Files
OmniRoute/open-sse/executors/cliproxyapi.ts
Diego Rodrigues de Sa e Souza f59f8daa94 Release v3.8.6 (#2804)
* fix(gemini): preserve structured tool calls for antigravity

* fix(gemini): parse prefixed textual tool calls

* fix(antigravity): preserve textual SSE tool calls

* fix(stream): normalize textual passthrough tool calls

* fix(stream): normalize split textual tool calls

* fix(stream): suppress malformed textual tool calls

* fix(stream): suppress compact malformed tool calls

* fix(stream): emit structured textual tool calls

* fix(stream): suppress unknown textual tool calls

* fix(stream): normalize responses textual tool calls

* chore: ignore .claude/settings.local.json (per-user Claude Code permissions)

* fix(opencode-go): route qwen3.x via claude messages + repair fixMissingToolResponses for Claude-shape upstreams (#2791)

Integrated into release/v3.8.6

* fix: resolve npm install warnings — remove dead deps, relax engine constraint (#2792)

Integrated into release/v3.8.6

* fix: register missing web-cookie validators (claude-web, gemini-web, copilot-web, t3-web) (#2793)

Integrated into release/v3.8.6

* fix: Error: Unable to inspect existing database #2771 (#2795)

Integrated into release/v3.8.6

* fix(oauth): repair Google loopback callback flow (#2796)

Integrated into release/v3.8.6

* feat(logs): add clean history button (#2799)

Integrated into release/v3.8.6

* [codex] home: restore settings-driven home layout and quota auto-refresh (#2800)

Integrated into release/v3.8.6

* fix(gemini): emit signaturelessToolCallMode:text for GEMINI format models (#2801)

Integrated into release/v3.8.6

* feat(modelSpecs): align opencode-go family with upstream provider limits (#2802)

Integrated into release/v3.8.6

* chore: apply unit test fixes, polyfills, and environment precedence fixes

* docs(agents): atualiza fluxos de release e triagem

Expande os workflows de release para incluir auditoria de segurança,
CHANGELOG completo por commits, quality gate obrigatório, homologação em
VPS local, publicação oficial, deploy em Akamai e validação de artefatos.

Reorganiza a triagem de features com arquivos permanentes por bucket,
suporte a itens em andamento, regra de reclaim após 15 dias e novo
tratamento para ideias viáveis catalogadas.

Corrige a orientação de revisão de discussões para usar a ordem
cronológica real dos comentários e respostas ao identificar a última
atividade.

* fix(lockout): classify Gemini Antigravity resource exhaustion as quota_exhausted

* fix(reasoning): gate replay by interleaved field

* docs(rule-16): permit human Co-authored-by, restrict only AI/bot trailers

Rule #16 previously banned all `Co-Authored-By` trailers absolutely.
That blocked the upstream-port workflows (`/port-upstream-features` and
`/port-upstream-issues`), which must credit human upstream PR authors
and issue reporters in OmniRoute commits.

Refine the rule to ban only AI/bot-attributed trailers (Claude, GPT,
Copilot, Bot; anthropic.com / openai.com / bot-owned noreply.github.com
emails) while allowing standard human `Co-authored-by: Name <email>`
attribution.

Sync the rule across the source CLAUDE.md, the E2E shakedown doc note,
and 41 i18n translations.

* fix(gitlawb): add specialty validators for connection test — bypass /models probe

GitLawB OpenGateway API (xiaomi-mimo compatible) does not expose a /models
endpoint, causing validateOpenAILikeProvider to 404 on the initial probe
and report 'Provider validation endpoint not supported'.

Add specialty validators for both gitlawb and gitlawb-gmi that follow the
same pattern as the existing xiaomi-mimo validator: skip GET /models,
validate directly via POST /chat/completions with a minimal test message.
Any 401/403 response means an invalid key; all other responses mean auth
is OK.

Fixes test-connection returning 404 for GitLawB providers.

* test(gitlawb): add 12 unit tests for gitlawb and gitlawb-gmi specialty validators

Covers success, auth failure (401/403), non-auth acceptance (400/422/429),
network errors, and custom baseUrl overrides for both providers.

* feat(gitlawb): serve models from static registry without API-unavailable warning

GitLawB's OpenGateway API does not expose a /models endpoint per
provider-path. Previously the models route fell through to the generic
fallback which returned static catalog models with the misleading
'API unavailable — using local catalog' warning.

Now gitlawb and gitlawb-gmi are handled as static model providers
(same pattern as reka and qwen OAuth) — models are served from the
provider registry without any warning, since all registered models
are functional via POST /chat/completions.

* refactor(gitlawb): extract shared opengateway validator factory, fix docs path in test

- Extract gitlawb/gitlawb-gmi validators into buildOpengatewayValidator factory
- Fix dockerignore-docs-coverage test: update stale docs/AUTO-COMBO.md -> docs/routing/AUTO-COMBO.md

* fix(reasoning): guard interleaved capability lookup

* feat(gitlawb): dynamic model fetch with gmi-cloud fallback

Hybrid approach:
- gitlawb (xiaomi-mimo): dynamic /models endpoint → 356 models
- gitlawb-gmi (gmi-cloud): 404 fallback → local catalog gracefully
Mimics Gitlawb/openclaude's model-routing pattern

* i18n(pt-BR): complete missing translations and sync with en.json

* feat(build): nix multi-OS package manager install (#2806)

Integrated into release/v3.8.6

* fix(i18n): translate 144 new __MISSING__ pt-BR strings (#2816)

Integrated into release/v3.8.6

* chore(docs): set coverage gate to 40/40/40/40 in CLAUDE.md

Aligns the documented coverage gate with the v3.8.6 release decision
(lowered from 75/75/75/70). Matches the threshold already set in
package.json by the large feature PRs (planos 11-22).

* fix(cli): respect PORT env var in serve command (#2845)

Integrated into release/v3.8.6.

* fix(deepseek-web): return 400 when client sends tools[] - chat.deepseek.com has no tool support (#2854)

Integrated into release/v3.8.6.

* fix(qoder): reject invalid/expired PATs returning Cosy 500 error (#2860)

Integrated into release/v3.8.6.

* fix(cli): register openclaw in tool-detector (#2833) (#2850)

Integrated into release/v3.8.6.

* fix(api): include noAuth providers in /v1/models catalog (#2798) (#2814)

Integrated into release/v3.8.6.

* fix(combo): resolve custom provider targets via combo name (#2778) (#2812)

Integrated into release/v3.8.6.

* fix(translator): strip safety_identifier in openai-responses cleanup (#2770) (#2809)

Integrated into release/v3.8.6.

* fix(quota): honor explicit per-connection preflight opt-out (#2831) (#2844)

Integrated into release/v3.8.6.

* fix(usage): un-invert GitHub Copilot Free/limited quota — limited_user_quotas is remaining (#2876) (#2881)

Integrated into release/v3.8.6.

* fix(nous-research): correct baseUrl to include /chat/completions (#2826) (#2835)

Integrated into release/v3.8.6.

* fix(opencode): qwen3.x max/plus models lack vision support (#2822) (#2836)

Integrated into release/v3.8.6.

* fix(translator): pass-through tool_search built-in tool type (#2766) (#2811)

Integrated into release/v3.8.6.

* fix(github): route claude-opus-4.6 via chat completions (#2821)

Integrated into release/v3.8.6.

* docs(oauth): add Windsurf login fix design (Phase 1 hotfix + Phase 2 Firebase OAuth)

Two-phase plan to fix the broken Windsurf OAuth flow:
- Phase 1: drop the dead app.devin.ai/editor/signin PKCE path, promote
  import-token from windsurf.com/show-auth-token as the primary path
- Phase 2: port Firebase OAuth + RegisterUser flow from
  fendoushaonian/WindSurf-gRPC-API for full browser-based automation

Spec only - no code changes yet.

* docs(plan): Phase 1 windsurf login hotfix implementation plan

10 tasks covering:
- TDD assertions for flowType + 410 Gone responses
- Provider switch to import_token
- Route handler retiring authorize/start-callback-server/poll-callback
- OAuthModal UI override
- i18n sync
- Verification + PR steps

* fix(cli): replace cli-table3 with hand-rolled formatter (#2752) (#2813)

Integrated into release/v3.8.6.

* fix(skills): skip interception for unregistered client-native tools (#2815) (#2817)

Integrated into release/v3.8.6.

* feat(sse): add RTK filters for kubectl, docker-build, composer, gh (#2824)

Integrated into release/v3.8.6.

* fix(geminiHelper): support rec.image content shape + warn on dropped remote URLs (refs #2807) (#2855)

Integrated into release/v3.8.6.

* fix(cli): allow nullable/optional apiKey in cliMitmStartSchema (#2857)

Integrated into release/v3.8.6.

* fix(combo): preserve system messages during context handoff summary generation (#2865)

Integrated into release/v3.8.6.

* fix: wire CLIProxyAPI fallback settings into chatCore routing engine (#2866)

Integrated into release/v3.8.6.

* fix(usage): add opencode quota fetcher (#2852) (#2867)

Integrated into release/v3.8.6.

* feat(claude): default xhigh support for newer Opus models (#2874)

Integrated into release/v3.8.6.

* fix(cli): restore omniroute logs command stream (#2756) (#2810)

Integrated into release/v3.8.6.

* fix(combo): normalize upstream Headers for Node 24 undici interop (#2751) (#2823)

Integrated into release/v3.8.6.

* Rename proxy log Public IP to Client IP (#2880)

Integrated into release/v3.8.6.

* fix(claude): preserve max effort for supported models (#2875)

Integrated into release/v3.8.6.

* fix(oauth): switch windsurf provider to import_token flow

The PKCE auth URL targeting app.devin.ai/editor/signin returns 404
post-rebrand. Until Phase 2 ports Firebase OAuth + RegisterUser, the
only supported path is import-token via windsurf.com/show-auth-token.

- windsurf.ts: drop buildAuthUrl, set flowType=import_token
- generateAuthData returns supported:false + helpful error for windsurf/devin-cli
- tests: assert flowType + disabled stub

* fix(oauth): return 410 Gone for retired windsurf/devin-cli PKCE actions

start-callback-server, authorize, and poll-callback (GET + POST) now
return 410 Gone with a pointer to /import-token. The 410 short-circuit
runs before auth so the response is honest about the action being
permanently gone, not gated. Codex PKCE flow unchanged.

Tests: 5 new assertions cover GET + POST 410 paths and a Codex
regression check.

* refactor(oauth): annotate retired PKCE fields in WINDSURF_CONFIG

No behaviour change - comment-only update documenting that authorizeUrl,
codeChallengeMethod, callbackPort, callbackPath, apiServerUrl, and
exchangePath are no longer consumed. Active fields (inferenceUrl,
showAuthTokenUrl, firebaseApiKey, ideName) called out separately.

* fix(cli,docs): use requireCliToolsAuth in logs route + document OPENCODE quota env

Post-merge contract fixes for v3.8.6:
- src/app/api/cli-tools/logs/route.ts (#2810) now uses the shared
  requireCliToolsAuth guard (param renamed req->request) to satisfy the
  cli-tools-auth-hardening contract test.
- Document OMNIROUTE_OPENCODE_QUOTA_URL (#2867) in docs/reference/ENVIRONMENT.md
  to satisfy the env/docs sync contract.

* fix(dashboard): force import-token panel for windsurf/devin-cli

Phase 1 hotfix: hide the 'Browser Login' tab and start in Paste API Key
mode. Removes windsurf/devin-cli from PKCE_CALLBACK_SERVER_PROVIDERS so
no callback server is started for them. Codex still uses the PKCE flow.

The 'Get token' link continues to point at windsurf.com/show-auth-token
via the existing supportsTokenPaste form copy.

* fix(oauth): windsurf import-token mapTokens signature mismatch

The route at `src/app/api/oauth/[provider]/[action]/route.ts` invokes
`providerData.mapTokens({ accessToken: token })` (object), matching the
cursor/kiro signature. The windsurf provider was declared with
`mapTokens(token: string)` instead, so the entire object was stored as
`accessToken`. When the connection record reached the SQLite layer it
crashed with:

  SQLite3 can only bind numbers, strings, bigints, buffers, and null

Fix by aligning windsurf's `mapTokens` signature with the route caller
and the cursor/kiro convention. Also dedupe a copy-pasted second
`if (action === "import-token")` block in the route handler — the
second block was unreachable but identical to the first.

Adds two regression tests asserting that
`provider.mapTokens({ accessToken })` returns a string `accessToken` for
both windsurf and devin-cli, so a future signature drift trips the gate
instead of the SQLite bind error in production.

* feat(compression): expand pt-BR pack with troglodita rules (15 → 49) (#2818)

Integrated into release/v3.8.6

* fix(sse): repair RTK engine defaults so dedup and direct calls work (#2825)

Integrated into release/v3.8.6

* fix(mcp): redirect console.log/warn to stderr in --mcp stdio mode (#2840)

Integrated into release/v3.8.6

* fix(gemini-cli): prefer real project IDs over default-project (#2841)

Integrated into release/v3.8.6

* fix(opencode-go): add provider limits quota fetcher (#2861)

Integrated into release/v3.8.6

* Audit & add web cookie providers: fix 4 missing registry entries + DuckDuckGo (#2862)

Integrated into release/v3.8.6

* fix(antigravity): harden signatureless tool history (#2878)

Integrated into release/v3.8.6

* fix: provider model sync pruning and dynamic antigravity MITM proxy mappings (#2886)

Integrated into release/v3.8.6

* feat(usage): per-API-key token limits scoped to model/provider/global (#2888)

Integrated into release/v3.8.6

* fix(audio): build multipart body manually to preserve Content-Type (#2842)

Integrated into release/v3.8.6

* refactor: remove agent skill documentation files and streamline maintenance workflows

* test(stabilization): resolve unit test failures in blackbox-web, schema-coercion, translator-helper-branches, usage-service-hardening, and audio-transcription

* fix(security): mitigate Socket.dev supply-chain findings + secrets opt-in + minimal build profile (#2863) (#2871)

Two real security gaps closed and four cosmetic Socket.dev fingerprints removed.
See docs/security/SOCKET_DEV_FINDINGS.md for the per-finding maintainer
attestation.

Real bugs fixed:
- cloudSync: HMAC verification of `X-Cloud-Sig` + opt-in
  `OMNIROUTE_CLOUD_SYNC_SECRETS=true` before overwriting `accessToken` /
  `refreshToken` / `providerSpecificData` from a remote response. Closes the
  silent-credential-swap surface (a misconfigured or hostile CLOUD_URL could
  previously replace local tokens unverified).
- Zed import: split into 2-step `/discover` + `/import` flow. `/import` now
  requires `confirmedAccounts: [{ service, account, fingerprint }]` and
  re-reads the keychain server-side to filter by fingerprint, so a tampered
  discover response cannot trick the endpoint into saving an unrelated token.

Cosmetic Socket.dev mitigations:
- runElevatedPowerShell writes the elevated payload to a per-call temp `.ps1`
  file (mode 0o600) and references it via `-File`. Removes the textbook
  `-EncodedCommand <base64utf16le>` pattern flagged as malware by Socket's AI
  classifier.
- Maintainer attestation `SECURITY-AUDITOR-NOTE:` blocks added at every
  flagged call site pointing to `docs/security/SOCKET_DEV_FINDINGS.md`.

Build-time hardening:
- `OMNIROUTE_BUILD_PROFILE=minimal` (`npm run build:secure`) physically
  removes the four sensitive modules from the standalone bundle via webpack
  `NormalModuleReplacementPlugin`. Stubs throw `FeatureDisabledError` at
  runtime. Intended for the `omniroute-secure` artifact.

Tests:
- 24 new unit tests in `tests/unit/security/` covering the wrapper builder,
  HMAC verification (4 cases), credential fingerprint determinism (5 cases),
  confirmedAccounts validation + fingerprint filtering (6 cases), and the
  minimal-build stubs (5 cases).

Docs:
- New `docs/security/SOCKET_DEV_FINDINGS.md` — per-finding attestation.
- New `socket.yml` — Socket.dev v2 config pointing at the attestation.
- Updated `SECURITY.md` — supply-chain scanner section.
- Updated `.env.example` — three new env vars documented.

Backwards compatibility:
- Cloud sync token overwrite is OFF by default. Users who relied on
  it must set `OMNIROUTE_CLOUD_SYNC_SECRETS=true`. Breaking change documented
  in CHANGELOG.
- Zed import 2-step is the new default; legacy 1-step preserved behind
  `OMNIROUTE_ZED_IMPORT_LEGACY_ONE_STEP=true` and will be removed in v3.9.

Closes #2863

* fix(security): redact public Firebase Web key from windsurf spec; doc SHA-256 cache-key rationale (#2894)

Two security-scanning findings on release/v3.8.6:

- Secret-scanning alert 7 (google_api_key): the windsurf login-fix design spec
  embedded the literal public Firebase Web API key on two lines. Firebase Web
  API keys are non-sensitive by design (they identify the project; access is
  gated by Firebase Security Rules + key restrictions), but the literal trips
  secret scanning. Redacted to a placeholder; the embedded default still goes
  through resolvePublicCred per rule #11.

- Code-scanning alert 261 (js/insufficient-password-hash): tokenCacheKey() uses
  SHA-256 to derive an in-memory cache key from the session token, not for
  password-at-rest storage. Added a comment documenting why CWE-916 KDFs do not
  apply (false positive).

* fix(ci): resolve release/v3.8.6 gate failures (docs-sync, any-budget, pack-artifact) (#2895)

* fix(ci): resolve release/v3.8.6 gate failures (docs-sync, any-budget, pack-artifact)

Three CI gates failed on release/v3.8.6 (run 26630300877):

- docs-sync: CHANGELOG had a spurious "## [3.8.6-patch]" section above
  "## [3.8.6]", so the latest release no longer matched package.json (3.8.6)
  and the 41 i18n CHANGELOG mirrors were flagged as missing that section.
  Fold the lone #2752 entry into [3.8.6] and drop the patch heading.
- any-budget:t11: open-sse/handlers/chatCore.ts regressed to 1 explicit `any`
  (budget 0). Type the persist callback arg as Record<string, unknown>, which
  matches runWithOnPersist's RefreshPersistFn contract exactly.
- pack-artifact: open-sse/utils/setupPolyfill.ts ships via package.json "files"
  (bin/omniroute.mjs imports it at startup) but was missing from the pack
  policy allowlist. Allow it and add a regression test.

* fix(security): redact public Firebase Web key from windsurf spec

Redact the literal public Firebase Web API key (secret-scanning #7) to a
placeholder, mirroring the redaction on release/v3.8.6 (PR #2894) and the
windsurf fix branch. Non-sensitive public Web key; trips secret scanning.

* feat(combo): Zero-Latency Combos (Hedging, Proactive Compression, Predictive TTFT) (#2868)

* feat(combo): implement zero-latency combo optimizations (hedging, proactive compression, predictive TTFT)

* fix(combo): fix predictive TTFT skip logic and unhandled promise rejections

---------

Co-authored-by: Automation <automation@omniroute>

* feat: implement automated skill workflows and update system configuration and validation schemas

* test: eliminate dynamic cast warnings in cloud-sync unit test

* test: isolate services-branch-hardening database directory to avoid concurrency issues

* feat(providers): add 7 new web-cookie providers + research catalog + discovery tool

New providers:
- huggingchat: free LLM chat via huggingface.co/chat (no subscription)
- phind: free dev-focused AI chat via phind.com/api/agent
- poe-web: multi-model chat via poe.com GraphQL (p-b cookie)
- venice-web: privacy-focused AI chat via venice.ai (session cookie)
- v0-vercel-web: Vercel v0 code gen via v0.dev (session cookie)
- kimi-web: Moonshot Kimi chat via kimi.moonshot.cn (session cookie)
- doubao-web: ByteDance Doubao chat via doubao.com (session cookie)

Additional:
- Research catalog: docs/research/UNLIMITED_LLM_ACCESS.md
- Discovery tool design + stub: src/lib/discovery/ + migration 073
- Unit tests: 33 tests for all 7 providers
- Shared helpers consolidated in error.ts (slop cleanup)
- All registered in WEB_COOKIE_PROVIDERS + providerRegistry + webSessionCredentials

Closes #2885

* fix(typecheck): resolve typecheck errors in combo spec and compression modules

* feat(api,oauth): add `agy` (Antigravity CLI) standalone provider with CLI token import (#2899)

Add a standalone OAuth provider `agy` (Antigravity CLI) next to gemini-cli/antigravity.
It reuses the antigravity inference backend (identical Google client_id +
daily-cloudcode-pa.googleapis.com endpoint, executor and token-refresh) but ships its own
model catalog — including the Claude models the backend exposes (claude-opus-4-6-thinking,
claude-sonnet-4-6) — its own account pool, and four ways to connect:

- token-file import (paste/upload the agy oauth token JSON)
- auto-detect a local CLI login (~/.gemini/antigravity-cli/antigravity-oauth-token)
- browser OAuth (via the shared OAuthModal Google loopback flow)
- bulk / ZIP import

New routes: POST /api/providers/agy-auth/{import,import-bulk,zip-extract,apply-local}.
Catalog pinned from the live :fetchAvailableModels endpoint. Docs (openapi.yaml,
ENVIRONMENT.md, .env.example, CHANGELOG) updated; new unit tests for registration,
the token parser, and route auth-hardening.

* fix(security): redact public Firebase Web key from windsurf spec (#2896)

Redact the literal public Firebase Web API key (secret-scanning #7) to a
placeholder. Firebase Web API keys are non-sensitive by design but the literal
trips GitHub secret scanning. Mirrors the redaction landed on release/v3.8.6
(PR #2894). Embedded default still flows through resolvePublicCred (rule #11).

* Pr 2871 (#2897)

* fix(security): mitigate Socket.dev supply-chain findings + secrets opt-in + minimal build profile (#2863)

Two real security gaps closed and four cosmetic Socket.dev fingerprints removed.
See docs/security/SOCKET_DEV_FINDINGS.md for the per-finding maintainer
attestation.

Real bugs fixed:
- cloudSync: HMAC verification of `X-Cloud-Sig` + opt-in
  `OMNIROUTE_CLOUD_SYNC_SECRETS=true` before overwriting `accessToken` /
  `refreshToken` / `providerSpecificData` from a remote response. Closes the
  silent-credential-swap surface (a misconfigured or hostile CLOUD_URL could
  previously replace local tokens unverified).
- Zed import: split into 2-step `/discover` + `/import` flow. `/import` now
  requires `confirmedAccounts: [{ service, account, fingerprint }]` and
  re-reads the keychain server-side to filter by fingerprint, so a tampered
  discover response cannot trick the endpoint into saving an unrelated token.

Cosmetic Socket.dev mitigations:
- runElevatedPowerShell writes the elevated payload to a per-call temp `.ps1`
  file (mode 0o600) and references it via `-File`. Removes the textbook
  `-EncodedCommand <base64utf16le>` pattern flagged as malware by Socket's AI
  classifier.
- Maintainer attestation `SECURITY-AUDITOR-NOTE:` blocks added at every
  flagged call site pointing to `docs/security/SOCKET_DEV_FINDINGS.md`.

Build-time hardening:
- `OMNIROUTE_BUILD_PROFILE=minimal` (`npm run build:secure`) physically
  removes the four sensitive modules from the standalone bundle via webpack
  `NormalModuleReplacementPlugin`. Stubs throw `FeatureDisabledError` at
  runtime. Intended for the `omniroute-secure` artifact.

Tests:
- 24 new unit tests in `tests/unit/security/` covering the wrapper builder,
  HMAC verification (4 cases), credential fingerprint determinism (5 cases),
  confirmedAccounts validation + fingerprint filtering (6 cases), and the
  minimal-build stubs (5 cases).

Docs:
- New `docs/security/SOCKET_DEV_FINDINGS.md` — per-finding attestation.
- New `socket.yml` — Socket.dev v2 config pointing at the attestation.
- Updated `SECURITY.md` — supply-chain scanner section.
- Updated `.env.example` — three new env vars documented.

Backwards compatibility:
- Cloud sync token overwrite is OFF by default. Users who relied on
  it must set `OMNIROUTE_CLOUD_SYNC_SECRETS=true`. Breaking change documented
  in CHANGELOG.
- Zed import 2-step is the new default; legacy 1-step preserved behind
  `OMNIROUTE_ZED_IMPORT_LEGACY_ONE_STEP=true` and will be removed in v3.9.

Closes #2863

* feat: implement automated skill workflows and update system configuration and validation schemas

* test: eliminate dynamic cast warnings in cloud-sync unit test

* test: isolate services-branch-hardening database directory to avoid concurrency issues

* chore(docs): refresh generated docs collection index

Update the generated Fumadocs browser collection mapping to keep
documentation imports in sync with the current docs structure.

* docs: update generated browser docs collection manifest

Refresh the generated Fumadocs browser collection mapping so the docs site can resolve the current documentation files correctly.

---------

Co-authored-by: OpenClaw <openclaw@kuzhomesrv.local>
Co-authored-by: Dmitry Kuznetsov <139351986+dmitry@users.noreply.local>
Co-authored-by: KuzyaBot <kuzya@local>
Co-authored-by: JeferssonLemes <jeferssondev@gmail.com>
Co-authored-by: Paijo <14921983+oyi77@users.noreply.github.com>
Co-authored-by: Markus Hartung <mail@hartmark.se>
Co-authored-by: akarray <akarray@users.noreply.github.com>
Co-authored-by: Apostol Apostolov <theapoapostolov@gmail.com>
Co-authored-by: Hernan Javier Ardila Sanchez <hjasgr@gmail.com>
Co-authored-by: Dmitry Kuznetsov <dmitry@kuznetsov.me>
Co-authored-by: Nikolay Alafuzov <alafuzov_nn@rusklimat.ru>
Co-authored-by: oyi77 <oyi77@users.noreply.github.com>
Co-authored-by: Ronaldo Davi <alltomatos@users.noreply.github.com>
Co-authored-by: levonk <277861+levonk@users.noreply.github.com>
Co-authored-by: Lenine Júnior <lenine@engrene.com.br>
Co-authored-by: Annas Alghoffar <aag.annas@gmail.com>
Co-authored-by: Tushar Agarwal <76201310+Tushar49@users.noreply.github.com>
Co-authored-by: GreatLiu <eurasiaxz@qq.com>
Co-authored-by: yuna amelia <230527278+yunaamelia@users.noreply.github.com>
Co-authored-by: Randi <55005611+rdself@users.noreply.github.com>
Co-authored-by: Container <78986709+disonjer@users.noreply.github.com>
Co-authored-by: nickwizard <35692452+nickwizard@users.noreply.github.com>
Co-authored-by: Rajvardhan Patil <rajvardhanpatil7890@gmail.com>
Co-authored-by: Raxxoor <manker_lol@hotmail.com>
Co-authored-by: Muhammad Mugni Hadi <mugnimaestra3@gmail.com>
Co-authored-by: mi <123757457+soyelmismo@users.noreply.github.com>
Co-authored-by: Automation <automation@omniroute>
2026-05-29 12:44:29 -03:00

434 lines
17 KiB
TypeScript

/**
* CLIProxyAPI Executor — routes requests to a local CLIProxyAPI instance.
*
* Always uses the OpenAI-compatible /v1/chat/completions endpoint. CLIProxyAPI
* internally detects Claude models and routes them through its Claude executor
* with full emulation (CCH signing, billing header, system prompt, uTLS,
* multi-account rotation, device profile learning, etc.).
*
* The UI toggle (cliproxyapiMode in providerSpecificData) controls WHETHER
* to use CLIProxyAPI as the backend, not the wire format. Response format
* is always OpenAI-compatible, so chatCore's SSE parsing works unchanged.
*
* Activation:
* 1. Per-provider upstream_proxy_config (mode=cliproxyapi or fallback)
* 2. Per-connection cliproxyapiMode toggle in providerSpecificData (UI)
*/
import {
BaseExecutor,
mergeUpstreamExtraHeaders,
mergeAbortSignals,
type ProviderCredentials,
} from "./base.ts";
import { HTTP_STATUS, FETCH_TIMEOUT_MS } from "../config/constants.ts";
const DEFAULT_PORT = 8317;
const DEFAULT_HOST = "127.0.0.1";
const HEALTH_CHECK_TIMEOUT_MS = 5000;
// Anthropic's reserved tool-name namespace: ^mcp_[^_].* triggers their
// server-side MCP connector billing gate, returning a misleading
// "out of extra usage" 400. Two-underscore (mcp__X) and capitalized
// (Mcp_X) variants pass cleanly.
const MCP_RESERVED_PREFIX_RE = /^mcp_(?=[^_])/;
function rewriteMcpToolName(name: string): string | null {
if (typeof name !== "string" || !MCP_RESERVED_PREFIX_RE.test(name)) return null;
return "M" + name.slice(1); // mcp_call → Mcp_call
}
/**
* Rewrite ^mcp_[^_] tool names on a body destined for Anthropic's
* /v1/messages. Returns a reverse map (rewritten → original) that the SSE
* response stream uses to restore the client's original names on tool_use
* blocks coming back.
*
* Non-mutating: replaces nested array elements with cloned objects rather
* than mutating in place, so the caller's input body is not affected. The
* outer `body` reference itself is the cloned `transformed` object from
* `transformRequest` — we mutate its top-level `tools`, `messages`, and
* `tool_choice` properties to point at the new clones.
*/
function applyMcpToolNameRewrite(body: Record<string, unknown>): Map<string, string> {
const reverseMap = new Map<string, string>();
const remember = (original: string, rewritten: string) => {
reverseMap.set(rewritten, original);
};
const tools = body.tools;
if (Array.isArray(tools)) {
body.tools = tools.map((tool) => {
if (!tool || typeof tool !== "object") return tool;
const t = tool as Record<string, unknown>;
const original = typeof t.name === "string" ? t.name : "";
const rewritten = rewriteMcpToolName(original);
if (rewritten) {
remember(original, rewritten);
return { ...t, name: rewritten };
}
return tool;
});
}
const messages = body.messages;
if (Array.isArray(messages)) {
body.messages = messages.map((msg) => {
if (!msg || typeof msg !== "object") return msg;
const m = msg as Record<string, unknown>;
const content = m.content;
if (!Array.isArray(content)) return msg;
let mutated = false;
const newContent = content.map((block) => {
if (!block || typeof block !== "object") return block;
const b = block as Record<string, unknown>;
if (b.type !== "tool_use") return block;
const original = typeof b.name === "string" ? b.name : "";
const rewritten = rewriteMcpToolName(original);
if (rewritten) {
mutated = true;
remember(original, rewritten);
return { ...b, name: rewritten };
}
return block;
});
return mutated ? { ...m, content: newContent } : msg;
});
}
const toolChoice = body.tool_choice;
if (toolChoice && typeof toolChoice === "object") {
const tc = toolChoice as Record<string, unknown>;
if (tc.type === "tool" && typeof tc.name === "string") {
const rewritten = rewriteMcpToolName(tc.name);
if (rewritten) {
const original = tc.name;
body.tool_choice = { ...tc, name: rewritten };
remember(original, rewritten);
}
}
}
return reverseMap;
}
// Cached URL from settings (loaded once, invalidated on settings change via clearCliproxyapiUrlCache)
let _cachedSettingsUrl: { url: string; ts: number } | null = null;
const URL_CACHE_TTL_MS = 60_000;
export function clearCliproxyapiUrlCache() {
_cachedSettingsUrl = null;
}
// Pre-load settings URL at module init so the sync path has a cache hit.
// This runs once when the executor module is first imported.
(async () => {
try {
const { getSettings } = await import("@/lib/db/settings");
const settings = await getSettings();
if (typeof settings.cliproxyapi_url === "string" && settings.cliproxyapi_url.trim()) {
_cachedSettingsUrl = { url: settings.cliproxyapi_url.trim(), ts: Date.now() };
}
} catch { /* env vars will be used as fallback */ }
})();
/**
* Resolve CLIProxyAPI base URL. Priority:
* 1. Settings table `cliproxyapi_url` (set via UI)
* 2. Environment variables CLIPROXYAPI_HOST / CLIPROXYAPI_PORT
* 3. Defaults (127.0.0.1:8317)
*/
async function resolveCliproxyapiBaseUrl(): Promise<string> {
// Check settings cache first
if (_cachedSettingsUrl && Date.now() - _cachedSettingsUrl.ts < URL_CACHE_TTL_MS) {
return _cachedSettingsUrl.url;
}
try {
const { getSettings } = await import("@/lib/db/settings");
const settings = await getSettings();
if (typeof settings.cliproxyapi_url === "string" && settings.cliproxyapi_url.trim()) {
const url = settings.cliproxyapi_url.trim();
_cachedSettingsUrl = { url, ts: Date.now() };
return url;
}
} catch { /* fall through to env vars */ }
const host = process.env.CLIPROXYAPI_HOST || DEFAULT_HOST;
const port = parseInt(process.env.CLIPROXYAPI_PORT || String(DEFAULT_PORT), 10);
const url = `http://${host}:${port}`;
_cachedSettingsUrl = { url, ts: Date.now() };
return url;
}
// Sync wrapper for backward compatibility (health checks, tests)
function resolveCliproxyapiBaseUrlSync(): string {
if (_cachedSettingsUrl && Date.now() - _cachedSettingsUrl.ts < URL_CACHE_TTL_MS) {
return _cachedSettingsUrl.url;
}
const host = process.env.CLIPROXYAPI_HOST || DEFAULT_HOST;
const port = parseInt(process.env.CLIPROXYAPI_PORT || String(DEFAULT_PORT), 10);
return `http://${host}:${port}`;
}
export { resolveCliproxyapiBaseUrl };
/**
* Check if a connection has CLIProxyAPI deep mode enabled via UI toggle.
* Used by chatCore's resolveExecutorWithProxy to decide routing.
*/
export function isCliproxyapiDeepModeEnabled(
providerSpecificData?: Record<string, unknown> | null
): boolean {
return providerSpecificData?.cliproxyapiMode === "claude-native";
}
export class CliproxyapiExecutor extends BaseExecutor {
private readonly upstreamBaseUrl: string;
constructor(baseUrl?: string) {
const effectiveBase = baseUrl ?? resolveCliproxyapiBaseUrlSync();
super("cliproxyapi", {
id: "cliproxyapi",
baseUrl: effectiveBase + "/v1/chat/completions",
headers: { "Content-Type": "application/json" },
});
this.upstreamBaseUrl = effectiveBase;
}
buildUrl(
_model: string,
_stream: boolean,
_urlIndex = 0,
_credentials: ProviderCredentials | null = null
): string {
// Default endpoint when called without body context (kept for back-compat).
// execute() picks the right endpoint from the body shape; see selectEndpoint().
return `${this.upstreamBaseUrl}/v1/chat/completions`;
}
/**
* Returns true when the body matches the Anthropic Messages wire shape.
*
* chatCore detects target=claude when the request comes from a Claude-source
* client (`/v1/messages`, Anthropic-version header, claude/* model). In that
* case no openai translation is applied and the executor sees the original
* Anthropic body: top-level `system` as an array of content blocks, and
* `messages[].content` as arrays. Routing those bodies to CPA's
* /v1/chat/completions causes CPA to emit OpenAI-style SSE chunks, which
* Anthropic SDK clients (Capy, claude-cli, etc.) cannot parse — the result
* looks like a 200 server-side with "0 chunks received" client-side.
*
* CPA exposes /v1/messages natively (claude executor with uTLS spoof,
* billing header, CCH signing, etc.) and emits proper Anthropic SSE:
* `event: message_start`, `content_block_delta`, etc.
*/
private isAnthropicShape(body: unknown): boolean {
if (!body || typeof body !== "object") return false;
const b = body as Record<string, unknown>;
// Strong signal: top-level `system` field is unique to the Anthropic
// Messages API. OpenAI Chat Completions encodes system as a role:"system"
// entry inside messages[], not at body level. Accept both string and
// array-of-content-blocks forms (Anthropic supports both per the docs).
if (b.system !== undefined) return true;
// Strong signal: top-level `thinking` field is Anthropic-only. OpenAI
// uses `reasoning` / `reasoning_effort`. Even adaptive/raw Capy shapes
// emit thinking, so this catches minimal Capy bodies (string content,
// no system block) that would otherwise miss the messages[0].content
// array check below.
if (b.thinking !== undefined) return true;
// Strong signal: top-level `metadata.user_id` is the CC wire-image
// identifier; OpenAI request bodies don't carry it.
if (
b.metadata &&
typeof b.metadata === "object" &&
(b.metadata as Record<string, unknown>).user_id !== undefined
)
return true;
// Strong signal: messages[0].content is an array of Anthropic content blocks
const msgs = b.messages;
if (Array.isArray(msgs) && msgs.length > 0) {
const first = msgs[0] as Record<string, unknown>;
if (Array.isArray(first?.content)) return true;
}
return false;
}
private selectEndpoint(body: unknown): string {
return this.isAnthropicShape(body) ? "/v1/messages" : "/v1/chat/completions";
}
buildHeaders(credentials: ProviderCredentials | null, stream = true): Record<string, string> {
const key = credentials?.apiKey || credentials?.accessToken;
const headers: Record<string, string> = {
"Content-Type": "application/json",
};
if (key) {
headers["Authorization"] = `Bearer ${key}`;
}
if (stream) {
headers["Accept"] = "text/event-stream";
}
return headers;
}
transformRequest(
model: string,
body: unknown,
_stream: boolean,
_credentials: ProviderCredentials | null
): unknown {
if (!body || typeof body !== "object") return body;
const transformed = { ...(body as Record<string, unknown>) };
if (transformed.model !== model) {
transformed.model = model;
}
// For Anthropic-shape bodies routed to CPA's /v1/messages, strip the
// Capy/Anthropic-SDK premium extras that Anthropic gates with
// "Extra usage is required" / "out of extra usage" (400). CPA does its
// own Claude Code wire-image cloak (CCH, billing header, uTLS, metadata
// user_id, system sentinel) downstream — but it forwards client extras
// like output_config.effort=xhigh which trigger the extras-billing gate.
//
// Mirrors the runtime "Patch I2/I4" effect previously applied via patch.mjs.
// Strips are no-op when fields are absent (OpenAI-shape passthrough).
if (this.isAnthropicShape(transformed)) {
delete transformed.output_config;
delete transformed.context_management;
delete transformed.client_info;
delete transformed.prompt_cache_key;
delete transformed.safety_identifier;
delete transformed.metadata;
// Conditional thinking strip: preserve Anthropic-valid shapes
// ({type:"enabled"|"disabled", budget_tokens:N}) that applyThinkingBudget
// already normalized. Strip non-Anthropic shapes (e.g. Capy's
// {type:"adaptive", display:"summarized"}) which trigger Anthropic 400
// "Extra usage required" / "out of extra usage". The `display` field is
// a Capy-specific hint Anthropic doesn't accept.
const thinking = transformed.thinking;
if (thinking && typeof thinking === "object") {
const t = thinking as Record<string, unknown>;
const validType = t.type === "enabled" || t.type === "disabled";
const hasValidBudget = typeof t.budget_tokens === "number" && t.budget_tokens >= 0;
const hasInvalidExtras = "display" in t;
if (!validType || !hasValidBudget || hasInvalidExtras) {
delete transformed.thinking;
}
}
// Rewrite tool names matching Anthropic's reserved ^mcp_[^_] namespace.
// Anthropic returns "out of extra usage" / "Extra usage required" 400
// when a client-declared tool name collides with their server-side MCP
// connector tools. Bisected character-by-character against the real
// Anthropic API via CPA (uTLS spoof, Claude OAuth):
// mcp_call, mcp_query, mcp_x, mcp_test → 400 (gate hit)
// Mcp_call, _mcp_call, mcp__call, mcp-call, mcpcall, my_mcp_call → 200
// The "Mcp_" capitalization is the smallest stable rewrite that
// preserves readability. The reverse map below is propagated to
// chatCore via body._toolNameMap, which the SSE passthrough stream
// uses (utils/stream.ts:restoreClaudePassthroughToolUseName) to
// rewrite tool_use.name back to the client's original namespace on
// the response side. Capy sees mcp_call back in tool_use blocks.
const toolNameMap = applyMcpToolNameRewrite(transformed);
if (toolNameMap.size > 0) {
transformed._toolNameMap = toolNameMap;
}
}
return transformed;
}
async execute(input: {
model: string;
body: unknown;
stream: boolean;
credentials: ProviderCredentials;
signal?: AbortSignal | null;
log?: any;
upstreamExtraHeaders?: Record<string, string> | null;
}) {
// Resolve URL dynamically so settings table cliproxyapi_url is respected.
// Uses 60s cache to avoid DB reads on every request.
const baseUrl = await resolveCliproxyapiBaseUrl();
const endpoint = this.selectEndpoint(input.body);
const url = `${baseUrl}${endpoint}`;
const shape = endpoint === "/v1/messages" ? "anthropic" : "openai";
const headers = this.buildHeaders(input.credentials, input.stream);
const transformedBody = this.transformRequest(
input.model,
input.body,
input.stream,
input.credentials
);
mergeUpstreamExtraHeaders(headers, input.upstreamExtraHeaders);
const timeoutSignal = AbortSignal.timeout(FETCH_TIMEOUT_MS);
const combinedSignal = input.signal
? mergeAbortSignals(input.signal, timeoutSignal)
: timeoutSignal;
input.log?.info?.("CPA", `CLIProxyAPI → ${url} (model: ${input.model}, shape: ${shape})`);
// _toolNameMap is an in-memory channel to chatCore for response-side
// tool name restoration; never send it over the wire.
const wireBody =
transformedBody && typeof transformedBody === "object"
? JSON.stringify(transformedBody, (key, value) =>
key === "_toolNameMap" ? undefined : value
)
: JSON.stringify(transformedBody);
const response = await fetch(url, {
method: "POST",
headers,
body: wireBody,
signal: combinedSignal,
});
if (response.status === HTTP_STATUS.RATE_LIMITED) {
input.log?.warn?.("CPA", `CLIProxyAPI rate limited: ${response.status}`);
}
return { response, url, headers, transformedBody };
}
/**
* Health check — verifies CLIProxyAPI is reachable.
*
* CPA 6.x doesn't expose a /health endpoint; previously we hit /health
* and got 404, which made the dashboard report "CLIProxyAPI not
* detected" even when the service was up and successfully serving
* /v1/messages. Probe /v1/models instead (returns 200 with the
* advertised model list), which is the closest thing CPA has to a
* liveness probe and works on every CPA version we've tested.
*/
async healthCheck(): Promise<{ ok: boolean; latencyMs: number; error?: string }> {
const start = Date.now();
try {
const baseUrl = await resolveCliproxyapiBaseUrl();
const res = await fetch(`${baseUrl}/v1/models`, {
signal: AbortSignal.timeout(HEALTH_CHECK_TIMEOUT_MS),
});
return {
ok: res.ok,
latencyMs: Date.now() - start,
...(!res.ok ? { error: `HTTP ${res.status}` } : {}),
};
} catch (err) {
return {
ok: false,
latencyMs: Date.now() - start,
error: err instanceof Error ? err.message : String(err),
};
}
}
}
export default CliproxyapiExecutor;