caution" },
-  { provider: "nvidia", modelId: "z-ai/glm-5.1", displayName: "GLM 5.1", monthlyTokens: 0, creditTokens: 0, freeType: "one-time-initial", poolKey: "nvidia", tos: "caution" },
   { provider: "nvidia", modelId: "z-ai/glm-5.2", displayName: "GLM 5.2", monthlyTokens: 0, creditTokens: 0, freeType: "one-time-initial", poolKey: "nvidia", tos: "caution" },
   { provider: "nvidia", modelId: "minimaxai/minimax-m2.7", displayName: "MiniMax M2.7", monthlyTokens: 0, creditTokens: 0, freeType: "one-time-initial", poolKey: "nvidia", tos: "caution" },
   { provider: "nvidia", modelId: "google/gemma-4-31b-it", displayName: "Gemma 4 31B", monthlyTokens: 0, creditTokens: 0, freeType: "one-time-initial", poolKey: "nvidia", tos: "caution" },
@@ -321,7 +320,6 @@ export const FREE_MODEL_BUDGETS: FreeModelBudget[] = [
   { provider: "nvidia", modelId: "qwen/qwen3.5-397b-a17b", displayName: "Qwen3.5-397B-A17B", monthlyTokens: 0, creditTokens: 0, freeType: "one-time-initial", poolKey: "nvidia", tos: "caution" },
   { provider: "nvidia", modelId: "qwen/qwen3.5-122b-a10b", displayName: "Qwen3.5-122B-A10B", monthlyTokens: 0, creditTokens: 0, freeType: "one-time-initial", poolKey: "nvidia", tos: "caution" },
   { provider: "nvidia", modelId: "stepfun-ai/step-3.5-flash", displayName: "Step 3.5 Flash", monthlyTokens: 0, creditTokens: 0, freeType: "one-time-initial", poolKey: "nvidia", tos: "caution" },
-  { provider: "nvidia", modelId: "deepseek-ai/deepseek-v4-pro", displayName: "DeepSeek V4 Pro", monthlyTokens: 0, creditTokens: 0, freeType: "one-time-initial", poolKey: "nvidia", tos: "caution" },
   { provider: "nvidia", modelId: "openai/gpt-oss-120b", displayName: "GPT OSS 120B", monthlyTokens: 0, creditTokens: 0, freeType: "one-time-initial", poolKey: "nvidia", tos: "caution" },
   { provider: "nvidia", modelId: "openai/gpt-oss-20b", displayName: "GPT OSS 20B", monthlyTokens: 0, creditTokens: 0, freeType: "one-time-initial", poolKey: "nvidia", tos: "caution" },
   { provider: "nvidia", modelId: "nvidia/nemotron-3-super-120b-a12b", displayName: "Nemotron 3 Super 120B A12B", monthlyTokens: 0, creditTokens: 0, freeType: "one-time-initial", poolKey: "nvidia", tos: "caution" },

open-sse/config/imageRegistry.ts

@@ -12,6 +12,10 @@ import { FREEPIK_IMAGE_PROVIDER } from "./providers/registry/freepik/index.ts";
 import { STABILITY_AI_IMAGE_MODELS } from "./providers/registry/stability-ai/imageModels.ts";
 import { GEMINI_IMAGEN_PROVIDER } from "./providers/registry/gemini/imageModels.ts";
 import { CHEAPERINFERENCE_IMAGE_PROVIDER } from "./providers/registry/cheaperinference/imageModels.ts";
+import {
+  ADOBE_FIREFLY_IMAGE_ROUTING_ALIASES,
+  toRegistryImageModels,
+} from "../services/adobeFireflyModels.ts";
 
 interface ImageModelEntry {
   id: string;
@@ -22,6 +26,8 @@ interface ImageModelEntry {
   imageRequired?: boolean;
   description?: string;
   isMarket?: boolean;
+  supportedSizes?: string[];
+  mediaCapabilities?: Record<string, unknown>;
 }
 
 interface ImageProviderConfig {
@@ -35,6 +41,7 @@ interface ImageProviderConfig {
   authHeader: string;
   format: string;
   models: ImageModelEntry[];
+  routingAliases?: readonly string[];
   supportedSizes: string[];
 }
 
@@ -46,6 +53,7 @@ interface ImageModelAliasEntry {
   inputModalities?: string[];
   imageRequired?: boolean;
   description?: string;
+  mediaCapabilities?: Record<string, unknown>;
 }
 
 interface ImageCatalogModelEntry {
@@ -55,6 +63,7 @@ interface ImageCatalogModelEntry {
   supportedSizes: string[];
   inputModalities: string[];
   description?: string;
+  mediaCapabilities?: Record<string, unknown>;
 }
 
 const IMAGE_MODEL_ALIASES: Record<string, ImageModelAliasEntry> = {
@@ -678,55 +687,9 @@ export const IMAGE_PROVIDERS: Record<string, ImageProviderConfig> = {
     authType: "apikey",
     authHeader: "bearer",
     format: "adobe-firefly-image",
-    models: [
-      {
-        id: "nano-banana-pro",
-        name: "Firefly Gemini 3.0 (Nano Banana Pro)",
-        inputModalities: ["text", "image"],
-      },
-      {
-        id: "nano-banana",
-        name: "Firefly Gemini 2.5 (Nano Banana)",
-        inputModalities: ["text", "image"],
-      },
-      {
-        id: "nano-banana-2",
-        name: "Firefly Gemini 3.1 (Nano Banana 2)",
-        inputModalities: ["text", "image"],
-      },
-      { id: "gpt-image-2", name: "Firefly GPT Image 2", inputModalities: ["text", "image"] },
-      { id: "gpt-image", name: "Firefly GPT Image 2", inputModalities: ["text", "image"] },
-      { id: "gpt-image-1.5", name: "Firefly GPT Image 1.5", inputModalities: ["text", "image"] },
-      { id: "flux-2", name: "Firefly Flux 2", inputModalities: ["text", "image"] },
-      { id: "flux-pro", name: "Firefly Flux 1.1 Pro", inputModalities: ["text", "image"] },
-      { id: "flux-ultra", name: "Firefly Flux 1.1 Ultra", inputModalities: ["text", "image"] },
-      { id: "seedream-4", name: "Firefly Seedream 4.0", inputModalities: ["text", "image"] },
-      {
-        id: "seedream-5-lite",
-        name: "Firefly Seedream 5.0 Lite",
-        inputModalities: ["text", "image"],
-      },
-      {
-        id: "runway-gen4-image",
-        name: "Firefly Runway Gen-4 Image",
-        inputModalities: ["text", "image"],
-      },
-      // Topaz Labs upscalers (inputMediaUseCase: ["upscaling"]).
-      // Served by firefly-3p /v2/3p-images/upsample — see config/upscaleRegistry.ts.
-      {
-        id: "topaz-standard",
-        name: "Firefly Topaz Upscale (Standard)",
-        inputModalities: ["image"],
-        imageRequired: true,
-      },
-      {
-        id: "topaz-bloom",
-        name: "Firefly Topaz Bloom (Creative Upscale)",
-        inputModalities: ["image"],
-        imageRequired: true,
-      },
-    ],
-    supportedSizes: ["1:1", "16:9", "9:16", "4:3", "3:4", "1024x1024", "1792x1024", "1024x1792"],
+    models: toRegistryImageModels(),
+    routingAliases: ADOBE_FIREFLY_IMAGE_ROUTING_ALIASES,
+    supportedSizes: [],
   },
 
   // Cheaper Inference (OSS-sponsor gateway). Declared AFTER adobe-firefly on
@@ -887,7 +850,7 @@ export function parseImageModel(modelStr) {
 
   // No provider prefix — try to find the model in every provider
   for (const [providerId, config] of Object.entries(IMAGE_PROVIDERS)) {
-    if (config.models.some((m) => m.id === modelStr)) {
+    if (config.routingAliases?.includes(modelStr) || config.models.some((m) => m.id === modelStr)) {
       return { provider: providerId, model: modelStr };
     }
   }
@@ -906,9 +869,10 @@ function imageProviderCatalogEntries(
     id: `${providerId}/${model.id}`,
     name: model.name,
     provider: providerId,
-    supportedSizes: config.supportedSizes,
+    supportedSizes: model.supportedSizes || config.supportedSizes,
     inputModalities: model.inputModalities || ["text"],
     description: model.description || undefined,
+    mediaCapabilities: model.mediaCapabilities,
   }));
 }
 

open-sse/config/nvidiaHostedModels.snapshot.json

@@ -1,5 +1,4 @@
 [
-  "deepseek-ai/deepseek-v4-pro",
   "google/gemma-4-31b-it",
   "minimaxai/minimax-m2.7",
   "mistralai/devstral-2-123b-instruct-2512",
@@ -13,6 +12,5 @@
   "qwen/qwen3.5-397b-a17b",
   "stepfun-ai/step-3.5-flash",
   "thinkingmachines/inkling",
-  "z-ai/glm-5.1",
   "z-ai/glm-5.2"
 ]

open-sse/config/providers/registry/gemini/web/index.ts

@@ -8,9 +8,32 @@ export const gemini_webProvider: RegistryEntry = {
   baseUrl: "https://gemini.google.com/app",
   authType: "apikey",
   authHeader: "cookie",
+  // #9356: `supportsReasoning: false` is a live-behavior statement, not a guess
+  // about the underlying Gemini model. The executor drives the gemini.google.com
+  // web UI by typing a prompt, so it has no thinking-budget control to set and
+  // never surfaces `reasoning_content` — agent routers reading /v1/models must
+  // not select these for reasoning work. `toolCalling: false` is the matching
+  // statement for native function calling; the prompt-emulation shim (#7286)
+  // stays available and is advertised separately as `toolCalling: "emulated"`
+  // on the provider constant (src/shared/constants/providers/web-cookie.ts).
   models: [
-    { id: "gemini-3.1-pro", name: "Gemini 3.1 Pro", toolCalling: false },
-    { id: "gemini-3.5-flash", name: "Gemini 3.5 Flash", toolCalling: false },
-    { id: "gemini-3.1-flash-lite", name: "Gemini 3.1 Flash-Lite", toolCalling: false },
+    {
+      id: "gemini-3.1-pro",
+      name: "Gemini 3.1 Pro",
+      toolCalling: false,
+      supportsReasoning: false,
+    },
+    {
+      id: "gemini-3.5-flash",
+      name: "Gemini 3.5 Flash",
+      toolCalling: false,
+      supportsReasoning: false,
+    },
+    {
+      id: "gemini-3.1-flash-lite",
+      name: "Gemini 3.1 Flash-Lite",
+      toolCalling: false,
+      supportsReasoning: false,
+    },
   ],
 };

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

@@ -32,8 +32,6 @@ export const nvidiaProvider: RegistryEntry = {
     { id: "qwen/qwen3.5-122b-a10b", name: "Qwen3.5-122B-A10B" },
     { id: "stepfun-ai/step-3.5-flash", name: "Step 3.5 Flash" },
     { id: "stepfun-ai/step-3.7-flash", name: "Step 3.7 Flash" },
-    { id: "deepseek-ai/deepseek-v4-pro", name: "DeepSeek V4 Pro", supportsReasoning: true },
-    { id: "deepseek-ai/deepseek-v4-flash", name: "DeepSeek V4 Flash", supportsReasoning: true },
     // Sweep 2026-06-19: verified present in the live NVIDIA NIM /v1/models catalog.
     { id: "moonshotai/kimi-k2.6", name: "Kimi K2.6" },
     { id: "openai/gpt-oss-120b", name: "GPT OSS 120B", toolCalling: false },

open-sse/config/videoRegistry.ts

@@ -5,14 +5,17 @@
  * Supports local providers plus hosted task-based APIs such as Runway.
  */
 
-import { parseModelFromRegistry, getAllModelsFromRegistry } from "./registryUtils.ts";
+import { parseModelFromRegistry } from "./registryUtils.ts";
 import { RUNWAYML_SUPPORTED_VIDEO_MODELS } from "./runway.ts";
 import { SEGMIND_VIDEO_MODELS } from "./providers/registry/segmind/videoModels.ts";
+import { toRegistryVideoModels } from "../services/adobeFireflyModels.ts";
 
 interface VideoModel {
   id: string;
   name: string;
   isMarket?: boolean;
+  supportedSizes?: string[];
+  mediaCapabilities?: Record<string, unknown>;
 }
 
 interface VideoProvider {
@@ -326,8 +329,7 @@ export const VIDEO_PROVIDERS: Record<string, VideoProvider> = {
   },
 
   // Adobe Firefly (unofficial) — same IMS/cookie credential as the image entry.
-  // Async 3P video generate + poll (Sora 2, Veo 3.1, Kling …). Fallback list
-  // from models/discovery capture (adobe/get_models.txt).
+  // Exact async video models and capabilities from the verified discovery snapshot.
   "adobe-firefly": {
     id: "adobe-firefly",
     alias: "firefly",
@@ -335,18 +337,7 @@ export const VIDEO_PROVIDERS: Record<string, VideoProvider> = {
     authType: "apikey",
     authHeader: "bearer",
     format: "adobe-firefly-video",
-    models: [
-      { id: "sora-2", name: "Firefly Sora 2" },
-      { id: "sora-2-pro", name: "Firefly Sora 2 Pro" },
-      { id: "veo-3.1", name: "Firefly Veo 3.1" },
-      { id: "veo-3.1-fast", name: "Firefly Veo 3.1 Fast" },
-      { id: "veo-3.1-ref", name: "Firefly Veo 3.1 Reference" },
-      { id: "kling-3", name: "Firefly Kling v3 Standard I2V" },
-      { id: "kling-v3-t2v", name: "Firefly Kling v3 Standard T2V" },
-      { id: "kling-v3-pro-i2v", name: "Firefly Kling v3 Pro I2V" },
-      { id: "luma-ray3", name: "Firefly Ray3" },
-      { id: "runway-gen4-turbo", name: "Firefly Runway Gen-4 Video" },
-    ],
+    models: toRegistryVideoModels(),
   },
 };
 
@@ -368,5 +359,17 @@ export function parseVideoModel(modelStr: string | null) {
  * Get all video models as a flat list
  */
 export function getAllVideoModels() {
-  return getAllModelsFromRegistry(VIDEO_PROVIDERS);
+  return Object.entries(VIDEO_PROVIDERS).flatMap(([providerId, config]) =>
+    [providerId, config.alias]
+      .filter((prefix): prefix is string => Boolean(prefix))
+      .flatMap((prefix) =>
+        config.models.map((model) => ({
+          id: `${prefix}/${model.id}`,
+          name: model.name,
+          provider: providerId,
+          supportedSizes: model.supportedSizes || [],
+          mediaCapabilities: model.mediaCapabilities,
+        }))
+      )
+  );
 }

open-sse/executors/codex.ts

@@ -309,6 +309,7 @@ export function stripStoredItemReferences(body: Record<string, unknown>): void {
 
 function stripOrphanedCodexFunctionCallOutputs(body: Record<string, unknown>): void {
   if (!Array.isArray(body.input)) return;
+  const input = body.input;
   // A previous_response_id delegates history resolution to the upstream
   // Responses service, so a matching function_call may legitimately live in
   // that remote response rather than in the local input array.
@@ -317,7 +318,7 @@ function stripOrphanedCodexFunctionCallOutputs(body: Record<string, unknown>): v
   const callIds = new Set<string>();
   let outputCount = 0;
 
-  for (const item of body.input) {
+  for (const item of input) {
     if (!item || typeof item !== "object" || Array.isArray(item)) continue;
     const record = item as Record<string, unknown>;
 
@@ -341,9 +342,7 @@ function stripOrphanedCodexFunctionCallOutputs(body: Record<string, unknown>): v
   }
 
   if (outputCount === 0) return;
-
-  const before = body.input.length;
-  body.input = body.input.filter((item) => {
+  const filteredInput = input.filter((item) => {
     if (!item || typeof item !== "object" || Array.isArray(item)) return true;
     const record = item as Record<string, unknown>;
     if (record.type === "function_call_output" && typeof record.call_id === "string") {
@@ -352,7 +351,8 @@ function stripOrphanedCodexFunctionCallOutputs(body: Record<string, unknown>): v
     return true;
   });
 
-  const removedCount = before - body.input.length;
+  const removedCount = input.length - filteredInput.length;
+  body.input = filteredInput;
   if (removedCount > 0) {
     console.debug(
       `[Codex] stripOrphanedCodexFunctionCallOutputs: removed ${removedCount} orphaned function_call_output item(s)`

open-sse/executors/codex/tools.ts

@@ -30,6 +30,114 @@ export function isCodexFreePlan(providerSpecificData: unknown): boolean {
   return typeof plan === "string" && plan.trim().toLowerCase() === "free";
 }
 
+type JsonRecord = Record<string, unknown>;
+
+const REDUNDANT_ONEOF_OBJECT_MAP_FIELDS = [
+  "properties",
+  "patternProperties",
+  "$defs",
+  "definitions",
+] as const;
+
+const REDUNDANT_ONEOF_ARRAY_SCHEMA_FIELDS = ["prefixItems", "oneOf", "anyOf", "allOf"] as const;
+
+const REDUNDANT_ONEOF_SINGLE_SCHEMA_FIELDS = [
+  "items",
+  "additionalProperties",
+  "not",
+  "if",
+  "then",
+  "else",
+] as const;
+
+const REDUNDANT_ONEOF_ANNOTATION_KEYS = new Set(["const", "description", "title", "$comment"]);
+
+/**
+ * Remove a redundant `oneOf` when it is fully covered by a sibling `enum`.
+ *
+ * The Codex private Responses endpoint (`chatgpt.com/backend-api/codex/responses`)
+ * intermittently returns a 502 `upstream_empty_response` when a tool parameter
+ * carries the JSON-Schema pattern `oneOf: [{const, ...annotations}]` together
+ * with a sibling `enum` whose value set exactly matches the `const` set. In that
+ * case `oneOf` adds no constraint beyond `enum`, so dropping it is semantically
+ * safe and eliminates the trigger.
+ *
+ * Only the exact-match redundant case is stripped. Bare `oneOf[const]` without
+ * a sibling `enum`, narrowing const sets, non-matching enums, type-discriminated
+ * `oneOf`, and `anyOf`/`allOf` are all preserved.
+ */
+export function stripRedundantOneOfConstEnum(schema: unknown): unknown {
+  if (Array.isArray(schema)) {
+    return schema.map((entry) => stripRedundantOneOfConstEnum(entry));
+  }
+  if (!isPlainObject(schema)) return schema;
+
+  const result: JsonRecord = { ...schema };
+
+  maybeStripRedundantOneOf(result);
+
+  for (const field of REDUNDANT_ONEOF_OBJECT_MAP_FIELDS) {
+    const map = result[field];
+    if (isPlainObject(map)) {
+      result[field] = Object.fromEntries(
+        Object.entries(map).map(([key, value]) => [key, stripRedundantOneOfConstEnum(value)])
+      );
+    }
+  }
+
+  for (const field of REDUNDANT_ONEOF_ARRAY_SCHEMA_FIELDS) {
+    if (Array.isArray(result[field])) {
+      result[field] = (result[field] as unknown[]).map((entry) =>
+        stripRedundantOneOfConstEnum(entry)
+      );
+    }
+  }
+
+  for (const field of REDUNDANT_ONEOF_SINGLE_SCHEMA_FIELDS) {
+    if (result[field] !== undefined) {
+      result[field] = stripRedundantOneOfConstEnum(result[field]);
+    }
+  }
+
+  return result;
+}
+
+function maybeStripRedundantOneOf(node: JsonRecord): void {
+  const branches = node.oneOf;
+  if (!Array.isArray(branches) || branches.length === 0) return;
+
+  const enumValues = Array.isArray(node.enum) ? node.enum : null;
+  if (!enumValues || enumValues.length === 0) return;
+
+  // Every branch must be {const, ...annotations only}.
+  const constValues: unknown[] = [];
+  for (const branch of branches) {
+    if (!isPlainObject(branch)) return;
+    const branchKeys = Object.keys(branch);
+    if (!branchKeys.includes("const")) return;
+    if (!branchKeys.every((key) => REDUNDANT_ONEOF_ANNOTATION_KEYS.has(key))) return;
+    constValues.push((branch as JsonRecord).const);
+  }
+
+  // Restrict to string consts and string enums (confirmed production shape).
+  if (!constValues.every((value) => typeof value === "string")) return;
+  if (!enumValues.every((value) => typeof value === "string")) return;
+
+  // All const values must be unique.
+  if (new Set(constValues).size !== constValues.length) return;
+
+  // The const set must exactly match the enum set.
+  const enumSet = new Set(enumValues);
+  if (enumSet.size !== constValues.length) return;
+  if (!constValues.every((value) => enumSet.has(value))) return;
+
+  delete node.oneOf;
+}
+
+function isPlainObject(value: unknown): value is JsonRecord {
+  return typeof value === "object" && value !== null && !Array.isArray(value);
+}
+
 export function normalizeCodexTools(
   body: Record<string, unknown>,
   options?: { dropImageGeneration?: boolean; preserveCustomTools?: boolean }
@@ -138,7 +246,9 @@ export function normalizeCodexTools(
     // Codex/OpenAI Responses API rejects `pattern` fields using regex lookaround
     // (e.g. `^(?=.*@).+$`) with a 400 "regex lookaround is not supported" error.
     // Strip those before the schema reaches upstream (9router#1556).
-    const sanitizedParameters = stripUnsupportedRegexPatterns(parameters);
+    const sanitizedParameters = stripRedundantOneOfConstEnum(
+      stripUnsupportedRegexPatterns(parameters)
+    );
 
     // Rewrite in-place to Responses format
     for (const key of Object.keys(tool)) {

open-sse/executors/commandCode.ts

@@ -420,7 +420,61 @@ type AggregateState = {
   usage: JsonRecord | null;
 };
 
+function firstRecord(record: JsonRecord, keys: readonly string[]): JsonRecord {
+  for (const key of keys) {
+    const value = record[key];
+    if (isRecord(value)) return value;
+  }
+  return {};
+}
+
+function firstNumber(record: JsonRecord, keys: readonly string[]): number | undefined {
+  for (const key of keys) {
+    const value = numberValue(record[key]);
+    if (value !== undefined) return value;
+  }
+  return undefined;
+}
+
+/** Keep earlier finish-step usage when the terminal finish event omits it. */
+function mergeCommandCodeUsage(previous: JsonRecord | null, next: unknown): JsonRecord | null {
+  if (!isRecord(next)) return previous;
+
+  const merged: JsonRecord = { ...(previous || {}), ...next };
+  for (const key of [
+    "inputTokenDetails",
+    "input_token_details",
+    "input_tokens_details",
+    "prompt_tokens_details",
+    "outputTokenDetails",
+    "output_token_details",
+    "output_tokens_details",
+    "completion_tokens_details",
+    "reasoningTokenDetails",
+    "reasoning_token_details",
+  ]) {
+    const before = isRecord(previous?.[key]) ? previous[key] : {};
+    const after = isRecord(next[key]) ? next[key] : {};
+    if (Object.keys(before).length > 0 || Object.keys(after).length > 0) {
+      merged[key] = { ...before, ...after };
+    }
+  }
+  return merged;
+}
+
+function rememberCommandCodeUsage(state: AggregateState, event: JsonRecord): void {
+  const usage =
+    event.type === "finish-step"
+      ? (event.usage ?? event.totalUsage)
+      : (event.totalUsage ?? event.usage);
+  state.usage = mergeCommandCodeUsage(state.usage, usage);
+}
+
 function applyEventToAggregate(event: JsonRecord, state: AggregateState): void {
+  // Some Command Code protocol revisions attach usage to the terminal payload
+  // without preserving the event type. Capture it before event-specific handling.
+  rememberCommandCodeUsage(state, event);
+
   switch (event.type) {
     case "text-delta":
       state.content += stringValue(event.text) || "";
@@ -440,9 +494,10 @@ function applyEventToAggregate(event: JsonRecord, state: AggregateState): void {
       });
       break;
     }
+    case "finish-step":
+      break;
     case "finish":
       state.finishReason = mapFinishReason(event.finishReason);
-      state.usage = isRecord(event.totalUsage) ? event.totalUsage : null;
       break;
   }
 }
@@ -460,30 +515,72 @@ function applyEventToAggregateOrThrow(event: JsonRecord, state: AggregateState):
 
 function usageFromCommandCode(usage: JsonRecord | null) {
   if (!usage) return undefined;
-  const details = isRecord(usage.inputTokenDetails) ? usage.inputTokenDetails : {};
-  const cacheRead = numberValue(details.cacheReadTokens) || 0;
-  const noCache = numberValue(details.noCacheTokens) || 0;
+  const inputDetails = firstRecord(usage, [
+    "inputTokenDetails",
+    "input_token_details",
+    "input_tokens_details",
+    "prompt_tokens_details",
+  ]);
+  const outputDetails = firstRecord(usage, [
+    "outputTokenDetails",
+    "output_token_details",
+    "output_tokens_details",
+    "completion_tokens_details",
+  ]);
+  const reasoningDetails = firstRecord(usage, [
+    "reasoningTokenDetails",
+    "reasoning_token_details",
+    "reasoning_tokens_details",
+  ]);
+  const cacheRead =
+    firstNumber(usage, [
+      "cachedInputTokens",
+      "cached_input_tokens",
+      "cacheReadInputTokens",
+      "cache_read_input_tokens",
+      "cacheReadTokens",
+      "cache_read_tokens",
+      "cached_tokens",
+    ]) ??
+    firstNumber(inputDetails, [
+      "cachedTokens",
+      "cached_tokens",
+      "cacheReadTokens",
+      "cache_read_tokens",
+    ]);
+  const noCache = firstNumber(inputDetails, ["noCacheTokens", "no_cache_tokens"]);
   // Command Code's totalUsage.inputTokens is the FULL prompt total and already
   // includes the cached portion (noCacheTokens + cacheReadTokens = inputTokens),
   // so we must NOT add cacheRead back — that would double-count. There is no
   // cache-write field in the upstream payload, so cache creation stays unset.
-  const inputTokens = numberValue(usage.inputTokens) || 0;
-  const prompt = inputTokens;
-  const completion = numberValue(usage.outputTokens) || 0;
+  const prompt =
+    firstNumber(usage, ["inputTokens", "input_tokens", "promptTokens", "prompt_tokens"]) ??
+    (noCache ?? 0) + (cacheRead ?? 0);
+  const reasoning =
+    firstNumber(usage, ["reasoningTokens", "reasoning_tokens"]) ??
+    firstNumber(outputDetails, ["reasoningTokens", "reasoning_tokens"]) ??
+    firstNumber(reasoningDetails, ["reasoningTokens", "reasoning_tokens"]);
+  const textOutput = firstNumber(outputDetails, ["textTokens", "text_tokens"]);
+  const completion =
+    firstNumber(usage, [
+      "outputTokens",
+      "output_tokens",
+      "completionTokens",
+      "completion_tokens",
+    ]) ?? (textOutput ?? 0) + (reasoning ?? 0);
+  const total = firstNumber(usage, ["totalTokens", "total_tokens"]) ?? prompt + completion;
   const result: JsonRecord = {
     prompt_tokens: prompt,
+    prompt_tokens_details: { cached_tokens: cacheRead ?? 0 },
     completion_tokens: completion,
-    total_tokens: prompt + completion,
+    completion_tokens_details: { reasoning_tokens: reasoning ?? 0 },
+    total_tokens: total,
   };
   // Surface the cache breakdown as informational fields so logUsage prints
   // `| cache_read=X | no_cache=Y` and appendRequestLog persists them. These are
   // NOT added to prompt_tokens (already included) — metering stays accurate.
-  if (cacheRead > 0) result.cache_read_input_tokens = cacheRead;
-  if (noCache > 0) result.no_cache_tokens = noCache;
-  // Keep reasoning_token_details (reasoningTokens) when present so stream.ts's
-  // extractUsage can surface it as reasoning_tokens.
-  const reasoningDetails = isRecord(usage.reasoningTokenDetails) ? usage.reasoningTokenDetails : {};
-  const reasoning = numberValue(reasoningDetails.reasoningTokens);
+  if (cacheRead !== undefined && cacheRead > 0) result.cache_read_input_tokens = cacheRead;
+  if (noCache !== undefined && noCache > 0) result.no_cache_tokens = noCache;
   if (reasoning !== undefined && reasoning > 0) result.reasoning_tokens = reasoning;
   return result;
 }
@@ -523,6 +620,7 @@ function createStreamResponse(
 
       const emitEvent = (event: unknown) => {
         if (!isRecord(event) || closed) return;
+        rememberCommandCodeUsage(state, event);
         if (!sentRole) {
           sentRole = true;
           controller.enqueue(sse(chatCompletionChunk(id, model, { role: "assistant" })));
@@ -562,9 +660,10 @@ function createStreamResponse(
           }
           case "reasoning-end":
             break;
+          case "finish-step":
+            break;
           case "finish": {
             state.finishReason = mapFinishReason(event.finishReason);
-            state.usage = isRecord(event.totalUsage) ? event.totalUsage : null;
             controller.enqueue(sse(chatCompletionChunk(id, model, {}, state.finishReason)));
             // Emit a standards-compliant usage-only chunk (choices: []) before
             // [DONE] when upstream reported usage. stream.ts's extractUsage

open-sse/executors/devin-agentic/serializer.ts

@@ -119,7 +119,8 @@ function serializeMessage(
       "unsupported_role"
     );
   }
-  const label = role === "assistant" ? "Assistant" : role === "system" ? "System" : "User";
+  // role was just narrowed to "user" | "assistant" by the guard above ("system" throws).
+  const label = role === "assistant" ? "Assistant" : "User";
   const content = record.content;
 
   if (typeof content === "string") return `[${label}]\n${content}`;

open-sse/executors/gemini-web.ts

@@ -14,9 +14,13 @@
  */
 
 import { BaseExecutor, type ExecuteInput } from "./base.ts";
-import { sanitizeErrorMessage } from "../utils/error.ts";
+import { buildErrorBody, sanitizeErrorMessage } from "../utils/error.ts";
 import { prepareToolMessages } from "../translator/webTools.ts";
 import { buildToolModeResponse } from "./chatgptWebTools.ts";
+import {
+  checkGeminiWebUnsupportedControls,
+  GEMINI_WEB_UNSUPPORTED_CONTROL_CODE,
+} from "./gemini-web/capabilities.ts";
 
 // ─── Constants ──────────────────────────────────────────────────────────────
 
@@ -406,6 +410,33 @@ export class GeminiWebExecutor extends BaseExecutor {
     const { model, body, stream, credentials, signal, log, onCredentialsRefreshed } = input;
     const requestBody = body as GeminiRequestBody;
 
+    // #9356: fail fast on controls this provider cannot honor (reasoning_effort
+    // above "minimal", forced tool_choice). Runs before the credential check and
+    // before Playwright launches — the request is unservable no matter which
+    // cookie is used, and answering 200 with ordinary prose made agents believe
+    // their reasoning/tool requirements had been met. See ./gemini-web/capabilities.ts.
+    const violation = checkGeminiWebUnsupportedControls(body as Record<string, unknown>);
+    if (violation) {
+      log?.warn?.(
+        "GEMINI-WEB",
+        `Rejected request: "${violation.param}" is not supported by this provider`
+      );
+      return {
+        response: new Response(
+          JSON.stringify(
+            buildErrorBody(400, violation.message, null, {
+              type: "invalid_request_error",
+              code: GEMINI_WEB_UNSUPPORTED_CONTROL_CODE,
+            })
+          ),
+          { status: 400, headers: { "Content-Type": "application/json" } }
+        ),
+        url: GEMINI_URL,
+        headers: {},
+        transformedBody: body,
+      };
+    }
+
     const cookie = resolveGeminiWebCookie(credentials);
     if (!cookie) {
       return {

open-sse/executors/gemini-web/capabilities.ts

@@ -0,0 +1,121 @@
+/**
+ * Request-contract guards for the Gemini Web executor (#9356).
+ *
+ * gemini-web is not an API client. It launches Playwright, types ONE flat
+ * prompt string into the gemini.google.com `.ql-editor` contenteditable,
+ * presses Enter, and captures the first `StreamGenerate` response off the page
+ * (see ../gemini-web.ts). There is no JSON request body on the wire, which
+ * makes two OpenAI controls structurally impossible to honor:
+ *
+ *   • `reasoning_effort` — no field exists to carry a thinking budget. Unlike
+ *     deepseek-web or perplexity-web, which post a real payload and can flip a
+ *     `thinking_enabled` flag or swap the model preference, there is nothing
+ *     here to set.
+ *   • forced `tool_choice` — the tools support gemini-web does have is the
+ *     prompt-emulation shim (`translator/webTools.ts`, #7286): it ASKS the
+ *     model, in prose, to answer with `<tool>{...}</tool>` and parses whatever
+ *     comes back. That is best-effort by construction. "required" / "any" /
+ *     a named function is a GUARANTEE, and a prompt cannot make one.
+ *
+ * Before this module both were accepted and quietly ignored, so an agent got a
+ * 200 with `finish_reason: "stop"`, no `reasoning_content`, and `tool_calls: []`
+ * and concluded its requirements had been met (#9356). Failing the request is
+ * the honest answer: the caller can drop the control, or route to a model that
+ * actually implements it.
+ *
+ * Deliberately NOT rejected — these are already satisfied or already work:
+ *   • `reasoning_effort: "none" | "minimal"` — asking for as little reasoning as
+ *     possible is something a non-thinking provider trivially complies with.
+ *   • `tool_choice: "auto" | "none"` and plain `tools[]` — the #7286 emulation
+ *     path, which several shipped combos depend on (#5240, #8488). Untouched.
+ *
+ * Pure and dependency-free so the whole contract is unit-testable without a
+ * browser.
+ */
+
+/** `error.code` on every compatibility rejection raised here. */
+export const GEMINI_WEB_UNSUPPORTED_CONTROL_CODE = "unsupported_control_for_provider";
+
+/** Effort levels a non-thinking provider already complies with. */
+const SATISFIED_EFFORT_LEVELS = new Set(["none", "minimal"]);
+
+/** `tool_choice` strings that demand a tool call rather than merely offering one. */
+const FORCING_TOOL_CHOICE_STRINGS = new Set(["required", "any"]);
+
+/** `tool_choice: { type }` values that pin the model to a specific/any tool. */
+const FORCING_TOOL_CHOICE_TYPES = new Set(["function", "tool", "any"]);
+
+export interface GeminiWebCapabilityViolation {
+  /** Which request field could not be honored. */
+  param: "reasoning_effort" | "tool_choice";
+  /** Client-facing explanation — already safe to put in a response body. */
+  message: string;
+}
+
+function normalizeString(value: unknown): string | null {
+  return typeof value === "string" && value.trim().length > 0 ? value.trim().toLowerCase() : null;
+}
+
+/**
+ * True when `tool_choice` demands a tool call. Covers the OpenAI strings
+ * ("required"), the Anthropic-flavored ones the translators also emit ("any"),
+ * and the object forms that name a function or force any tool. "auto" / "none"
+ * and every unrecognized shape are treated as non-forcing — this guard only
+ * blocks contracts it is certain gemini-web cannot keep.
+ */
+export function isForcingToolChoice(toolChoice: unknown): boolean {
+  const asString = normalizeString(toolChoice);
+  if (asString) return FORCING_TOOL_CHOICE_STRINGS.has(asString);
+
+  if (toolChoice && typeof toolChoice === "object" && !Array.isArray(toolChoice)) {
+    const type = normalizeString((toolChoice as Record<string, unknown>).type);
+    return type !== null && FORCING_TOOL_CHOICE_TYPES.has(type);
+  }
+
+  return false;
+}
+
+/** True when `reasoning_effort` asks for MORE thinking than "none at all". */
+export function requestsThinkingBudget(reasoningEffort: unknown): boolean {
+  const effort = normalizeString(reasoningEffort);
+  if (effort === null) return false;
+  return !SATISFIED_EFFORT_LEVELS.has(effort);
+}
+
+/**
+ * Inspect an OpenAI-shaped request body for controls gemini-web cannot honor.
+ * Returns the first violation found, or `null` when the request is servable.
+ *
+ * `reasoning_effort` is checked before `tool_choice` only for determinism; a
+ * request carrying both is rejected either way.
+ */
+export function checkGeminiWebUnsupportedControls(
+  body: Record<string, unknown> | null | undefined
+): GeminiWebCapabilityViolation | null {
+  if (!body || typeof body !== "object") return null;
+
+  if (requestsThinkingBudget(body.reasoning_effort)) {
+    return {
+      param: "reasoning_effort",
+      message:
+        'Model provider "gemini-web" does not support "reasoning_effort". It drives the ' +
+        "gemini.google.com web UI through a typed prompt and has no thinking-budget control " +
+        'to set, so any effort above "minimal" would be silently ignored. Remove ' +
+        '"reasoning_effort" (or send "none"/"minimal") or route to a reasoning-capable model.',
+    };
+  }
+
+  if (isForcingToolChoice(body.tool_choice)) {
+    return {
+      param: "tool_choice",
+      message:
+        'Model provider "gemini-web" cannot guarantee a forced tool call. Its tool support is ' +
+        "prompt-emulated — the model is asked to emit a tool block and may answer with prose " +
+        'instead — so "tool_choice" values that require one ("required", "any", or a named ' +
+        'function) cannot be honored. Use "auto" to keep best-effort tool calling, or route to ' +
+        "a model with native function calling.",
+    };
+  }
+
+  return null;
+}

open-sse/executors/kiro.ts

@@ -61,7 +61,7 @@ type KiroStreamState = {
   contextUsagePercentage?: number;
   hasContextUsage?: boolean;
   hasMeteringEvent?: boolean;
-  usage?: UsageSummary;
+  usage?: Partial<UsageSummary>;
   hasReasoningContent?: boolean;
   reasoningChunkCount?: number;
   // Inline-thinking splitter state (populated only when thinkingExpected=true).
@@ -185,8 +185,7 @@ function resolveKiroMaxInputTokens(model: string): number {
  * inflate `total_tokens`.
  */
 function ensureKiroUsage(state: KiroStreamState, model: string) {
-  if (state.usage) return;
-
+  if (state.usage?.total_tokens !== undefined) return;
   const estimatedOutputTokens =
     state.totalContentLength && state.totalContentLength > 0
       ? Math.max(1, Math.floor(state.totalContentLength / 4))
@@ -198,11 +197,11 @@ function ensureKiroUsage(state: KiroStreamState, model: string) {
       : 0;
 
   if (estimatedTotalTokens <= 0 && estimatedOutputTokens <= 0) return;
-
   // Without a percentage there is no total to split, so the output estimate is
   // all that is known and stands on its own.
   if (estimatedTotalTokens <= 0) {
     state.usage = {
+      ...state.usage,
       prompt_tokens: 0,
       completion_tokens: estimatedOutputTokens,
       total_tokens: estimatedOutputTokens,
@@ -213,6 +212,7 @@ function ensureKiroUsage(state: KiroStreamState, model: string) {
   const promptTokens = Math.max(0, estimatedTotalTokens - estimatedOutputTokens);
 
   state.usage = {
+    ...state.usage,
     prompt_tokens: promptTokens,
     completion_tokens: estimatedOutputTokens,
     total_tokens: promptTokens + estimatedOutputTokens,

open-sse/executors/lmarena/response.ts

@@ -7,6 +7,8 @@ import { isCloudflareChallenge } from "../../services/lmarenaTlsClient.ts";
 import { markLMArenaCatalogModelDead } from "./models.ts";
 import { parseArenaSSE } from "./stream.ts";
 
+const encoder = new TextEncoder();
+
 export function errorResponse(
   status: number,
   message: string,
@@ -165,7 +167,7 @@ function baseChunk(model: string) {
 }
 
 function enqueueSse(controller: ReadableStreamDefaultController, chunk: Record<string, unknown>) {
-  controller.enqueue(new TextEncoder().encode(`data: ${JSON.stringify(chunk)}\n\n`));
+  controller.enqueue(encoder.encode(`data: ${JSON.stringify(chunk)}\n\n`));
 }
 
 function emitStopAndDone(controller: ReadableStreamDefaultController, model: string) {
@@ -173,7 +175,8 @@ function emitStopAndDone(controller: ReadableStreamDefaultController, model: str
     ...baseChunk(model),
     choices: [{ index: 0, delta: {}, finish_reason: "stop" }],
   });
-  controller.enqueue(new TextEncoder().encode("data: [DONE]\n\n"));
+
+  controller.enqueue(encoder.encode("data: [DONE]\n\n"));
   controller.close();
 }
 

open-sse/executors/raycast.ts

@@ -28,7 +28,14 @@ export class RaycastExecutor extends BaseExecutor {
     return RAYCAST_CHAT_URL;
   }
 
-  buildHeaders(credentials: ProviderCredentials, payload?: string): Record<string, string> {
+  // Not a BaseExecutor.buildHeaders override: Raycast signs headers over the exact
+  // request payload (2nd param is the body string, not the base's `stream` boolean),
+  // and execute() below is fully custom — keep it as a distinct helper so a
+  // polymorphic buildHeaders(credentials, true) call can never land here.
+  private buildRaycastRequestHeaders(
+    credentials: ProviderCredentials,
+    payload?: string
+  ): Record<string, string> {
     const body = payload || "{}";
     return buildRaycastHeaders(body, credentials as JsonRecord);
   }
@@ -44,7 +51,11 @@ export class RaycastExecutor extends BaseExecutor {
       return {
         response: new Response(
           JSON.stringify({
-            error: { message: sanitizeErrorMessage(message), type: "invalid_request_error", code: "" },
+            error: {
+              message: sanitizeErrorMessage(message),
+              type: "invalid_request_error",
+              code: "",
+            },
           }),
           { status: 400, headers: { "Content-Type": "application/json" } }
         ),
@@ -54,7 +65,7 @@ export class RaycastExecutor extends BaseExecutor {
       };
     }
 
-    const headers = this.buildHeaders(credentials as ProviderCredentials, payload);
+    const headers = this.buildRaycastRequestHeaders(credentials as ProviderCredentials, payload);
     mergeUpstreamExtraHeaders(headers, upstreamExtraHeaders as Record<string, string> | null);
 
     let raycastResponse: Response;

open-sse/handlers/imageGeneration/providers/adobeFirefly.ts

@@ -16,11 +16,11 @@ import {
   AdobeFireflyError,
   adobeFireflyGenerateImage,
   adobeFireflyImageTimeoutMs,
-  adobeFireflyMaxImageRefs,
   resolveAdobeAccessToken,
-  resolveAdobeSourceImageIds,
+  resolveAdobeSourceImageReferences,
   resolveAdobeImageModel,
 } from "../../../services/adobeFireflyClient.ts";
+import { getAdobeReferenceUploadLimit } from "../../../services/adobeFireflyModels.ts";
 import { isAdobeFireflyUpscaleModel } from "../../../services/adobeFireflyUpscale.ts";
 import { handleAdobeFireflyImageUpscale } from "../../imageUpscale/adobeFirefly.ts";
 
@@ -90,7 +90,8 @@ export async function handleAdobeFireflyImageGeneration({
 
     // Keep the raw credential blob for Cookie + sherlockToken (x-arp-session-id).
     // JWT may be embedded in the same paste as cookies (HAR / multi-line).
-    const psd = (credentials as { providerSpecificData?: { cookie?: string } })?.providerSpecificData;
+    const psd = (credentials as { providerSpecificData?: { cookie?: string } })
+      ?.providerSpecificData;
     const sessionCookie =
       (typeof psd?.cookie === "string" && psd.cookie.trim()) ||
       (typeof credentials?.apiKey === "string" && credentials.apiKey.trim()) ||
@@ -98,15 +99,11 @@ export async function handleAdobeFireflyImageGeneration({
         ? credentials.accessToken
         : undefined);
 
-    // Cap uploads by model family. gpt-image: 2 subject refs max (3–4+ stalls colligo → 504).
-    // nano: 4 general refs for multi-panel composition.
-    const { id: resolvedId } = resolveAdobeImageModel(model);
-    const maxRefs = adobeFireflyMaxImageRefs(resolvedId);
-
-    const sourceImageIds = await resolveAdobeSourceImageIds({
+    const { spec } = resolveAdobeImageModel(model);
+    const references = await resolveAdobeSourceImageReferences({
       accessToken,
       body,
-      max: maxRefs,
+      max: getAdobeReferenceUploadLimit(spec, "image"),
       sessionCookie,
       prompt,
       fetchImpl,
@@ -121,13 +118,13 @@ export async function handleAdobeFireflyImageGeneration({
           : undefined;
     const timeoutMs = adobeFireflyImageTimeoutMs({
       timeoutMs: explicitTimeout,
-      refCount: sourceImageIds.length,
+      refCount: references.length,
     });
 
     log?.info?.(
       "IMAGE",
       `${provider}/${model} (adobe-firefly) | prompt: "${prompt.slice(0, 60)}${prompt.length > 60 ? "..." : ""}"` +
-        (sourceImageIds.length ? ` | refs: ${sourceImageIds.length}/${maxRefs}` : "") +
+        (references.length ? ` | refs: ${references.length}` : "") +
         ` | pollTimeoutMs=${timeoutMs}`
     );
 
@@ -139,9 +136,8 @@ export async function handleAdobeFireflyImageGeneration({
       aspectRatio: body.aspect_ratio ?? body.aspectRatio ?? body.size,
       quality: body.quality,
       seed: Number.isFinite(seed as number) ? (seed as number) : undefined,
-      negativePrompt:
-        typeof body.negative_prompt === "string" ? body.negative_prompt : undefined,
-      sourceImageIds: sourceImageIds.length ? sourceImageIds : undefined,
+      negativePrompt: typeof body.negative_prompt === "string" ? body.negative_prompt : undefined,
+      references: references.length ? references : undefined,
       sessionCookie,
       timeoutMs,
       fetchImpl,

open-sse/handlers/videoGeneration/adobeFireflyHandler.ts

@@ -10,9 +10,10 @@ import {
   AdobeFireflyError,
   adobeFireflyGenerateVideo,
   resolveAdobeAccessToken,
-  resolveAdobeSourceImageIds,
+  resolveAdobeSourceImageReferences,
   resolveAdobeVideoModel,
 } from "../../services/adobeFireflyClient.ts";
+import { getAdobeReferenceUploadLimit } from "../../services/adobeFireflyModels.ts";
 
 function normalizePositiveNumber(value: unknown, fallback: number): number {
   const n = Number(value);
@@ -55,7 +56,8 @@ export async function handleAdobeFireflyVideoGeneration({
           ? Number(body.seed)
           : undefined;
     // Keep raw paste for Cookie + sherlockToken (x-arp-session-id).
-    const psd = (credentials as { providerSpecificData?: { cookie?: string } })?.providerSpecificData;
+    const psd = (credentials as { providerSpecificData?: { cookie?: string } })
+      ?.providerSpecificData;
     const sessionCookie =
       (typeof psd?.cookie === "string" && psd.cookie.trim()) ||
       (typeof credentials?.apiKey === "string" && credentials.apiKey.trim()) ||
@@ -63,13 +65,11 @@ export async function handleAdobeFireflyVideoGeneration({
         ? credentials.accessToken
         : undefined);
 
-    // Kling i2v / Veo ref / Sora frame: upload reference images first.
-    const { id: videoModelId } = resolveAdobeVideoModel(String(model));
-    const maxFrames = videoModelId.includes("kling") || videoModelId.includes("sora") ? 2 : 3;
-    const sourceImageIds = await resolveAdobeSourceImageIds({
+    const { spec } = resolveAdobeVideoModel(String(model));
+    const references = await resolveAdobeSourceImageReferences({
       accessToken,
       body,
-      max: maxFrames,
+      max: getAdobeReferenceUploadLimit(spec, "image"),
       sessionCookie,
       prompt,
       fetchImpl,
@@ -79,7 +79,7 @@ export async function handleAdobeFireflyVideoGeneration({
     log?.info?.(
       "VIDEO",
       `${provider}/${model} (adobe-firefly) | prompt: "${prompt.slice(0, 60)}${prompt.length > 60 ? "..." : ""}"` +
-        (sourceImageIds.length ? ` | frames: ${sourceImageIds.length}` : "")
+        (references.length ? ` | refs: ${references.length}` : "")
     );
 
     const result = await adobeFireflyGenerateVideo({
@@ -99,7 +99,7 @@ export async function handleAdobeFireflyVideoGeneration({
             ? body.negativePrompt
             : undefined,
       generateAudio: body.generate_audio !== false && body.generateAudio !== false,
-      sourceImageIds: sourceImageIds.length ? sourceImageIds : undefined,
+      references: references.length ? references : undefined,
       sessionCookie,
       timeoutMs,
       fetchImpl,

open-sse/services/__tests__/tierResolver.test.ts

@@ -199,10 +199,10 @@ describe("TierResolver", () => {
       ]);
       // Observable effect of the cache: the duplicate resolves to the same tier and only
       // ONE entry is memoized (getTierStats counts cache entries, not classify calls).
-      assert.equal(results.length, 2);
-      assert.equal(results[0].tier, results[1].tier);
+      expect(results).toHaveLength(2);
+      expect(results[0].tier).toBe(results[1].tier);
       const stats = getTierStats();
-      assert.equal(stats.free + stats.cheap + stats.premium, 1);
+      expect(stats.free + stats.cheap + stats.premium).toBe(1);
     });
   });
 

open-sse/services/adobeFireflyModels.ts

@@ -1,328 +1,590 @@
 /**
- * Adobe Firefly model catalog: live discovery + static fallback from browser capture.
+ * Adobe Firefly model discovery and normalized media capabilities.
  *
- * Live: POST firefly-3p.ff.adobe.io/v2/models/discovery (needs valid IMS token).
- * Fallback: curated rows from adobe/get_models.txt (2026-07 Firefly SPA capture) so
- * Media/Models still list usable ids when discovery fails or credentials are missing.
+ * The live discovery schema is authoritative. The generated snapshot is used only
+ * when a request cannot perform authenticated discovery (for example /v1/models).
  */
 
-import {
-  type AdobeFireflyDiscoveredModel,
-  discoverAdobeFireflyModels,
-  resolveAdobeAccessToken,
-} from "./adobeFireflyClient.ts";
+import { ADOBE_FIREFLY_DISCOVERY_SNAPSHOT } from "./adobeFireflyModelSnapshot.ts";
+
+export type AdobeFireflyModality = "image" | "video" | "audio" | "unknown";
+
+export interface AdobeFireflyDiscoveredModel {
+  modelId: string;
+  modelVersion: string;
+  displayName: string;
+  modality: AdobeFireflyModality;
+  enabled: boolean;
+  providerName?: string;
+  releaseReadiness?: string;
+  healthStatus?: string;
+  inputMediaUseCases: string[];
+  requestSchema?: Record<string, unknown>;
+  backingModel?: string;
+}
+
+export interface AdobeFireflyReferenceInputCapability {
+  mediaType: string;
+  usageType: string;
+  minItems: number;
+  maxItems: number | null;
+  maxFileSizeBytes: number | null;
+}
+
+export interface AdobeFireflyMediaCapabilities {
+  inputMediaUseCases: string[];
+  schemaProperties: string[];
+  requiredProperties: string[];
+  referenceInputs: AdobeFireflyReferenceInputCapability[];
+  maxReferenceItems: number | null;
+  supportedSizes: string[];
+  supportedAspectRatios: string[];
+  supportedResolutions: string[];
+  supportedDurations: number[];
+  durationMin: number | null;
+  durationMax: number | null;
+  durationDefault: number | null;
+  outputCountMin: number | null;
+  outputCountMax: number | null;
+  promptMaxLength: number | null;
+  releaseReadiness: string;
+  healthStatus: string;
+}
 
 export interface AdobeFireflyCatalogModel {
-  /** OpenAI-style id without provider prefix, e.g. nano-banana-pro or flux-fluxPro */
+  /** Stable API id without the provider prefix. */
   id: string;
   name: string;
   modality: "image" | "video";
-  /** Upstream wire modelId for generate-async */
   upstreamModelId: string;
-  /** Upstream wire modelVersion for generate-async */
   upstreamModelVersion: string;
-  inputModalities?: string[];
+  providerName: string;
+  backingModel: string;
+  inputModalities: string[];
+  capabilities: AdobeFireflyMediaCapabilities;
 }
 
-/**
- * Static fallback built from adobe/get_models.txt discovery response.
- * Friendly aliases first (Media page defaults), then popular upstream families.
- */
-export const ADOBE_FIREFLY_FALLBACK_MODELS: AdobeFireflyCatalogModel[] = [
-  // ── Friendly aliases (handler resolveAdobeImageModel / resolveAdobeVideoModel) ──
-  {
-    id: "nano-banana-pro",
-    name: "Gemini 3.0 (Nano Banana Pro)",
-    modality: "image",
-    upstreamModelId: "gemini-flash",
-    upstreamModelVersion: "nano-banana-2",
-    inputModalities: ["text", "image"],
-  },
-  {
-    id: "nano-banana",
-    name: "Gemini 2.5 (Nano Banana)",
-    modality: "image",
-    upstreamModelId: "gemini-flash",
-    upstreamModelVersion: "nano-banana",
-    inputModalities: ["text", "image"],
-  },
-  {
-    id: "nano-banana-2",
-    name: "Gemini 3.1 (Nano Banana 2)",
-    modality: "image",
-    upstreamModelId: "gemini-flash",
-    upstreamModelVersion: "nano-banana-3",
-    inputModalities: ["text", "image"],
-  },
-  {
-    id: "gpt-image-2",
-    name: "GPT Image 2",
-    modality: "image",
-    upstreamModelId: "gpt-image",
-    upstreamModelVersion: "2",
-    inputModalities: ["text", "image"],
-  },
-  {
-    id: "gpt-image",
-    name: "GPT Image 2",
-    modality: "image",
-    upstreamModelId: "gpt-image",
-    upstreamModelVersion: "2",
-    inputModalities: ["text", "image"],
-  },
-  {
-    id: "gpt-image-1.5",
-    name: "GPT Image 1.5",
-    modality: "image",
-    upstreamModelId: "gpt-image",
-    upstreamModelVersion: "1.5",
-    inputModalities: ["text", "image"],
-  },
-  {
-    id: "sora-2",
-    name: "Sora 2",
-    modality: "video",
-    upstreamModelId: "sora",
-    upstreamModelVersion: "sora-2",
-  },
-  {
-    id: "sora-2-pro",
-    name: "Sora 2 Pro",
-    modality: "video",
-    upstreamModelId: "sora",
-    upstreamModelVersion: "sora-2-pro",
-  },
-  {
-    id: "veo-3.1",
-    name: "Veo 3.1",
-    modality: "video",
-    upstreamModelId: "veo",
-    upstreamModelVersion: "3.1-generate",
-  },
-  {
-    id: "veo-3.1-fast",
-    name: "Veo 3.1 Fast",
-    modality: "video",
-    upstreamModelId: "veo",
-    upstreamModelVersion: "3.1-fast-generate",
-  },
-  {
-    id: "veo-3.1-ref",
-    name: "Veo 3.1 Reference",
-    modality: "video",
-    upstreamModelId: "veo",
-    upstreamModelVersion: "3.1-generate",
-  },
-  {
-    id: "kling-3",
-    name: "Kling Video v3 Standard Image to Video",
-    modality: "video",
-    upstreamModelId: "kling",
-    upstreamModelVersion: "kling_v3_standard_i2v",
-  },
-  // ── Additional image families from discovery capture ──
-  {
-    id: "flux-2",
-    name: "Flux 2",
-    modality: "image",
-    upstreamModelId: "flux",
-    upstreamModelVersion: "2",
-    inputModalities: ["text", "image"],
-  },
-  {
-    id: "flux-pro",
-    name: "Flux 1.1 Pro",
-    modality: "image",
-    upstreamModelId: "flux",
-    upstreamModelVersion: "fluxPro",
-    inputModalities: ["text", "image"],
-  },
-  {
-    id: "flux-ultra",
-    name: "Flux 1.1 Ultra",
-    modality: "image",
-    upstreamModelId: "flux",
-    upstreamModelVersion: "fluxUltra",
-    inputModalities: ["text", "image"],
-  },
-  {
-    id: "seedream-4",
-    name: "Seedream 4.0",
-    modality: "image",
-    upstreamModelId: "seedream",
-    upstreamModelVersion: "seedream_v4",
-    inputModalities: ["text", "image"],
-  },
-  {
-    id: "seedream-5-lite",
-    name: "Seedream 5.0 Lite",
-    modality: "image",
-    upstreamModelId: "seedream",
-    upstreamModelVersion: "seedream_v5_lite",
-    inputModalities: ["text", "image"],
-  },
-  {
-    id: "runway-gen4-image",
-    name: "Runway Gen-4 Image",
-    modality: "image",
-    upstreamModelId: "runway-gen4-image",
-    upstreamModelVersion: "gen4_image",
-    inputModalities: ["text", "image"],
-  },
-  // ── Additional video families ──
-  {
-    id: "kling-v3-t2v",
-    name: "Kling Video v3 Standard Text to Video",
-    modality: "video",
-    upstreamModelId: "kling",
-    upstreamModelVersion: "kling_v3_standard_t2v",
-  },
-  {
-    id: "kling-v3-pro-i2v",
-    name: "Kling Video v3 Pro Image to Video",
-    modality: "video",
-    upstreamModelId: "kling",
-    upstreamModelVersion: "kling_v3_pro_i2v",
-  },
-  {
-    id: "luma-ray3",
-    name: "Ray3",
-    modality: "video",
-    upstreamModelId: "luma",
-    upstreamModelVersion: "3.0-ray",
-  },
-  {
-    id: "runway-gen4-turbo",
-    name: "Runway Gen-4 Video",
-    modality: "video",
-    upstreamModelId: "runway",
-    upstreamModelVersion: "gen4_turbo",
-  },
-];
+export interface AdobeFireflyImageModelSpec extends AdobeFireflyCatalogModel {
+  modality: "image";
+  /** Payload dialect observed for this model family. */
+  family: "gemini" | "gpt-image" | "generic";
+}
 
-/** Stable slug for upstream modelId + modelVersion (catalog id when not a friendly alias). */
+export interface AdobeFireflyVideoModelSpec extends AdobeFireflyCatalogModel {
+  modality: "video";
+  defaultDuration: number;
+  defaultResolution: string;
+}
+
+interface MergedObjectSchema {
+  properties: Record<string, Record<string, unknown>>;
+  required: string[];
+}
+
+function asRecord(value: unknown): Record<string, unknown> {
+  return value && typeof value === "object" && !Array.isArray(value)
+    ? (value as Record<string, unknown>)
+    : {};
+}
+
+function asStringArray(value: unknown): string[] {
+  return Array.isArray(value)
+    ? value.map((item) => String(item)).filter((item) => item.length > 0)
+    : [];
+}
+
+function finiteInteger(value: unknown): number | null {
+  return Number.isInteger(value) ? (value as number) : null;
+}
+
+/** Merge object properties/required keys contributed through JSON Schema allOf. */
+export function mergeAdobeObjectSchema(schema: unknown): MergedObjectSchema {
+  const merged: MergedObjectSchema = { properties: {}, required: [] };
+  const visit = (value: unknown) => {
+    const node = asRecord(value);
+    const properties = asRecord(node.properties);
+    for (const [key, property] of Object.entries(properties)) {
+      merged.properties[key] = asRecord(property);
+    }
+    merged.required.push(...asStringArray(node.required));
+    if (Array.isArray(node.allOf)) node.allOf.forEach(visit);
+  };
+  visit(schema);
+  merged.required = [...new Set(merged.required)];
+  return merged;
+}
+
+function schemaBranches(schema: unknown): Record<string, unknown>[] {
+  const root = asRecord(schema);
+  if (Object.keys(root).length === 0) return [];
+  return [
+    root,
+    ...(Array.isArray(root.anyOf) ? root.anyOf.map(asRecord) : []),
+    ...(Array.isArray(root.oneOf) ? root.oneOf.map(asRecord) : []),
+  ];
+}
+
+function enumStrings(schema: unknown): string[] {
+  return [
+    ...new Set(
+      schemaBranches(schema)
+        .flatMap((branch) => (Array.isArray(branch.enum) ? branch.enum : []))
+        .filter((value): value is string => typeof value === "string")
+    ),
+  ];
+}
+
+function integerBranch(schema: unknown): Record<string, unknown> {
+  return schemaBranches(schema).find((branch) => branch.type === "integer") || {};
+}
+
+/** Stable, collision-resistant public id for an exact upstream model/version pair. */
 export function slugifyAdobeModel(modelId: string, modelVersion: string): string {
-  const mid = String(modelId || "")
-    .trim()
-    .toLowerCase()
-    .replace(/[^a-z0-9]+/g, "-")
-    .replace(/^-|-$/g, "");
-  const ver = String(modelVersion || "")
-    .trim()
-    .toLowerCase()
-    .replace(/[^a-z0-9.]+/g, "-")
-    .replace(/^-|-$/g, "");
-  if (!ver || ver === "default" || ver === mid) return mid || "model";
-  return `${mid}-${ver}`;
+  const slug = (value: string, allowDot = false) =>
+    String(value || "")
+      .trim()
+      .toLowerCase()
+      .replace(allowDot ? /[^a-z0-9.]+/g : /[^a-z0-9]+/g, "-")
+      .replace(/^-|-$/g, "");
+  const family = slug(modelId);
+  // Adobe still uses `kling_v3_omni*` internally, while discovery exposes these
+  // products to users as Kling O3. Never leak the obsolete/internal "omni" name
+  // into the public API catalog; the untouched upstream version stays in the spec.
+  const publicVersion =
+    family === "kling" ? modelVersion.replace(/^kling_v3_omni/i, "kling_o3") : modelVersion;
+  const version = slug(publicVersion, true);
+  if (!version || version === "default" || version === family) return family || "model";
+  return `${family}-${version}`;
 }
 
-/** Map discovery rows → catalog entries (image/video only). */
-export function mapDiscoveredToCatalog(
-  rows: AdobeFireflyDiscoveredModel[]
-): AdobeFireflyCatalogModel[] {
-  const out: AdobeFireflyCatalogModel[] = [];
-  const seen = new Set<string>();
+/** Parse POST /v2/models/discovery without discarding its resolved request schema. */
+export function parseAdobeModelsDiscovery(body: unknown): AdobeFireflyDiscoveredModel[] {
+  const root = asRecord(body);
+  const families = Array.isArray(root.models) ? root.models : [];
+  const rows: AdobeFireflyDiscoveredModel[] = [];
 
-  // Prefer friendly aliases when upstream matches known fallback rows.
-  for (const fb of ADOBE_FIREFLY_FALLBACK_MODELS) {
-    const hit = rows.find(
-      (r) =>
-        r.modelId === fb.upstreamModelId &&
-        r.modelVersion === fb.upstreamModelVersion &&
-        (r.modality === fb.modality || r.modality === "unknown")
-    );
-    if (hit && !seen.has(fb.id)) {
-      seen.add(fb.id);
-      out.push({
-        ...fb,
-        name: hit.displayName || fb.name,
+  for (const familyValue of families) {
+    const family = asRecord(familyValue);
+    const modelId = String(family.modelId || "").trim();
+    if (!modelId) continue;
+    for (const [modelVersion, versionValue] of Object.entries(asRecord(family.modelVersions))) {
+      const version = asRecord(versionValue);
+      if (version.enabled === false) continue;
+      const outputModalities = asStringArray(version.outputModality).map((item) =>
+        item.toLowerCase()
+      );
+      const modality: AdobeFireflyModality = outputModalities.includes("image")
+        ? "image"
+        : outputModalities.includes("video")
+          ? "video"
+          : outputModalities.includes("audio")
+            ? "audio"
+            : "unknown";
+      rows.push({
+        modelId,
+        modelVersion,
+        displayName: String(
+          version.modelDisplayName || version.modelCaiDisplayName || modelVersion
+        ),
+        modality,
+        enabled: version.enabled !== false,
+        providerName:
+          typeof family.acModelFamilyProviderDisplayName === "string"
+            ? family.acModelFamilyProviderDisplayName
+            : undefined,
+        releaseReadiness:
+          typeof version.releaseReadiness === "string" ? version.releaseReadiness : undefined,
+        healthStatus: typeof version.healthStatus === "string" ? version.healthStatus : undefined,
+        inputMediaUseCases: asStringArray(version.inputMediaUseCase),
+        requestSchema: asRecord(version.requestSchema),
+        backingModel:
+          typeof version.bksGenerationModel === "string" ? version.bksGenerationModel : undefined,
+      });
+    }
+  }
+  return rows;
+}
+
+function normalizeCapabilities(row: AdobeFireflyDiscoveredModel): AdobeFireflyMediaCapabilities {
+  const schema = mergeAdobeObjectSchema(row.requestSchema);
+  const referenceSchema = asRecord(schema.properties.referenceBlobs);
+  const referenceInputs: AdobeFireflyReferenceInputCapability[] = [];
+  const mediaCapabilities = Array.isArray(referenceSchema["x-capabilities"])
+    ? referenceSchema["x-capabilities"]
+    : [];
+  for (const mediaValue of mediaCapabilities) {
+    const media = asRecord(mediaValue);
+    const maxFileSizeBytes = finiteInteger(media.maxFileSizeBytes);
+    const usageConstraints = Array.isArray(media.usageConstraints) ? media.usageConstraints : [];
+    for (const usageValue of usageConstraints) {
+      const usage = asRecord(usageValue);
+      if (usage.deprecated === true) continue;
+      const usageType = String(usage.usageType || "");
+      const mediaType = String(media.mediaType || "");
+      if (!usageType || !mediaType) continue;
+      referenceInputs.push({
+        mediaType,
+        usageType,
+        minItems: finiteInteger(usage.minItems) ?? 0,
+        maxItems: finiteInteger(usage.maxItems),
+        maxFileSizeBytes,
       });
     }
   }
 
-  for (const r of rows) {
-    if (r.modality !== "image" && r.modality !== "video") continue;
-    const id = slugifyAdobeModel(r.modelId, r.modelVersion);
-    if (seen.has(id)) continue;
-    // Skip if already covered by a friendly alias with same upstream
-    if (
-      out.some(
-        (o) =>
-          o.upstreamModelId === r.modelId && o.upstreamModelVersion === r.modelVersion
+  const supportedSizes = [
+    ...new Set(
+      schemaBranches(schema.properties.size)
+        .flatMap((branch) => (Array.isArray(branch.enum) ? branch.enum : []))
+        .map(asRecord)
+        .filter((size) => finiteInteger(size.width) !== null && finiteInteger(size.height) !== null)
+        .map((size) => `${size.width}x${size.height}`)
+    ),
+  ];
+  const supportedAspectRatios = [
+    ...new Set(
+      schemaBranches(schema.properties.generationSettings).flatMap((branch) =>
+        enumStrings(asRecord(asRecord(branch.properties).aspectRatio))
       )
-    ) {
-      continue;
-    }
-    seen.add(id);
-    out.push({
-      id,
-      name: r.displayName || id,
-      modality: r.modality,
-      upstreamModelId: r.modelId,
-      upstreamModelVersion: r.modelVersion,
-      inputModalities: r.modality === "image" ? ["text", "image"] : ["text"],
-    });
-  }
-
-  return out;
-}
-
-export function getAdobeFireflyFallbackCatalog(modality?: "image" | "video"): AdobeFireflyCatalogModel[] {
-  if (!modality) return [...ADOBE_FIREFLY_FALLBACK_MODELS];
-  return ADOBE_FIREFLY_FALLBACK_MODELS.filter((m) => m.modality === modality);
-}
-
-/**
- * Live discovery when credentials resolve; otherwise static fallback from get_models capture.
- */
-export async function resolveAdobeFireflyCatalog(opts: {
-  credentials?: {
-    apiKey?: string;
-    accessToken?: string;
-    providerSpecificData?: Record<string, unknown> | null;
-  } | null;
-  modality?: "image" | "video";
-  fetchImpl?: typeof fetch;
-}): Promise<{ models: AdobeFireflyCatalogModel[]; source: "api" | "fallback" }> {
-  const fetchImpl = opts.fetchImpl || fetch;
-  try {
-    if (opts.credentials) {
-      const token = await resolveAdobeAccessToken(opts.credentials, fetchImpl);
-      const discovered = await discoverAdobeFireflyModels(token, fetchImpl);
-      let catalog = mapDiscoveredToCatalog(discovered);
-      if (opts.modality) catalog = catalog.filter((m) => m.modality === opts.modality);
-      if (catalog.length > 0) return { models: catalog, source: "api" };
-    }
-  } catch {
-    // fall through to static catalog
-  }
+    ),
+  ];
+  const duration = integerBranch(schema.properties.duration);
+  const outputCount = integerBranch(schema.properties.n);
+  const prompt =
+    schemaBranches(schema.properties.prompt).find((branch) => branch.type === "string") || {};
 
   return {
-    models: getAdobeFireflyFallbackCatalog(opts.modality),
-    source: "fallback",
+    inputMediaUseCases: [...row.inputMediaUseCases],
+    schemaProperties: Object.keys(schema.properties),
+    requiredProperties: [...schema.required],
+    referenceInputs,
+    maxReferenceItems: finiteInteger(referenceSchema.maxItems),
+    supportedSizes,
+    supportedAspectRatios,
+    supportedResolutions: enumStrings(schema.properties.resolution),
+    supportedDurations: [
+      ...new Set(
+        schemaBranches(schema.properties.duration)
+          .flatMap((branch) => (Array.isArray(branch.enum) ? branch.enum : []))
+          .filter((value): value is number => Number.isInteger(value))
+      ),
+    ],
+    durationMin: finiteInteger(duration.minimum),
+    durationMax: finiteInteger(duration.maximum),
+    durationDefault: finiteInteger(duration.default),
+    outputCountMin: finiteInteger(outputCount.minimum),
+    outputCountMax: finiteInteger(outputCount.maximum),
+    promptMaxLength: finiteInteger(prompt.maxLength),
+    releaseReadiness: row.releaseReadiness || "",
+    healthStatus: row.healthStatus || "",
   };
 }
 
-/** Registry-shaped models for imageRegistry / videoRegistry. */
-export function toRegistryImageModels(
-  models: AdobeFireflyCatalogModel[] = getAdobeFireflyFallbackCatalog("image")
-): Array<{ id: string; name: string; inputModalities?: string[] }> {
-  return models
-    .filter((m) => m.modality === "image")
-    .map((m) => ({
-      id: m.id,
-      name: m.name.startsWith("Firefly ") ? m.name : `Firefly ${m.name}`,
-      inputModalities: m.inputModalities || ["text", "image"],
-    }));
+function isCallableGenerationModel(row: AdobeFireflyDiscoveredModel): boolean {
+  if (row.modality !== "image" && row.modality !== "video") return false;
+  if (!mergeAdobeObjectSchema(row.requestSchema).properties.prompt) return false;
+  const excluded = new Set(["upscaling", "sharpening", "denoising"]);
+  return !row.inputMediaUseCases.some((value) => excluded.has(value.toLowerCase()));
 }
 
-export function toRegistryVideoModels(
-  models: AdobeFireflyCatalogModel[] = getAdobeFireflyFallbackCatalog("video")
-): Array<{ id: string; name: string }> {
-  return models
-    .filter((m) => m.modality === "video")
-    .map((m) => ({
-      id: m.id,
-      name: m.name.startsWith("Firefly ") ? m.name : `Firefly ${m.name}`,
-    }));
+function deriveInputModalities(capabilities: AdobeFireflyMediaCapabilities): string[] {
+  return ["text", ...new Set(capabilities.referenceInputs.map((reference) => reference.mediaType))];
+}
+
+function semanticCatalogKey(model: AdobeFireflyCatalogModel): string {
+  return JSON.stringify({
+    backingModel: model.backingModel,
+    name: model.name,
+    modality: model.modality,
+    capabilities: model.capabilities,
+  });
+}
+
+/** Normalize and de-duplicate callable image/video rows from live discovery. */
+export function mapDiscoveredToCatalog(
+  rows: AdobeFireflyDiscoveredModel[]
+): AdobeFireflyCatalogModel[] {
+  const output: AdobeFireflyCatalogModel[] = [];
+  const seen = new Set<string>();
+  for (const row of rows) {
+    if (!isCallableGenerationModel(row)) continue;
+    const capabilities = normalizeCapabilities(row);
+    const model: AdobeFireflyCatalogModel = {
+      id: slugifyAdobeModel(row.modelId, row.modelVersion),
+      name: row.displayName,
+      modality: row.modality as "image" | "video",
+      upstreamModelId: row.modelId,
+      upstreamModelVersion: row.modelVersion,
+      providerName: row.providerName || "",
+      backingModel: row.backingModel || "",
+      inputModalities: deriveInputModalities(capabilities),
+      capabilities,
+    };
+    const key = semanticCatalogKey(model);
+    if (seen.has(key)) continue;
+    seen.add(key);
+    output.push(model);
+  }
+  return output;
+}
+
+function snapshotCatalog(): AdobeFireflyCatalogModel[] {
+  return ADOBE_FIREFLY_DISCOVERY_SNAPSHOT.map((model) => {
+    const capabilities: AdobeFireflyMediaCapabilities = {
+      inputMediaUseCases: [...model.inputMediaUseCases],
+      schemaProperties: [...model.schemaProperties],
+      requiredProperties: [...model.requiredProperties],
+      referenceInputs: model.referenceInputs.map((reference) => ({ ...reference })),
+      maxReferenceItems: model.maxReferenceItems,
+      supportedSizes: [...model.supportedSizes],
+      supportedAspectRatios: [...model.supportedAspectRatios],
+      supportedResolutions: [...model.supportedResolutions],
+      supportedDurations: [...model.supportedDurations],
+      durationMin: model.durationMin,
+      durationMax: model.durationMax,
+      durationDefault: model.durationDefault,
+      outputCountMin: model.outputCountMin,
+      outputCountMax: model.outputCountMax,
+      promptMaxLength: model.promptMaxLength,
+      releaseReadiness: model.releaseReadiness,
+      healthStatus: model.healthStatus,
+    };
+    return {
+      id: model.id,
+      name: model.name,
+      modality: model.modality,
+      upstreamModelId: model.upstreamModelId,
+      upstreamModelVersion: model.upstreamModelVersion,
+      providerName: model.providerName,
+      backingModel: model.backingModel,
+      inputModalities: deriveInputModalities(capabilities),
+      capabilities,
+    };
+  });
+}
+
+export const ADOBE_FIREFLY_FALLBACK_MODELS: AdobeFireflyCatalogModel[] = snapshotCatalog();
+
+export function getAdobeFireflyFallbackCatalog(
+  modality?: "image" | "video"
+): AdobeFireflyCatalogModel[] {
+  return ADOBE_FIREFLY_FALLBACK_MODELS.filter((model) => !modality || model.modality === modality);
+}
+
+function imageFamily(model: AdobeFireflyCatalogModel): AdobeFireflyImageModelSpec["family"] {
+  if (model.upstreamModelId === "gemini-flash") return "gemini";
+  if (model.upstreamModelId === "gpt-image" || model.upstreamModelId === "gpt-4o-image") {
+    return "gpt-image";
+  }
+  return "generic";
+}
+
+export const ADOBE_FIREFLY_IMAGE_MODELS: Record<string, AdobeFireflyImageModelSpec> =
+  Object.fromEntries(
+    getAdobeFireflyFallbackCatalog("image").map((model) => [
+      model.id,
+      { ...model, modality: "image" as const, family: imageFamily(model) },
+    ])
+  );
+
+function defaultDuration(model: AdobeFireflyCatalogModel): number {
+  const caps = model.capabilities;
+  return caps.durationDefault ?? caps.supportedDurations[0] ?? caps.durationMin ?? 5;
+}
+
+function defaultResolution(model: AdobeFireflyCatalogModel): string {
+  if (model.capabilities.supportedSizes.some((value) => value.includes("1920x1080"))) {
+    return "1080p";
+  }
+  return "720p";
+}
+
+export const ADOBE_FIREFLY_VIDEO_MODELS: Record<string, AdobeFireflyVideoModelSpec> =
+  Object.fromEntries(
+    getAdobeFireflyFallbackCatalog("video").map((model) => [
+      model.id,
+      {
+        ...model,
+        modality: "video" as const,
+        defaultDuration: defaultDuration(model),
+        defaultResolution: defaultResolution(model),
+      },
+    ])
+  );
+
+const LEGACY_MODEL_ALIASES: Record<string, string> = {
+  "nano-banana": "gemini-flash-nano-banana",
+  "nano-banana-pro": "gemini-flash-nano-banana-2",
+  "nano-banana-2": "gemini-flash-nano-banana-3",
+  "gpt-image": "gpt-image-2",
+  "gpt-image-2": "gpt-image-2",
+  "gpt-image-1.5": "gpt-image-1.5",
+  "flux-2": "flux-2",
+  "flux-pro": "flux-fluxpro",
+  "flux-ultra": "flux-fluxultra",
+  "seedream-4": "seedream-seedream-v4",
+  "seedream-5-lite": "seedream-seedream-v5-lite",
+  "runway-gen4-image": "runway-gen4-image",
+  "veo-3.1": "veo-3.1-generate",
+  "veo-3.1-fast": "veo-3.1-fast-generate",
+  "luma-ray3": "luma-3.0-ray",
+  "runway-gen4-turbo": "runway-gen4-turbo",
+  // Backward compatibility only; the catalog advertises the exact discovered id.
+  "kling-3": "kling-kling-v3-standard-i2v",
+};
+
+// Preserve established API aliases when (and only when) they resolve to a model
+// that is present in the verified discovery snapshot. These keys are not listed.
+for (const [alias, target] of Object.entries(LEGACY_MODEL_ALIASES)) {
+  const imageTarget = ADOBE_FIREFLY_IMAGE_MODELS[target];
+  if (imageTarget) ADOBE_FIREFLY_IMAGE_MODELS[alias] = imageTarget;
+  const videoTarget = ADOBE_FIREFLY_VIDEO_MODELS[target];
+  if (videoTarget) ADOBE_FIREFLY_VIDEO_MODELS[alias] = videoTarget;
+}
+
+/** Backward-compatible request ids. Kept out of every advertised model catalog. */
+export const ADOBE_FIREFLY_IMAGE_ROUTING_ALIASES = Object.freeze(
+  Object.entries(LEGACY_MODEL_ALIASES)
+    .filter(([, target]) => Boolean(ADOBE_FIREFLY_IMAGE_MODELS[target]))
+    .map(([alias]) => alias)
+);
+
+function normalizeRequestedId(model: string): string {
+  return String(model || "")
+    .trim()
+    .toLowerCase()
+    .replace(/^adobe-firefly\//, "")
+    .replace(/^firefly\//, "");
+}
+
+function resolveCatalogId(model: string): string {
+  const requested = normalizeRequestedId(model);
+  return LEGACY_MODEL_ALIASES[requested] || requested;
+}
+
+export function resolveAdobeImageModel(model: string): {
+  id: string;
+  spec: AdobeFireflyImageModelSpec;
+} {
+  const id = resolveCatalogId(model);
+  const spec = ADOBE_FIREFLY_IMAGE_MODELS[id];
+  if (!spec) {
+    throw new Error(
+      `Unknown Adobe Firefly image model: ${normalizeRequestedId(model) || "(empty)"}`
+    );
+  }
+  return { id, spec };
+}
+
+export function resolveAdobeVideoModel(model: string): {
+  id: string;
+  spec: AdobeFireflyVideoModelSpec;
+} {
+  const id = resolveCatalogId(model);
+  const spec = ADOBE_FIREFLY_VIDEO_MODELS[id];
+  if (!spec) {
+    throw new Error(
+      `Unknown Adobe Firefly video model: ${normalizeRequestedId(model) || "(empty)"}`
+    );
+  }
+  return { id, spec };
+}
+
+export function toRegistryImageModels(): Array<{
+  id: string;
+  name: string;
+  inputModalities: string[];
+  imageRequired?: boolean;
+  supportedSizes: string[];
+  mediaCapabilities: Record<string, unknown>;
+}> {
+  const generated = getAdobeFireflyFallbackCatalog("image").map((model) => ({
+    id: model.id,
+    name: `Firefly ${model.name}`,
+    inputModalities: model.inputModalities,
+    supportedSizes: model.capabilities.supportedSizes,
+    mediaCapabilities: toAdobeMediaCapabilitiesApi(model),
+  }));
+  // Upscaling uses a distinct Firefly endpoint and is not returned by the image
+  // generation discovery schema. Keep its two supported Topaz models visible in
+  // the same provider catalog so image clients can select them deliberately.
+  return [
+    ...generated,
+    {
+      id: "topaz-standard",
+      name: "Firefly Topaz Upscale (Standard)",
+      inputModalities: ["image"],
+      imageRequired: true,
+      supportedSizes: [],
+      mediaCapabilities: { input_media_use_cases: ["upscaling"] },
+    },
+    {
+      id: "topaz-bloom",
+      name: "Firefly Topaz Bloom (Creative Upscale)",
+      inputModalities: ["image"],
+      imageRequired: true,
+      supportedSizes: [],
+      mediaCapabilities: { input_media_use_cases: ["upscaling"] },
+    },
+  ];
+}
+
+export function toRegistryVideoModels(): Array<{
+  id: string;
+  name: string;
+  supportedSizes: string[];
+  mediaCapabilities: Record<string, unknown>;
+}> {
+  return getAdobeFireflyFallbackCatalog("video").map((model) => ({
+    id: model.id,
+    name: `Firefly ${model.name}`,
+    supportedSizes: model.capabilities.supportedSizes,
+    mediaCapabilities: toAdobeMediaCapabilitiesApi(model),
+  }));
+}
+
+/** JSON-safe extension emitted by /v1/models. */
+export function toAdobeMediaCapabilitiesApi(
+  model: AdobeFireflyCatalogModel
+): Record<string, unknown> {
+  const caps = model.capabilities;
+  return {
+    upstream_model_id: model.upstreamModelId,
+    upstream_model_version: model.upstreamModelVersion,
+    provider_name: model.providerName,
+    release_readiness: caps.releaseReadiness,
+    health_status: caps.healthStatus,
+    input_media_use_cases: caps.inputMediaUseCases,
+    reference_inputs: caps.referenceInputs.map((reference) => ({
+      media_type: reference.mediaType,
+      usage_type: reference.usageType,
+      min_items: reference.minItems,
+      max_items: reference.maxItems,
+      max_file_size_bytes: reference.maxFileSizeBytes,
+    })),
+    max_reference_items: caps.maxReferenceItems,
+    supported_sizes: caps.supportedSizes,
+    supported_aspect_ratios: caps.supportedAspectRatios,
+    supported_resolutions: caps.supportedResolutions,
+    supported_durations: caps.supportedDurations,
+    duration_min: caps.durationMin,
+    duration_max: caps.durationMax,
+    duration_default: caps.durationDefault,
+    output_count_min: caps.outputCountMin,
+    output_count_max: caps.outputCountMax,
+    prompt_max_length: caps.promptMaxLength,
+  };
+}
+
+export function getAdobeReferenceUploadLimit(
+  model: AdobeFireflyCatalogModel,
+  mediaType: string
+): number {
+  if (model.capabilities.maxReferenceItems !== null) {
+    return Math.max(1, Math.min(32, model.capabilities.maxReferenceItems));
+  }
+  const declaredTotal = model.capabilities.referenceInputs
+    .filter((reference) => reference.mediaType === mediaType)
+    .reduce((total, reference) => total + (reference.maxItems ?? 0), 0);
+  return Math.max(1, Math.min(32, declaredTotal || 1));
 }

open-sse/services/backgroundTaskDetector.ts

@@ -190,15 +190,31 @@ export function getBackgroundTaskReason(
   const messages = toMessageArray(typedBody.messages ?? typedBody.input ?? []);
   if (!Array.isArray(messages) || messages.length === 0) return null;
 
-  // Find system message
+  // Derive system content from messages array (OpenAI format) or top-level
+  // system field (Anthropic format).
   const systemMsg = messages.find(
     (message: BackgroundMessage) => message.role === "system" || message.role === "developer"
   );
-  if (!systemMsg) return null;
-
-  const systemContent =
-    typeof systemMsg.content === "string" ? systemMsg.content.toLowerCase() : "";
-
+  let systemContent = "";
+  if (systemMsg && typeof systemMsg.content === "string") {
+    systemContent = systemMsg.content.toLowerCase();
+  } else if (!systemMsg) {
+    // Anthropic top-level system field: string or array of text blocks
+    const raw = (typedBody as Record<string, unknown>).system;
+    if (typeof raw === "string") {
+      systemContent = raw.toLowerCase();
+    } else if (Array.isArray(raw)) {
+      systemContent = raw
+        .map((part) =>
+          part && typeof (part as { text?: unknown }).text === "string"
+            ? (part as { text: string }).text
+            : ""
+        )
+        .filter(Boolean)
+        .join(" ")
+        .toLowerCase();
+    }
+  }
   if (!systemContent) return null;
 
   // Check against detection patterns

open-sse/services/combo/comboStructure.ts

@@ -18,6 +18,7 @@ import { getHiddenModelsByProvider } from "../../../src/lib/db/models";
 import { getComboModelString, normalizeComboStep } from "../../../src/lib/combos/steps.ts";
 import { getProviderByAlias, getProviderById } from "../../../src/shared/constants/providers.ts";
 import { estimateTokens } from "../contextManager.ts";
+import { containsMediaKind } from "../../utils/mediaParts.ts";
 import { getResolvedModelCapabilities } from "../modelCapabilities.ts";
 import { parseModel, stripContextWindowSuffix } from "../model.ts";
 import { dedupeTargetsByExecutionKey, isRecord } from "./comboData.ts";
@@ -483,21 +484,15 @@ function estimateRequestInputTokens(body: Record<string, unknown>): number {
   return Object.keys(estimatePayload).length > 0 ? estimateTokens(estimatePayload) : 0;
 }
 
-function valueContainsImagePart(value: unknown, depth = 0): boolean {
-  if (depth > 8 || value === null || value === undefined) return false;
-  if (typeof value === "string") return value.startsWith("data:image/");
-  if (Array.isArray(value)) return value.some((entry) => valueContainsImagePart(entry, depth + 1));
-  if (!isRecord(value)) return false;
-
-  const type = typeof value.type === "string" ? value.type.toLowerCase() : null;
-  if (type === "image" || type === "image_url" || type === "input_image") return true;
-  if ("image_url" in value || "input_image" in value) return true;
-
-  const source = isRecord(value.source) ? value.source : null;
-  const mediaType = typeof source?.media_type === "string" ? source.media_type.toLowerCase() : "";
-  if (mediaType.startsWith("image/")) return true;
-
-  return Object.values(value).some((entry) => valueContainsImagePart(entry, depth + 1));
+function valueContainsImagePart(value: unknown): boolean {
+  // Delegates to the unified media detector (open-sse/utils/mediaParts.ts) —
+  // single source of truth shared with the vision-bridge guardrail. The
+  // detector keeps this filter's legacy permissive matches (image-ish `type`
+  // in any casing, bare `image_url`/`input_image` keys, source.media_type
+  // image/*, bare data:image strings, recursion capped at depth 8) via
+  // "image_indicator" parts. containsMediaKind short-circuits on the first
+  // hit — this runs on every request, so no full-part collection here.
+  return containsMediaKind([{ content: [value] }], "image");
 }
 
 export function deriveRequestCompatibilityRequirements(

open-sse/services/combo/validateQuality.ts

@@ -190,10 +190,26 @@ function isRecord(value: unknown): value is Record<string, unknown> {
   return !!value && typeof value === "object" && !Array.isArray(value);
 }
 
+/**
+ * Whether an `error` field carries a real failure signal. A key-presence check
+ * (`!= null`) false-positives on benign values some backends emit on every
+ * chunk (`{}`, `""`, `false`, `0`) — e.g. tool-call turns where a chunk with
+ * real tool_calls content also carries `"error": {}`. Only substantive values
+ * are treated as upstream failures.
+ */
+function isSubstantiveError(value: unknown): boolean {
+  if (value === null || value === undefined) return false;
+  if (typeof value === "string") return value.trim().length > 0;
+  if (typeof value === "object" && !Array.isArray(value)) {
+    return Object.keys(value as Record<string, unknown>).length > 0;
+  }
+  return value === true;
+}
+
 function isStreamingUpstreamError(parsed: unknown, eventType: string): boolean {
   if (eventType === "response.failed" || eventType === "error") return true;
   if (!isRecord(parsed)) return false;
-  if (parsed.error != null) return true;
+  if (isSubstantiveError(parsed.error)) return true;
 
   const nestedResponse = isRecord(parsed.response) ? parsed.response : null;
   return nestedResponse?.status === "failed" && nestedResponse.error != null;

open-sse/services/compression/rules/ru/ultra.json

@@ -32,7 +32,7 @@
       "replacement": " ",
       "context": "all",
       "category": "ultra",
-      "minIntensity": "notes"
+      "minIntensity": "ultra"
     },
     {
       "name": "ultra_lowercase",
@@ -40,7 +40,7 @@
       "replacement": " $1",
       "context": "all",
       "category": "ultra",
-      "minIntensity": "notes"
+      "minIntensity": "ultra"
     }
   ]
 }

open-sse/transformer/responsesTransformer.ts

@@ -37,6 +37,102 @@ async function getPath() {
   return _path || null;
 }
 
+type UsageRecord = Record<string, unknown>;
+
+function usageRecord(value: unknown): UsageRecord {
+  return typeof value === "object" && value !== null && !Array.isArray(value)
+    ? (value as UsageRecord)
+    : {};
+}
+
+function usageNumber(value: unknown): number | undefined {
+  return typeof value === "number" && Number.isFinite(value) ? value : undefined;
+}
+
+function usageDetails(record: UsageRecord, ...keys: string[]): UsageRecord {
+  for (const key of keys) {
+    const value = usageRecord(record[key]);
+    if (Object.keys(value).length > 0) return value;
+  }
+  return {};
+}
+
+/** Normalize Chat Completions and Responses usage into the Responses API shape. */
+function normalizeResponsesUsage(previous: unknown, raw: unknown): UsageRecord | null {
+  const source = usageRecord(raw);
+  if (Object.keys(source).length === 0) return usageRecord(previous);
+
+  const before = usageRecord(previous);
+  const beforeInputDetails = usageDetails(before, "input_tokens_details", "prompt_tokens_details");
+  const beforeOutputDetails = usageDetails(
+    before,
+    "output_tokens_details",
+    "completion_tokens_details"
+  );
+  const inputDetails = usageDetails(
+    source,
+    "input_tokens_details",
+    "prompt_tokens_details",
+    "inputTokenDetails",
+    "input_token_details"
+  );
+  const outputDetails = usageDetails(
+    source,
+    "output_tokens_details",
+    "completion_tokens_details",
+    "outputTokenDetails",
+    "output_token_details",
+    "reasoningTokenDetails",
+    "reasoning_token_details"
+  );
+
+  const inputTokens =
+    usageNumber(source.input_tokens) ??
+    usageNumber(source.prompt_tokens) ??
+    usageNumber(source.inputTokens) ??
+    usageNumber(source.promptTokens) ??
+    usageNumber(before.input_tokens) ??
+    usageNumber(before.prompt_tokens) ??
+    0;
+  const cachedTokens =
+    usageNumber(source.cache_read_input_tokens) ??
+    usageNumber(source.cached_input_tokens) ??
+    usageNumber(source.cachedInputTokens) ??
+    usageNumber(source.cached_tokens) ??
+    usageNumber(inputDetails.cached_tokens) ??
+    usageNumber(inputDetails.cachedTokens) ??
+    usageNumber(inputDetails.cacheReadTokens) ??
+    usageNumber(beforeInputDetails.cached_tokens) ??
+    0;
+  const outputTokens =
+    usageNumber(source.output_tokens) ??
+    usageNumber(source.completion_tokens) ??
+    usageNumber(source.outputTokens) ??
+    usageNumber(source.completionTokens) ??
+    usageNumber(before.output_tokens) ??
+    usageNumber(before.completion_tokens) ??
+    0;
+  const reasoningTokens =
+    usageNumber(source.reasoning_tokens) ??
+    usageNumber(source.reasoningTokens) ??
+    usageNumber(outputDetails.reasoning_tokens) ??
+    usageNumber(outputDetails.reasoningTokens) ??
+    usageNumber(beforeOutputDetails.reasoning_tokens) ??
+    0;
+  const totalTokens =
+    usageNumber(source.total_tokens) ??
+    usageNumber(source.totalTokens) ??
+    inputTokens + outputTokens;
+
+  return {
+    input_tokens: inputTokens,
+    input_tokens_details: { cached_tokens: cachedTokens },
+    output_tokens: outputTokens,
+    output_tokens_details: { reasoning_tokens: reasoningTokens },
+    total_tokens: totalTokens,
+  };
+}
+
 // Create log directory for responses (Node.js only)
 export function createResponsesLogger(model, logsDir = null) {
   // Skip logging in worker environment (no fs)
@@ -477,10 +573,11 @@ export function createResponsesApiTransformStream(
             continue;
           }
 
+          if (parsed.usage) {
+            state.usage = normalizeResponsesUsage(state.usage, parsed.usage);
+          }
+
           if (!parsed.choices?.length) {
-            if (parsed.usage) {
-              state.usage = parsed.usage;
-            }
             // #6906: trailing usage-only chunk after finish_reason already deferred
             // completion — send it now with the usage just captured above.
             if (state.awaitingTrailingUsage && !state.completedSent) {

open-sse/translator/index.ts

@@ -166,6 +166,29 @@ function isReasoningOnlyReplayTarget(provider: unknown, model: unknown): boolean
   );
 }
 
+/**
+ * Upstreams that reject an ABSENT reasoning_content on replay turns, so the
+ * placeholder must survive the cache miss.
+ *
+ * #9573/#9610 removed the placeholder globally because the model echoed it as
+ * its own reasoning and stopped (empty turns). That holds for DeepSeek, where
+ * an absent field was verified to be accepted — but Xiaomi MiMo still 400s
+ * ("Param Incorrect: The reasoning_content in the thinking mode must be passed
+ * back to the API", 9router#1321/#1337), so omitting the field there trades one
+ * live bug for another. Keep the placeholder only for those providers; the echo
+ * that comes back is still stripped on the way in by
+ * isInternalReasoningPlaceholder(), so it never re-poisons cache or history.
+ */
+function requiresReasoningContentPresence(provider: unknown, model: unknown): boolean {
+  const normalizedProvider = String(provider ?? "")
+    .trim()
+    .toLowerCase();
+  const normalizedModel = String(model ?? "")
+    .trim()
+    .toLowerCase();
+  return normalizedProvider === "xiaomi-mimo" || /(^|\/)mimo/i.test(normalizedModel);
+}
+
 /** @param options.normalizeToolCallId - When true, use 9-char tool call ids (e.g. Mistral); when false, leave ids as-is */
 /** @param options.preserveDeveloperRole - undefined/true: keep developer for OpenAI format (default); false: map to system */
 /** @param options.preserveCacheControl - When true, preserve client-side cache_control markers (for Claude Code, etc.) */
@@ -372,7 +395,9 @@ export function translateRequest(
       preserveReasoningContent: isReasoner,
       // Per-provider/model preserveVideoUrl flag from compat overrides.
       // Falls back to true for moonshot/kimi when unset (legacy behavior).
-      preserveVideoUrl: getModelPreserveVideoUrl(normalizedProvider, options?.model ?? "") ?? (normalizedProvider === "moonshot" || normalizedProvider === "kimi"),
+      preserveVideoUrl:
+        getModelPreserveVideoUrl(normalizedProvider, normalizedModel) ??
+        (normalizedProvider === "moonshot" || normalizedProvider === "kimi"),
     });
   }
 
@@ -573,7 +598,11 @@ export function translateRequest(
       // the field instead; providers that genuinely enforce the contract
       // (kimi-coding, moonshot authentic-reasoning) have their own paths above.
       if ((hasToolCalls || shouldReplayReasoningOnly) && !msg.reasoning_content) {
-        delete msg.reasoning_content;
+        if (requiresReasoningContentPresence(normalizedProvider, normalizedModel)) {
+          msg.reasoning_content = NON_ANTHROPIC_THINKING_PLACEHOLDER;
+        } else {
+          delete msg.reasoning_content;
+        }
       }
     }
   } else if (

open-sse/translator/request/openai-responses.ts

@@ -724,7 +724,7 @@ export function openaiResponsesToOpenAIRequest(
     const reasoningRec = toRecord(root.reasoning);
     const effort = toString(reasoningRec.effort);
     if (effort && result.reasoning_effort === undefined) {
-      result.reasoning_effort = normalizeResponsesReasoningEffort(effort, model);
+      result.reasoning_effort = normalizeResponsesReasoningEffort(effort, model ?? root.model);
     }
     if (
       credentialRecord._copilotClient === true &&

open-sse/translator/response/gemini-to-claude.ts

@@ -112,7 +112,7 @@ export function geminiToClaudeResponse(chunk, state) {
         // When the toolNameMap provides a match (e.g., lowercase "bash" → "Bash"),
         // use it directly without passing through normalizeToolName(), which would
         // reverse TitleCase back to lowercase via REVERSE_MAP (#9568).
-        const restoredToolName = mappedName || normalizeToolName(rawToolName);
+        const restoredToolName = mappedName ?? normalizeToolName(rawToolName);
         const idx = state.contentBlockIndex++;
         const toolId = fc.id || `toolu_${Date.now()}_${idx}`;
 

open-sse/translator/response/openai-responses.ts

@@ -451,11 +451,22 @@ function closeMessage(state, emit, idx) {
   }
 }
 
+// Tool calls sit after reasoning (if any) AND after a text message (if one was
+// actually emitted this turn) — a model commonly emits a short preamble before
+// calling a tool (e.g. "Kör nu, på riktigt — apply_patch..."), and that message
+// claims the same reasoningIndex+1 slot the old per-call math (`reasoningIndex
+// + 1 + tcIdx`) assumed was free for tcIdx=0. Not accounting for the message
+// item collided the tool call's added/delta/done events onto the same
+// output_index as the just-closed message, which a client keying per-item
+// state by output_index can silently drop (live incident 2026-08-08).
+function toolCallOutputIndexBase(state) {
+  const msgIdx = state.reasoningId ? normalizeOutputIndex(state.reasoningIndex) + 1 : 0;
+  return state.msgItemAdded[msgIdx] ? msgIdx + 1 : msgIdx;
+}
+
 function emitToolCall(state, emit, tc) {
   const tcIdx = tc.index ?? 0;
-  const outputIndex = state.reasoningId
-    ? normalizeOutputIndex(state.reasoningIndex) + 1 + normalizeOutputIndex(tcIdx)
-    : normalizeOutputIndex(tcIdx);
+  const outputIndex = toolCallOutputIndexBase(state) + normalizeOutputIndex(tcIdx);
   const newCallId = tc.id;
   const funcName = tc.function?.name;
 
@@ -536,9 +547,7 @@ function emitToolCall(state, emit, tc) {
 function closeToolCall(state, emit, idx, recordAsCompleted = true) {
   const callId = state.funcCallIds[idx];
   if (callId && !state.funcItemDone[idx]) {
-    const normalizedIndex = state.reasoningId
-      ? normalizeOutputIndex(state.reasoningIndex) + 1 + normalizeOutputIndex(idx)
-      : normalizeOutputIndex(idx);
+    const normalizedIndex = toolCallOutputIndexBase(state) + normalizeOutputIndex(idx);
     const args = state.funcArgsBuf[idx] || "{}";
     const toolName = state.funcNames[idx] || "";
     const isCustomTool =
@@ -874,6 +883,7 @@ function openaiResponsesToOpenAIResponseStream(chunk, state) {
     if (state.currentToolCallId) state.toolCallIdsSeen.add(state.currentToolCallId);
 
     const toolName = normalizeToolName(item.name);
+    state.currentToolName = toolName; // track for schema lookup at done time
     if (!toolName) {
       // Some Responses providers briefly emit placeholder/empty tool names.
       // Defer emission until output_item.done in case the final name is populated there.
@@ -919,26 +929,9 @@ function openaiResponsesToOpenAIResponseStream(chunk, state) {
     state.currentToolCallArgsBuffer = (state.currentToolCallArgsBuffer || "") + argsDelta;
     if (state.currentToolCallDeferred) return null;
 
-    return {
-      id: state.chatId,
-      object: "chat.completion.chunk",
-      created: state.created,
-      model: state.model || "gpt-4",
-      choices: [
-        {
-          index: 0,
-          delta: {
-            tool_calls: [
-              {
-                index: state.toolCallIndex,
-                function: { arguments: argsDelta },
-              },
-            ],
-          },
-          finish_reason: null,
-        },
-      ],
-    };
+    // #9168: buffer arguments until output_item.done for schema-aware null normalization
+    // Previously emitted raw null values for optional enum fields (e.g. isolation: null).
+    return null;
   }
 
   // Function call done — emit args chunk from item.arguments when no deltas were received,
@@ -1011,6 +1004,35 @@ function openaiResponsesToOpenAIResponseStream(chunk, state) {
     if (item.arguments != null && !buffered) {
       const argsToEmit = stripEmptyOptionalToolArgs(item.arguments, toolName, toolSchema);
 
+      const argsStr = typeof argsToEmit === "string" ? argsToEmit : JSON.stringify(argsToEmit);
+      if (argsStr) {
+        return {
+          id: state.chatId,
+          object: "chat.completion.chunk",
+          created: state.created,
+          model: state.model || "gpt-4",
+          choices: [
+            {
+              index: 0,
+              delta: {
+                tool_calls: [
+                  {
+                    index: currentIndex,
+                    function: { arguments: argsStr },
+                  },
+                ],
+              },
+              finish_reason: null,
+            },
+          ],
+        };
+      }
+    } else if (buffered) {
+      // #9168: deltas were buffered — normalize against the original client schema
+      // and emit the cleaned arguments once, stripping optional null values that
+      // would otherwise reach the client raw.
+      const argsToEmit = stripEmptyOptionalToolArgs(buffered, toolName, toolSchema);
+
       const argsStr = typeof argsToEmit === "string" ? argsToEmit : JSON.stringify(argsToEmit);
       if (argsStr) {
         return {

open-sse/utils/cursorAgentProtobuf.ts

@@ -19,6 +19,11 @@
 import zlib from "node:zlib";
 import crypto from "node:crypto";
 import { decodeNativeTodoWriteCompletion } from "./cursorAgentProtobuf/nativeTodoWrite.ts";
+import {
+  cursorImageAttachmentPath,
+  encodeSelectedImageBody,
+  type EncodedImage,
+} from "./cursorAgentProtobuf/imageEncoding.ts";
 import {
   WT_VARINT,
   WT_LEN,
@@ -63,25 +68,8 @@ const UM_MESSAGE_ID = 2; // UserMessage.message_id
 const UM_SELECTED_CONTEXT = 3; // UserMessage.selected_context (empty placeholder required)
 const UM_MODE = 4; // UserMessage.mode (cursor-agent sends 1)
 
-// ─── Vision input (image) field numbers ────────────────────────────────────
-// Pinned from cursor-agent's agent.v1 protobuf descriptor (bundle version
-// 2026.06.02-8c11d9f, cross-checked against composer-api's older-endpoint
-// encoder for shape). Images attach to the current UserMessage through its
-// selected_context (field 3): UserMessage.selected_context is a SelectedContext
-// whose `selected_images` (field 1) is a repeated SelectedImage. Each
-// SelectedImage carries the raw bytes inline in its `data_or_blob_id` oneof
-// (the `data` case, field 8) — cursor-agent's CLI instead sends a local file
-// `path`, which a proxy cannot use, so we inline the bytes like composer-api.
 const SC_SELECTED_IMAGES = 1; // SelectedContext.selected_images [repeated SelectedImage]
 
-const SI_UUID = 2; // SelectedImage.uuid
-const SI_DIMENSION = 4; // SelectedImage.dimension (SelectedImage.Dimension)
-const SI_MIME_TYPE = 7; // SelectedImage.mime_type
-const SI_DATA = 8; // SelectedImage.data (oneof data_or_blob_id) — inline image bytes
-
-const DIM_WIDTH = 1; // SelectedImage.Dimension.width (int32)
-const DIM_HEIGHT = 2; // SelectedImage.Dimension.height (int32)
-
 const RM_MODEL_ID = 1; // RequestedModel.model_id
 const RM_PARAMETERS = 3; // RequestedModel.parameters [repeated]
 
@@ -413,58 +401,15 @@ export type AgentRunInput = {
   // which the executor's processFrame replies to with the stored bytes.
   systemPrompt?: string;
   blobStore?: Map<string, Buffer>;
-  // Vision input: images attached to the current user turn. Encoded inline as
-  // SelectedContext.selected_images[] (see encodeSelectedImageBody). Empty /
-  // undefined keeps the request byte-identical to the text-only path.
+  // Vision input: images attached to the current user turn. Encoded as
+  // SelectedContext.selected_images[] via blobIdWithData (see
+  // encodeSelectedImageBody). Empty / undefined keeps the request
+  // byte-identical to the text-only path.
   images?: EncodedImage[];
 };
 
-/**
- * A resolved image ready to embed in a cursor request. `data` is the raw
- * decoded image bytes (already SSRF-checked / size-capped by the executor's
- * resolveCursorImages helper). `mimeType` (e.g. "image/png") helps cursor
- * decode the inline bytes; `width`/`height` populate the optional Dimension
- * sub-message when cheaply known; `uuid` is a stable per-image id.
- */
-export type EncodedImage = {
-  data: Buffer;
-  mimeType?: string;
-  width?: number;
-  height?: number;
-  uuid: string;
-};
-
-/**
- * Encode the body of a SelectedImage message (no outer field tag — the caller
- * wraps it via encodeMessage(SC_SELECTED_IMAGES, [body])). Sets the inline
- * `data` oneof case plus uuid, optional dimension, and mime_type. Fields are
- * written in ascending field-number order (canonical protobuf layout).
- */
-export function encodeSelectedImageBody(img: EncodedImage): Buffer {
-  const parts: Buffer[] = [encodeString(SI_UUID, img.uuid)];
-  if (
-    typeof img.width === "number" &&
-    typeof img.height === "number" &&
-    Number.isFinite(img.width) &&
-    Number.isFinite(img.height) &&
-    img.width > 0 &&
-    img.height > 0
-  ) {
-    parts.push(
-      encodeMessage(SI_DIMENSION, [
-        encodeUInt32Field(DIM_WIDTH, Math.floor(img.width)),
-        encodeUInt32Field(DIM_HEIGHT, Math.floor(img.height)),
-      ])
-    );
-  }
-  if (img.mimeType) {
-    parts.push(encodeString(SI_MIME_TYPE, img.mimeType));
-  }
-  // data_or_blob_id oneof = data (inline bytes) — field 8, written last to
-  // keep ascending field order.
-  parts.push(encodeBytes(SI_DATA, img.data));
-  return Buffer.concat(parts);
-}
+export { cursorImageAttachmentPath, encodeSelectedImageBody };
+export type { EncodedImage };
 
 /**
  * Convert OpenAI tool definitions to cursor McpToolDefinition bodies. Used
@@ -493,12 +438,15 @@ export function encodeAgentRunRequest(input: AgentRunInput): Buffer {
   // UserMessage { text, message_id, selected_context, mode=1 }.
   // selected_context is normally an empty placeholder (required by the server
   // even when empty — see below), but when the turn carries vision input we
-  // populate its selected_images[] with the inline-encoded images. The
-  // empty-images path produces byte-identical output to the text-only request.
+  // populate its selected_images[] with blobIdWithData-encoded images (and
+  // store the bytes in blobStore for getBlob). The empty-images path produces
+  // byte-identical output to the text-only request.
   const selectedContextParts: Buffer[] = [];
   if (input.images && input.images.length > 0) {
     for (const img of input.images) {
-      selectedContextParts.push(encodeMessage(SC_SELECTED_IMAGES, [encodeSelectedImageBody(img)]));
+      selectedContextParts.push(
+        encodeMessage(SC_SELECTED_IMAGES, [encodeSelectedImageBody(img, input.blobStore)])
+      );
     }
   }
   // The empty selected_context placeholder and mode=1 match cursor-agent's

open-sse/utils/cursorAgentProtobuf/imageEncoding.ts

@@ -0,0 +1,80 @@
+import crypto from "node:crypto";
+import {
+  encodeBytes,
+  encodeMessage,
+  encodeString,
+  encodeUInt32Field,
+} from "./wire.ts";
+
+const SI_UUID = 2;
+const SI_PATH = 3;
+const SI_DIMENSION = 4;
+const SI_MIME_TYPE = 7;
+const SI_BLOB_ID_WITH_DATA = 9;
+
+const SIBD_BLOB_ID = 1;
+const SIBD_DATA = 2;
+
+const DIM_WIDTH = 1;
+const DIM_HEIGHT = 2;
+
+export type EncodedImage = {
+  data: Buffer;
+  mimeType?: string;
+  width?: number;
+  height?: number;
+  uuid: string;
+};
+
+export function cursorImageAttachmentPath(uuid: string, mimeType?: string): string {
+  const normalized = (mimeType || "").toLowerCase();
+  const ext =
+    normalized === "image/jpeg" || normalized === "image/jpg"
+      ? "jpg"
+      : normalized === "image/gif"
+        ? "gif"
+        : normalized === "image/webp"
+          ? "webp"
+          : "png";
+  return `attachment-${uuid}.${ext}`;
+}
+
+export function encodeSelectedImageBody(
+  img: EncodedImage,
+  blobStore?: Map<string, Buffer>
+): Buffer {
+  const blobId = crypto.createHash("sha256").update(img.data).digest();
+  if (blobStore) {
+    blobStore.set(blobId.toString("hex"), img.data);
+  }
+
+  const parts: Buffer[] = [
+    encodeString(SI_UUID, img.uuid),
+    encodeString(SI_PATH, cursorImageAttachmentPath(img.uuid, img.mimeType)),
+  ];
+  if (
+    typeof img.width === "number" &&
+    typeof img.height === "number" &&
+    Number.isFinite(img.width) &&
+    Number.isFinite(img.height) &&
+    img.width > 0 &&
+    img.height > 0
+  ) {
+    parts.push(
+      encodeMessage(SI_DIMENSION, [
+        encodeUInt32Field(DIM_WIDTH, Math.floor(img.width)),
+        encodeUInt32Field(DIM_HEIGHT, Math.floor(img.height)),
+      ])
+    );
+  }
+  if (img.mimeType) {
+    parts.push(encodeString(SI_MIME_TYPE, img.mimeType));
+  }
+  parts.push(
+    encodeMessage(SI_BLOB_ID_WITH_DATA, [
+      encodeBytes(SIBD_BLOB_ID, blobId),
+      encodeBytes(SIBD_DATA, img.data),
+    ])
+  );
+  return Buffer.concat(parts);
+}

open-sse/utils/cursorImages.ts

@@ -2,8 +2,8 @@
  * Image resolution + security for Cursor vision input.
  *
  * Turns OpenAI `image_url` parts (base64 `data:` URIs or remote `http(s)`
- * URLs) into decoded bytes ready to inline into a cursor SelectedImage
- * (see ../utils/cursorAgentProtobuf.ts::encodeSelectedImageBody).
+ * URLs) into decoded, JPEG-prepped bytes ready for SelectedImage
+ * `blobIdWithData` encoding (see cursorAgentProtobuf.ts).
  *
  * Security (OmniRoute hard rules):
  *  - SSRF: remote fetches go through the repo's canonical outbound guard
@@ -12,9 +12,9 @@
  *    cloud-metadata hostnames. Client-supplied image URLs are always held to
  *    the strict public-only policy (never gated by the private-URL toggle that
  *    admin-configured provider URLs use).
- *  - Size cap: each image must decode to <= 1 MiB (matches composer-api).
- *    Enforced both before base64 decode (cheap pre-check) and while streaming
- *    a remote body (so a hostile server can't stream gigabytes).
+ *  - Size caps: inbound decode/fetch is bounded (16 MiB) so large clipboard
+ *    PNGs can shrink via JPEG soft-cap prep; the final wire image must be
+ *    <= 1 MiB. Soft target is ~100 KiB JPEG for reliable Cursor hydration.
  *  - Content type: data URIs and URL responses must be `image/*`.
  *  - Errors throw `CursorImageError` with a clean, path-free message; the
  *    executor routes it through the sanitized 400 path (hard rule #12).
@@ -23,6 +23,7 @@
 import crypto from "node:crypto";
 import dns from "node:dns";
 import { isIP } from "node:net";
+import sharp from "sharp";
 import {
   parseAndValidatePublicUrl,
   isPrivateHost,
@@ -30,14 +31,47 @@ import {
 } from "@/shared/network/outboundUrlGuard";
 import type { EncodedImage } from "./cursorAgentProtobuf.ts";
 
-// 1 MiB per image — matches composer-api's MAX_CURSOR_IMAGE_BYTES. Large
-// enough for a typical screenshot, small enough to bound request size and
-// memory.
+/** Final per-image byte cap after prep (composer-api / wire bound). */
 export const MAX_CURSOR_IMAGE_BYTES = 1024 * 1024;
 
-// Upper bound on the number of images per request. Each image triggers (at
-// most) one remote fetch, so an unbounded count is a DoS vector; 12 is well
-// above any realistic vision prompt.
+/**
+ * Inbound decode/fetch bomb ceiling before JPEG prep. Large clipboard PNGs may
+ * exceed {@link MAX_CURSOR_IMAGE_BYTES} raw but shrink under the wire cap after
+ * re-encode.
+ */
+export const MAX_CURSOR_IMAGE_DECODE_BYTES = 16 * 1024 * 1024;
+
+/**
+ * Soft target for Cursor vision hydration. Prefer JPEG at or under this size.
+ */
+export const CURSOR_VISION_SOFT_MAX_BYTES = 100 * 1024;
+
+/** Soft target when the client requests `detail: original` or `high`. */
+export const CURSOR_VISION_SOFT_MAX_BYTES_HIGH = 256 * 1024;
+
+/** Longest edge after Cursor vision prep. */
+export const CURSOR_VISION_MAX_EDGE = 2000;
+
+/** Decode bomb: reject images whose sniffed longest edge exceeds this. */
+export const MAX_CURSOR_IMAGE_DECODE_EDGE = 8192;
+
+/** Decode bomb: reject images whose sniffed pixel count exceeds this. */
+export const MAX_CURSOR_IMAGE_PIXELS = 25_000_000;
+
+const CURSOR_VISION_JPEG_QUALITIES_DEFAULT = [85, 70, 55, 40] as const;
+const CURSOR_VISION_JPEG_QUALITIES_HIGH = [90, 80, 65, 50] as const;
+const CURSOR_VISION_SOFT_MIN_EDGE = 256;
+const CURSOR_VISION_SOFT_SHRINK = 0.85;
+
+const CURSOR_VISION_PASSTHROUGH_MIME = new Set([
+  "image/jpeg",
+  "image/jpg",
+  "image/png",
+  "image/gif",
+  "image/webp",
+]);
+
+/** Upper bound on images attached to one Cursor turn. */
 export const MAX_CURSOR_IMAGES = 12;
 
 // Wall-clock cap for a single remote image fetch. A malformed env value
@@ -64,6 +98,25 @@ export class CursorImageError extends Error {
   }
 }
 
+function estimatedBase64DecodedBytes(payload: string): number {
+  return Math.floor((payload.length * 3) / 4);
+}
+
+function isHighDetail(detail: string | undefined): boolean {
+  const normalized = (detail || "").toLowerCase();
+  return normalized === "high" || normalized === "original";
+}
+
+function softMaxBytesForDetail(detail: string | undefined): number {
+  return isHighDetail(detail) ? CURSOR_VISION_SOFT_MAX_BYTES_HIGH : CURSOR_VISION_SOFT_MAX_BYTES;
+}
+
+function jpegQualitiesForDetail(detail: string | undefined): readonly number[] {
+  return isHighDetail(detail)
+    ? CURSOR_VISION_JPEG_QUALITIES_HIGH
+    : CURSOR_VISION_JPEG_QUALITIES_DEFAULT;
+}
+
 function decodeDataUrl(url: string): { data: Buffer; mimeType: string } {
   // data:[<mediatype>][;base64],<data>
   const comma = url.indexOf(",");
@@ -86,16 +139,21 @@ function decodeDataUrl(url: string): { data: Buffer; mimeType: string } {
 
   // Reject on the raw payload length BEFORE the regex/normalize pass, so an
   // arbitrarily large data URL can't burn CPU on the whitespace strip. Base64
-  // expands ~4:3, so 2x the byte cap is a safe upper bound on the encoded text.
-  if (payload.length > MAX_CURSOR_IMAGE_BYTES * 2) {
-    throw new CursorImageError("Image input is too large (max 1 MiB). Resize and retry.");
+  // expands ~4:3, so 2x the decode ceiling is a safe upper bound on the text.
+  if (payload.length > MAX_CURSOR_IMAGE_DECODE_BYTES * 2) {
+    throw new CursorImageError("Image input is too large to process safely.");
   }
 
   const normalized = payload.replace(/\s/g, "");
-  // Cheap pre-check: 4 base64 chars -> 3 bytes. Reject obviously oversized
-  // payloads before allocating the decode buffer.
-  if (Math.floor((normalized.length * 3) / 4) > MAX_CURSOR_IMAGE_BYTES) {
-    throw new CursorImageError("Image input is too large (max 1 MiB). Resize and retry.");
+  if (normalized.length === 0) {
+    throw new CursorImageError("Image data URL contains invalid base64 data.");
+  }
+  // Reject lenient Buffer.from acceptances (wrong alphabet, bad padding).
+  if (normalized.length % 4 !== 0 || !/^[A-Za-z0-9+/]+={0,2}$/.test(normalized)) {
+    throw new CursorImageError("Image data URL contains invalid base64 data.");
+  }
+  if (estimatedBase64DecodedBytes(normalized) > MAX_CURSOR_IMAGE_DECODE_BYTES) {
+    throw new CursorImageError("Image input is too large to process safely.");
   }
 
   let data: Buffer;
@@ -104,11 +162,16 @@ function decodeDataUrl(url: string): { data: Buffer; mimeType: string } {
   } catch {
     throw new CursorImageError("Image data URL contains invalid base64 data.");
   }
-  // Buffer.from(base64) silently drops invalid trailing chars; guard against a
-  // payload that decoded to nothing despite being non-empty.
-  if (normalized.length > 0 && data.length === 0) {
+  if (data.length === 0) {
     throw new CursorImageError("Image data URL contains invalid base64 data.");
   }
+  // Round-trip guard: Node can silently drop trailing garbage.
+  if (data.toString("base64").replace(/=+$/, "") !== normalized.replace(/=+$/, "")) {
+    throw new CursorImageError("Image data URL contains invalid base64 data.");
+  }
+  if (data.length > MAX_CURSOR_IMAGE_DECODE_BYTES) {
+    throw new CursorImageError("Image input is too large to process safely.");
+  }
   return { data, mimeType };
 }
 
@@ -216,10 +279,10 @@ async function fetchImageBytes(url: string): Promise<{ data: Buffer; mimeType: s
       // Reject early on an oversized Content-Length, then still cap during read
       // (the header is advisory / may be absent).
       const declaredLen = Number(response.headers.get("content-length") || "0");
-      if (Number.isFinite(declaredLen) && declaredLen > MAX_CURSOR_IMAGE_BYTES) {
-        throw new CursorImageError("Image input is too large (max 1 MiB). Resize and retry.");
+      if (Number.isFinite(declaredLen) && declaredLen > MAX_CURSOR_IMAGE_DECODE_BYTES) {
+        throw new CursorImageError("Image input is too large to process safely.");
       }
-      const data = await readCapped(response, MAX_CURSOR_IMAGE_BYTES);
+      const data = await readCapped(response, MAX_CURSOR_IMAGE_DECODE_BYTES);
       return { data, mimeType };
     } finally {
       clearTimeout(timer);
@@ -249,7 +312,7 @@ async function readCapped(response: Response, cap: number): Promise<Buffer> {
   const pushCapped = (chunk: Uint8Array) => {
     total += chunk.byteLength;
     if (total > cap) {
-      throw new CursorImageError("Image input is too large (max 1 MiB). Resize and retry.");
+      throw new CursorImageError("Image input is too large to process safely.");
     }
     chunks.push(Buffer.from(chunk));
   };
@@ -284,22 +347,311 @@ async function readCapped(response: Response, cap: number): Promise<Buffer> {
   // Last resort: buffer then cap-check (only exotic non-stream bodies).
   const buf = Buffer.from(await response.arrayBuffer());
   if (buf.length > cap) {
-    throw new CursorImageError("Image input is too large (max 1 MiB). Resize and retry.");
+    throw new CursorImageError("Image input is too large to process safely.");
   }
   return buf;
 }
 
+/** Magic-byte format sniff (independent of declared MIME). */
+export function sniffCursorImageFormat(
+  data: Uint8Array
+): "png" | "jpeg" | "gif" | "webp" | undefined {
+  if (
+    data.byteLength >= 8 &&
+    data[0] === 0x89 &&
+    data[1] === 0x50 &&
+    data[2] === 0x4e &&
+    data[3] === 0x47 &&
+    data[4] === 0x0d &&
+    data[5] === 0x0a &&
+    data[6] === 0x1a &&
+    data[7] === 0x0a
+  ) {
+    return "png";
+  }
+  if (
+    data.byteLength >= 6 &&
+    data[0] === 0x47 &&
+    data[1] === 0x49 &&
+    data[2] === 0x46 &&
+    data[3] === 0x38
+  ) {
+    return "gif";
+  }
+  if (data.byteLength >= 4 && data[0] === 0xff && data[1] === 0xd8) return "jpeg";
+  if (
+    data.byteLength >= 12 &&
+    data[0] === 0x52 &&
+    data[1] === 0x49 &&
+    data[2] === 0x46 &&
+    data[3] === 0x46 &&
+    data[8] === 0x57 &&
+    data[9] === 0x45 &&
+    data[10] === 0x42 &&
+    data[11] === 0x50
+  ) {
+    return "webp";
+  }
+  return undefined;
+}
+
+/**
+ * Sniff PNG/JPEG/GIF/WebP dimensions from raw bytes when the header is present.
+ * Best-effort only — unknown formats return undefined (dimension is optional).
+ */
+export function sniffCursorImageDimensions(
+  data: Uint8Array
+): { width: number; height: number } | undefined {
+  // PNG: signature + IHDR chunk (width/height at bytes 16..23)
+  if (
+    data.byteLength >= 24 &&
+    data[0] === 0x89 &&
+    data[1] === 0x50 &&
+    data[2] === 0x4e &&
+    data[3] === 0x47 &&
+    data[4] === 0x0d &&
+    data[5] === 0x0a &&
+    data[6] === 0x1a &&
+    data[7] === 0x0a
+  ) {
+    const width = ((data[16]! << 24) | (data[17]! << 16) | (data[18]! << 8) | data[19]!) >>> 0;
+    const height = ((data[20]! << 24) | (data[21]! << 16) | (data[22]! << 8) | data[23]!) >>> 0;
+    if (width > 0 && height > 0) return { width, height };
+  }
+  // GIF: "GIF8" + width/height as little-endian u16 at bytes 6..9
+  if (
+    data.byteLength >= 10 &&
+    data[0] === 0x47 &&
+    data[1] === 0x49 &&
+    data[2] === 0x46 &&
+    data[3] === 0x38
+  ) {
+    const width = data[6]! | (data[7]! << 8);
+    const height = data[8]! | (data[9]! << 8);
+    if (width > 0 && height > 0) return { width, height };
+  }
+  // WebP: RIFF....WEBP + VP8X / VP8 / VP8L
+  if (
+    data.byteLength >= 30 &&
+    data[0] === 0x52 &&
+    data[1] === 0x49 &&
+    data[2] === 0x46 &&
+    data[3] === 0x46 &&
+    data[8] === 0x57 &&
+    data[9] === 0x45 &&
+    data[10] === 0x42 &&
+    data[11] === 0x50
+  ) {
+    const fourcc = String.fromCharCode(data[12]!, data[13]!, data[14]!, data[15]!);
+    if (fourcc === "VP8X") {
+      const width = 1 + (data[24]! | (data[25]! << 8) | (data[26]! << 16));
+      const height = 1 + (data[27]! | (data[28]! << 8) | (data[29]! << 16));
+      if (width > 0 && height > 0) return { width, height };
+    } else if (fourcc === "VP8 ") {
+      if (data[23] === 0x9d && data[24] === 0x01 && data[25] === 0x2a) {
+        const width = (data[26]! | (data[27]! << 8)) & 0x3fff;
+        const height = (data[28]! | (data[29]! << 8)) & 0x3fff;
+        if (width > 0 && height > 0) return { width, height };
+      }
+    } else if (fourcc === "VP8L" && data[20] === 0x2f) {
+      const raw = data[21]! | (data[22]! << 8) | (data[23]! << 16) | (data[24]! << 24);
+      const width = (raw & 0x3fff) + 1;
+      const height = ((raw >> 14) & 0x3fff) + 1;
+      if (width > 0 && height > 0) return { width, height };
+    }
+  }
+  // JPEG: scan for SOF0/SOF2 marker with dimensions
+  if (data.byteLength >= 4 && data[0] === 0xff && data[1] === 0xd8) {
+    let offset = 2;
+    while (offset + 8 < data.byteLength) {
+      if (data[offset] !== 0xff) break;
+      const marker = data[offset + 1]!;
+      // Standalone markers (TEM, RSTn, SOI, EOI) carry no length payload.
+      if (marker === 0x01 || (marker >= 0xd0 && marker <= 0xd9)) {
+        offset += 2;
+        continue;
+      }
+      const length = (data[offset + 2]! << 8) | data[offset + 3]!;
+      if (marker === 0xc0 || marker === 0xc2) {
+        const height = (data[offset + 5]! << 8) | data[offset + 6]!;
+        const width = (data[offset + 7]! << 8) | data[offset + 8]!;
+        if (width > 0 && height > 0) return { width, height };
+        break;
+      }
+      if (length < 2) break;
+      offset += 2 + length;
+    }
+  }
+  return undefined;
+}
+
+type PreparedImage = {
+  data: Buffer;
+  mimeType: string;
+  width?: number;
+  height?: number;
+};
+
+/**
+ * Re-encode toward a JPEG under the soft vision cap when sharp can decode the
+ * payload. Fail-closed with CursorImageError on unsupported MIME, decode bombs,
+ * or undecodable bytes. After the quality ladder, edges shrink iteratively
+ * until the soft byte cap is met (or the min edge floor is hit).
+ */
+export async function prepareCursorImageForWire(input: {
+  data: Buffer;
+  mimeType: string;
+  detail?: string;
+}): Promise<PreparedImage> {
+  const mime = input.mimeType.toLowerCase();
+  const softMax = softMaxBytesForDetail(input.detail);
+  const qualities = jpegQualitiesForDetail(input.detail);
+  const lowestQuality = qualities[qualities.length - 1]!;
+
+  if (!CURSOR_VISION_PASSTHROUGH_MIME.has(mime)) {
+    throw new CursorImageError("Image input type is unsupported.");
+  }
+
+  const format = sniffCursorImageFormat(input.data);
+  const sniffed = sniffCursorImageDimensions(input.data);
+  if (sniffed) {
+    const edge = Math.max(sniffed.width, sniffed.height);
+    const pixels = sniffed.width * sniffed.height;
+    if (edge > MAX_CURSOR_IMAGE_DECODE_EDGE || pixels > MAX_CURSOR_IMAGE_PIXELS) {
+      throw new CursorImageError("Image input dimensions are too large.");
+    }
+  }
+
+  // Soft-cap skip: already soft-capped JPEG that has a real SOF (not SOI-only).
+  const declaredJpeg = mime === "image/jpeg" || mime === "image/jpg";
+  const alreadySmallJpeg =
+    declaredJpeg && format === "jpeg" && sniffed !== undefined && input.data.byteLength <= softMax;
+  if (alreadySmallJpeg) {
+    return {
+      data: input.data,
+      mimeType: "image/jpeg",
+      width: sniffed!.width,
+      height: sniffed!.height,
+    };
+  }
+
+  try {
+    // Force a full decode before accepting passthrough / encode.
+    await sharp(input.data, { failOn: "error" }).resize(1, 1).jpeg({ quality: 1 }).toBuffer();
+
+    // Passthrough only when declared MIME matches actual JPEG magic.
+    if (declaredJpeg && format === "jpeg" && input.data.byteLength <= softMax) {
+      const dims = sniffed ?? (await sharp(input.data).metadata());
+      const width = typeof dims.width === "number" ? dims.width : undefined;
+      const height = typeof dims.height === "number" ? dims.height : undefined;
+      return {
+        data: input.data,
+        mimeType: "image/jpeg",
+        ...(width && height && width > 0 && height > 0 ? { width, height } : {}),
+      };
+    }
+
+    const meta = await sharp(input.data).metadata();
+    const width = typeof meta.width === "number" ? meta.width : 0;
+    const height = typeof meta.height === "number" ? meta.height : 0;
+    if (width > 0 && height > 0) {
+      const edge = Math.max(width, height);
+      if (edge > MAX_CURSOR_IMAGE_DECODE_EDGE || width * height > MAX_CURSOR_IMAGE_PIXELS) {
+        throw new CursorImageError("Image input dimensions are too large.");
+      }
+    }
+
+    let targetW = width;
+    let targetH = height;
+    if (width > 0 && height > 0 && Math.max(width, height) > CURSOR_VISION_MAX_EDGE) {
+      const scale = CURSOR_VISION_MAX_EDGE / Math.max(width, height);
+      targetW = Math.max(1, Math.round(width * scale));
+      targetH = Math.max(1, Math.round(height * scale));
+    }
+
+    const encodeAt = async (w: number, h: number, quality: number): Promise<Buffer> => {
+      let pipeline = sharp(input.data, { failOn: "error" });
+      if (w > 0 && h > 0 && (w !== width || h !== height)) {
+        pipeline = pipeline.resize(w, h);
+      }
+      return pipeline.jpeg({ quality, mozjpeg: true }).toBuffer();
+    };
+
+    let best: Buffer | undefined;
+    for (const quality of qualities) {
+      const encoded = await encodeAt(targetW, targetH, quality);
+      if (!best || encoded.byteLength < best.byteLength) best = encoded;
+      if (encoded.byteLength <= softMax) {
+        const outDims = sniffCursorImageDimensions(encoded);
+        return {
+          data: encoded,
+          mimeType: "image/jpeg",
+          ...(outDims ?? (targetW > 0 && targetH > 0 ? { width: targetW, height: targetH } : {})),
+        };
+      }
+    }
+
+    while (
+      best &&
+      best.byteLength > softMax &&
+      targetW > 0 &&
+      targetH > 0 &&
+      Math.max(targetW, targetH) > CURSOR_VISION_SOFT_MIN_EDGE
+    ) {
+      const nextW = Math.max(1, Math.round(targetW * CURSOR_VISION_SOFT_SHRINK));
+      const nextH = Math.max(1, Math.round(targetH * CURSOR_VISION_SOFT_SHRINK));
+      if (Math.max(nextW, nextH) < CURSOR_VISION_SOFT_MIN_EDGE) {
+        const scale = CURSOR_VISION_SOFT_MIN_EDGE / Math.max(targetW, targetH);
+        targetW = Math.max(1, Math.round(targetW * scale));
+        targetH = Math.max(1, Math.round(targetH * scale));
+      } else {
+        targetW = nextW;
+        targetH = nextH;
+      }
+      const encoded = await encodeAt(targetW, targetH, lowestQuality);
+      if (!best || encoded.byteLength < best.byteLength) best = encoded;
+      if (encoded.byteLength <= softMax) {
+        const outDims = sniffCursorImageDimensions(encoded);
+        return {
+          data: encoded,
+          mimeType: "image/jpeg",
+          ...(outDims ?? { width: targetW, height: targetH }),
+        };
+      }
+      if (Math.max(targetW, targetH) <= CURSOR_VISION_SOFT_MIN_EDGE) break;
+    }
+
+    if (best) {
+      const outDims = sniffCursorImageDimensions(best);
+      return {
+        data: best,
+        mimeType: "image/jpeg",
+        ...(outDims ?? (targetW > 0 && targetH > 0 ? { width: targetW, height: targetH } : {})),
+      };
+    }
+
+    if (declaredJpeg && format !== "jpeg") {
+      throw new CursorImageError("Image input is not a valid JPEG.");
+    }
+    throw new CursorImageError("Image input could not be prepared for Cursor vision.");
+  } catch (err) {
+    if (err instanceof CursorImageError) throw err;
+    throw new CursorImageError("Image input is undecodable or unsupported.");
+  }
+}
+
 /**
  * Resolve OpenAI `image_url` URLs (data: or http(s):) into EncodedImage[]
- * ready to inline into a cursor request. Each image gets a stable random uuid.
- * Throws CursorImageError (clean message, sanitizable) on any invalid /
- * oversized / blocked input.
+ * ready for SelectedImage blobIdWithData encoding. Each image gets a stable
+ * random uuid. Throws CursorImageError (clean message, sanitizable) on any
+ * invalid / oversized / blocked / undecodable input.
  */
-export async function resolveCursorImages(imageUrls: string[]): Promise<EncodedImage[]> {
+export async function resolveCursorImages(
+  imageUrls: string[],
+  options?: { detail?: string }
+): Promise<EncodedImage[]> {
   if (imageUrls.length > MAX_CURSOR_IMAGES) {
-    throw new CursorImageError(
-      `Too many images in one request (max ${MAX_CURSOR_IMAGES}).`
-    );
+    throw new CursorImageError(`Too many images in one request (max ${MAX_CURSOR_IMAGES}).`);
   }
   const out: EncodedImage[] = [];
   for (const url of imageUrls) {
@@ -314,10 +666,27 @@ export async function resolveCursorImages(imageUrls: string[]): Promise<EncodedI
     if (!data.length) {
       throw new CursorImageError("Image input is empty.");
     }
-    if (data.length > MAX_CURSOR_IMAGE_BYTES) {
+    if (data.length > MAX_CURSOR_IMAGE_DECODE_BYTES) {
+      throw new CursorImageError("Image input is too large to process safely.");
+    }
+
+    const prepared = await prepareCursorImageForWire({
+      data,
+      mimeType,
+      detail: options?.detail,
+    });
+    if (prepared.data.length > MAX_CURSOR_IMAGE_BYTES) {
       throw new CursorImageError("Image input is too large (max 1 MiB). Resize and retry.");
     }
-    out.push({ data, mimeType, uuid: crypto.randomUUID() });
+
+    out.push({
+      data: prepared.data,
+      mimeType: prepared.mimeType,
+      uuid: crypto.randomUUID(),
+      ...(typeof prepared.width === "number" && typeof prepared.height === "number"
+        ? { width: prepared.width, height: prepared.height }
+        : {}),
+    });
   }
   return out;
 }
@@ -327,17 +696,11 @@ export async function resolveCursorImages(imageUrls: string[]): Promise<EncodedI
  * Returns the raw url strings (data: or http(s):) in order. Non-image parts
  * are ignored. A plain string content has no images.
  */
-export function extractImageUrls(
-  content: unknown
-): string[] {
+export function extractImageUrls(content: unknown): string[] {
   if (!Array.isArray(content)) return [];
   const urls: string[] = [];
   for (const part of content) {
-    if (
-      part &&
-      typeof part === "object" &&
-      (part as { type?: unknown }).type === "image_url"
-    ) {
+    if (part && typeof part === "object" && (part as { type?: unknown }).type === "image_url") {
       const imageUrl = (part as { image_url?: unknown }).image_url;
       if (typeof imageUrl === "string") {
         urls.push(imageUrl);

open-sse/utils/mediaParts.ts

@@ -0,0 +1,250 @@
+/**
+ * Unified media-part detection for request messages.
+ * Single source of truth shared by the vision/audio bridge guardrails (src/)
+ * and the combo compatibility filter (open-sse/) — the two previously kept
+ * divergent copies (guardrail missed input_image; combo saw it).
+ */
+export type MediaKind = "image" | "audio";
+
+export interface MediaPart {
+  kind: MediaKind;
+  /** URL, data URI, or base64 payload reference for the media content. */
+  ref: string;
+  /**
+   * Location of the top-level content part this hit belongs to. For nested
+   * hits (`nested: true`) these indexes point at the CONTAINER part — the
+   * entry of `message.content` under which the media was found — not at the
+   * media object itself.
+   */
+  messageIndex: number;
+  partIndex: number;
+  /**
+   * True when the media was found below the top level of the content part
+   * (inside another object/array, e.g. an image nested in an audio payload
+   * or a data URI inside a text field). Splice-style consumers can only
+   * replace top-level parts, so they must skip nested hits.
+   */
+  nested: boolean;
+  /** Original wire shape, for callers that need format-specific handling. */
+  shape:
+    | "image_url"
+    | "image_base64"
+    | "image_source_url"
+    | "input_image"
+    | "data_uri_string"
+    | "input_audio"
+    | "audio_url"
+    /** Audio detected via `source.media_type: audio/*` (no explicit type). */
+    | "audio_source"
+    /**
+     * Combo-parity indicator: the value looks like an image part (image-ish
+     * `type` in any casing, a bare `image_url`/`input_image` key, or a
+     * `source.media_type` of image/*) but carries no extractable ref — `ref`
+     * may be "". Boolean callers (combo compatibility filter) count it;
+     * ref-consuming callers (vision bridge) must skip empty refs.
+     */
+    | "image_indicator";
+}
+
+const MAX_DEPTH = 8;
+
+interface DetectCtx {
+  out: MediaPart[];
+  messageIndex: number;
+  partIndex: number;
+  /** When set, `found` flips true on the first part of this kind (early exit). */
+  stopAtKind?: MediaKind;
+  found?: boolean;
+}
+
+/** Extract a URL from either a bare string or a `{ url }` object. */
+function urlFrom(raw: unknown): string | undefined {
+  if (typeof raw === "string") return raw;
+  const url = (raw as Record<string, unknown> | undefined)?.url;
+  return typeof url === "string" ? url : undefined;
+}
+
+function pushPart(
+  ctx: DetectCtx,
+  kind: MediaKind,
+  ref: string,
+  shape: MediaPart["shape"],
+  depth: number
+): void {
+  ctx.out.push({
+    kind,
+    ref,
+    messageIndex: ctx.messageIndex,
+    partIndex: ctx.partIndex,
+    nested: depth > 0,
+    shape,
+  });
+  if (ctx.stopAtKind === kind) ctx.found = true;
+}
+
+/** Strict image shapes with an extractable ref. Returns true when one was pushed. */
+function inspectImageShapes(
+  obj: Record<string, unknown>,
+  type: string | undefined,
+  ctx: DetectCtx,
+  depth: number
+): boolean {
+  if (type === "image_url" || type === "input_image") {
+    const url = urlFrom(obj.image_url);
+    if (url) {
+      pushPart(ctx, "image", url, type === "input_image" ? "input_image" : "image_url", depth);
+      return true;
+    }
+  }
+  if (type === "image") {
+    const source = obj.source as Record<string, unknown> | undefined;
+    if (source?.type === "base64" && typeof source.data === "string") {
+      const media = typeof source.media_type === "string" ? source.media_type : "image/png";
+      pushPart(ctx, "image", `data:${media};base64,${source.data}`, "image_base64", depth);
+      return true;
+    }
+    // Non-empty url required: an empty `source.url` is not an extractable image
+    // (mirrors the guardrail's historical `if (url)` guard).
+    if (source?.type === "url" && typeof source.url === "string" && source.url) {
+      pushPart(ctx, "image", source.url, "image_source_url", depth);
+      return true;
+    }
+  }
+  return false;
+}
+
+/**
+ * Audio shapes. Returns true when a part was pushed (at most one per object).
+ * Callers must NOT early-return on audio: the same object can also carry
+ * image indicators or nest image parts inside its payload.
+ */
+function inspectAudioShapes(
+  obj: Record<string, unknown>,
+  type: string | undefined,
+  mediaType: unknown,
+  ctx: DetectCtx,
+  depth: number
+): boolean {
+  if (type === "input_audio") {
+    const audio = obj.input_audio as Record<string, unknown> | undefined;
+    if (typeof audio?.data === "string") {
+      pushPart(ctx, "audio", audio.data, "input_audio", depth);
+      return true;
+    }
+  }
+  if (type === "audio_url") {
+    const url = urlFrom(obj.audio_url);
+    if (url) {
+      pushPart(ctx, "audio", url, "audio_url", depth);
+      return true;
+    }
+  }
+  if (typeof mediaType === "string" && mediaType.startsWith("audio/")) {
+    const data = (obj.source as Record<string, unknown>).data;
+    if (typeof data === "string") {
+      pushPart(ctx, "audio", data, "audio_source", depth);
+      return true;
+    }
+  }
+  return false;
+}
+
+/**
+ * Combo-parity image indicators: the legacy valueContainsImagePart
+ * (comboStructure) matched image-ish `type` names case-insensitively, bare
+ * `image_url`/`input_image` keys, and `source.media_type` image/* — all
+ * without needing an extractable ref. Emit an indicator part (ref
+ * best-effort, possibly "") so boolean callers keep seeing those requests as
+ * vision requests. Returns true when one was pushed.
+ */
+function inspectImageIndicators(
+  obj: Record<string, unknown>,
+  type: string | undefined,
+  mediaType: unknown,
+  ctx: DetectCtx,
+  depth: number
+): boolean {
+  const lowerType = type?.toLowerCase();
+  const looksLikeImage =
+    lowerType === "image" ||
+    lowerType === "image_url" ||
+    lowerType === "input_image" ||
+    "image_url" in obj ||
+    "input_image" in obj;
+  const imageMediaType =
+    typeof mediaType === "string" && mediaType.toLowerCase().startsWith("image/");
+  if (!looksLikeImage && !imageMediaType) return false;
+  pushPart(ctx, "image", urlFrom(obj.image_url ?? obj.input_image) ?? "", "image_indicator", depth);
+  return true;
+}
+
+function inspect(value: unknown, ctx: DetectCtx, depth: number): void {
+  if (ctx.found || depth > MAX_DEPTH || value == null) return;
+  if (typeof value === "string") {
+    if (value.startsWith("data:image/")) pushPart(ctx, "image", value, "data_uri_string", depth);
+    return;
+  }
+  if (Array.isArray(value)) {
+    for (const entry of value) {
+      inspect(entry, ctx, depth + 1);
+      if (ctx.found) return;
+    }
+    return;
+  }
+  if (typeof value !== "object") return;
+  const obj = value as Record<string, unknown>;
+  const type = typeof obj.type === "string" ? obj.type : undefined;
+
+  if (inspectImageShapes(obj, type, ctx, depth)) return;
+
+  const mediaType = (obj.source as Record<string, unknown> | undefined)?.media_type;
+  // Audio does not early-return: the same object can also carry image
+  // indicators (bare `image_url`/`input_image` keys the legacy combo filter
+  // matched) or nest image parts inside its payload.
+  inspectAudioShapes(obj, type, mediaType, ctx, depth);
+  if (ctx.found) return;
+  if (inspectImageIndicators(obj, type, mediaType, ctx, depth)) return;
+  for (const nested of Object.values(obj)) {
+    inspect(nested, ctx, depth + 1);
+    if (ctx.found) return;
+  }
+}
+
+export function detectMediaParts(
+  messages: ReadonlyArray<{ role?: string; content?: unknown }> | undefined | null
+): MediaPart[] {
+  const out: MediaPart[] = [];
+  if (!Array.isArray(messages)) return out;
+  for (let messageIndex = 0; messageIndex < messages.length; messageIndex++) {
+    const content = messages[messageIndex]?.content;
+    if (!Array.isArray(content)) continue;
+    for (let partIndex = 0; partIndex < content.length; partIndex++) {
+      inspect(content[partIndex], { out, messageIndex, partIndex }, 0);
+    }
+  }
+  return out;
+}
+
+/**
+ * Early-exit presence check: returns true as soon as the FIRST part of the
+ * requested kind is found, without collecting the full part list or finishing
+ * the traversal. Prefer this on hot paths (e.g. the combo compatibility
+ * filter runs on every request) over `detectMediaParts(...).some(...)`.
+ */
+export function containsMediaKind(
+  messages: ReadonlyArray<{ role?: string; content?: unknown }> | undefined | null,
+  kind: MediaKind
+): boolean {
+  if (!Array.isArray(messages)) return false;
+  const out: MediaPart[] = [];
+  for (let messageIndex = 0; messageIndex < messages.length; messageIndex++) {
+    const content = messages[messageIndex]?.content;
+    if (!Array.isArray(content)) continue;
+    for (let partIndex = 0; partIndex < content.length; partIndex++) {
+      const ctx: DetectCtx = { out, messageIndex, partIndex, stopAtKind: kind };
+      inspect(content[partIndex], ctx, 0);
+      if (ctx.found) return true;
+    }
+  }
+  return false;
+}

open-sse/utils/sseHeartbeat.ts

@@ -79,7 +79,8 @@ export function sseCommentsEnabled(): boolean {
   if (typeof process === "undefined") return true;
   const v = process.env.OMNIROUTE_SSE_COMMENTS;
   if (v === undefined || v === "") return true;
-  return v.trim().toLowerCase() !== "off";
+  const normalized = v.trim().toLowerCase();
+  return normalized !== "off" && normalized !== "false" && normalized !== "0" && normalized !== "no";
 }
 
 export function createSseHeartbeatTransform({

open-sse/utils/stream.ts

@@ -25,6 +25,7 @@ import {
 } from "./streamHelpers.ts";
 import { calculateCost } from "@/lib/usage/costCalculator";
 import { buildOmniRouteSseMetadataComment } from "@/domain/omnirouteResponseMeta";
+import { sseCommentsEnabled } from "./sseHeartbeat.ts";
 import {
   createStructuredSSECollector,
   buildStreamSummaryFromEvents,
@@ -1001,6 +1002,11 @@ export function createSSEStream(options: StreamOptions = {}) {
     controller: TransformStreamDefaultController,
     finalUsage: UsageTokenRecord | Record<string, unknown> | null | undefined
   ) => {
+    // Skip SSE metadata comment lines when OMNIROUTE_SSE_COMMENTS is disabled
+    // (e.g., "off", "false", "0", "no"). Strict OpenAI-compatible clients that
+    // JSON.parse every SSE line will crash on `: x-omniroute-*` comment lines.
+    if (!sseCommentsEnabled()) return;
+
     const costUsd = finalUsage ? await calculateCost(provider, model, finalUsage) : 0;
     const comment = buildOmniRouteSseMetadataComment({
       provider,
@@ -2463,8 +2469,18 @@ export function createSSEStream(options: StreamOptions = {}) {
                   status: 200,
                   usage,
                   responseBody,
+                  // #9315 switched the summary to the accumulated responseBody to avoid
+                  // stale/truncated event data — but responseBody here is synthesized in
+                  // chat-completion shape, which loses the Responses API `response` object.
+                  // Keep the events-derived summary for OPENAI_RESPONSES only.
                   providerPayload: providerPayloadCollector.build(
-                    responseBody,
+                    sourceFormat === FORMATS.OPENAI_RESPONSES
+                      ? buildStreamSummaryFromEvents(
+                          providerPayloadCollector.getEvents(),
+                          sourceFormat,
+                          model
+                        )
+                      : responseBody,
                     { includeEvents: false }
                   ),
                   clientPayload: clientPayloadCollector.build(responseBody, {
@@ -2734,8 +2750,16 @@ export function createSSEStream(options: StreamOptions = {}) {
                 status: 200,
                 usage: state?.usage,
                 responseBody,
+                // Same OPENAI_RESPONSES carve-out as the passthrough branch above —
+                // the synthesized chat-shaped responseBody drops the `response` object.
                 providerPayload: providerPayloadCollector.build(
-                  responseBody,
+                  targetFormat === FORMATS.OPENAI_RESPONSES
+                    ? buildStreamSummaryFromEvents(
+                        providerPayloadCollector.getEvents(),
+                        targetFormat,
+                        model
+                      )
+                    : responseBody,
                   { includeEvents: false }
                 ),
                 clientPayload: clientPayloadCollector.build(responseBody, {

package-lock.json

@@ -27,12 +27,11 @@
         "@xyflow/react": "^12.11.1",
         "axios": "^1.16.1",
         "bcryptjs": "^3.0.3",
-        "better-sqlite3": "^13.0.2",
         "bottleneck": "^2.19.5",
         "clsx": "^2.1.1",
         "commander": "^15.0.0",
         "csv-stringify": "^6.7.0",
-        "dompurify": "^3.4.12",
+        "dompurify": "^3.4.13",
         "express": "^5.2.1",
         "fetch-socks": "^1.3.3",
         "fflate": "^0.8.3",
@@ -74,6 +73,7 @@
         "recharts": "^3.8.1",
         "safe-regex": "^2.1.1",
         "selfsigned": "^5.5.0",
+        "sharp": "^0.35.3",
         "smol-toml": "1.7.1",
         "socks": "^2.8.7",
         "sql.js": "^1.14.1",
@@ -104,7 +104,7 @@
         "@testing-library/jest-dom": "^6.9.1",
         "@testing-library/react": "^16.3.2",
         "@types/better-sqlite3": "^7.6.13",
-        "@types/bun": "*",
+        "@types/bun": "latest",
         "@types/node": "^26.1.0",
         "@types/react": "^19.2.15",
         "@types/react-dom": "^19.2.3",
@@ -463,9 +463,9 @@
       }
     },
     "node_modules/@apidevtools/json-schema-ref-parser/node_modules/js-yaml": {
-      "version": "4.3.0",
-      "resolved": "https://registry.npmjs.org/js-yaml/-/js-yaml-4.3.0.tgz",
-      "integrity": "sha512-1td788aAnnZ5qs7V2QIRl1owjtYpbKt749Y3xauqQgwIIGF/xXWz1wMTEBx5O3LK3lXLVuqXPdPxj2BoFHaW9Q==",
+      "version": "4.3.1",
+      "resolved": "https://registry.npmjs.org/js-yaml/-/js-yaml-4.3.1.tgz",
+      "integrity": "sha512-CY6crGq313MX8GkwvB7tzgp99vjQxY1++5y10/BKN/GUfHqWaOGQMNZkBvqSzsZKWk/ijwHlWzzkLulsGHhjWQ==",
       "dev": true,
       "funding": [
         {
@@ -3076,9 +3076,9 @@
       }
     },
     "node_modules/@eslint/eslintrc/node_modules/js-yaml": {
-      "version": "4.3.0",
-      "resolved": "https://registry.npmjs.org/js-yaml/-/js-yaml-4.3.0.tgz",
-      "integrity": "sha512-1td788aAnnZ5qs7V2QIRl1owjtYpbKt749Y3xauqQgwIIGF/xXWz1wMTEBx5O3LK3lXLVuqXPdPxj2BoFHaW9Q==",
+      "version": "4.3.1",
+      "resolved": "https://registry.npmjs.org/js-yaml/-/js-yaml-4.3.1.tgz",
+      "integrity": "sha512-CY6crGq313MX8GkwvB7tzgp99vjQxY1++5y10/BKN/GUfHqWaOGQMNZkBvqSzsZKWk/ijwHlWzzkLulsGHhjWQ==",
       "dev": true,
       "funding": [
         {
@@ -3561,7 +3561,6 @@
       "resolved": "https://registry.npmjs.org/@img/colour/-/colour-1.1.0.tgz",
       "integrity": "sha512-Td76q7j57o/tLVdgS746cYARfSyxk8iEfRxewL9h4OMzYhbW4TAcppl0mT4eyqXddh6L/jwoM75mo7ixa/pCeQ==",
       "license": "MIT",
-      "optional": true,
       "engines": {
         "node": ">=18"
       }
@@ -12640,9 +12639,9 @@
       }
     },
     "node_modules/@yarnpkg/parsers/node_modules/js-yaml": {
-      "version": "4.3.0",
-      "resolved": "https://registry.npmjs.org/js-yaml/-/js-yaml-4.3.0.tgz",
-      "integrity": "sha512-1td788aAnnZ5qs7V2QIRl1owjtYpbKt749Y3xauqQgwIIGF/xXWz1wMTEBx5O3LK3lXLVuqXPdPxj2BoFHaW9Q==",
+      "version": "4.3.1",
+      "resolved": "https://registry.npmjs.org/js-yaml/-/js-yaml-4.3.1.tgz",
+      "integrity": "sha512-CY6crGq313MX8GkwvB7tzgp99vjQxY1++5y10/BKN/GUfHqWaOGQMNZkBvqSzsZKWk/ijwHlWzzkLulsGHhjWQ==",
       "dev": true,
       "funding": [
         {
@@ -16980,9 +16979,9 @@
       "license": "MIT"
     },
     "node_modules/dompurify": {
-      "version": "3.4.12",
-      "resolved": "https://registry.npmjs.org/dompurify/-/dompurify-3.4.12.tgz",
-      "integrity": "sha512-zQvGet8Z2sWbQhCmfFz/T5QWH2oBmjnqK3qvOjaqaNLrLEF912WamU+ohnTp0TCep/MFVHpdJuCZEdFOdTnEFg==",
+      "version": "3.4.13",
+      "resolved": "https://registry.npmjs.org/dompurify/-/dompurify-3.4.13.tgz",
+      "integrity": "sha512-2vmYIoqjze2d+kakP8S/nS5shfsl587kzwEjcGlTdiksUVgFHnFCsLYDVj/JNqJVOQZGSYBTmuycv0PodwmnMQ==",
       "license": "(MPL-2.0 OR Apache-2.0)",
       "optionalDependencies": {
         "@types/trusted-types": "^2.0.7"
@@ -25197,9 +25196,9 @@
       }
     },
     "node_modules/lockfile-lint/node_modules/js-yaml": {
-      "version": "4.3.0",
-      "resolved": "https://registry.npmjs.org/js-yaml/-/js-yaml-4.3.0.tgz",
-      "integrity": "sha512-1td788aAnnZ5qs7V2QIRl1owjtYpbKt749Y3xauqQgwIIGF/xXWz1wMTEBx5O3LK3lXLVuqXPdPxj2BoFHaW9Q==",
+      "version": "4.3.1",
+      "resolved": "https://registry.npmjs.org/js-yaml/-/js-yaml-4.3.1.tgz",
+      "integrity": "sha512-CY6crGq313MX8GkwvB7tzgp99vjQxY1++5y10/BKN/GUfHqWaOGQMNZkBvqSzsZKWk/ijwHlWzzkLulsGHhjWQ==",
       "dev": true,
       "funding": [
         {
@@ -26006,9 +26005,9 @@
       }
     },
     "node_modules/mermaid": {
-      "version": "11.16.0",
-      "resolved": "https://registry.npmjs.org/mermaid/-/mermaid-11.16.0.tgz",
-      "integrity": "sha512-Zvm3kbstgdpvIJPPItlL7fppIZ3kibvc1oZIGxdvk9t6UFz6flv+Jw7FtRGKwfcI8OckmH04LqG6LlS6X4B1pA==",
+      "version": "11.16.1",
+      "resolved": "https://registry.npmjs.org/mermaid/-/mermaid-11.16.1.tgz",
+      "integrity": "sha512-TQsq6u22fAn3rek5VOubrhKPo1g5hwC3FXUN9hiyupTckcYiGuuKGkNQrKYwGJkXUxZdojwRG46gsSCFZMDp4g==",
       "license": "MIT",
       "dependencies": {
         "@braintree/sanitize-url": "^7.1.2",
@@ -27529,9 +27528,9 @@
       "optional": true
     },
     "node_modules/nanoid": {
-      "version": "3.3.16",
-      "resolved": "https://registry.npmjs.org/nanoid/-/nanoid-3.3.16.tgz",
-      "integrity": "sha512-bzlKTyNJ7+LdGIIwy8ijFpIqEQIvafahV7eYykJ8Cvh42EdJeODoJ6gUJXpQJvej1BddH8OqTXZNE/KfbWAu8Q==",
+      "version": "3.3.18",
+      "resolved": "https://registry.npmjs.org/nanoid/-/nanoid-3.3.18.tgz",
+      "integrity": "sha512-DTg4MJbGMWkfi6VZFdNt2/caMbQy4Ou+Op/hJQvGEWcnVfoA1QA+xzRKAzw9jD6+GVOOeYr/mIcuDSdug6F6+w==",
       "funding": [
         {
           "type": "github",
@@ -32779,7 +32778,6 @@
       "resolved": "https://registry.npmjs.org/sharp/-/sharp-0.35.3.tgz",
       "integrity": "sha512-ej0zVHuZGHCiABXcNxeYhpRnPNPAcvbG8RMdBAhDAxLKkCRVSpK3Iyu7qbqw3JMzoj0REeM6f3tJLtVwl0023Q==",
       "license": "Apache-2.0",
-      "optional": true,
       "dependencies": {
         "@img/colour": "^1.1.0",
         "detect-libc": "^2.1.2",
@@ -32829,7 +32827,6 @@
       "resolved": "https://registry.npmjs.org/semver/-/semver-7.8.5.tgz",
       "integrity": "sha512-Y7/KDsb8LjooZpwaqGyulO6DQlksgCncchHGk+sZIY4SBvUocMBEFH5Ur1fI4dV+Jvl0w6cjvucaIi40puRioA==",
       "license": "ISC",
-      "optional": true,
       "bin": {
         "semver": "bin/semver.js"
       },
@@ -36609,9 +36606,9 @@
       }
     },
     "node_modules/xmlbuilder2/node_modules/js-yaml": {
-      "version": "4.3.0",
-      "resolved": "https://registry.npmjs.org/js-yaml/-/js-yaml-4.3.0.tgz",
-      "integrity": "sha512-1td788aAnnZ5qs7V2QIRl1owjtYpbKt749Y3xauqQgwIIGF/xXWz1wMTEBx5O3LK3lXLVuqXPdPxj2BoFHaW9Q==",
+      "version": "4.3.1",
+      "resolved": "https://registry.npmjs.org/js-yaml/-/js-yaml-4.3.1.tgz",
+      "integrity": "sha512-CY6crGq313MX8GkwvB7tzgp99vjQxY1++5y10/BKN/GUfHqWaOGQMNZkBvqSzsZKWk/ijwHlWzzkLulsGHhjWQ==",
       "dev": true,
       "funding": [
         {
@@ -36951,12 +36948,7 @@
     },
     "open-sse": {
       "name": "@omniroute/open-sse",
-      "version": "3.8.50",
-      "dependencies": {
-        "@toon-format/toon": "^4.1.0",
-        "safe-regex": "^2.1.1",
-        "smol-toml": "1.7.1"
-      }
+      "version": "3.8.50"
     }
   }
 }

package.json

@@ -34,8 +34,11 @@
     "scripts/dev/tls-options.mjs",
     "scripts/check/check-supported-node-runtime.ts",
     "scripts/dev/sync-env.mjs",
-    "scripts/build/native-binary-compat.mjs",
+    "scripts/build/assembleStandalone.mjs",
+    "scripts/build/backendOnlyPages.mjs",
     "scripts/build/build-next-isolated.mjs",
+    "scripts/build/build-tproxy-native.mjs",
+    "scripts/build/native-binary-compat.mjs",
     "scripts/build/runtime-env.mjs",
     "README.md",
     "LICENSE",
@@ -266,7 +269,7 @@
     "clsx": "^2.1.1",
     "commander": "^15.0.0",
     "csv-stringify": "^6.7.0",
-    "dompurify": "^3.4.12",
+    "dompurify": "^3.4.13",
     "express": "^5.2.1",
     "fetch-socks": "^1.3.3",
     "fflate": "^0.8.3",
@@ -308,6 +311,7 @@
     "recharts": "^3.8.1",
     "safe-regex": "^2.1.1",
     "selfsigned": "^5.5.0",
+    "sharp": "^0.35.3",
     "smol-toml": "1.7.1",
     "socks": "^2.8.7",
     "sql.js": "^1.14.1",
@@ -415,7 +419,6 @@
     "unrs-resolver": true
   },
   "overrides": {
-    "dompurify": "^3.4.12",
     "fast-xml-parser": "^5.10.1",
     "sharp": "^0.35.0",
     "postcss": "^8.5.18",
@@ -431,7 +434,7 @@
     "fast-uri": "^3.1.5",
     "body-parser": "^2.3.0",
     "@yarnpkg/parsers": {
-      "js-yaml": "^4.2.0"
+      "js-yaml": "^4.3.1"
     },
     "jsdom": {
       "undici": "^7.29.0"
@@ -446,11 +449,27 @@
     "promptfoo": {
       "js-yaml": "^5.2.2",
       "@apidevtools/json-schema-ref-parser": {
-        "js-yaml": "^4.2.0"
+        "js-yaml": "^4.3.1"
       },
       "undici": "^7.29.0"
     },
     "socket.io-parser": "^4.2.7",
-    "tar": "^7.5.21"
+    "tar": "^7.5.21",
+    "nanoid": "^3.3.17",
+    "@eslint/eslintrc": {
+      "js-yaml": "^4.3.1"
+    },
+    "lockfile-lint": {
+      "js-yaml": "^4.3.1"
+    },
+    "xmlbuilder2": {
+      "js-yaml": "^4.3.1"
+    },
+    "monaco-editor": {
+      "dompurify": "^3.4.13"
+    },
+    "@apidevtools/json-schema-ref-parser": {
+      "js-yaml": "^4.3.1"
+    }
   }
 }

scripts/build/assembleStandalone.mjs

@@ -89,6 +89,15 @@ const NATIVE_ASSET_ENTRIES = [
     src: ["node_modules", "better-sqlite3", "build"],
     dest: ["node_modules", "better-sqlite3", "build"],
   },
+  {
+    // #8847: Bun (and npx -g global installs) resolve better-sqlite3's native
+    // binary from prebuilds/ instead of build/Release/, so the compiled build/
+    // copy alone leaves a hollow package that falls back to sql.js (OOM under
+    // Bun). Ship the prebuilds alongside the compiled binary.
+    label: "better-sqlite3 prebuilds (Bun / global installs)",
+    src: ["node_modules", "better-sqlite3", "prebuilds"],
+    dest: ["node_modules", "better-sqlite3", "prebuilds"],
+  },
   {
     // TPROXY IP_TRANSPARENT addon (Fase 3 / Epic A). Built by build-tproxy-native
     // before assembly; Linux-only + opt-in, so the source is absent on non-Linux

scripts/build/colocateOptionals.mjs

@@ -47,7 +47,8 @@
  */
 
 import { cpSync, existsSync, mkdirSync, readFileSync } from "node:fs";
-import { dirname, join } from "node:path";
+import { createRequire } from "node:module";
+import { dirname, join, sep } from "node:path";
 
 /**
  * Entry packages of the SLM optional stack (the closure roots). `@huggingface/transformers` is
@@ -96,6 +97,33 @@ export function computeDependencyClosure(nodeModulesDir, seeds = SEED_PACKAGES)
   return closure;
 }
 
+/**
+ * A package in the target tree counts as PRESENT only when its entrypoint
+ * resolves from inside that tree — the same contract the Dockerfile's
+ * post-build guard enforces. Next's file tracing can materialize a package
+ * PARTIALLY (the package.json lands, the files its `main` points at do not),
+ * and a directory-level `existsSync` check then skips the package forever
+ * while the runtime dies with "Cannot find module <pkg>/dist/index.js".
+ *
+ * @param {string} targetNodeModulesDir
+ * @param {string} name
+ * @returns {boolean}
+ */
+function isPackageIntact(targetNodeModulesDir, name) {
+  if (!existsSync(join(targetNodeModulesDir, name))) return false;
+  try {
+    const probe = createRequire(
+      join(targetNodeModulesDir, "__colocate_probe__.js")
+    );
+    const resolved = probe.resolve(name);
+    // A resolution that walked past the target into an ancestor tree does not
+    // prove the target copy is usable.
+    return resolved.startsWith(targetNodeModulesDir + sep);
+  } catch {
+    return false;
+  }
+}
+
 /**
  * Co-locate the SLM optional dependency closure from `<rootDir>/node_modules`
  * into a standalone bundle's `node_modules`.
@@ -142,11 +170,12 @@ export function colocateLlmlinguaOptionals({
 
   const closure = computeDependencyClosure(rootNm, seeds);
 
-  // Check the complete closure rather than only the entry package. A partially
-  // populated bundle must still receive any missing transitive dependencies.
+  // Check the complete closure rather than only the entry package, and judge
+  // presence by entrypoint integrity — a partially traced directory (see
+  // isPackageIntact) must still receive its missing files.
   if (
     closure.length > 0 &&
-    closure.every((name) => existsSync(join(targetNm, name)))
+    closure.every((name) => isPackageIntact(targetNm, name))
   ) {
     return { skipped: true, reason: "already co-located" };
   }
@@ -155,11 +184,18 @@ export function colocateLlmlinguaOptionals({
 
   for (const name of closure) {
     const dest = join(targetNm, name);
-    if (existsSync(dest)) continue;
+    if (isPackageIntact(targetNm, name)) continue;
 
     try {
       mkdirSync(dirname(dest), { recursive: true });
-      cpSync(join(rootNm, name), dest, { recursive: true });
+      // force:false merges into a partially traced directory: files the trace
+      // already materialized are kept, missing ones (the package payload) are
+      // filled in from the root tree.
+      cpSync(join(rootNm, name), dest, {
+        recursive: true,
+        force: false,
+        errorOnExist: false,
+      });
       copied++;
     } catch (err) {
       log(

scripts/build/pack-artifact-policy.ts

@@ -93,6 +93,10 @@ export const PACK_ARTIFACT_ROOT_ALLOWED_EXACT_PATHS: string[] = [
   // runtime; shipped via package.json "files", so it must be allowed here.
   "bin/aliasResolverHook.mjs",
   "bin/mcp-server.mjs",
+  // #9281: stdout/stderr console guard preloaded via `node --import` by
+  // bin/mcp-server.mjs before the MCP entry's module graph evaluates — without it
+  // the published CLI's `omniroute --mcp` crashes on the pathToFileURL() import.
+  "bin/mcpStdioConsoleGuard.mjs",
   "bin/nodeRuntimeSupport.mjs",
   "bin/omniroute.mjs",
   "bin/reset-password.mjs",
@@ -117,6 +121,9 @@ export const PACK_ARTIFACT_ROOT_ALLOWED_EXACT_PATHS: string[] = [
   // shipped via package.json "files", so it must be allowed in the tarball.
   "open-sse/utils/setupPolyfill.ts",
   "package.json",
+  "scripts/build/assembleStandalone.mjs",
+  "scripts/build/backendOnlyPages.mjs",
+  "scripts/build/build-tproxy-native.mjs",
   "scripts/build/build-next-isolated.mjs",
   "scripts/check/check-supported-node-runtime.ts",
   "scripts/build/native-binary-compat.mjs",
@@ -183,6 +190,10 @@ export const PACK_ARTIFACT_REQUIRED_PATHS: string[] = [
   "bin/cli/utils/storageKeyProvision.mjs",
   "bin/cli/utils/versionFastPath.mjs",
   "bin/mcp-server.mjs",
+  // #9281: stdout/stderr console guard preloaded via `node --import` by
+  // bin/mcp-server.mjs before the MCP entry's module graph evaluates — without it
+  // the published CLI's `omniroute --mcp` crashes on the pathToFileURL() import.
+  "bin/mcpStdioConsoleGuard.mjs",
   "bin/nodeRuntimeSupport.mjs",
   "bin/omniroute.mjs",
   // #7808: aliasResolver + its hook file. bin/omniroute.mjs imports

scripts/dev/generate-adobe-firefly-snapshot.mjs

@@ -0,0 +1,207 @@
+#!/usr/bin/env node
+
+import fs from "node:fs";
+import path from "node:path";
+import { createHash } from "node:crypto";
+
+function usage() {
+  console.error(
+    "Usage: node scripts/dev/generate-adobe-firefly-snapshot.mjs <discovery.json> <output.ts>"
+  );
+  process.exit(2);
+}
+
+const [, , inputArg, outputArg] = process.argv;
+if (!inputArg || !outputArg) usage();
+
+const inputPath = path.resolve(inputArg);
+const outputPath = path.resolve(outputArg);
+const inputBytes = fs.readFileSync(inputPath);
+const sourceHash = createHash("sha256").update(inputBytes).digest("hex");
+const root = JSON.parse(inputBytes.toString("utf8"));
+
+function mergeObjectSchema(schema) {
+  const merged = { properties: {}, required: [] };
+  const visit = (node) => {
+    if (!node || typeof node !== "object") return;
+    if (node.properties && typeof node.properties === "object") {
+      Object.assign(merged.properties, node.properties);
+    }
+    if (Array.isArray(node.required)) merged.required.push(...node.required);
+    if (Array.isArray(node.allOf)) node.allOf.forEach(visit);
+  };
+  visit(schema);
+  merged.required = [...new Set(merged.required)];
+  return merged;
+}
+
+function branches(schema) {
+  if (!schema || typeof schema !== "object") return [];
+  return [schema, ...(schema.anyOf || []), ...(schema.oneOf || [])];
+}
+
+function stringEnums(schema) {
+  return [
+    ...new Set(
+      branches(schema)
+        .flatMap((branch) => (Array.isArray(branch.enum) ? branch.enum : []))
+        .filter((value) => typeof value === "string")
+    ),
+  ];
+}
+
+function integerSchema(schema) {
+  return branches(schema).find((branch) => branch.type === "integer") || {};
+}
+
+function publicModelId(modelId, modelVersion) {
+  const slug = (value, allowDot = false) =>
+    String(value || "")
+      .trim()
+      .toLowerCase()
+      .replace(allowDot ? /[^a-z0-9.]+/g : /[^a-z0-9]+/g, "-")
+      .replace(/^-|-$/g, "");
+  const family = slug(modelId);
+  const publicVersion =
+    family === "kling" ? String(modelVersion).replace(/^kling_v3_omni/i, "kling_o3") : modelVersion;
+  const version = slug(publicVersion, true);
+  if (!version || version === "default" || version === family) return family || "model";
+  return `${family}-${version}`;
+}
+
+function normalizeModel(family, modelVersion, version) {
+  const schema = mergeObjectSchema(version.requestSchema);
+  const properties = schema.properties;
+  const referenceSchema = properties.referenceBlobs || {};
+  const referenceInputs = [];
+  for (const media of referenceSchema["x-capabilities"] || []) {
+    for (const usage of media.usageConstraints || []) {
+      if (usage.deprecated === true) continue;
+      referenceInputs.push({
+        mediaType: String(media.mediaType || ""),
+        usageType: String(usage.usageType || ""),
+        minItems: Number.isInteger(usage.minItems) ? usage.minItems : 0,
+        maxItems: Number.isInteger(usage.maxItems) ? usage.maxItems : null,
+        maxFileSizeBytes: Number.isInteger(media.maxFileSizeBytes) ? media.maxFileSizeBytes : null,
+      });
+    }
+  }
+
+  const supportedSizes = [
+    ...new Set(
+      branches(properties.size)
+        .flatMap((branch) => (Array.isArray(branch.enum) ? branch.enum : []))
+        .filter(
+          (size) =>
+            size &&
+            Number.isInteger(size.width) &&
+            size.width > 0 &&
+            Number.isInteger(size.height) &&
+            size.height > 0
+        )
+        .map((size) => `${size.width}x${size.height}`)
+    ),
+  ];
+  const supportedAspectRatios = [
+    ...new Set(
+      branches(properties.generationSettings).flatMap((branch) =>
+        stringEnums(branch?.properties?.aspectRatio)
+      )
+    ),
+  ];
+  const duration = integerSchema(properties.duration);
+  const supportedDurations = [
+    ...new Set(
+      branches(properties.duration)
+        .flatMap((branch) => (Array.isArray(branch.enum) ? branch.enum : []))
+        .filter(Number.isInteger)
+    ),
+  ];
+  const prompt = branches(properties.prompt).find((branch) => branch.type === "string") || {};
+  const outputCount = integerSchema(properties.n);
+
+  return {
+    id: publicModelId(family.modelId, modelVersion),
+    name: String(version.modelDisplayName || version.modelCaiDisplayName || modelVersion),
+    modality: version.outputModality[0],
+    upstreamModelId: family.modelId,
+    upstreamModelVersion: modelVersion,
+    providerName: String(family.acModelFamilyProviderDisplayName || ""),
+    releaseReadiness: String(version.releaseReadiness || ""),
+    healthStatus: String(version.healthStatus || ""),
+    inputMediaUseCases: (version.inputMediaUseCase || []).map(String),
+    schemaProperties: Object.keys(properties),
+    requiredProperties: schema.required,
+    referenceInputs,
+    maxReferenceItems: Number.isInteger(referenceSchema.maxItems) ? referenceSchema.maxItems : null,
+    supportedSizes,
+    supportedAspectRatios,
+    supportedResolutions: stringEnums(properties.resolution),
+    supportedDurations,
+    durationMin: Number.isInteger(duration.minimum) ? duration.minimum : null,
+    durationMax: Number.isInteger(duration.maximum) ? duration.maximum : null,
+    durationDefault: Number.isInteger(duration.default) ? duration.default : null,
+    outputCountMin: Number.isInteger(outputCount.minimum) ? outputCount.minimum : null,
+    outputCountMax: Number.isInteger(outputCount.maximum) ? outputCount.maximum : null,
+    promptMaxLength: Number.isInteger(prompt.maxLength) ? prompt.maxLength : null,
+    backingModel: String(version.bksGenerationModel || ""),
+  };
+}
+
+const rawModels = [];
+for (const family of Array.isArray(root.models) ? root.models : []) {
+  for (const [modelVersion, version] of Object.entries(family.modelVersions || {})) {
+    if (!version || version.enabled === false) continue;
+    const modality = Array.isArray(version.outputModality)
+      ? version.outputModality.map((value) => String(value).toLowerCase())[0]
+      : "";
+    if (modality !== "image" && modality !== "video") continue;
+
+    const schema = mergeObjectSchema(version.requestSchema);
+    if (!schema.properties.prompt) continue;
+    const useCases = (version.inputMediaUseCase || []).map((value) => String(value).toLowerCase());
+    if (useCases.some((value) => ["upscaling", "sharpening", "denoising"].includes(value))) {
+      continue;
+    }
+    rawModels.push(normalizeModel(family, modelVersion, version));
+  }
+}
+
+// Discovery currently repeats a few exact aliases (for example flux/fluxPro and
+// fluxPro/1.1). Keep the first canonical wire pair and suppress duplicate cards.
+const seen = new Set();
+const models = [];
+for (const model of rawModels) {
+  const semanticKey = JSON.stringify({
+    backingModel: model.backingModel,
+    name: model.name,
+    modality: model.modality,
+    schemaProperties: model.schemaProperties,
+    requiredProperties: model.requiredProperties,
+    referenceInputs: model.referenceInputs,
+    maxReferenceItems: model.maxReferenceItems,
+    supportedSizes: model.supportedSizes,
+    supportedAspectRatios: model.supportedAspectRatios,
+    supportedResolutions: model.supportedResolutions,
+    supportedDurations: model.supportedDurations,
+    durationMin: model.durationMin,
+    durationMax: model.durationMax,
+  });
+  if (seen.has(semanticKey)) continue;
+  seen.add(semanticKey);
+  models.push(model);
+}
+
+const source = `/**
+ * Generated from Adobe Firefly POST /v2/models/discovery with resolveSchema=true.
+ * Source SHA-256: ${sourceHash}
+ * Regenerate with scripts/dev/generate-adobe-firefly-snapshot.mjs; do not edit by hand.
+ * The generated literal stays compact to satisfy the repository's line-count gate.
+ */
+// prettier-ignore
+export const ADOBE_FIREFLY_DISCOVERY_SNAPSHOT = ${JSON.stringify(models)} as const;
+`;
+
+fs.mkdirSync(path.dirname(outputPath), { recursive: true });
+fs.writeFileSync(outputPath, source, "utf8");
+console.log(`Wrote ${models.length} models to ${outputPath}`);

scripts/quality/validate-release-green.mjs

@@ -609,14 +609,6 @@ async function main() {
         args: ["run", "check:pack-artifact"],
         timeout: 20 * 60 * 1000,
       });
-      // WS1.2 (#7065 class): boot the REAL packed tarball from a clean install —
-      // the runtime gate structure checks cannot provide. Reuses the same dist/ build.
-      slow.push({
-        id: "pack-boot",
-        label: "Tarball boot-smoke (installed CLI serves /health)",
-        args: ["run", "check:pack-boot"],
-        timeout: 15 * 60 * 1000,
-      });
     }
     slow.forEach((g) => announce(`${g.label} [parallel]`));
     const slowResults = await Promise.all(
@@ -633,6 +625,41 @@ async function main() {
         detail: code === 0 ? "pass" : firstFailureLine(out),
       });
     });
+
+    if (WITH_BUILD) {
+      // WS1.2 (#7065 class): boot the REAL packed tarball from a clean install.
+      // check:pack-artifact is the builder for dist/ when staging is absent, so the
+      // boot smoke MUST run after it completes. Running both in the parallel wave
+      // races check:pack-boot against dist/server.js creation on clean worktrees.
+      const packArtifactIndex = slow.findIndex((g) => g.id === "pack-artifact");
+      const packArtifactResult = slowResults[packArtifactIndex];
+      const bootLabel = "Tarball boot-smoke (installed CLI serves /health)";
+
+      if (!packArtifactResult || packArtifactResult.code !== 0) {
+        const out = "skipped because package-artifact did not produce a valid dist/ build";
+        saveGateLog("pack-boot", out);
+        record({
+          id: "pack-boot",
+          label: bootLabel,
+          kind: "hard",
+          ok: false,
+          detail: out,
+        });
+      } else {
+        announce(bootLabel);
+        const { code, out } = await runAsync(npmCmd, ["run", "check:pack-boot"], {
+          timeout: 15 * 60 * 1000,
+        });
+        saveGateLog("pack-boot", out);
+        record({
+          id: "pack-boot",
+          label: bootLabel,
+          kind: "hard",
+          ok: code === 0,
+          detail: code === 0 ? "pass" : firstFailureLine(out),
+        });
+      }
+    }
   } else if (WITH_BUILD) {
     // --with-build without the suites (--quick): still verify the package artifact.
     const { code, out } = await runAsync(npmCmd, ["run", "check:pack-artifact"], {

src/app/(dashboard)/dashboard/media-providers/components/LlmChatCard.tsx

@@ -134,7 +134,7 @@ export function LlmChatCard({
 }: Props) {
   const t = useTranslations("miniPlayground");
   const { keys } = useApiKey();
-  const { models } = useProviderModels(providerId);
+  const { models, loading, error, retry } = useProviderModels(providerId);
 
   const [internalSelectedKey, setInternalSelectedKey] = useState<string>("");
   const [internalModel, setInternalModel] = useState<string>(initialModel ?? "");
@@ -392,15 +392,31 @@ export function LlmChatCard({
             <select
               value={model || firstModel}
               onChange={(e) => setModel(e.target.value)}
-              className="min-w-0 flex-1 rounded-md border border-border bg-bg-subtle text-xs px-2 py-1 text-text-main focus:outline-none focus:ring-1 focus:ring-primary"
+              disabled={loading}
+              className="min-w-0 flex-1 rounded-md border border-border bg-bg-subtle text-xs px-2 py-1 text-text-main focus:outline-none focus:ring-1 focus:ring-primary disabled:opacity-60"
             >
-              {modelOptions.length === 0 && <option value="">{initialModel || "—"}</option>}
+              {modelOptions.length === 0 && !loading && <option value="">{initialModel || "—"}</option>}
+              {loading && <option value="">{t("loading") ?? "Loading…"}</option>}
               {modelOptions.map((m) => (
                 <option key={m.id} value={m.id}>
                   {m.id}
                 </option>
               ))}
             </select>
+            {error && (
+              <span className="text-xs text-red-500 flex items-center gap-1" role="alert">
+                <span className="truncate max-w-[180px]" title={String(error)}>
+                  {String(error)}
+                </span>
+                <button
+                  type="button"
+                  onClick={retry}
+                  className="shrink-0 text-xs text-primary hover:text-primary-strong underline"
+                >
+                  {t("retry") ?? "Retry"}
+                </button>
+              </span>
+            )}
           </div>
           {/* Key select */}
           {keys.length > 0 && (

src/app/(dashboard)/dashboard/providers/hooks/useProviderModels.ts

@@ -1,6 +1,6 @@
 "use client";
 
-import { useState, useEffect } from "react";
+import { useState, useEffect, useCallback, useRef } from "react";
 
 export interface ProviderModel {
   id: string;
@@ -18,6 +18,8 @@ interface UseProviderModelsResult {
   models: ProviderModel[];
   loading: boolean;
   error: string | null;
+  /** Re-runs the model fetch for the current provider. Useful for a Retry action. */
+  retry: () => void;
 }
 
 /**
@@ -32,15 +34,14 @@ export function useProviderModels(providerId: string): UseProviderModelsResult {
   const [models, setModels] = useState<ProviderModel[]>([]);
   const [loading, setLoading] = useState<boolean>(true);
   const [error, setError] = useState<string | null>(null);
+  // Cancels any in-flight load (component unmount or a retry superseding the
+  // previous request) so a stale response never overwrites a newer one.
+  const cleanupRef = useRef<(() => void) | null>(null);
 
-  useEffect(() => {
-    if (!providerId) {
-      setLoading(false);
-      return;
-    }
-
+  const load = useCallback(() => {
+    cleanupRef.current?.();
     let cancelled = false;
-    const load = async () => {
+    const run = async () => {
       setLoading(true);
       setError(null);
       try {
@@ -109,11 +110,33 @@ export function useProviderModels(providerId: string): UseProviderModelsResult {
         if (!cancelled) setLoading(false);
       }
     };
-    void load();
-    return () => {
+    void run();
+    const cleanup = () => {
       cancelled = true;
     };
+    cleanupRef.current = cleanup;
+    return cleanup;
   }, [providerId]);
 
-  return { models, loading, error };
+  useEffect(() => {
+    if (!providerId) {
+      setLoading(false);
+      return;
+    }
+    return load();
+  }, [providerId, load]);
+
+  // Release the current in-flight cleanup on unmount so no state updates leak.
+  useEffect(() => {
+    return () => {
+      cleanupRef.current?.();
+    };
+  }, []);
+
+  const retry = useCallback(() => {
+    if (!providerId) return;
+    load();
+  }, [providerId, load]);
+
+  return { models, loading, error, retry };
 }

src/app/(dashboard)/dashboard/radar/page.tsx

@@ -6,6 +6,7 @@ import { notFound } from "next/navigation";
 import Link from "next/link";
 import { Card } from "@/shared/components";
 import { shouldAutoSyncOnOpen } from "@/lib/radar/autoSync";
+import { isValidSupporterKeyFormat } from "@/lib/radar/supporterKey";
 
 // ---------------------------------------------------------------------------
 // Types
@@ -132,6 +133,18 @@ export default function RadarPage() {
   const [contributorClaimUrl, setContributorClaimUrl] = useState<string | null>(null);
   const [supporterPlansUrl, setSupporterPlansUrl] = useState<string | null>(null);
 
+  // Paste-key activation — the primary path on the activation screen: paste
+  // an already-obtained supporter key (`omr_` + 40 hex) to activate opt-in
+  // AND the supporter tier in a single POST. `hasSupporterKey`/
+  // `supporterKeyMasked` mirror GET /api/radar/settings so a key set out of
+  // band (e.g. a direct curl call before this UI existed) shows the masked
+  // form instead of an empty input — the raw key is never displayed.
+  const [keyInput, setKeyInput] = useState("");
+  const [keySubmitting, setKeySubmitting] = useState(false);
+  const [hasSupporterKey, setHasSupporterKey] = useState(false);
+  const [supporterKeyMasked, setSupporterKeyMasked] = useState<string | null>(null);
+  const [showKeyForm, setShowKeyForm] = useState(false);
+
   // Fetch catalog
   const fetchCatalog = useCallback(async () => {
     setLoading(true);
@@ -189,6 +202,10 @@ export default function RadarPage() {
       if (!settingsRes.ok) throw new Error(`HTTP ${settingsRes.status}`);
       const settingsData = await settingsRes.json();
       setOptIn(settingsData.optIn === true);
+      setHasSupporterKey(settingsData.hasSupporterKey === true);
+      setSupporterKeyMasked(
+        typeof settingsData.supporterKeyMasked === "string" ? settingsData.supporterKeyMasked : null,
+      );
       // F4/T7 — best-effort: keep whatever we already had if the field is
       // absent (older cached response shape), never fall back to a literal.
       if (typeof settingsData.contributorClaimUrl === "string") {
@@ -278,6 +295,40 @@ export default function RadarPage() {
     }
   }, [t, handleSync]);
 
+  // Activate with a pasted supporter key — the primary path on this screen.
+  // Submitting a key both sets it AND opts in, in a single POST (pasting a
+  // key unlocks the activation screen). Client-side format validation is a
+  // UX nicety only — the server (Zod) always revalidates.
+  const handleSubmitKey = useCallback(async () => {
+    setError("");
+    const trimmed = keyInput.trim();
+    if (!isValidSupporterKeyFormat(trimmed)) {
+      setError(t("keyInvalidFormatError"));
+      return;
+    }
+    setKeySubmitting(true);
+    try {
+      const res = await fetch("/api/radar/settings", {
+        method: "POST",
+        headers: { "Content-Type": "application/json" },
+        body: JSON.stringify({ optIn: true, supporterKey: trimmed }),
+      });
+      if (!res.ok) throw new Error(`HTTP ${res.status}`);
+      const data = await res.json();
+      setOptIn(true);
+      setHasSupporterKey(true);
+      setSupporterKeyMasked(typeof data.supporterKey === "string" ? data.supporterKey : null);
+      setKeyInput("");
+      setShowKeyForm(false);
+      // After activation, trigger a sync so the live tier catalog loads.
+      await handleSync();
+    } catch (err) {
+      setError(err instanceof Error ? err.message : t("activationFailed"));
+    } finally {
+      setKeySubmitting(false);
+    }
+  }, [keyInput, t, handleSync]);
+
   // Determine effective state
   const flagOn = optIn !== false || entries.length > 0 || meta !== null;
   const pageState = resolveRadarPageState(
@@ -359,6 +410,51 @@ export default function RadarPage() {
                     <span>{t("privacyLocalOnly")}</span>
                   </div>
                 </div>
+
+                {/* Paste-key activation — primary path: pasting an already-obtained
+                    supporter key both sets it AND opts in (unlocks this screen).
+                    The raw key is NEVER displayed — once set, only the masked
+                    form (supporterKeyMasked) is shown, with a "change key" escape
+                    hatch to paste a new one. */}
+                <div className="w-full flex flex-col gap-3 text-left">
+                  <p className="text-sm font-medium">{t("keySectionTitle")}</p>
+                  {hasSupporterKey && !showKeyForm ? (
+                    <div className="flex items-center justify-between gap-3 p-3 rounded-lg border border-border">
+                      <span className="font-mono text-sm text-text-muted">
+                        {supporterKeyMasked}
+                      </span>
+                      <button
+                        type="button"
+                        onClick={() => setShowKeyForm(true)}
+                        className="text-sm text-violet-400 hover:underline shrink-0"
+                      >
+                        {t("changeKeyButton")}
+                      </button>
+                    </div>
+                  ) : (
+                    <div className="flex flex-col sm:flex-row gap-2">
+                      <input
+                        type="text"
+                        value={keyInput}
+                        onChange={(e) => setKeyInput(e.target.value)}
+                        placeholder="omr_..."
+                        aria-label={t("keySectionTitle")}
+                        className="flex-1 px-3 py-2 text-sm font-mono rounded-lg border border-border bg-transparent focus:outline-none focus:ring-2 focus:ring-violet-500"
+                      />
+                      <button
+                        type="button"
+                        onClick={handleSubmitKey}
+                        disabled={keySubmitting || keyInput.trim().length === 0}
+                        className="px-6 py-2 bg-violet-500 hover:bg-violet-600 text-white font-medium rounded-lg transition-colors disabled:opacity-50 shrink-0"
+                      >
+                        {keySubmitting ? t("activating") : t("activateWithKeyButton")}
+                      </button>
+                    </div>
+                  )}
+                </div>
+
+                <div className="w-full h-px bg-border" />
+
                 <button
                   onClick={handleActivate}
                   disabled={activating}

src/app/api/memory/[id]/route.ts

@@ -1,6 +1,10 @@
 import { NextResponse } from "next/server";
 import { requireManagementAuth } from "@/lib/api/requireManagementAuth";
-import { memoryManager } from "@/lib/memory/manager";
+// Import through the module index, NOT "@/lib/memory/manager" directly: the index's
+// import-time side effect is what calls memoryManager.register(sqliteBackend). Importing
+// the bare manager gives an EMPTY registry, so every handler here threw
+// `Primary backend "sqlite" not registered` and returned 500 (#8752).
+import { memoryManager } from "@/lib/memory";
 import { validateBody, isValidationFailure } from "@/shared/validation/helpers";
 import { MemoryUpdatePutSchema } from "@/shared/schemas/memory";
 import { sanitizeErrorMessage } from "@omniroute/open-sse/utils/error.ts";