Repository navigation
feat(providers): restore per-connection upstream timeout tier - #10885
Merged
diegosouzapw merged 65 commits intoAug 21, 2026
Merged
diegosouzapw merged 65 commits into
diegosouzapw merged 65 commits into
Conversation
providerSpecificData.timeoutMs now preempts the model/provider registry tiers and the global FETCH_TIMEOUT_MS in upstream start-timeout resolution, and the combo per-target timeout ceiling follows the selected connection.
…CT_REQUIRED (cheaper-inference#10470) Obrigado — follow-up bem feito do cheaper-inference#10424: um 422 GCP_PROJECT_REQUIRED (BYOP — a conta Google precisa trazer seu próprio GCP Project) é específico da CONTA, não do provider inteiro, então o fast-fail anterior falhava mesmo quando uma conta antigravity irmã saudável poderia atender o request. Agora rotaciona automaticamente para a conta irmã (excluindo a conta BYOP por 24h) e só surfaça o erro acionável quando não há irmã disponível. Estado de rotação rastreado separado de maxAttempts, então falhas normais do antigravity nunca ganham uma segunda chance (sem dispatch duplo). Validação (worktree própria a partir de origin/release/v3.8.50, merge limpo — auto-merge em chatCore.ts, 0 conflitos reais): - typecheck:core limpo, complexity/cognitive-complexity dentro do baseline - tests/unit/antigravity-byop-account-rotation.test.ts + error-classifier.test.ts — 36/36 passando
…ty, confidence, TTFT/ITL, end-to-end test (cheaper-inference#10881) Obrigado — feature substancial e bem estruturada: separa qualidade operacional (comportamento de wire: 4xx/5xx, 429, respostas malformadas, stream interrompido) de qualidade semântica (só setada por avaliadores externos, nunca inferida do sucesso HTTP), com confidence/sample-awareness para não deixar poucos sucessos de sorte dominarem o ranking. Instrumentação de streaming (TTFT/ITL) threaded até RoutingEvent, endpoint de explicabilidade, e teste E2E determinístico cobrindo degradação→recuperação→blip. Validação (worktree própria a partir de origin/release/v3.8.50, merge limpo, 0 conflitos): - typecheck:core limpo, complexity/cognitive-complexity dentro do baseline - 59/59 testes passando (mlx-provider, routing-adaptive-e2e, routing-events(-concurrency), routing-otel, routing-quality, routing-scoring-quality, stream-timing, auto-combo-scoring-clamp)
…guard (cheaper-inference#10887) Merged — clean single-commit extraction from cheaper-inference#9115's genuinely new content (see PR body for the full extraction rationale: 66-commit branch, only 1 commit matched the stated scope). typecheck/file-size/changelog gates clean, 19/19 unit + 14/14 integration tests passing.
…ce#10197) (cheaper-inference#10363) Obrigado — feature real e bem verificada: POST /v1/images/edits rejeitava o provider built-in openrouter mesmo ele suportando edição por imagem de referência via sua Image API unificada. Traduz a imagem de entrada para o formato input_references documentado do OpenRouter e despacha para /api/v1/images, removendo o prefixo do provider do model id antes de encaminhar. Nota: o contribuidor não conseguiu rodar o teste localmente (better-sqlite3 ausente no ambiente dele) — rodei aqui. Validação (worktree própria a partir de origin/release/v3.8.50, merge limpo, 0 conflitos): - typecheck:core limpo, complexity/cognitive-complexity dentro do baseline - tests/unit/10197-openrouter-image-edits-route.test.ts — 3/3 passando (forward bem-sucedido, credenciais ausentes 401, rate-limit)
Merged — clean extraction from cheaper-inference#10358's genuinely new content (see PR body for the rationale: an unrelated .planning/codebase/ scaffolding dump was dropped). typecheck/file-size/changelog/provider-consistency gates clean, 18/18 tests passing.
…heaper-inference#10900) Merged — reimplementation extracting the non-conflicting Recent Requests panel + excludeTests allowlist fix from cheaper-inference#8450 (see PR body for the full scoping rationale, including why the topology UX rework was deliberately excluded — it contradicts the already-shipped cheaper-inference#8428). typecheck/file-size/changelog/complexity/cognitive-complexity/i18n-coverage gates all clean, 2/2 unit + 1/1 vitest passing.
…ference#10904) Merged — extraction of the one still-uncovered fix from cheaper-inference#8634 (the other two items — mode "search"→CONCISE downgrade, pplx-opus generation — were already applied on this release tip). typecheck/file-size/changelog/complexity/cognitive-complexity gates all clean, 32/32 tests passing.
…rect-dispatcher-timeout-10214 fix(network): bound direct-path response-start timeout and retry on fresh socket (cheaper-inference#10214)
…e#10673) Obrigado — bug real: GET /v1/files aceitava limit negativo sem validação (`-5 || 20` avalia truthy em -5, então Math.min(-5, 10000) = -5 passava direto). Agora valida integer/positivo/tamanho e retorna 400 estruturado para valores inválidos, preservando o default 20 e o máximo 10.000. Validação (worktree combinado a partir de origin/release/v3.8.50, 0 conflitos): - typecheck:core limpo, complexity/cognitive-complexity dentro do baseline - tests/integration/files-api-limit-validation.test.ts — 5/5 passando - tests/integration/files-api.test.ts — 12/12 passando (sem regressão) - tests/unit/batch_api.test.ts teve 1 falha, confirmada DRIFT pré-existente idêntica no tip puro do release (não relacionada, timing de cancelamento de batch)
…eToGeminiRequest (cheaper-inference#10658) Obrigado — bug real: a tradução direta claudeToGeminiRequest emitia mensagens consecutivas do mesmo role em contents[], o que a API do Gemini rejeita com HTTP 400 (turnos alternados user/model são obrigatórios). Traz claudeToGeminiRequest à paridade com openaiToGeminiRequest reutilizando mergeConsecutiveSameRoleContents. Validação (worktree combinado a partir de origin/release/v3.8.50, merge limpo, 0 conflitos): - typecheck:core limpo, complexity/cognitive-complexity dentro do baseline - tests/unit/claude-to-gemini-consecutive-roles.test.ts — 7/7 passando - tests/unit/claude-to-gemini-budget-tokens-zero-6813.test.ts — 2/2 passando (sem regressão)
…ols (cheaper-inference#10668) Obrigado — PR muito bem documentado e verificado. Adiciona o gateway TabiToken (Anthropic-first, /v1/messages, x-api-key) e estende hcnsec de 1 para 4 protocolos (Chat, Responses, Anthropic Messages, Gemini). AlternateFormat ganha o hook urlBuilder opcional (necessário para o path model-scoped do Gemini), compartilhado com o provider gemini nativo em vez de duplicado. Reconciliado nesta sessão contra o release tip atualizado (base drift real: 343→345 canônicos entre quando o PR foi criado e o merge, mais os PRs cheaper-inference#10673/cheaper-inference#10658 mergeados nesse meio-tempo). Conflitos em contagens de providers (docs, file-size baseline, teste de partição) resolvidos additivamente. Validação (reconciliação a partir de origin/release/v3.8.50): - typecheck:core limpo, complexity/cognitive-complexity dentro do baseline - npm run check:provider-consistency — OK (266 REGISTRY entries, 346 providers canônicos, 0 exceções) - 40/40 testes passando (newapi-gateway-providers, hcnsec-provider, providers-constants-split, alternate-formats)
…ngEfforts (cheaper-inference#10788) glm-5.1, glm-5.2, deepseek-v4-pro and deepseek-v4-flash declared supportsReasoning:true but no supportedThinkingEfforts, so the catalog's appendSyncedEffortVariants() pass (which only synthesizes -low/-high/-max ids from an already-populated capabilities.effort_tiers) never exposed a selectable effort tier for them, unlike gpt-oss:20b/120b. Add the documented low/medium/high/max vocabulary (see supportsMaxEffortForProvider's isOllamaCloud comment in reasoningEffort.ts).
…00 (cheaper-inference#10849) v1SearchSchema.provider was a hard-coded z.enum that rejected any id outside its list before the route's own resolveSearchProvider() check ever ran, so unknown/short-alias provider ids (grok, brave, serper, ...) always surfaced a generic "Invalid request" instead of the informative "Unknown search provider: <id>" message. Relax the schema to a free-form string and let resolveSearchProvider() own runtime validation (as it already did for ids that passed the enum). Also extend SEARCH_PROVIDER_ALIASES with short-form aliases mirroring the existing jina/jina-ai pattern (brave, serper, perplexity, exa, tavily, google-pse, linkup, ollama, searchapi, youcom, searxng, zai, duckduckgo), and surface the first Zod validation issue's field name instead of the generic message for other still-invalid fields (e.g. search_type).
…apshot (cheaper-inference#10156) Live SSE frames for a phase:"commentary" message were already dropped per cheaper-inference#6199, but the terminal response.completed.response.output array was forwarded verbatim whenever the upstream echoed the same item back non-empty, since backfillResponsesCompletedOutput only fills an empty array. Reuse the existing isResponsesCommentaryMessageItem predicate to filter the terminal snapshot's output array (and, defensively, the backfill buffer it can be seeded from) so both representations agree. Regression test added to tests/unit/responses-commentary-passthrough-6199.test.ts reproducing the exact upstream shape from the issue.
…5-antigravity-multiaccount-quota fix(domain): treat unreported Antigravity quota fraction as unknown, not exhausted (cheaper-inference#10095)
…6-responses-commentary-sse fix(sse): strip commentary items from Responses response.completed snapshot (cheaper-inference#10156)
…2-playground-endpoint-selector fix: route Playground ChatTab Send to the selected endpoint (cheaper-inference#10592)
…7-combo-log-error-body fix: log upstream error body in COMBO per-target failure warnings (cheaper-inference#10597)
…5-rtk-unconditional-stats fix: skip expensive RTK compression stats computation on no-op runs (cheaper-inference#10765)
…8-opencode-effort-tiers fix(open-sse): declare Ollama Cloud reasoning models' supportedThinkingEfforts (cheaper-inference#10788)
…eaper-inference#6793) (cheaper-inference#10531) Adds native support for Freebuff (Codebuff CLI free-tier gateway): executor with upstream session acquisition, agent-run lifecycle (START/FINISH), canonical system-prompt injection, and model→agent mapping for 9 free models; registry entry, dashboard branding/icons, and API-key validation. Closes cheaper-inference#6793. Validated in an isolated worktree boarded onto origin/release/v3.8.50 (0 conflicts after resolving generated provider-count drift, 15 files): - 9/9 focused tests pass (freebuff-provider, providers-constants-split). - Dropped one out-of-scope, unrelated hunk in scripts/build/build-next-isolated.mjs (build-memory heap tuning) that had nothing to do with the Freebuff provider — kept the branch scoped to its stated purpose. - check-changelog-integrity: OK. - file-size: gateways.ts and AddApiKeyModal.tsx crossed the frozen cap by +15/+5 lines (irreducible catalog-entry + credential-hint additions) — rebaselined with justification, pushed fix-in-place to the PR branch. - typecheck:core: clean. - check-complexity / check-cognitive-complexity: OK, both under baseline. Co-authored-by: adrianaryaputra <adrianaryaputra@users.noreply.github.com>
…instead of hanging silently (cheaper-inference#10463) Guarantees combo/target loops always terminate with an actionable error instead of hanging when an upstream hangs — fixes 10 silent-stop gaps (G1–G10): combo-loop safety timer + abort on hang → 504, unexpected task throw → 502, round-robin safety net, chaos all-panel-failure visibility, ReDoS-safe eval regex, autoRefreshDaemon swallowed-error logging, batch-item wall-clock timeout. Validated in an isolated worktree boarded onto origin/release/v3.8.50 (2 real conflicts in combo.ts and chaosEngine.ts, both additive features from concurrently-merged PRs landing at the same insertion point — resolved by combining both, verified no variable-shadowing/scoping issues): - 10/10 new regression tests pass (combo-silent-stop-gaps.test.ts): hung upstream → 504, unexpected throw → 502, chaos all-fail logged, ReDoS regex rejected, batch timeout. - 163/163 broader focused tests pass (combo-routing-engine, combo-target-timeout-runner, pipeline-router, chaos-executor, batch-processor ×2, service-batch-processor, evalrunner-builtinsuites-split). - check-file-size, check-changelog-integrity: OK. - typecheck:core: clean. - check-complexity / check-cognitive-complexity: OK, both under baseline. Co-authored-by: herjarsa <herjarsa@users.noreply.github.com>
…eaper-inference#10354) Adds a snapshot-generation button to each Auto-Combo catalog card: computeSnapshotWeights() scores candidates (taskFit/stability/tierPriority/costInv) at combo-creation time instead of the previous hardcoded weight:1, and the new POST /api/combos/duplicate endpoint materializes any auto/* template into a persistent, editable static combo with normalized weights. Closes cheaper-inference#10231. Validated in an isolated worktree boarded onto origin/release/v3.8.50 (0 conflicts, 51 files): - 19/19 focused tests pass (snapshot-weights, combos-duplicate-route, combos-duplicate-resolution-audit) — covers auth gate (401/403), input validation (400/422), success shape, weight normalization, naming/dedup, and error-response sanitization (no stack traces). - check-file-size, check-changelog-integrity: OK. - typecheck:core: clean. - check-complexity / check-cognitive-complexity: OK, both under baseline. Co-authored-by: swingtempo <swingtempo@users.noreply.github.com>
…heaper-inference#10810) Adds Zed Hosted Models to OAUTH_TEST_CONFIG so the dashboard connection test no longer reports "Provider test not supported" for that provider. Validated in an isolated worktree boarded onto origin/release/v3.8.50 (0 conflicts, 1 file): - tests/unit/oauth-test-config-8408.test.ts (including its "every OAuth provider ID has an OAUTH_TEST_CONFIG entry" check) — 3/3 pass. - check-file-size, check-changelog-integrity: OK. - typecheck:core: clean. - check-complexity / check-cognitive-complexity: OK, both under baseline. Co-authored-by: Hsia97 <Hsia97@users.noreply.github.com>
…heaper-inference#10874) OpenCode Zen serves muse-spark-1.2 and muse-spark-1.2-contributor-free only on the OpenAI Responses API endpoint, not /chat/completions. Declares targetFormat: "openai-responses" for both so requests route correctly instead of returning null/empty content. Closes cheaper-inference#10867. Validated in an isolated worktree boarded onto origin/release/v3.8.50 (0 conflicts, 1 file): - Added a TDD regression test (tests/unit/opencode-muse-spark-responses-10867.test.ts) since the PR had none — confirmed RED against origin/release/v3.8.50 (entries absent) and GREEN on this branch, pushed fix-in-place. - check-file-size, check-changelog-integrity: OK. - typecheck:core: clean. - check-complexity / check-cognitive-complexity: OK, both under baseline. Co-authored-by: zoser69 <zoser69@users.noreply.github.com>
…r-inference#10803) Re-runs hoistLeadingSystemMessage on the final outbound array at translateRequest's single return, instead of only pre-translation. claudeToOpenAI (and the Responses source path, which never ran the pre-translation hoist at all since `messages` doesn't exist yet there) re-introduces/normalizes a leading system message after the hoist already ran, so a strict provider (e.g. vLLM/Qwen3, xiaomi-mimo) could still receive a non-compliant array and 400 with "System message must be at the beginning." Validated live against a vLLM/Qwen3 endpoint (documented in the PR) plus in an isolated worktree boarded onto origin/release/v3.8.50 (0 conflicts, 2 files): - 39/39 focused tests pass (probe-7293-strict-system-hoist including the new Claude-source regression case, memory-system-first-6135, claude-system-role-cache-boundary, memory-cache-safe-injection). - check-file-size, check-changelog-integrity: OK. - typecheck:core: clean. - check-complexity / check-cognitive-complexity: OK, both under baseline. Co-authored-by: Kizuno18 <Kizuno18@users.noreply.github.com>
…ent card) (cheaper-inference#10839) Adds A2A v1.0 client compatibility: aliases the renamed v1.0 method names (SendMessage → message/send, SendStreamingMessage → message/stream) and reshapes the synchronous reply into the v1.0 SendMessageResponse shape (task.status.message.parts[].text, task.artifacts) for requests that arrive via a v1.0 method — v0.3 callers keep the exact existing response. Also serves a v1.0 Agent Card at /.well-known/agent-card.json declaring both protocol versions on the same JSON-RPC endpoint. The PR had no tests and the author noted they couldn't build/typecheck locally — validated in an isolated worktree boarded onto origin/release/v3.8.50 (0 conflicts, 2 files): - Wrote a TDD regression test (tests/unit/a2a-v1-compat-10839.test.ts, 4 tests) exercising the real POST handler end-to-end: v1.0 SendMessage aliasing + response reshaping, v0.3 message/send keeping its existing shape, SendStreamingMessage no longer 404ing, and the new agent-card.json route declaring both 1.0/0.3 interfaces. Pushed fix-in-place. - 20/20 existing A2A tests still pass (a2a-auth-timing-safe, a2a-enabled-route, a2a-tasks-auth, t09-a2a-lifecycle, agent-card-route) — no regressions. - check-file-size, check-changelog-integrity: OK. - typecheck:core: clean (the author's local-build concern didn't reproduce). - check-complexity / check-cognitive-complexity: OK, both under baseline. Co-authored-by: wpec <wpec@users.noreply.github.com>
…Results boolean guard (cheaper-inference#10805) Two fixes: (1) createDisconnectAwareStream now distinguishes graceful max_tokens truncation (partial content already reached the client, upstream closes without a terminal marker → clean stop, no error) from a real empty-content failure (still surfaces the 502). Fixes cheaper-inference#7699, keeps cheaper-inference#8649 intact. (2) liteEngine's compressToolResults now requires an explicit boolean before overriding step config, instead of letting a malformed value leak through the `??` chain. Validated in an isolated worktree boarded onto origin/release/v3.8.50 (0 conflicts, 4 files): - 68/68 tests pass (silent-sse-close-7699, compression/lite, empty-stream-no-content-8649, stream-handler). - check-file-size, check-changelog-integrity: OK. - typecheck:core: clean. - check-complexity / check-cognitive-complexity: OK, both under baseline. Note: the empty-content Claude error message text changed from "Upstream stream ended without a terminal marker" to "Provider returned empty content" (matches the OpenAI/Responses branch wording) — intentional, documented in the PR. Co-authored-by: minhlongs <minhlongs@users.noreply.github.com>
Requests AgentUrlConfig from Cursor with each selected account token and selects the account's actual server-assigned Agent endpoint (agentUrl/agentnUrl) instead of a fixed global/us host, which fails for teams pinned to a different region. Caches validated endpoints by connection+token. Closes cheaper-inference#10802. Validated in an isolated worktree boarded onto origin/release/v3.8.50 (0 conflicts, 4 files): - 18/18 focused tests pass (cursor-agent-host, cursor-apikey-provider). - provider-translate-path-golden.test.ts initially failed — traced to a pre-existing base-red (stale golden snapshot left by an earlier freebuff merge, cheaper-inference#10531, unrelated to this PR) and confirmed it reproduces on the pure release tip without this PR's changes. Fixed directly on release/v3.8.50 (mechanical key-ordering regen, values unchanged) rather than folding it into this PR's scope; green after merging that fix in. - check-file-size, check-changelog-integrity: OK. - typecheck:core: clean. - check-complexity / check-cognitive-complexity: OK, both under baseline. - Author additionally validated live: a real Cursor request selected agentn.us.api5.cursor.sh and returned HTTP 200/PING. Co-authored-by: tuandinh0801 <tuandinh0801@users.noreply.github.com>
…-inference#10854) Dashboard-installed SkillsMP and skills.sh skills now store under the canonical global skill scope and merge into every API-key-scoped lookup, so marketplace installs actually reach API keys instead of staying invisible outside the installing session. Tenant-owned skill overrides stay isolated; existing skillsmp/skillssh rows are recognized without a migration, with canonical rows preferred on identity overlap. Closes cheaper-inference#9716. Validated in an isolated worktree boarded onto origin/release/v3.8.50 (0 conflicts, 9 files): - 94/94 skills-*.test.ts tests pass, including the tenant-isolation regression coverage in skills-injection.test.ts (global skills reach a different API key without leaking another tenant's skills). - check-file-size, check-changelog-integrity: OK. - typecheck:core: clean. - check-complexity / check-cognitive-complexity: OK, both under baseline. Co-authored-by: kriptoburak <kriptoburak@users.noreply.github.com>
…rence#10830) Adds a Grok Build card to the CLI Code dashboard: configures the main model and per-subagent model overrides, surgically upserts only the [model.omniroute] section of config.toml (preserving unrelated user sections), resolves config.toml via GROK_HOME or the CLI config home, and omits API keys from settings responses. Closes cheaper-inference#10829. Validated in an isolated worktree boarded onto origin/release/v3.8.50 (0 conflicts, 14 files): - 6/48 focused node:test cases in tests/integration/cli-settings-grok-build.test.ts initially failed with 422 instead of 200/409 — traced to this route being the first CLI-tool settings route to actually pass a targetPath into the shared guardCliConfigWrite/ensureCliConfigWriteAllowed container-ephemeral-write guard (every other cli-tools/*-settings route calls it with no path, which always no-ops). This devbox genuinely runs inside Docker with no bind mount for the test's tmpdir fixtures, so the guard correctly refused — a real environment-dependent gap in the PR's own test setup (missing OMNIROUTE_ALLOW_CONTAINER_CONFIG_WRITE=1, the pattern already used by tests/unit/cli-tools-apply-container-422.test.ts for this exact guard). Fixed by setting that flag in the test file's setup, pushed fix-in-place. - 61/61 tests pass after the fix (grok-build-config, check-tool-config-status, all-statuses-route, cli-settings-grok-build, plus 13/13 UI tests via vitest for GrokBuildToolCard/ToolDetailClient). - Verified the "omit API keys from settings responses" claim: GET applies omitApiKeys() to both config/settings before responding. - check-file-size, check-changelog-integrity: OK. - typecheck:core: clean. - check-complexity / check-cognitive-complexity: OK, both under baseline. Co-authored-by: tuandinh0801 <tuandinh0801@users.noreply.github.com>
…-inference#10928) Bump mecânico e verificado: `github/codeql-action/upload-sarif` v4.37.6 → v4.37.7, 1 linha em `docker-publish.yml`. Diff conferido linha a linha. CI vermelho neste PR é o base-red já rastreado em cheaper-inference#9985 (correção em andamento via PR cheaper-inference#10778, outra sessão) — não é defeito deste bump. Obrigado, dependabot!
…heaper-inference#10929) Bump mecânico e verificado: `github/codeql-action/analyze` v4.37.6 → v4.37.7, 1 linha (SHA pinado) em `codeql.yml`. Diff conferido linha a linha. CI vermelho neste PR é o base-red já rastreado em cheaper-inference#9985 (correção em andamento via PR cheaper-inference#10778, outra sessão) — não é defeito deste bump. Obrigado, dependabot!
…eaper-inference#10930) Bump mecânico e verificado: `github/codeql-action/init` v4.37.6 → v4.37.7, 1 linha (SHA pinado) em `codeql.yml` — sequencial após cheaper-inference#10929 no mesmo arquivo, sem conflito real. Diff conferido linha a linha. CI vermelho neste PR é o base-red já rastreado em cheaper-inference#9985 (correção em andamento via PR cheaper-inference#10778, outra sessão) — não é defeito deste bump. Obrigado, dependabot!
) Bump mecânico e verificado (dev-only): `concurrently` 10.0.4→10.0.5 (patch, correção Windows non-ASCII) e `ctrf` 0.2.1→0.3.0 (minor). `npm install` + `typecheck:core` + `npm run lint` limpos no worktree combinado com cheaper-inference#10931. CI vermelho é o base-red já rastreado em cheaper-inference#9985 (correção em andamento via cheaper-inference#10778, outra sessão), não defeito deste bump. Obrigado, dependabot!
Bump mecânico e verificado (produção): aws-sdk client-bedrock-runtime, jose, next-intl (patches) + onnxruntime-node 1.24.3→1.27.0 + @atjsh/llmlingua-2 2.0.5→3.0.0 (MAJOR). Validação extra pro major: os 41 testes da suíte llmlingua (tests/unit/compression/llmlingua-*.test.ts + docker-llmlingua-optionals-9166) passaram, incluindo os casos GATED que exercitam a lib real v3.0.0 — API do factory/promptCompressor compatível. `typecheck:core` + `npm run lint` limpos. CI vermelho é o base-red já rastreado em cheaper-inference#9985 (correção em andamento via cheaper-inference#10778, outra sessão), não defeito deste bump. Obrigado, dependabot!
…ce#10884) Validado no worktree combinado do lote (`.claude/worktrees/batch-round2-0821`): typecheck:core, lint, changelog-integrity, file-size, complexity e cognitive-complexity todos verdes; testes focados (71 casos citados na PR + suíte combo-builder-effort-variants/model-discovery-reasoning-levels) passando. CI vermelho neste PR é o base-red já rastreado em cheaper-inference#9985. Obrigado!
…r-inference#10883) Validado no worktree combinado do lote: typecheck:core, lint, gates de qualidade e testes focados (opencode-go-catalog-alignment + opencode-go-effort-aliases-8353, incluindo os novos casos muse-spark-1.2-contributor-*) todos verdes. CI vermelho neste PR é o base-red já rastreado em cheaper-inference#9985. Obrigado!
…ference#10872) Validado no worktree combinado do lote: typecheck:core, lint, gates de qualidade e o novo teste OpenClawToolCard-secret-ref-apikey.test.tsx (via vitest) verdes. Correção real e bem isolada de um crash client-side (`e.apiKey.slice is not a function`). CI vermelho neste PR é o base-red já rastreado em cheaper-inference#9985. Obrigado!
Validado no worktree combinado do lote: typecheck:core, lint, gates de qualidade e os 13 testes unitários + 1 de integração (cline-task-id-propagation) todos verdes. Correção legítima de identidade de tarefa fabricada. CI vermelho neste PR é o base-red já rastreado em cheaper-inference#9985. Obrigado!
…ixes (cheaper-inference#10807) Validado no worktree combinado do lote: typecheck:core, lint, gates de qualidade e os novos testes gcf-numeric-domain/gcf-count-mismatch verdes (mais os já existentes do codec GCF). Fix de losslessness bem documentado e cirúrgico. CI vermelho neste PR é o base-red já rastreado em cheaper-inference#9985. Obrigado!
…, creds, cache key, CC vision) (cheaper-inference#10859) Validado no worktree combinado do lote: typecheck:core, lint, gates de qualidade (file-size rebaselineado com justificativa — crescimento legítimo em modelCapabilities.ts/commandCode.ts) e os 97+9 testes focados (vision-bridge, command-code vision, model-select-field-catalog-vision) todos verdes. Duas correções reais (cheaper-inference#10808/cheaper-inference#10809) bem documentadas. CI vermelho neste PR é o base-red já rastreado em cheaper-inference#9985. Obrigado!
…path (cheaper-inference#8982) (cheaper-inference#10798) Reconciliado com a release (conflito mecânico em stryker.conf.json — registro de teste que já existia na tip, apenas resolvido mantendo a entrada) e revalidado: 12/12 testes do arquivo log-level.test.ts passando (incluindo os 4 novos deste PR). CI vermelho é o base-red já rastreado em cheaper-inference#9985. Obrigado!
diegosouzapw
merged commit Aug 21, 2026
fdeac49
into
cheaper-inference:release/v3.8.50
4 of 7 checks passed
5 tasks done
muhamadgalihsaputra
pushed a commit
to niyatna/NiyatnaRoute
that referenced
this pull request
Sep 27, 2026
…r-inference#10885) Reconciliado com a release (conflito aditivo em targetTimeoutRunner.ts — combina o warning G3 já mergeado com a resolução de effectiveTimeoutMs deste PR) e revalidado: 41/41 testes focados passando (upstream-timeout-connection-tier, combo-target-timeout-runner, provider-specific-data-schema). CI vermelho é o base-red já rastreado em cheaper-inference#9985. Obrigado!
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
#6354 shipped per-provider upstream timeouts for the slow-reasoning models, but the knobs all landed on the registry side — provider configs, model entries. An operator who runs this server behind a firewall, or who feeds it a slow model through a connection that is not the registry's, had nowhere to set a timeout of their own. The only remaining escape was the global
FETCH_TIMEOUT_MS, which drags every route along with it.This PR restores that operator-owned tier, at the connection level: a connection's
providerSpecificData.timeoutMsnow preempts the model tier, the provider tier, and the global default when the upstream start-timeout is resolved. A connection with no timeout set is untouched — resolution falls through exactly as before, and the fallback value stays at the global default.The combo path gets the same rule. A per-target timeout ceiling exists there (the "budget" that bounds how long a target may run); it now follows the connection that actually serves the request instead of the registry values alone. So a request routed through a slow connection gets the operator's ceiling, and a request on a normal connection keeps the old behavior.
Related Issues
Validation
npm run lintTests Added Or Updated
tests/unit/provider-specific-data-schema.test.ts— 3 new cases:timeoutMsaccepts an integer second, rejects0/-1/fractional/string/out-of-range, andnullis normalized away like the other house fields.tests/unit/upstream-timeout-connection-tier.test.ts(new) — 8 cases: the resolver floors and rejects garbage; the connection tier preempts model/provider/global; an invalid connection timeout falls through to the next tier; the executor actually aborts on the connection timeout before the global one fires; the two chatCore call sites extract the field from real executor credentials; the combo helper returnsundefined(i.e. falls back to the existing ceiling) when the target has no connection, an unknown connection, or a connection without a timeout.tests/unit/combo-target-timeout-runner.test.ts— the runner tests now cover both a resolved connection timeout and a resolver that returnsundefined; a non-finite resolver result falls back to the configured ceiling rather than aborting instantly.Run:
node --import tsx/esm --test tests/unit/provider-specific-data-schema.test.ts tests/unit/upstream-timeout-connection-tier.test.ts tests/unit/combo-target-timeout-runner.test.ts— 41/41.npm run lint: clean.npm run typecheck:core: clean.Coverage Notes
The new behavior lives in three production files:
src/shared/validation/providerSpecificData.ts(the zod schema and normalization),open-sse/handlers/chatCore/upstreamTimeouts.ts(resolution order, plus the two call sites inopen-sse/handlers/chatCore.ts), andopen-sse/services/combo/(the per-target ceiling). The schema cases exercise the validation and normalization branches directly; the tier tests drive the resolver and both executor call sites with real credentials; the combo tests cover the resolver fallback and the runner's ceiling application. The one path not reachable from a unit test is the DB hit in the combo helper (the connection cache is module-private) — the helper is pinned by the fallback-contract tests, and a connection that misses the cache returnsundefined, which is the safe direction.Reviewer Notes
MAX_PROVIDER_SPECIFIC_TIMEOUT_MS) — the resolver rejects anything above it, and the executor clamps defensively in case a future caller skips the resolver. The clamp is not a second source of truth; the resolver remains the gate.timeoutMson a connection is optional andnull-normalized likequotaPerUnit; nothing breaks for connections that never set it.setTimeout(NaN), which would have aborted the request instantly.changelog.d/features/per-connection-upstream-timeout.md.