There was an asymmetry in the CLIENT_API policy: with REQUIRE_API_KEY
off, a request with no bearer was allowed as anonymous, but a request
with an *invalid* bearer was rejected with 401 "Invalid API key". That
surprises CLI integrations (Codex Desktop auto-config, Hermes Agent)
that ship a stale Bearer in their saved config — they'd see a 401 even
though the operator explicitly opted out of auth.
Fix: when validateApiKey fails and REQUIRE_API_KEY != "true", log a
single warning carrying the masked key id (last-4) and fall through to
anonymous. When REQUIRE_API_KEY is "true", the strict 401 path is
preserved.
The warning preserves observability:
[clientApiPolicy] invalid bearer presented to /api/v1/responses
but REQUIRE_API_KEY=false — falling through to anonymous
(key_id=key_XYZW)
Tests are added in a standalone file because the existing
client-api-policy.test.ts shares a DB-backed setup with a pre-existing
SQLite migration race (5/7 of its tests already failed on baseline).
The new file mocks validateApiKey via a require-resolve interceptor so
the policy's invalid-bearer branch is exercised without touching SQLite.
Reported by @k00shi on the Codex Desktop auto-config path.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
- Force skills: "" and params.stream: true in Command Code wrapper
- Align validation probe payload with upstream-required shape
- Default validation model to deepseek/deepseek-v4-flash
Co-authored-by: ddarkr <ddarkr@users.noreply.github.com>
- management.ts: replace === with timingSafeEqual for CLI token comparison
- machineToken.ts: salt upgraded to omniroute-cli-auth-v1; OMNIROUTE_CLI_SALT env
var honoured for rotation; full 64-char SHA-256 hex token
- tray.ps1: accept .png via GDI+ Bitmap->Icon handle; Windows tray works without .ico
- tray.ts: getIconPath() tries icon.ico then icon.png on Windows
- compression/types.ts: DEFAULT_CAVEMAN_CONFIG.preservePatterns filled with
six defaults (fenced code, inline code, URLs, paths, error lines, stack traces)
- CLAUDE.md: Hard Rule #15 — spawn-capable routes must use isLocalOnlyPath()
- .env.example + docs/reference/ENVIRONMENT.md: document OMNIROUTE_CLI_SALT
- docs/security/CLI_TOKEN.md: new (was referenced in changelog but missing)
- docs/security/ROUTE_GUARD_TIERS.md: new (was referenced in changelog but missing)
- tests/unit/lib/machineToken.test.ts: updated for 64-char token; added
OMNIROUTE_CLI_SALT env-var rotation test
Adds bin/cli/runtime/sqliteRuntime.mjs that resolves better-sqlite3 from:
(1) bundled optionalDependency, (2) ~/.omniroute/runtime/ install,
(3) lazy npm install into runtime dir, (4) node:sqlite stdlib (Node >=22.5),
(5) bundled sql.js WASM. Each native binary is validated against expected
platform magic bytes (ELF/Mach-O/PE) before load.
Adds bin/cli/runtime/magicBytes.mjs with validateBinaryMagic() helper
(9 tests). Adds bin/cli/runtime/index.mjs as warmUpRuntimes() orchestrator.
Adds scripts/postinstall.mjs warm-up hook (non-fatal, skipped in CI).
Integrates it as the last step of scripts/build/postinstall.mjs.
Extends src/lib/db/core.ts with ensureDbInitialized() (async, idempotent)
and getDriverInfo() so the startup orchestrator can await the resolver
before any DB access, enabling graceful degradation without crashing the
process on missing better-sqlite3.
Solves Windows EBUSY error on 'npm install -g omniroute@latest' while the
previous version is still running, and works in environments without C++
build tools or with unreachable npm registry.
Documents OMNIROUTE_SKIP_POSTINSTALL in .env.example and ENVIRONMENT.md.
Ref: 9router/cli/hooks/sqliteRuntime.js (pattern origin).
Exports SHARED_BOUNDARIES constant (ported from 9router cavemanPrompts.js)
that instructs the model to write normally for security warnings,
irreversible confirmations, and multi-step sequences, then resume terse
style. Appended to all 6 languages x 3 intensity levels. Polished full/ultra
English prompts with article-drop and short-synonym guidance.
Fixes alreadyApplied check order so SHARED_BOUNDARIES keywords in the
injected system prompt don't trigger a false-positive bypass on re-injection.
Tier 1 — LOCAL_ONLY: /api/mcp/ and /api/cli-tools/runtime/ are
restricted to loopback regardless of auth (prevents CVE-class exposure
of process-spawning endpoints via tunnels or LAN access).
Tier 2 — ALWAYS_PROTECTED: /api/shutdown and /api/settings/database
always require auth, even when requireLogin=false.
Tier 3 — MANAGEMENT: existing behaviour (auth bypassed when
requireLogin=false). IPv6 loopback [::1] correctly parsed.
Generates a deterministic HMAC-SHA256(key=rawMachineId, msg=salt) token
in src/lib/machineToken.ts. The management authz policy now accepts this
token via x-omniroute-cli-token when Host is loopback, letting the local
CLI process call management APIs without requiring a user login session.
`omniroute config token` prints the token for manual use.
When users configured `visionBridgeModel: "gemini/gemini-2.0-flash"` (or
any non-Anthropic prefix like `openrouter/...`, `google/...`), every
request failed with `Vision API error 401: You didn't provide an API
key` from OpenAI. The helper hardcoded `https://api.openai.com/v1` as
the base URL and `OPENAI_API_KEY` as the auth header for any model
that wasn't `anthropic/*`, so users without an OpenAI key (or who
wanted to use Gemini/OpenRouter/OmniRoute self-loop) had no path that
worked.
This change adds two env vars:
- VISION_BRIDGE_BASE_URL — alternate OpenAI-compatible base URL.
Priority: VISION_BRIDGE_BASE_URL → legacy OpenAI URL env →
api.openai.com (default).
- VISION_BRIDGE_API_KEY — alternate API key for that endpoint.
Priority: explicit caller arg → VISION_BRIDGE_API_KEY →
per-provider env (Anthropic/Google/OpenAI) → OpenAI fallback.
Anthropic models (anthropic/*) keep their dedicated `x-api-key` path
with the Anthropic env key unchanged — the override only affects the
OpenAI-compat branch, since the wire format differs.
Operators now have stable paths to:
- Route through OmniRoute itself (any registered model works):
VISION_BRIDGE_BASE_URL=http://localhost:20128/v1
VISION_BRIDGE_API_KEY=sk-<omniroute-key>
- Use Google's Gemini OpenAI-compat endpoint directly:
VISION_BRIDGE_BASE_URL=https://generativelanguage.googleapis.com/v1beta/openai
- Use OpenRouter directly:
VISION_BRIDGE_BASE_URL=https://openrouter.ai/api/v1
Reported by @kapustacool-lgtm. Documented in `.env.example` and
`docs/reference/ENVIRONMENT.md`. 11 unit tests cover env precedence
and the Anthropic-bypass guarantee.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Cross-platform system tray for `omniroute --tray`: Windows uses a
PowerShell NotifyIcon script (zero native binary, AV-safe); macOS/Linux
use systray2 lazy-installed at runtime into ~/.omniroute/runtime/.
Includes per-platform autostart (LaunchAgent / .desktop / registry) and
`omniroute config tray <enable|disable>` CLI command.
DISPLAY added to env-sync allowlist as an OS/X11 variable, not an
OmniRoute config variable.
Adds /skills/omniroute*/SKILL.md following the Anthropic skill manifest
spec (frontmatter name/description + self-contained body): chat, image,
tts, stt, embeddings, web-search, web-fetch, mcp, a2a + entry point.
External agents (Claude Desktop, ChatGPT, Cursor, Cline) can fetch one
raw GitHub URL to learn how to call OmniRoute — zero-friction onboarding.
OmniRoute-specific differentiators (mcp + a2a skills) extend the 9router
pattern with 2 extra manifests not present in the reference.
Adds structural lint test (tests/unit/docs/skillManifestsLint.test.ts)
enforcing frontmatter, env-var references, and trigger-phrase quality.
Adds "AI Agent Skills" section to README.md root.
Ref: 9router/skills/ pattern (adapted).
Adds compression engine that collapses repeated sibling lines (≥30 items
with same indent + role prefix) into head + summary + tail, preserves
[ref=eXX] anchors required by Playwright/computer-use MCPs, and
hard-truncates oversized text with a navigation hint footer.
Reduces token usage 60-80% on browser snapshots/accessibility trees from
external MCP servers (playwright-mcp, chrome-mcp). Configurable via
settings.compression.mcpAccessibility namespace.
Changes:
- open-sse/services/compression/engines/mcpAccessibility/: new engine
(constants.ts, collapseRepeated.ts, index.ts with smartFilterText)
- open-sse/services/compression/types.ts: re-exports McpAccessibilityConfig
- src/lib/db/compression.ts: getMcpAccessibilityConfig/setMcpAccessibilityConfig
- src/lib/db/migrations/056_mcp_accessibility_compression.sql: default settings
- open-sse/mcp-server/server.ts: apply filter to all tool result text blocks
- tests/unit/compression/mcpAccessibility.test.ts: 4 unit tests
- tests/unit/mcp/serverSmartFilter.test.ts: 4 integration tests (DB getter/setter)
Ref: 9router/src/lib/mcp/stdioSseBridge.js:14-90 (algorithm origin).
Five issues raised in the v3.8.0 release review, all release-blocking:
P1 — open-sse/services/tokenRefresh.ts
Read Windsurf Firebase API key from WINDSURF_CONFIG.firebaseApiKey
(resolvePublicCred wrapper) instead of process.env directly. Without
this, the literal removal from .env.example silently broke browser-flow
Windsurf/Devin token refresh.
P1 — open-sse/translator/request/openai-to-kiro.ts
Mark synthetic "(empty)" turns injected for assistant-first chats as
non-enumerable __synthetic and skip them when deriving conversationId
via uuidv5. Prevents unrelated chats from colliding on the same upstream
Kiro/AWS Builder ID context.
P2 — open-sse/utils/publicCreds.ts
Harden decodePublicCred against raw credential overrides outside
RAW_VALUE_PATTERN: strict-base64 alphabet check + printable-plain check
on the decoded result. Buffer.from(v, "base64") is lenient and was
silently mangling unrecognized raw values.
P2 — src/sse/services/auth.ts
Gate the x-api-key fallback on the anthropic-version header. Without
this scoping, local-mode requests with placeholder x-api-key from
non-Anthropic clients were rejected as Invalid API key even with
REQUIRE_API_KEY=false.
P2 — src/app/api/providers/[id]/test/route.ts
Move Qoder OAuth+PAT disambiguation BEFORE the CLI-runtime early-return
that was making the new message branch unreachable for the target
scenario from #2247.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Prevent raw exception messages from leaking stack frames or absolute
paths in the console logs and token health endpoints.
Also harden the i18n mirror move script by replacing shell-based git
commands with execFileSync and a safer fallback for untracked files.
Add regression coverage for clearing provider-scoped synced available models without affecting other providers.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Closes#2213
Conflict resolution: kept both images field (HEAD) and origin field (PR),
kept toolsAttached return value (HEAD) while incorporating all 8mbe
improvements (schema normalization, synthetic user guard, orphaned tool
results, alternating role enforcement). All 16 tests pass.
Blackbox's `/api/chat` now rejects requests whose `validated` field
doesn't match the frontend `tk` token (exported from app.blackbox.ai's
Next.js bundle), returning HTTP 403 even when the session cookie is
valid and the subscription is active. The previous executor sent a
random UUID, which works only until Blackbox enforces the check.
This change:
- Adds `resolveBlackboxValidatedToken()` that returns
`BLACKBOX_WEB_VALIDATED_TOKEN` when set, otherwise falls back to the
legacy random UUID (no regression for users who already work).
- Detects 403 responses whose body indicates a token-specific failure
("invalid validated token", "validation token", etc.) and replaces
the generic "cookie expired" message with explicit guidance to set
BLACKBOX_WEB_VALIDATED_TOKEN. The cookie-expired path is preserved
for non-token 401/403.
- Documents the env var in `.env.example` and
`docs/reference/ENVIRONMENT.md` (env-doc-sync check passes).
Deliberately NOT included: runtime scraping of Blackbox's Next.js
chunks to auto-extract `tk`. That coupling to their bundle hash would
silently break on every frontend deploy — the env override is the
stable path for operators who have already resolved the token.
Reported by @kazimshah39 with detailed root-cause analysis.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
When a Qoder connection lands in OAuth/CLI-flavored mode but the user
has pasted a Personal Access Token, the provider test route surfaces
"Local CLI runtime is not installed" plus a cascading 401 from
DashScope. Neither error tells the user "you picked the wrong auth
mode, switch to API Key".
The runtime check now detects this state (Qoder + non-apikey authType +
a token present on the connection or providerSpecificData) and surfaces
a single actionable message: "Qoder OAuth/Local CLI mode is selected
but the Qoder CLI is not detected. If you have a Personal Access Token,
switch this connection to API Key auth instead."
Non-Qoder providers and Qoder in real OAuth/CLI mode without a token
still get the original generic message.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
When OmniRoute translates Claude responses to the OpenAI format,
prompt_tokens was summing input + cache_read + cache_creation. Anthropic
pads short prompts up to a 1024-token minimum to create a cache, so:
Request: {"messages":[{"role":"user","content":"hi"}]}
Claude usage: input=8, cache_read=4, cache_creation=2000
Dashboard "Total In": 8
HTTP response prompt_tokens: 2008 (8 + 4 + 2000)
That 250x inflation broke downstream billing systems (Sub2API, NewAPI,
OneAPI) that trust prompt_tokens as the source of truth.
This change:
- prompt_tokens = input_tokens + cache_read_input_tokens (matches what
the dashboard reports as "Total In", preserves issue #1426's intent
that cache reads are billable input)
- cache_creation_tokens stays visible in
prompt_tokens_details.cache_creation_tokens for auditing — it just no
longer inflates the headline number
Reported by @downdawn with a complete root-cause trace.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Anthropic-native clients (Claude Code, @anthropic-ai/sdk) authenticate
via x-api-key per the Messages API contract. extractApiKey only read
Authorization: Bearer, so:
- usage_history.api_key_id was NULL for all x-api-key traffic (~50% of
real-world traffic invisible in Costs/Analytics)
- api_keys.last_used_at never updated for keys delivered via x-api-key
- per-key policies (allowedModels, budget, rateLimits, accessSchedule,
expiresAt) were bypassed for every Anthropic-native client
The fallback honors x-api-key (case-insensitive) when no Authorization:
Bearer is present. Bearer still wins when both are set, preserving
back-compat for clients that already send both. Reported by
@Forcerecon with a complete repro and validation plan.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Embed Gemini, Antigravity and Windsurf public OAuth/Firebase identifiers
(extracted from upstream CLI binaries) through a XOR-masked byte sequence
in open-sse/utils/publicCreds.ts instead of source literals, so pattern
scanners (GitHub Secret Scanning, Semgrep) stop raising false positives on
every release. decodePublicCred passes raw values through unchanged for
users who already have plaintext in their .env (no migration needed).
- New utils/publicCreds.ts with decode/encode + tests
- Replace 6 hardcoded Google client_id/secret in oauth.ts + providerRegistry
- Drop literals from .env.example (comment-only documentation)
- Sanitize error messages inside buildErrorBody so every caller (incl.
createErrorResult) is covered; cursor.ts now reuses the shared helper
- Cover the new helpers with unit tests (publicCreds + error sanitization)
Resolves the open code-scanning js/stack-trace-exposure findings and the
secret-scanning Google API Key alert without breaking existing setups.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
- Normalize tool schemas: strip additionalProperties and empty required arrays
- Merge consecutive assistant messages and adjacent user turns after role normalization
- Prepend synthetic user message when conversation starts with assistant
- Convert orphaned toolResults to inline text when assistant with toolUses is missing
- Enforce strictly alternating user/assistant roles in history
- Use deterministic uuidv5 conversationId based on first message for session caching
- Ensure origin field is present on all userInputMessage entries
After cherry-picking PR #2231, the function signature changed from
positional (provider, model) to object ({ provider, model }). Fixes the
2 pre-existing tests that still used the old positional style.
Fixes Anthropic HTTP 400 errors (~49/h on claude-opus-4-7) by preserving
the latest assistant message's thinking blocks verbatim instead of
rewriting them to redacted_thinking.
Co-authored-by: NomenAK <anton@nomenak.dev>
- antigravity: AntigravityCredentials.projectId widened to string|null
to match base ProviderCredentials shape post-#2227 squash merge.
- responses-handler: heartbeat assertion updated for #2233's new
openai-responses-in-progress shape (was: keepalive comment).
- search-registry: expected count is now 12 (ollama-search +
zai-search both landed in this release).
After merging PRs #2221 (ModelSync shared loopback readiness gate + IPv4 force)
and #2219 (Antigravity loadCodeAssist bootstrap + fetchAvailableModels fallback)
into release/v3.8.0, two test suites needed updates to match the new routing:
- tests/unit/model-sync-route.test.ts:
* resetStorage() now calls __resetLoopbackReadinessForTests() so the
module-level __loopbackReadyPromise cache does not leak between tests.
* Every fetch mock now answers the /__readiness_probe__/ URL with 404 so
the gate opens immediately (any HTTP response satisfies the probe).
* Self-fetch target URL assertions updated from http://localhost/...
to http://127.0.0.1:20128/... per PR #2221's IPv4-force.
- tests/unit/provider-models-route.test.ts:
* The Antigravity discovery-retry test now treats loadCodeAssist calls as
non-fatal failures so the discovery path is still exercised.
* The expected discovery URL sequence is updated to the new
fetchAvailableModels-first order introduced by PR #2219.
