mirror of
https://github.com/diegosouzapw/OmniRoute.git
synced 2026-08-16 04:03:02 +03:00
c4ee73dee074100384bf41fcdef576ff5a9bf7d0
573 Commits
| Author | SHA1 | Message | Date | |
|---|---|---|---|---|
adevwithpurpose
|
c4ee73dee0 |
fix(antigravity): strip trailing model turn for native Gemini requests too
Newer Gemini endpoints reject a request ending on a model turn with HTTP 400 'Requests ending with a model turn are not supported' — the same rejection class Claude hits via Vertex. transformRequest() previously wired stripTrailingAntigravityAssistantTurn() only into the isClaude branch, so native Gemini models routed through Antigravity kept a trailing role:model entry and hit the 400. Extend the guarded strip (never empties contents) to native Gemini models too, gated by upstreamModel including "gemini". The Claude path is untouched (byte-identical), preserving PR #6114's live validation against Vertex Claude. Flips tests/unit/antigravity-claude-prefill-strip.test.ts test (b), which previously asserted the buggy pass-through, and adds (b2) for the gemini-3-flash-agent tier. Closes #10104 |
||
|
Xiangzhe
|
8417ace4b3 |
feat(codex): add OAuth fingerprint convergence modes (#10243)
* fix(deps): bump nanoid, dompurify for 2 new Dependabot alerts (#189, #190) Bumps: nanoid ^3.3.17 (was transitive, now overridden), dompurify ^3.4.13 (with monaco-editor scoped override). Closes Dependabot #189, #190. Remaining #182-#188 (js-yaml + mermaid) already closed by #9651 merge — awaiting Dependabot re-scan. npm audit → 0 vulnerabilities. * fix(repo): harden .gitignore to also ignore a _tasks symlink (/_tasks) _tasks is a SEPARATE nested git repo (gitignored). The pattern _tasks/ (trailing slash) ignores only a directory, not a SYMLINK named _tasks. A self-referential _tasks symlink can slip in via git add -A and, once pulled, checkout materializes it over the real _tasks repo (destroying plans/specs/hands-off). Anchored /_tasks ignores the symlink too, preventing re-capture. * feat(codex): converge OAuth fingerprints * test(codex): preserve identity assertions * fix(codex): preserve explicit off identity * fix(codex): close fingerprint transport gaps --------- Co-authored-by: diegosouzapw <diegosouzapw@users.noreply.github.com> Co-authored-by: Diego Rodrigues de Sa e Souza <diegosouza.pw@outlook.com> Co-authored-by: diegosouzapw <diegosouza.pw@gmail.com> Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com> |
||
|
Jeyhun F. Aslanov
|
05b1311884 |
fix(sse): extract perplexity-web answers from workflow_block (#10259)
Perplexity moved the answer text out of `markdown_block` into
`workflow_block` (`intended_usage: "workflow_root"`), streaming it as
RFC-6902 patches whose `field` is `"workflow_block"` and whose paths
address `/steps/<n>/items/<m>/payload/text_payload/chunks/<k>`.
`extractContent` recognised neither shape. Two independent guards dropped
every answer frame:
- `isAnswerTextUsage("workflow_root")` is false, so the block loop
`continue`d before any accumulation.
- the diff guard skipped every patch whose `field !== "markdown_block"`.
The stream therefore ran to `COMPLETED` with an empty accumulator and the
executor surfaced `Provider returned empty content` (502) even though the
upstream SSE carried the full answer. Every model was affected — the
carrying block is model-independent — so the provider was unusable.
Adds `workflow_block` to `PplxBlock`, an `applyWorkflowDiff` patch
applier for the streaming path, and `applyWorkflowBlock` for a
materialized block on the terminal frame. Answer tracks are keyed per
step+item so concurrent items cannot overwrite each other's chunk
indices, and only `variant: "answer"` payloads are accumulated — search
queries, sources and "thinking" items stay out of the message.
Fixtures in the regression test are trimmed from a live capture
(pplx-auto, mode=copilot); replaying the full 96 KB capture through the
patched extractor yields the complete 247-char answer over 7 incremental
deltas, against an empty string before the fix.
Co-authored-by: Jeyhun F. Aslanov <jeyhun.f.aslanov@Jeyhuns-MacBook-Pro.local>
|
||
|
adevwithpurpose
|
ce4abd7ef4 | fix(opencode): force CLI User-Agent when CLI identity synthesis is enabled (#10222) | ||
|
Xiangzhe
|
99d19f8f35 |
fix(kimi): normalize MFJS tool schemas (#10079)
Remove unsupported root-level anyOf constraints only on Kimi and Moonshot OpenAI tool requests while preserving nested schemas and caller-owned inputs. Mark Kimi Web models as unable to execute function tools so combo routing filters them correctly. |
||
|
Arnav Rastogi
|
48124fca5a |
fix(zed-hosted): send the provider wire values cloud.zed.dev accepts (#10051)
Every zed-hosted completion failed with
500 {"error":{"message":"[500]: An internal server error occurred."}}
for every model id, including deliberately invalid ones.
Root cause: ZED_PROVIDER held display-cased names ("Anthropic", "OpenAi",
"Google", "XAi"), and normalizeZedProvider's return value is serialized
straight into the `provider` field of the POST /completions envelope. Zed
matches that field exactly and fails the request before looking at the model,
which is why the model id never mattered.
Verified live against cloud.zed.dev with an otherwise identical request:
{"provider":"anthropic",...} -> 200
{"provider":"Anthropic",...} -> 500 {"message":"An internal server error occurred."}
{"provider":"open_ai",...} -> reaches the OpenAI request parser
{"provider":"openai",...} -> 500 (same internal error)
The spellings now follow Zed's own GET /models catalog, which reports
`anthropic`, `open_ai` and `google`. That also makes normalizeZedProvider
identity on catalog values instead of corrupting a value Zed just supplied —
previously it accepted the correct lowercase input and re-cased it into the
form that 500s.
`x_ai` follows the same underscore convention; this account's catalog exposes
no xAI models, so that one spelling is by convention rather than observation.
The constant is module-local and every branch compares against it, so internal
dispatch (initProviderState / convertProviderEvent / buildProviderRequest) is
unaffected. Two existing tests asserted the display-cased values and one passed
"Anthropic" to wrapZedCompletionStream directly; all are updated to the wire
values the executor now produces.
Co-authored-by: root <root@srv1710948.hstgr.cloud>
|
||
|
zannen7
|
cc41503c4a |
fix(chatgpt-web): preserve native max thinking effort (#10077)
* fix(chatgpt-web): preserve max thinking effort * fix(chatgpt-web): allow native max effort * test(chatgpt-web): cover native max effort * docs(changelog): record ChatGPT Web max effort fix |
||
|
NorD
|
d085a0e693 |
fix(providers): xai-oauth chat→responses body + missing breaker import (#10165) (#10170)
- Import isProviderBreakerFailureStatus in chat.ts (ReferenceError on cooldown path) - Tag xai-oauth/grok-4.5 with targetFormat openai-responses - XaiExecutor converts messages/max_tokens/response_format before /v1/responses - normalizeOpenAIResponsesRequest safety net for chat-shaped bodies Fixes #10165 Co-authored-by: nordz0r <nordz0r@users.noreply.github.com> |
||
|
Chewji
|
f2d94957c8 | fix(ollama-cloud): map xhigh reasoning effort to max (#10160) | ||
|
Jefferson Alves
|
10c622afa6 |
fix(providers): default missing cache_control.ttl to 1h on the native Claude OAuth path (#10221)
Real Claude Code (and CC-protocol-compatible clients) commonly send
`cache_control: { type: "ephemeral" }` with no `ttl`. On the native
Claude OAuth path (provider `claude`/`cc`) the outbound anthropic-beta
set always includes extended-cache-ttl-2025-04-11, so requesting the 1h
TTL is always valid here — but Anthropic only honors it when `ttl` is
explicit; an absent `ttl` silently falls back to the platform default of
5 minutes even though the 1h beta was negotiated.
Practical effect: any pause longer than 5 minutes between turns forces a
full prefix rewrite (tens of thousands of tokens for a typical Claude
Code system+tools prefix) instead of a cache hit, burning through the
subscription's rate limit far faster than native (direct-to-Anthropic)
usage for the same workload.
Adds `normalizeCacheControlTtl()` to claudeCodeConstraints.ts (same
module as the sibling cache_control helpers enforceCacheControlLimit /
ensureCacheControlOnLastUserMessage) and calls it right after the
billing-header system-block manipulation in base.ts, immediately before
the request is signed and sent. Never touches a cache_control that
already specifies a ttl.
Measured before/after with a real Claude Code CLI session through this
path (system + tools prefix ~46k tokens):
before: cache writes always land in ephemeral_5m_input_tokens; a >5min
gap between turns forces a full rewrite (cache_read resets to 0)
after: cache writes land in ephemeral_1h_input_tokens; a >6min gap
survives (cache_read stays intact)
--no-verify note: local pre-commit's check:docs-sync fails on this branch
tip ("CHANGELOG.md first section must be Unreleased") for reasons
unrelated to this diff (pre-existing state of release/v3.8.50 mid-cycle,
CHANGELOG.md untouched by this change). Added the required changelog.d
fragment per CONTRIBUTING.md regardless.
Co-authored-by: Jefferson Alves <jefferson@rastrosystem.com.br>
|
||
|
Diego Rodrigues de Sa e Souza
|
d259d9fcba |
fix(ci): clear base-reds on release/v3.8.50 (round 3) (#10213)
* fix(ci): clear base-reds on release/v3.8.50 (round 3) - CHANGELOG.md: restore the top [Unreleased] section dropped by the #10189 reconcile (docs-sync gate: first section must be Unreleased) - env-doc-sync: document CONDUCTOR_ORCHESTRATOR_TOKEN + CONDUCTOR_SPOKESPERSON_URL in .env.example/ENVIRONMENT.md; allowlist the CI-only GITHUB_STEP_SUMMARY and TS7_BASE_REF (ts7 ratchet signals); drop a stray merge artifact line - providers: restore the audited chatanywhere metadata entry that base-reds round 2 dropped together with its duplicate — the provider was half-wired (registry+endpoint without APIKEY metadata), which is what the wave3 test catches; re-pin providers-constants-split at the measured 228 - docs counts: 338 -> 339 (today's +2 void-ai/helixmind, -1 Puter) via gen:provider-reference + README/AGENTS/llm.txt/package.json/diagrams/i18n mirrors - file-size ratchet: annotated rebaseline for the two pre-existing drifts (ModelSelectModal 1138, gateways 1250) following the 2026-08-11 precedent Refs #9985 * fix(ci): base-reds round 3b — stale sibling tests + mode-pack weight contract - check-docs-counts-sync.test.ts: drop the imports/subtests of the four helpers #10196 removed from the gate script (readMcpFactsFromSource, listLocalizedDocs, makeRequiredCountsValidator, checkFreeTierInventory) — the new-API tests that #10196 added stay; the file now loads again under the node runner - quota-connection-recovery.test.ts: convert from vitest APIs to node:test — the file lives in tests/unit/*.test.ts (node-runner glob) and the vitest runtime crashes when imported outside vitest, killing the whole shard entry - modePacks.ts: re-normalize all six mode packs to sum 1.0 — #8940 added sessionAvailability: 0.05 to every pack without rebalancing (1.05 total); ratios preserved exactly (÷1.05), so post-normalizeScoringWeights behavior is unchanged; restores the declared sum-to-1.0 contract the 4235 test pins Refs #9985 * fix(ci): base-reds round 3c — vitest siblings, weights default, secrets FP, mutation tap - DistributeProxiesButton.test.tsx: wrap renders in NextIntlClientProvider — #9245 localized the component (useTranslations) and left the test without the intl context, failing all 14 cases - scoring.ts: re-normalize DEFAULT_WEIGHTS to sum 1.0 (same #8940 class as the mode packs — sessionAvailability added without rebalancing; ratios preserved) - .gitleaks.toml: generalize the kimi sponsor-banner localStorage-key allowlist to -v\d+ — #10200 bumped v1→v2 and the stale regex regressed the secrets ratchet with a false positive - stryker.conf.json: register 6 covering unit tests in tap.testFiles (4 modules) so their mutant kills count — unblocks check:mutation-test-coverage --strict Refs #9985 * fix(ci): base-reds round 3d — inspector factor gap, stale registry/gap tests, i18n key sync - comboScoringInspector: add cacheAffinity/sessionAvailability/connectionDensity to FACTOR_KEYS + the factor-key type — calculateScore() weighs them but the breakdown omitted them, so the explained contributions never summed to the reported score (inspector bug, red on the pure tip) - combo-scoring-inspector.test: make the explicit-weights override sum-neutral (±0.05 shift) so it stays valid for any DEFAULT_WEIGHTS values — the hardcoded override only summed to 1.0 against the pre-#8940 defaults, which is also why explicit weights silently fell back to 'default' on the tip - unorouter-registry.test: align to the canonical .com host (api.unorouter.ai 301-redirects there, verified live) and to wave4's live model discovery (passthrough, no static seed) — the .ai/auto-model expectations were stale - check-migration-numbering.test: 147 left KNOWN_GAPS when 147_api_keys_model_access_mode.sql landed — assert absent (same as 143) - i18n: sync-ui pass — 35,914 missing UI keys stamped as __MISSING__ placeholders across 42 locales (mechanical; greens the pt-BR key-presence integrity test; coverage pct unchanged by design — translation is a separate workstream) Refs #9985 * fix(ci): base-reds round 3e — 2 real defects + 14 stale sibling tests (waves A-E) Real defects fixed: - src/lib/db/apiKeys.ts: #9313's empty-allowlist early return bypassed the group permission check, silently disabling group deny rules (#8817) for every key without a per-key allowlist; fall-through restored, restricted+[] deny-all kept - open-sse/utils/proxyFetch.ts: #10032 re-appended the raw transport error to the propagated message, reintroducing the proxy user:password leak #9837 closed; new redactProxyDetailsInMessage() keeps the reason, redacts URL/credentials - .github/workflows/quality.yml: #10134 added the TS7 ratchet as a separate blocking step AFTER the aggregated gates — the exact #8542 masking mechanism; folded into the non-fail-fast loop (still blocking, still PR-only) ⚠️ CI edit, gate-strengthening — explicit owner sign-off requested on the PR - src/i18n/messages/ko.json: 3 machine-mistranslation regressions caught by the #8244 glossary checker (장애인→비활성화됨, 양말5://→socks5://, 비클로드→Claude가 아닌) Stale sibling tests aligned to deliberately-moved contracts (each cites its mover): request-log-detail-layout + -stream (#9245 intl provider), repro-8542 pin update, quality-rail-gate-membership (#10134 shape), agentSkills-routes 45→46 (#9058), cloudflare-ai-catalog-8717 (#8804 supersedes #8808), executor-xai (#9994), vision-bridge-claude-wire (#9463 minimax→openai), sse-auth forced-pin (#8893), tls-proxy-context (strengthened leak guards), rate-limit-local-error-classification (#9164/#9342), minimax-thinking-signature (#9463), codebuddy-cn (#9723 +1 test), github-copilot-custom-model (#9050), providers-g4f-batch3 (#9584), synced-capability-warmup (#9199, stricter), sidebar-tools-group (#8221), oauth-modal-grok-cli-paste (#9245); agentSkills/catalog.ts comment 45→46; file-size rebaseline for proxyFetch (+19, annotated) Refs #9985 * fix(ci): base-reds round 3f — waves F-J: 9 more real defects + stale sibling sweep Real production defects fixed (all red on the pure tip, each with its origin): - routeGuard.ts: #8949 accidentally DELETED the /api/providers/[id]/login local-only pattern — the route spawns a browser, so the loopback gate for a process-spawning route was gone (Hard Rules #15/#17); restored (314 guard tests green) - agentSkills generator: #9058's category dispatch gave the config category an empty body, wiping skills/config-codex-cli/SKILL.md at the #10131 sync; fixed + SKILL.md regenerated via the official generator - imageRegistry: #9982 broke same-provider bare aliasing (antigravity preview id sent upstream unresolved); new resolveSameProviderBareAlias() keeps the fal cross-provider fix intact - imageRegistry: #9982's prefix strip handed the bare nano-banana ids to fal-ai, violating the pinned 2026-07-31 operator decision (adobe-firefly owns them); fal entries made prefix-only (dispatch already re-prefixes) - mediaGeneration/fal.ts: the missing-credential 401 guard was lost when #10198 deleted the superseded falHandler — tests were hitting the live network - bottleneckPatch/rateLimitManager: #9041's merge clobbered #9604, resurrecting the Bottleneck v2.19.5 heartbeat bug (reservoir never refills); patched the library defect at the root and re-aligned chat-rate-limit-body-lock to the working reservoir contract - processSupervisor.mjs: #9761 regressed the Node spawn to bare "node" (the #9156 launchd bug) and dropped #9209's ipv4first args; both restored - openai-responses/pureHelpers: #9423's Agent null-sentinel was unreachable on the schemaless JSON-string path; gate extended - i18n en.json: #8222's regen reverted the #9976 unclosed-tag fix and #8559's combo-cooldown copy; #9038 shipped 40 t() calls with no messages (runtime MISSING_MESSAGE); all restored/added + official sync-ui stamps, and vi's zero-marker policy re-established via the sanctioned translation backend Stale sibling tests aligned (movers cited inline): chat-helpers (#9447), executor-antigravity (#9351), video-fal-grok (#9982), visionBridge (#9759), web-session-credentials (#8974), production-build-module-integrity (positive anchor added), agentSkills-generator/skillManifestsLint/skills-injection/ agentSkillTools-mcp/listCapabilities-a2a (#9058), memory-settings (#10010), model-catalog-policy-invalidation (#8906), model-alias-seed (#9485), reactive-context-compaction (#8949), combo-provider-wildcard (broken upsert helper), oauth-google-loopback (43-locale resurrected-key removal) Validation: 501/501 across the 47 touched test files; typecheck:core, lint, file-size, docs-sync all green. Refs #9985 * fix(ci): base-reds round 3g — wave K/L: 4 more real defects + stale alignments Real defects: - base/reasoningEffort.ts: the stale duplicate cherry-pick #9612 re-added the codex minimal→low rewrite that #9883 had deliberately removed (OMP minimal passthrough); block removed again - cursorImages.ts: #9840 wired prepareCursorImageForWire (sharp re-encode, fail-closed) into the SHARED resolveCursorImages, breaking zai-web and conol-web image uploads (HTTP 400 'undecodable'); new prepareForWire opt-out, Cursor default path unchanged (8 cursor suites green) - modelCapabilities/snapshot: catalog prepare still issued 323 per-model reads of model_context_overrides + max_input_tokens overrides, violating #9199's bulk-load contract; both now resolve from the snapshot single pass - v1-models-discovery-conformance: re-pinned to the bounded 30s SWR window (#9199/#10198) — the old 'stale-first regardless of age' contract is gone Stale tests aligned (movers cited inline): codex-tools-strict-default (#9828 redundant-oneOf strip), devin-providers (#9245 i18n), db-migrationrunner- constants-split (147→151 renumber #8228), gitlab-duo-oauth-setup (#9245), chatcore-extracted-modules (#9161 outbound-protocol keying) compression-api CI failures were cascade artifacts of codex-tools-strict-default failing in the same force-exit shard process — no own defect (171/171 local). Refs #9985 * fix(test): compression-api — register both describes before the runner starts The DATA_DIR setup + route/db top-level awaits sat BETWEEN the two describes; under --test-force-exit (the CI unit-runner flag) the process exits once the already-registered tests finish, so on slow CI machines the whole second describe died as 'Promise resolution is still pending' — the recurring CI-only shard-2 failure that never reproduced locally without the flag. Moved to the top of the file; 10/10 under --test-force-exit locally. Refs #9985 * fix(quality): freeze modelCapabilities.ts at 1006 (annotated) — snapshot routing growth Refs #9985 * fix(quality): move the modelCapabilities freeze into the frozen map (nested schema) Refs #9985 * fix(i18n): translate all 39,718 pending UI keys across 42 locales (owner-approved) Mass-translated every __MISSING__ placeholder via the official i18n:sync-ui --translate-markers pipeline (operator backend), restoring i18nUiCoverage to the 100 baseline (was 89.9 after the merge-storm UI landings + the 42 keys #9038 never shipped). Post-pass repairs, all caught by the existing gates: - glossary: retired renderings the machine reintroduced normalized again (提供商→提供者 zh-CN/zh-TW, 鏈接→連結, 文檔→文件, 調用→呼叫, 供應商→提供者, 響應→回應, 不活躍→未啟用 zh-TW; 클로드→Claude, 옴니루트→OmniRoute ko); DATA_DIR forbidden rendering avoided via 数据文件夹 rephrase - ICU integrity: 120 values with renamed/dropped {params} repaired (39 positional renames, 81 reset to the en source — functional over fluent) Validation: glossary/pt-BR/vi/deno-relay/settings-keys/value-drift/google- loopback suites 76/76; placeholder diff en×42 locales = 0; worst-locale coverage = 100.0%. Refs #9985 --------- Co-authored-by: backryun <bakryun0718@proton.me> |
||
|
Diego Rodrigues de Sa e Souza
|
f1eb0b8357 |
refactor(providers): remove the Puter provider at its owner's request (#10210)
Remove the Puter provider (id `puter`, alias `pu`) entirely, at the request of Puter's owner, Nariman Jelveh: - registry entry (open-sse/config/providers/registry/puter/) and PuterExecutor (open-sse/executors/puter.ts), with their registrations - API-key preset card (gateways.ts), provider icon and public SVG asset - 33 free-model catalog entries (pool `puter`) - authHint i18n key across all 43 UI locales - credential-requirement frozen-list entry and related comments - docs: ARCHITECTURE, CODEBASE_DOCUMENTATION, FREE_TIERS (removal note), PROVIDER_REFERENCE regenerated (337 providers), translated doc mirrors, llm.txt + its 42 i18n mirrors, README/AGENTS/package.json counts (338→337 providers, 144→145 migrations) and the 5 canonical SVGs - migration 152 cleans up stored puter connections/keys/custom models; historical usage records are preserved (same principle as migration 151) - regression guard: tests/unit/puter-provider-removed.test.ts; puter fixtures in shared tests swapped for neutral providers; translate-path golden snapshot regenerated Historical CHANGELOG mentions are intentionally preserved; the removal carries its own CHANGELOG entry. Co-authored-by: backryun <bakryun0718@proton.me> |
||
|
Diego Rodrigues de Sa e Souza
|
ecc89eef14 |
feat(providers): integrate audited free-tier gateways (#9210)
* feat(providers): add Zylo UnoRouter and Poolside registries * feat(providers): integrate audited free-tier gateways * feat: add wave2 free-tier provider registries * feat(providers): add Mixlayer Speka and TokenReply registries * feat: add wave 2 free-tier provider registries * fix: align meganova provider slug * feat(providers): integrate wave2 free-tier gateways * feat(providers): add Wave 3-A free-tier registries * feat(providers): add HelyxAI Auriko and Poixe registries * feat(providers): add Naga AI and Chat Oripe registries * feat(providers): integrate wave3 free-tier gateways * feat(providers): add FreeInference registry * feat(providers): add Free.ai registry * feat(providers): integrate wave4 free-tier gateways * docs: synchronize provider and free-tier inventories * refactor(providers): split audited gateway catalog * feat(providers): add audited Void AI and HelixMind gateways * feat(providers): finalize audited free-tier integration * test(providers): update APIKEY split count to 229 after rebase onto release/v3.8.50 The rebase merged the release catalog (201 APIKEY providers) with the PR's 28 free-tier additions, yielding 229 total. Correct the characterization count so the partition assertion reflects the true merged state. --------- Co-authored-by: diegosouzapw <diegosouzapw@users.noreply.github.com> Co-authored-by: backryun <bakryun0718@proton.me> |
||
|
Diego Rodrigues de Sa e Souza
|
f6ccd3cf9f |
fix(quality): green release/v3.8.50 base-reds round 2 (#9985) (#10131)
* fix(quality): green release/v3.8.50 base-reds round 2 — gateways/conol/deepai corruption, migrations, docs, ratchets, dashboard-typecheck Base-red fix for issue #9985 after the 2026-08-11 merge storm (99 PRs). Real defects fixed: - gateways.ts: close regolo entry (was swallowing naga-ac + chatanywhere from #9421), drop stale duplicate chatanywhere entry (#9594) - conol-web + deepai registry: correct ../shared import depth + deepai executor:default - modelSelectModalHelpers: close isProviderModelHidden (#9011) - driverFactory.test.ts: restore eaten test-closing brace (#9173) - usageTracking: remove duplicate cache_* props - modelCapability{Overrides,ResolutionSnapshot,Capabilities}: max_token -> max_output_tokens (#9199 vs #8908) + test align - videoGeneration: drop duplicate handleFalVideoGeneration import (mediaGeneration/fal canonical, #9982) - responseSanitizer: cast input_tokens_details before .cached_tokens access - EditConnectionModal: missing alibaba code fields, hoist validationPsd, providerPageHelpers Badge variant union - FreeBudgetCard: t() -> labels.noApiKey - peerRouting + cliRuntime: ProcessEnv typing - image-combo.test.ts: type any -> unknown - fal.test.ts: moved to tests/unit/services (collected path) 14 tests green - remove duplicate 143_job_registry.sql (146 canonical), KNOWN_GAPS fix Docs/ratchets (owner-authorized rebaselines, annotated): - CHANGELOG 3.8.50 living section restored + 42 i18n mirrors - MCP-SERVER.md 104->105 tools + i18n - ENVIRONMENT.md/.env.example: ADOBE_FIREFLY_CHROME_HEADED + DEBUG_CLAUDE_NONSTREAM - fabricated-docs allowlist: TELEGRAM proposal env vars - file-size: 5 grown files + proxyFetch 1207->1220 - dead-code 230->248, codeql 2->9 (drift from merged PRs, not this PR) - untrack _tasks symlink; agent-skills-sync --apply (config-codex-cli) * fix(changelog): reformat two feature fragments to the bullet convention (#9239, #9490) * fix(quality): prune stale ESLint suppressions (base-red) * fix(quality): resolve open-sse type errors + catalog/build regressions (base-red round 3) Storm-merge splices repaired in the base-fix PR #10131: - doctor.ts: AppConfig missing brokerSocketPath - conol-web.ts: Buffer not assignable to BodyInit (Uint8Array) - tinycms.ts: TinyCmsExecutor.execute return matches BaseExecutor (response/url/transformedBody) - tinycmsSigner.ts: encodeInto never-narrowing guard + dead wasm URL fallback (Turbopack) - virtualFactory.ts: options slot for resolutionSnapshot - bottleneckPatch.ts: insufficient-overlap casts (as unknown as) - imageCombo.ts: narrow handleImageGeneration union result - browser-worker.ts: AppConfig + turn.capabilities splice - conolDiscovery.ts: getProviderOutboundGuard from Policy module - catalog.ts: drop removed SWR hooks (getCatalogStaleWhileRevalidateMs + accessors), CatalogCachePolicy -> inline settings, resolve 4-arg call - catalogCache.ts: remove dead inFlight/promise refs - chat.ts: add isProviderBreakerFailureStatus import - model-catalog-cache-swr-8728.test.ts: align to #9199 new API (policy injection removed) * fix(quality): align UI test fixtures to current component contracts (base-red vitest) - setup-wizard: provide required serverState prop (component gained it in a merged PR) - grok-device-oauth-modal: next-intl stub resolves grok flow keys to EN labels - provider-quota-widget: label now inline (PR #8916 removed AutoRefreshButtonLabel extraction) — test the widget - use-provider-connections-cursor-refresh + phase1f: match /api/providers?provider=<id> query form; hoist heavy dynamic imports to module scope (timeout flake) - home-topology: mock next/navigation useRouter (component added node-click navigation) - cooling/lobe/AutoComboCatalog: raise cold-import describe timeouts to 30-60s - request-logger-*: align to current detail-view contract * fix(search): guard params.token undefined in serper headers (typecheck base-red) * fix(search): guard token headers + non-null providerConfig (typecheck base-red) * fix(changelog): restore base CHANGELOGs eaten by merge auto-resolve (43 files) --------- Co-authored-by: diegosouzapw <diegosouzapw@users.noreply.github.com> Co-authored-by: backryun <bakryun0718@proton.me> |
||
|
Diego Rodrigues de Sa e Souza
|
85c4292ce3 |
fix(security): correct XML double-unescape and non-CSPRNG nonce from CodeQL sweep (#10154)
* fix(security): correct XML double-unescape and non-CSPRNG nonce from CodeQL sweep Two real defects surfaced by the 2026-08-12 code-scanning triage. decodeXmlText decoded `&` before `"`/`'`, so `&quot;` — the encoding of the literal text `"` — collapsed to `"` in a second pass. The decoded values feed the workspace-root trust comparison in parseTrustedCodexEnvironment, so an encoded path could decode into a different path than the client declared. Decoding `&` last fixes it. The tinycms nonce fell back to `Date.now()` plus a non-cryptographic PRNG when `crypto.randomUUID` was absent. That nonce is signed into the provider's anti-replay payload, so the fallback produced a predictable value silently. It is now always `randomUUID()` from node:crypto, which is present on every supported runtime. Both guards are mutation-validated: reverting either fix makes the new test fail. Refs #9985 * fix(security): reword tinycms nonce comment so the CSPRNG regression test holds --------- Co-authored-by: backryun <bakryun0718@proton.me> |
||
|
Diego Rodrigues de Sa e Souza
|
c19da4db46 |
fix(security): resolve open CodeQL alerts (#10188)
Alert 806 (js/insecure-randomness, open-sse/executors/tinycms.ts): the TinyCMS nonce is signed into x-secure-signature and reused as x-secure-nonce / x-session-id, so the Math.random() fallback made a signed request predictable and replayable. Use randomUUID() from node:crypto unconditionally. Alert 811 (js/double-escaping, chatgpt-web adapters/environment.ts): decodeXmlText() decoded & before " / ', so the bare & it produced was re-consumed and the text was unescaped twice (&quot; collapsed to "). These values become the trusted Codex sandbox cwd / workspace_roots, so the double-unescape silently rewrote the workspace boundary. Decode & last. Alerts 813/814 (js/incomplete-url-substring-sanitization, test files): replace the includes() URL checks with exact comparisons (new URL(url).hostname === ... and an explicit === over the recorded URL array). Both assertions get strictly tighter. Regression guards: tests/unit/tinycms-secure-nonce-randomness.test.ts and tests/unit/chatgpt-web-environment-double-unescape.test.ts, both failing before the fix and passing after. Co-authored-by: backryun <bakryun0718@proton.me> |
||
|
Diego Rodrigues de Sa e Souza
|
37d70785c7 |
fix(build): repair broken production build, red lint gate and SWR crash (#10198)
The Build CI job is advisory, so eight module-level defects from eight different PRs accumulated on release/v3.8.50 until `npm run build` failed with 7 Turbopack errors and `npm run lint` with 14. Build (link-time): - modelSelectModalHelpers.ts: a lost `}` swallowed PROVIDER_TEST_CHUNK_SIZE into isProviderModelHidden's body (#9011). - videoGeneration.ts: handleFalVideoGeneration imported twice; the standalone falHandler.ts is superseded by the provider-neutral mediaGeneration/fal.ts and is removed here (#9982 over #9969). - catalog.ts: re-exported and called the injectable SWR policy that #9199 deliberately replaced with a fixed 30s bound. Fixed on the consumer side — restoring the accessor would resurrect the unbounded window #9199 removed after measuring a 41s catalog build in production. - tinycmsSigner.ts: generated wasm-bindgen glue kept a sidecar `new URL('wasm_signer_bg.wasm', import.meta.url)` that no file backs; Turbopack resolves it statically. The module ships inlined as WASM_BASE64 and the only caller always passes it explicitly (#8736/#10087). - conolDiscovery.ts: imported getProviderOutboundGuard from outboundUrlGuard, which does not export it. Fixed on the consumer side: outboundUrlGuard.ts is loaded by the packaged CLI without a tsconfig, so it must stay free of `@/`-aliased imports (#7682). Runtime (the build never caught this one): - catalogCache.ts::scheduleBackgroundRefresh had two dangling statements referencing undeclared `inFlight`/`promise`, so EVERY stale-while-revalidate read threw a ReferenceError. Surfaced by realigning the #8728 suite, which #9199 left asserting a removed contract. Lint: - driverFactory.test.ts: a case inserted between the preceding test's `finally` and its `});` left the file unparseable, so the SQLite driver-cascade suite (26 tests) had not run since 2026-08-11 (#9173). - providerModelsConfig.ts: imported an executor directly, crossing the G14 boundary; routed through a new open-sse/services/zaiWebCredentials.ts (#8451). - image-combo.test.ts: 11 `any` violations, now typed (#9499). Validation: npm run build exit 0, npm run lint clean, typecheck:core clean, 41/41 tests green across the affected suites. Refs #9011 #9982 #9199 #8728 #8736 #10087 #8974 #9173 #8451 #9499 Co-authored-by: backryun <bakryun0718@proton.me> |
||
|
backryun
|
bd472200d5 |
[v3.8.50] Fix Z.ai web browser transport and model capabilities (#8451)
* fix: complete Z.ai web browser transport * refactor: address Z.ai review feedback * test(zai-web): reconcile the #8014 endpoint guard with the chats/new + signed flow Rebasing onto release/v3.8.49 pulled in #8503, which repointed CHAT_URL to /api/v2/chat/completions and added an endpoint probe. This branch already targets v2, so the executor conflict resolved to this branch's superset (NEW_CHAT_URL + signature constants alongside the same v2 CHAT_URL). The two tests needed adapting, because #8503's assertions assume the pre-rework flow: - executor-zai-web.test.ts: the completion URL now carries the request signature as a query string, so an exact-equality check on the endpoint can never match. Assert the v2 prefix instead. - zai-web-chat-endpoint-8014-probe.test.ts: the probe drove the executor with a bare cookie credential and no captcha proof, which now routes through the browser transport — fetch was never called and the probe captured nothing. Supplied a direct-path credential, and matched on pathname across all requests (the executor also probes the homepage for the frontend version and calls /api/v1/chats/new first). The guard's intent is unchanged and slightly strengthened: it now asserts no request reaches the stale unversioned path and that exactly one completions request is issued, against v2. 54/54 across the zai suites; typecheck:core and eslint clean. * fix(zai-web): surface upstream error frames instead of finishing empty Reported on this PR: HTTP 200, `out=0`, stream "complete", no content and no diagnosis. Cause. HTTP-level failures are already handled — fetchUpstream turns any !ok response into a makeErrorResult with the sanitized body. The gap is a 200 whose SSE body carries an error payload: parseZaiFrame returns null for it, drainSseDeltas drops it, and buildZaiStreamingBody then closes with an empty assistant message + stop + [DONE]. The caller reads that as a successful empty completion, so a rejected signature, an expired captcha and a stale token all look identical — which is why this had to be diagnosed by reading code rather than logs. Hard Rule #6. Fix. parseZaiFrame now classifies an affirmatively error-shaped frame (`error` at the top level or under `data`, string or {detail|message|msg}) as a terminal delta, checked before the delta paths so it cannot fall through to the "no usable delta" null. The stream emits it as `[Z.ai error] <message>`, matching the mid-stream convention the other web executors already use (zed-hosted's createErrorChunk) — the 200 is on the wire, so the status cannot change, but the caller must not be left reading a blank success. Content streamed before the failure is preserved. Message goes through sanitizeErrorMessage (Rule #12). Deliberately NOT changed: a contentless frame still parses to null. That is live-validated behaviour, not an oversight — z.ai emits phase frames with no delta_content, and executor-zai-web.test.ts pins it ("returns null for frames with no usable delta"). Treating "nothing parseable arrived" as a failure would invent policy on top of an observed protocol and risk false errors on the happy path, so this only adds recognition of explicit error frames. Tests (TDD, RED then GREEN): zai-web-silent-empty-repro.test.ts — 7 cases. Error frame classified and terminal; surfaced through the stream with the upstream's own text; surfaced after partial content without losing it; plus a REGRESSION GUARD that contentless/phase-only frames are still skipped, and two controls that the happy path and reasoning-only output are untouched. The guard and controls passed before the fix; the four error cases did not. 94/94 across the zai + stream suites; typecheck:core, eslint and check:file-size clean. * refactor(sse): extract the zai-web transports so the complexity ratchet holds The v3.8.49 merge-train rebaseline (#8686) set the ceiling to the tip's own measurement, leaving zero headroom, so this branch's +5 cyclomatic / +3 cognitive own-growth had nowhere to sit once rebased onto it. Eight violations, all in code this branch introduces, resolved by extraction — no behaviour change: - `execute` (152 lines, complexity 25, cognitive 20) now delegates to `resolveZaiRequest()` for the four client-error rejections and to a `fetchViaSignedApi()` method for the CAPTCHA/signature path, so it reads as "validate, pick a transport, shape the response". - `fetchThroughBrowser` (126 lines, cognitive 16) hands its image decoding to `resolveZaiBrowserAttachments()`, its Playwright options to `buildZaiBrowserChatOptions()`, and its call-log payload to `buildZaiBrowserAuditBody()`. - `configureZaiBrowserEffort` (cognitive 35 — the worst of the set) repeated a wrap-and-relabel try/catch four times inside an if/else. `runStage`, which already existed one function below, is now module-scoped and reused, and the toggle collapses to `checked !== config.enabled` (same four cases). - `validateWebCookieProvider` (complexity 19) moves its can-we-probe-this cascade into `resolveWebCookieProbe()`, which returns either a rejection or the URL + headers to use. - `acquireBrowserContext`'s creation closure (complexity 17) hands cookie and localStorage seeding to `seedContextSession()`. That last extraction also clears a violation that predates this branch — `acquireBrowserContext` was already over the 80-line ceiling — so cyclomatic lands at 2187 against a baseline of 2188. Verified: check:complexity-ratchets green both metrics; typecheck:core clean; ESLint clean on all four files; 85 tests across the zai-web, web-cookie validation, browser-pool and model-test-runner suites pass. * fix(zai-web): surface upstream errors on the non-streaming path collectZaiNonStreaming ignored delta.error — a 200 whose SSE body carries an error frame (rejected signature, expired captcha, stale token) came back as a successful empty completion. Now it throws on an error frame, matching the streaming path's [Z.ai error] convention; the caller's existing try/catch returns makeErrorResult(502) instead of an empty 200. Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com> --------- Co-authored-by: backryun <busan011@ormbiz.co.kr> Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com> |
||
|
backryun
|
acc066db3f |
[v3.8.50] feat(devin-desktop): replace public Windsurf provider (#8228)
* feat(devin-desktop): replace public Windsurf provider * fix(migrations): renumber Devin Desktop migration to 151 (avoid 147 collision) 147_windsurf_to_devin_desktop.sql collided with the released 147_api_keys_model_access_mode.sql — getMigrationFiles throws "Migration version collision detected" on every DB start. Base occupies slots up to 150, so renumber the new migration to 151 and point the windsurf→devin RENAMED_MIGRATION_COMPATIBILITY entries (and tests) at it. 147 is freed in KNOWN_GAPS since 147_api_keys now owns the slot. Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com> --------- Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com> |
||
|
backryun
|
5f5c0bc94a | fix(tinycms): align executor and signer contracts (#10087) | ||
|
Bob.Hou
|
ca6e944bb1 |
fix(antigravity): propagate switchAuth signal from 429 engine to retry guard (#9351)
When Google returns a 429 with no parseable retry hint, decide429 correctly classifies it as short_cooldown_switch_auth (switch accounts). But the executor discarded that decision, keeping only retryMs=60000. The retry guard then slept 60s against the same URL/account up to 3 times because 60000 <= LONG_RETRY_THRESHOLD_MS (inclusive boundary). Plumb a switchAuth boolean through tryResolveRetryFromErrorBody so the retry guard can decline the sleep branch and fall through to URL/account fallback immediately. Signed-off-by: Minxi Hou <houminxi@gmail.com> |
||
|
Jan Leon
|
a99c795a67 |
Add native ChatGPT Web provider for Codex clients (#8949)
* Bypass proxy compaction for native Codex context
* Add native ChatGPT Web provider pipeline
* Add managed browser and tunnel deployment
* Add ChatGPT Web setup and doctor UI
* Document and test ChatGPT Web integration
* fix(security): register chatgpt-web-codex-doctor in LOCAL_ONLY_API_PATTERNS
The diagnostic route under /api/providers/{id}/chatgpt-web-codex-doctor
was not registered in the spawn-capable route guard. Adding it for
parity with the existing /login pattern.
Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>
* fix(providers): route chatgpt-web-codex admin routes through a service boundary
The provider CRUD/doctor routes imported chatgpt-web-codex helpers
(finalizeValidatedChatGptWebCodexSecrets, encode/decodeChatGptWebCodexSecrets,
getChatGptWebCodexDoctorStatus) directly from open-sse/executors/**, which
no-restricted-imports (EXECUTOR_IMPORT_RESTRICTION) forbids for src/app/**
files — executor implementations must stay behind an open-sse handler or
service boundary.
Add open-sse/services/chatgptWebCodexAdmin.ts as a thin re-export boundary
(mirroring the existing tokenRefresh.ts re-export pattern) and import from
there instead. No behavior change.
Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>
---------
Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>
|
||
|
Jonathan Bailey
|
84e83e2f19 |
feat(providers): add DeepSeek V4 thinking effort aliases (#9485)
* feat(providers): add DeepSeek V4 thinking effort aliases * docs(changelog): add DeepSeek effort alias entry * fix(catalog): scope effort-tier fallback to declared models and harden resolver Addresses reviewer findings on #9485: - CRITICAL #1: catalog no longer synthesizes unresolvable effort aliases for static reasoning models without declared tiers (cheaperinference, cline, etc.) - CRITICAL #2: tiered static models survive synced-coverage suppression so normal installs with synced DeepSeek base models still expose aliases - WARNING #3: registry suffix resolution short-circuits when the raw id matches a direct custom or synced model, preserving custom apiFormat/targetFormat - WARNING #4: empty synced effort array no longer erases the registry fallback - WARNING #5: isFlash check is robust to suffixed/prefixed model ids - Added regression tests for blast radius, custom-model shadowing, none-path, and suffixed isFlash * fix(combos): expose static registry effort tiers in Combo Builder (#9485) Static provider registry models (e.g. DeepSeek V4 Flash/Pro) declare supportedThinkingEfforts, but buildModelOptions() only ran appendSyncedEffortVariants() over DB-synced rows. Synced metadata for a DeepSeek connection can omit supportedThinkingEfforts, so the catalog/ Playground surfaced the declared aliases while the Combo Builder picker showed only the bare base ids. Feed builtInModels with declared effort tiers through the same appendSyncedEffortVariants() utility used for synced rows, inheriting the base entry's contextLength/outputTokenLimit/supportedEndpoints/ supportsThinking and preserving its source. DeepSeek is not skipped by shouldExposeSyncedEffortVariants(), so Flash (none/low/high/max) and Pro (none/high/max) aliases now appear in the Combo Builder for any connection whose synced rows omit effort metadata. Regression test seeds a DeepSeek connection with effort-less synced rows and asserts the exact alias sets, source preservation, and metadata inheritance. |
||
|
NOXX - Commiter
|
f5ce51a9ff |
feat(providers): add Conol (conol.ai) web session provider (#8974)
* feat(providers): add Conol web support
* fix(conol): preserve sessions and image turns
* fix(conol): pin session model and effort via /model endpoint
Conol ignores agentModel/agentEffort on POST /api/sessions, so every
session silently ran on the downgraded account default (the create
response reports modelDowngraded: true / effectiveModel).
Sessions are now created empty and configured out-of-band against
POST /api/sessions/{id}/model before the first turn is submitted, in the
order the web client uses: modelPreset, then agentModel, then agentEffort.
The ordering is load-bearing because the model call resets agentEffort to
null server-side.
Effort now defaults to xhigh when the caller does not pin one via the
-<effort> model suffix, and is clamped onto the ladder each model actually
advertises, so xhigh degrades to high on claude-sonnet-5 and is skipped
entirely for models without an effort ladder such as openrouter/fusion.
Model and effort are also dropped from the session binding key so switching
models re-pins the existing session instead of stranding it and losing the
conversation history. Re-pinning only happens on an actual change, so
steady-state follow-ups cost no extra round trips.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
|
||
|
Ababil
|
55c2b35eb7 |
fix(codebuddy-cn): replace agent system prompts to bypass Tencent content filter (#9723)
* fix(codebuddy-cn): replace agent system prompts to bypass Tencent content filter Tencent's content filter flags CLI agent system prompts (e.g. 'You are Claude Code, Anthropic's official CLI...') as prompt injection / sensitive content and rejects the entire request with error: 抱歉，系统检测到您当前输入的信息存在敏感内容，我无法响应您的请求 This patch adds detection and replacement logic to the CodeBuddyCnExecutor: - Regex-based identity marker detection (Claude Code, Cursor, Windsurf, Cline, Aider, Copilot, Cody, etc.) + length catch-all (>2000 chars) - Handles both top-level 'system' field (Anthropic format) and messages array with role:'system' (OpenAI format) - Preserves original content shape (string vs typed content blocks) - Strips oversized tool descriptions (>64KB) that can also trigger the filter - Replaces with neutral prompt, leaving legitimate user prompts untouched Based on approach from rafilajhh/9router commit 7f7d7ce. * test(codebuddy-cn): add regression coverage for system prompt replacement Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com> --------- Co-authored-by: diegosouzapw <diegosouzapw@users.noreply.github.com> Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com> |
||
|
Aman
|
1423259fec | fix(opencode): fallback unsupported DeepSeek json schema output (#9992) | ||
|
backryun
|
861011c02b | fix(types): preserve GHE Copilot executor configuration (#9988) | ||
|
backryun
|
16c566d146 | fix(copilot-web): restore browser authentication (#9984) | ||
|
Vasily Larin
|
16bf95fe33 |
fix(executors): strip redundant oneOf matching sibling enum (#9828)
* fix(executors): strip redundant oneOf matching sibling enum
The Codex private Responses endpoint intermittently returns a 502 upstream_empty_response for tool parameters that combine oneOf:[{const,...}] with a sibling enum containing the same value set.
When the const and enum sets match exactly, oneOf adds no constraint beyond enum. Add stripRedundantOneOfConstEnum to normalizeCodexTools to remove only this semantically redundant form.
The schema-aware recursive walker requires non-empty, unique string const branches containing annotations only, string enum values, and an exact set match. It preserves bare oneOf[const], narrowing or non-matching sets, type-discriminated oneOf, empty oneOf, non-string values, and anyOf/allOf.
Run the normalization after stripUnsupportedRegexPatterns and before assigning tool.parameters. Add focused regression coverage for matching, non-matching, nested, immutable, and Chat-to-Responses cases.
* docs(changelog): update PR number in changelog fragment
|
||
|
Chewji
|
029a43359c |
fix(command-code): include tool call arguments (#9821)
* fix(command-code): normalize malformed tool call arguments and fix test assertion handling * fix(command-code): resolve toolName from assistant calls and update version header to 1.15.1 * refactor(command-code): consolidate pre-pass message tool metadata extraction and add unknown fallback test * fix(command-code): fallback unnamed tool calls to unknown to satisfy upstream name validation * fix(db): rename 139_job_registry -> 143 to avoid collision with 139_ccr_blocks release/v3.8.50 owns version 139 (ccr_blocks, #9061). The #9631 job registry cherry-pick ( |
||
|
Mo'men Qatr
|
5106dd4ceb | fix: pass max reasoning effort through by default, add global model registry fallback (#8057) (#9612) | ||
|
Diego Rodrigues de Sa e Souza
|
c61cdc30aa |
fix(providers): switch minimax from claude to openai format so images work (#9463)
* fix(providers): switch minimax from claude to openai format so images work The Anthropic-compatible /anthropic/v1/messages endpoint rejects image input with 403. MiniMax's OpenAI-compatible /v1/chat/completions endpoint supports image_url natively for MiniMax-M3. - minimax + minimax-cn: format claude→openai, baseUrl→/v1/chat/completions - Remove Anthropic-Version header + ?beta=true suffix (not needed for openai) - Remove minimax/minimax-cn from ?beta=true executor case - Update cache-control tests (openai format uses different caching path) - Fix reasoning-split test names (no longer claude format) TDD: 2 registry tests assert format=openai (red→green). Refs: Hermes Agent #15715, MiniMax OpenAI-compatible API docs. * fix(sse): re-align stream-readiness-policy tests with minimax's openai format PR #9463 switched minimax/minimax-cn from claude to openai format so images work. The stream-readiness bump for Claude-format replicas is keyed off the registry's format field (single source of truth), so minimax legitimately falls out of that group now. Swap the "Claude-format replica" test fixtures to agentrouter (still format: "claude") and add explicit coverage that minimax no longer gets the claude_format_heavy_reasoning bump. --------- Co-authored-by: diegosouzapw <diegosouzapw@users.noreply.github.com> |
||
|
ikelvingo
|
f0d976c341 |
[v3.8.50] fix: passthrough non-standard cache token fields for DeepSeek / MiniMax / Bedrock across streaming, non-streaming, and Dashboard paths (#8591)
* fix(#8171): map DeepSeek prompt_cache_hit_tokens into prompt_tokens_details.cached_tokens DeepSeek native API returns cache stats in flat top-level fields (prompt_cache_hit_tokens / prompt_cache_miss_tokens) instead of the standard prompt_tokens_details.cached_tokens. The usage sanitizer (sanitizeUsage / sanitizeResponsesUsage) was stripping these non-standard fields, so clients never received real cache hit counts even when the upstream served cached responses. Changes: - sanitizeUsage(): map prompt_cache_hit_tokens into prompt_tokens_details.cached_tokens when the latter is unset - sanitizeResponsesUsage(): same mapping for input_tokens_details - filterUsageForFormat(): add prompt_cache_hit_tokens and prompt_cache_miss_tokens to the default format allow list so they survive field-level filtering * fix: passthrough non-standard cache token fields for DeepSeek / MiniMax / Bedrock across streaming, non-streaming, and Dashboard paths * fix(sse): shrink cache-hit token passthrough to fit file-size gate PR #8591 added a DeepSeek/MiniMax/Bedrock flat cache-hit-token -> nested prompt_tokens_details.cached_tokens mapping (#8171) that grew responseSanitizer.ts and stream.ts past their frozen file-size baselines. - Extract the chat-completions/Responses-API mapping logic into a new leaf module (responseSanitizer/cacheHitTokens.ts). - Move the streaming-path rebuild into filterUsageForFormat() (usageTracking.ts), the single conversion chokepoint both stream.ts call sites already used, eliminating the duplicated stream.ts patch entirely. - Rebaseline responseSanitizer.ts by the 2 lines that remain irreducible (the mandatory ES import for the extracted helper). Behavior verified unchanged via the existing response-sanitizer and stream-handler unit suites. Co-authored-by: ikelvingo <ikelvingo@users.noreply.github.com> --------- Co-authored-by: diegosouzapw <diegosouzapw@users.noreply.github.com> Co-authored-by: ikelvingo <ikelvingo@users.noreply.github.com> |
||
|
jhordanjw123
|
b4fc835d25 |
[v3.8.50] feat(providers): add support for TinyCMS Web (#8736)
* feat(providers): add support for TinyCMS Web including WASM-based cryptographic signing and Proof-of-Work emulation * feat(providers): add unit tests, ESLint suppressions, and fix hardcoded userid for TinyCMS Web - Add unit tests for WASM init, UUID validation, challenge flow (15 tests) - Add WASM source comment explaining binary origin - Replace hardcoded userid with dynamic provider-specific data - Add ESLint suppressions for no-explicit-any in WASM bridge code - Add explanatory comments for DOM shim (runtime WASM-bindgen, not test mocks) Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com> * refactor(providers): extract TinyCMS DOM shims into an explicit setup function tinycmsSigner.ts installed its window/document/HTMLCanvasElement/ CanvasRenderingContext2D shims for the wasm-bindgen glue as a module-load side effect. That meant merely importing the module (even transitively, e.g. through the provider registry from an unrelated test) mutated global state for the rest of the test process. Extract the shim installation into setupDomMocks(), which returns a restore callback: - initTinyCmsWasm() calls it once before instantiating the WASM module (production path — unchanged behavior, still automatic). - tests/unit/provider-tinycms-web.test.ts now calls it explicitly in a `before` hook and restores the previous globals in `after`, so the shims never leak into other test files. As a side effect, replacing five separate `as any` casts with a single typed `global as Record<string, any>` handle drops the file's no-explicit-any count from 5 to 1; eslint-suppressions.json updated to match. Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com> * docs(providers): regenerate PROVIDER_REFERENCE.md for tinycms-web Mechanical `npm run gen:provider-reference` run after merging release/ v3.8.50 into this branch — the generated table was stale for both the new tinycms-web entry this PR adds and the release's own cheaperinference addition. Total providers 290 -> 292, Web Cookie Providers 31 -> 32. Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com> --------- Co-authored-by: diegosouzapw <diegosouzapw@users.noreply.github.com> Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com> |
||
|
Chewji
|
e58c5ee060 |
fix(command-code): preserve literal max effort for command-code provider (#9257)
* fix(command-code): preserve literal max effort for command-code provider * test(command-code): type the new sanitizeReasoningEffortForProvider assertions The 3 new command-code reasoning-effort test cases cast the function's unknown return value with `as any`, which pushes the file's frozen no-explicit-any suppression count (48) to 51 and trips the "No new ESLint warnings" gate. Use a minimal EffortCarrierResult shape instead of any, matching the fields the assertions actually read. Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com> * test(v1-models): type the API key lookup in the #9320 auth-leak regression test The release-tip test file added by #9320 used `(k: any)` in an Array.find callback, which is not covered by config/quality/eslint-suppressions.json (the file was added after the suppressions snapshot was frozen). That leaves the "No new ESLint warnings" gate red for any branch that merges this exact release/v3.8.50 tip, unrelated to this PR's own diff. Fixing it here with a minimal derived type (Awaited<ReturnType<typeof getApiKeys>>[number]) unblocks the gate without touching the frozen suppressions baseline. Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com> --------- Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com> |
||
|
Will Gordon
|
57744aeb14 |
feat(cursor): proactively renews Cursor sessions and fixes manual refresh (#9173)
* refactor(cursor): extracts token extraction into shared lib
Moves tryIdeAuth/tryAgentAuth and supporting helpers out of the
auto-import route into src/lib/cursor/tokenExtractor.ts, and adds
an agent-cli-state.json fallback candidate path to tryAgentAuth
(alongside the existing auth.json candidate) so the extraction
logic can be reused by the upcoming renewal orchestrator.
* feat(cursor): adds cursor-agent-backed token renewal orchestrator
Builds the renewal orchestrator in src/lib/cursor/renewal.ts: a
bounded, unattended-safe --list-models nudge, a side-effect-free
status availability check, an in-flight spawn lock keyed by
command, and renewCursorConnection() which nudges cursor-agent
then independently re-scrapes the IDE and cursor-agent credential
sources to detect whichever refreshed. Extends cursorAgent.ts's
binary resolution and spawn helper with fixed-paths-only mode and
a SIGKILL follow-up for background use. Adds a generic keyed-mutex
utility (src/shared/utils/keyedMutex.ts) for serializing a
connection's renew-then-persist cycle, and forwards a busy-timeout
through driverFactory's node:sqlite fallback path.
* feat(cursor): proactively renews Cursor sessions in the sweep
Adds src/lib/tokenHealthCheckCursor.ts, sweep-side glue that calls
the renewal orchestrator and persists the result, wired into
tokenHealthCheck.ts's checkConnection() via a new Cursor-specific
branch placed ahead of the generic no-refresh-token fallthrough.
Carves out a non-terminal exception for a Cursor connection that
already landed at testStatus "expired" via the request-time 401
path, excluding permanently-dead account_deactivated connections.
Extends buildRefreshFailureUpdate() with an overrides param so
Cursor's failure path can use a distinct, non-terminal errorCode
instead of the generic refresh_failed/expired taxonomy.
* feat(cursor): adds local-only manual refresh route
Adds POST /api/providers/[id]/refresh-cursor, a dedicated
loopback-only route that calls the renewal orchestrator on demand
for a single Cursor connection, bounded by a 30s per-connection
cooldown. Classifies the new route in LOCAL_ONLY_API_PATTERNS and
closes the manage-scope-bypass gap for dynamic-segment spawn-capable
routes under /api/providers/ via a new SPAWN_CAPABLE_PATTERNS /
SPAWN_CAPABLE_PATTERN_ANCESTORS mechanism, which also retroactively
covers the pre-existing /login route. The existing shared
/api/providers/[id]/refresh route is untouched and stays
remote-reachable for every other provider.
* feat(cursor): surfaces a dismissible cursor-agent nudge
Adds GET /api/providers/cursor/agent-availability, a credential-free
LOCAL_ONLY route returning only { cursorAgentAvailable: boolean },
backed by a 5-minute cached wrapper around the renewal orchestrator's
existing availability check. Surfaces a dismissible dashboard banner
on the Cursor provider page suggesting cursor-agent installation
when it isn't detected, following the existing dismissible-banner
convention. Also fixes a pre-existing bracket character in a
routeGuard.ts comment that was silently truncating
check-openapi-security-tiers.mjs's view of LOCAL_ONLY_API_PREFIXES.
* fix(cursor): wires manual refresh button to the new route
Branches handleRefreshToken to call the dedicated Cursor refresh
route instead of the generic /refresh route, which silently 502s
for Cursor connections today since they carry no refresh token.
Every other provider's refresh behavior is unaffected. Adds the
cursorSessionUnchanged i18n key and syncs it (plus a pre-existing,
unrelated 28-key backlog) across all 42 locale files.
* fix(cursor): addresses Phase 4/4.5 review findings
Restores the legacy stdout/stderr auth-pattern fallback in
checkCursorAgentAvailability() that the plan's Task 2 Step 4
required but the implementation had dropped. Threads an optional
deps parameter through checkCursorConnectionIfNeeded() so its
error branch is reachable in tests, and switches both it and the
manual-refresh route to exhaustive switch statements over the
renewal result. Adds a short-lived host-keyed dedup cache around
tryIdeAuth() so multiple due Cursor connections sharing a host
don't each open the same state.vscdb file in one sweep tick.
Adds opportunistic eviction to the manual-refresh cooldown map,
an outer try/catch to the availability route for defense-in-depth
consistency with the plan's other routes, and corrects a stale
JSDoc claim about the /login route's auth check. Documents the
now-empirically-confirmed agent-cli-state.json schema mismatch
found while validating against a real cursor-agent install.
* docs(cursor): adds changelog fragments for the renewal plan
Adds one fragment per user-facing outcome per changelog.d/README.md's
convention for a PR that both fixes and adds. PR number placeholder
to be filled in once the PR is opened.
* fix(i18n): translates the new Cursor keys into Vietnamese
The i18n:sync-ui run in an earlier commit left __MISSING__
sentinels for the 4 new Cursor keys in every locale, but
Vietnamese has a dedicated completeness test requiring zero
internal missing markers. Provides real translations for
cursorSessionUnchanged, cursorAgentNudgeTitle,
cursorAgentNudgeBody, and cursorAgentNudgeDismiss.
* fix(cursor): addresses quality-gate Layer 1.5 findings
Restores a comment that misrepresented execFile's actual argv shape
after an earlier bracket-removal fix, this time avoiding literal
closing-bracket characters entirely so the openapi checker's naive
array parser can't be broken by either version. Bounds the sweep-
and manual-route-triggered tryIdeAuth() busy-timeout to 250ms
(down from the interactive auto-import path's 2000ms), since both
share the main event loop with all other in-flight requests and
should fail fast on a WAL-lock collision rather than block the
whole instance for up to ~4s. Has the manual refresh route bypass
the sweep's IDE-auth dedup cache so a click always sees a fresh
read, consistent with this plan's existing "manual actions never
see stale cached data" convention. Documents the previously-missing
agent-availability route in ROUTE_GUARD_TIERS.md's spawn-capable
table.
* fix(cursor): adds SIGKILL follow-up to the status-check spawn
Matches the nudge spawn's existing SIGTERM+SIGKILL pattern so an
unresponsive cursor-agent status check can't leak a lingering
process if it ignores SIGTERM.
* docs(cursor): fills in the PR number for changelog fragments
Renames the 3 changelog.d fragments to their PR-numbered filenames and replaces the (#PR) placeholder with #9173, now that the PR exists.
* fix(cursor): corrects changelog fragments to reference PR #9173
The prior commit only staged the git mv rename — a git add invocation with a stale (pre-rename) pathspec aborted before the actual (#PR) -> (#9173) content edit was staged, so the rename landed without the fix it was meant to carry. This captures the actual content change.
* docs(cursor): regenerates the agent-skills catalog for the new route
check:agent-skills-sync (CI's Merge integrity gate) requires SKILL.md files to stay in sync with the live route catalog. Adding /api/providers/cursor/agent-availability in an earlier commit needed a regen this branch never ran.
* chore(quality): rebaselines file-size caps grown by agentrouter merges
Two already-merged agentrouter commits (
|
||
|
Jan Leon
|
32f8340719 |
Fix custom tool output pairing during context compression (#8933)
* Fix custom tool output pairing during compression (#8932) * Bypass proxy compaction for native Codex context * fix(sse): extract Codex tool-call output repair to leaf module for file-size gate repairMissingCodexToolCallOutputs (added by #8932 for custom_tool_call pairing) pushed codex.ts past the frozen file-size baseline. Extract it to open-sse/executors/codex/toolCallRepair.ts, leaving only the wiring call in codex.ts. Rebaseline the test file's genuine +41 line growth from #8932's new custom_tool_call_output coverage. Co-authored-by: JxnLexn <JxnLexn@users.noreply.github.com> --------- Co-authored-by: diegosouzapw <diegosouzapw@users.noreply.github.com> Co-authored-by: JxnLexn <JxnLexn@users.noreply.github.com> |
||
|
Andrew B.
|
9a99a39b33 |
fix(antigravity): quota-aware account selection and projectId persistence (#8891)
* fix(antigravity): per-model quota + 30min credits_exhausted reprobe - accountFallback.ts: hasPerModelQuota() now treats antigravity/agy as per-model quota. A single-model 429 no longer cascades to all models in the provider. - connectionRecovery.ts: credits_exhausted removed from terminal set; isCreditsExhaustedReprobeCandidate() with 30min default. Loads active+inactive rows so inactive credits_exhausted accounts can recover. - tests/unit/quota-connection-recovery.test.ts: 6 cases covering pure helpers + tick wiring. * fix(antigravity): persist projectId and prefer healthy accounts Save Cloud Code projectId after runtime discovery, skip accounts missing projectId when alternatives exist, and mark missing_project_id on 422. * fix(antigravity): skip quota-exhausted models during account selection Avoid repeatedly dispatching to Antigravity models that already report exhausted quota, reducing wasted upstream calls and combo fallback latency. --------- Co-authored-by: hermes <hermes@nous.local> |
||
|
WITALO ROCHA
|
016c3fdb77 |
fix(github): honor per-model targetFormat override for Copilot custom models (#8713)
GithubExecutor.buildUrl() only consulted the static PROVIDER_MODELS registry
via getModelTargetFormat("gh", model), so a custom Copilot model (e.g.
gpt-5.6-terra/gpt-5.6-luna) with its dashboard "Target Format" set to
OpenAI Responses API always still routed to /chat/completions and got
rejected upstream with "model ... is not accessible via the
/chat/completions endpoint" — the setting had no effect on real routing.
chatCore already resolves the correct per-request targetFormat (including
the custom-model override) via resolveChatCoreTargetFormat(), but that value
was never threaded past chatCore into the executor's own URL-building
decision. Mirrors the zai/glm-coding-apikey fix (#7364) for the identical
class of bug: chatCore/executionCredentials.ts now surfaces the resolved
override onto providerSpecificData.targetFormat when it resolves to
openai-responses for the github provider, and GithubExecutor.buildUrl()
prefers that value over the static registry lookup when present.
Verified: 6 new regression tests plus all 95 pre-existing github/executor
tests green.
Co-authored-by: Wital <wital@example.com>
|
||
|
SB Yoon
|
0dbc34ea56 |
feat(cursor): exclusive live listing + verbatim AgentRun model ids (#9911)
* feat(cursor): prefer live synced catalog for listing and Test All When an active synced Cursor catalog exists, list only live models plus injected auto routers (and customs). Keep the static registry as offline fallback. Co-authored-by: Cursor <cursoragent@cursor.com> * fix(cursor): send live-catalog model ids verbatim on AgentRun Skip #7289 effort/reasoning splits when the exact id is in the active synced Cursor catalog so AgentRun does not rewrite flattened live ids into missing bases that return AI Model Not Found. Also wires auto-cost/balance/intelligence to default + optimization for the injected routers. Co-authored-by: Cursor <cursoragent@cursor.com> --------- Co-authored-by: Cursor <cursoragent@cursor.com> Co-authored-by: yansigit <yansigit@users.noreply.github.com> |
||
|
Vasily Larin
|
40f9709071 |
fix(executors): preserve non-strict Codex tool semantics (#9931)
* fix(executors): preserve non-strict Codex tool semantics * docs(changelog): add Codex strict semantics fix |
||
|
backryun
|
a13107731f | fix(types): type Copilot WebSocket construction (#9974) | ||
|
diegosouzapw
|
b7bad4006b | fix(release): clear remaining Adobe and typecheck gates | ||
|
Diego Rodrigues de Sa e Souza
|
05ab06f3a1 |
fix: address self-review findings (#9900)
Co-authored-by: Matt Van Horn <455140+mvanhorn@users.noreply.github.com> |
||
|
Diego Rodrigues de Sa e Souza
|
740e16c3e2 |
cherry-pick(pr-9601): feat(responses): add encrypted reasoning replay opt-in (#9876)
* feat(codex): add encrypted reasoning replay opt-in * feat(responses): generalize encrypted reasoning replay * docs: clarify encrypted reasoning provider scope * fix(ui): group reasoning replay with connection controls * fix(logs): omit encrypted reasoning payloads * fix(chat): reduce file size Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com> * fix(chat): reduce combined file size Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com> --------- Co-authored-by: jackjinke <jack.kejin@gmail.com> |
||
|
Diego Rodrigues de Sa e Souza
|
3a66761cb7 |
fix: pass max reasoning effort through by default, add global model registry fallback (#8057) (#9883)
Co-authored-by: Mo'men Qatr <momen.qatr04@eng-st.cu.edu.eg> |
||
|
Diego Rodrigues de Sa e Souza
|
356fd5d606 |
fix(executors): repair DuckDuckGo AI Chat challenge solver (418 ERR_CHALLENGE) (#9866)
Every duckduckgo-web chat request failed with HTTP 418 ERR_CHALLENGE while
duck.ai worked normally in a browser from the same IP. Ground truth was
established by driving a real headful Chromium at duck.ai from that IP (it
returned 200), so the environment was never the problem — the anti-abuse
challenge solver was. Six independent defects were found; the first alone
disabled the solver completely.
1. Module syntax inside the vm sandbox source.
CHALLENGE_STUBS is executed with vm.runInContext, which compiles in SCRIPT
mode. A refactor mass-added `export` to the five `function` declarations
inside that template literal (they read as ordinary top-level TS functions),
so every solve threw SyntaxError. The executor swallows solve failures and
posts the raw unsolved challenge, which upstream answers with 418.
2. Double-escaped regex in a String.raw template.
`\\s` in __parseCssDisplay reached the sandbox as a literal backslash, so the
display regex never matched and a getComputedStyle probe silently read empty.
3. buildHtmlLookup undercounted descendants by one.
`count` backs el.querySelectorAll('*').length; that returns DESCENDANTS and
countHtmlElements already skips the #document-fragment root, so the `- 1` was
wrong. Chromium reports 3 for '<li><div></li><li></div'; we reported 2, and a
variant multiplies innerHTML.length by that count.
4. Browser-fidelity probes.
Newer challenge variants assert JS/DOM invariants a flat stub cannot satisfy:
real prototype chains (HTMLDivElement -> HTMLElement -> Element), NodeList
identity, a live body.children HTMLCollection, native-code toString, and
sloppy-mode `this === window`. Nine of thirteen failed. Notably Math must NOT
be sealed — Chromium reports Object.isSealed(Math) === false, and sealing it
made our vector differ by one.
5. The solved payload dropped meta.origin / meta.stack / meta.duration.
The duck.ai bundle always sends all three; captured browser requests confirm
it. Without them upstream returns 418 even when every client_hash is correct.
6. reasoningEffort is now mandatory on duckchat/v1/chat.
An otherwise byte-identical payload returns 200 with the field and 400
ERR_BAD_REQUEST without it (A/B verified live, repeated).
Also removes the throwaway "seed" chat POST that ran before every real request.
It existed to coax a usable challenge out of the upstream while the solver was
broken; it only doubled chat calls against an IP-rate-limited endpoint, showing
up as spurious 429 ERR_RATE_LIMIT.
Verification: the solver now reproduces real Chromium's probe vectors exactly
for all 8 captured challenge variants, and the executor returns 200 end-to-end
live (non-streaming, streaming, claude-haiku-4-5, and a math prompt returning
"42").
Tests: tests/unit/duckduckgo-challenge-solver-regression.test.ts (32 tests) and
tests/unit/duckduckgo-reasoning-effort-required.test.ts (5 tests), backed by
tests/fixtures/duckduckgo/challenge-variants.json — real captured challenge
programs plus the probe vectors a real browser produced for them, so the suite
asserts against recorded browser behaviour rather than our own output. Each fix
was confirmed to fail its test when individually reverted.
Co-authored-by: Mynacol <git@mynacol.xyz>
|
||
|
Diego Rodrigues de Sa e Souza
|
a7d2dba1eb |
fix(types): narrow DeepSeek tool calls (#9860)
Co-authored-by: backryun <bakryun0718@proton.me> |
||
|
Diego Rodrigues de Sa e Souza
|
a102a2d773 |
maint: final follow-up cherry-pick #9783 (#9904)
* fix(deps): bump transitive deps for 6 Dependabot + remaining audit vulns on main Same overrides as #9464 (ip-address, hono, fast-uri, socket.io-parser, undici) applied directly to main. Also covers brace-expansion (scoped), js-yaml v4 copies, and mermaid. npm audit: 6→0 vulnerabilities. Closes Dependabot #161-#166. * fix(deps): bump nanoid, dompurify for 2 new Dependabot alerts (#189, #190) Bumps: nanoid ^3.3.17 (was transitive, now overridden), dompurify ^3.4.13 (with monaco-editor scoped override). Closes Dependabot #189, #190. Remaining #182-#188 (js-yaml + mermaid) already closed by #9651 merge — awaiting Dependabot re-scan. npm audit → 0 vulnerabilities. * fix(repo): harden .gitignore to also ignore a _tasks symlink (/_tasks) _tasks is a SEPARATE nested git repo (gitignored). The pattern _tasks/ (trailing slash) ignores only a directory, not a SYMLINK named _tasks. A self-referential _tasks symlink can slip in via git add -A and, once pulled, checkout materializes it over the real _tasks repo (destroying plans/specs/hands-off). Anchored /_tasks ignores the symlink too, preventing re-capture. * fix(translator): keep Responses namespace identity across the hub-and-spoke pivot Step 1 of the pivot (openai-responses -> openai) flattens namespace sub-tools to a qualified wire name (#8295) and records the `{namespace, name}` pair on a non-enumerable `_toolNameMap`. Step 2 (openai -> target) returns a brand-new object, so the property was dropped for every non-OpenAI target. chatCore then handed `null` to the #7936 response seam and namespace sub-tool calls reached the client under their flattened name, which Codex rejects with `unsupported call: <name>` — the symptom #7936 was opened to fix. Copying `_toolNameMap` through is not viable: openai-to-claude and openai-to-gemini publish their own `Map<string, string>` alias map on that same property during step 2, so it carries two incompatible types. This adds a dedicated `_namespaceToolIdentityMap`, propagated by translateRequest across the pivot; chatCore prefers it and falls back to `_toolNameMap` for the non-pivot producers. Both keys are stripped from the cliproxyapi wire body. Fixes #9780 * fix(chat): reduce file size Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com> * fix(chat): reduce combined file size Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com> * fix(chat): reduce combined file size Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com> --------- Co-authored-by: diegosouzapw <diegosouzapw@users.noreply.github.com> Co-authored-by: VXNCXNX <vincent@preuve.ai> |
||
|
Diego Rodrigues de Sa e Souza
|
5926d35758 |
cherry-pick(pr-9787): fix(sse): apply Azure param rules on azure-ai and clamp gpt-4o-mini output tokens (#9855)
* fix(sse): apply Azure request-param rules on the azure-ai wire path
Azure rejects several stock Chat Completions params on its newer deployments
and returns HTTP 400 rather than ignoring them:
max_tokens -> 'max_tokens' is not supported with this model.
Use 'max_completion_tokens' instead.
reasoning_effort -> Function tools with reasoning_effort are not supported.
Those rules lived inline in AzureOpenAIExecutor, so they only covered the
azure-openai provider. azure-ai (Azure AI Foundry) had no executor entry and
fell through to the bare DefaultExecutor, so the SAME Azure deployment
succeeded on one connection and 400'd on the other. Every agentic client sends
tools on every turn, so azure-ai failed on the first request.
Extract the rules to open-sse/executors/azureParamRules.ts, add an
AzureAiExecutor that inherits DefaultExecutor's azure-ai URL/header/apiType
handling unchanged and applies the shared rules, and register it for azure-ai.
Also widen the deployment pattern to cover gpt-chat-latest: it is a moving
alias that resolves to a GPT-5-era model and rejects max_tokens, but carries no
version number for the token-boundary pattern to key on. Verified against the
base regex - gpt-chat-latest did not match, which is exactly the observed 400.
Regression guard: tests/unit/azure-param-rules.test.ts, including an assertion
that getExecutor("azure-ai") no longer resolves to a bare DefaultExecutor.
* fix(sse): clamp Azure gpt-4o-mini completion tokens to its 16384 ceiling
Azure gpt-4o-mini deployments accept at most 16384 completion tokens and 400 on
anything larger:
max_tokens is too large: 32000. This model supports at most 16384 completion
tokens, whereas you provided 32000.
The 32000 is OmniRoute's own doing: adjustMaxTokens raises any smaller
max_tokens to DEFAULT_MIN_TOKENS (32000) whenever tools are present, to avoid
truncated tool arguments. That floor has no upper bound, so an agentic client
asking for far less still trips the model ceiling on its first turn.
Add scoped maxOutputCap rules in paramSupport.ts for both Azure wire paths.
PROVIDER_MAX_TOKENS is the wrong lever here - it is provider-wide, and the same
Azure resource also serves GPT-5 deployments with a much higher ceiling.
Regression guard: tests/unit/azure-max-output-clamp.test.ts, which also pins
that the clamp does not leak to gpt-5.1 or to gpt-4o-mini on other providers.
---------
Co-authored-by: Mihaly Bodo <michael@proton-quantum.com>
|