When no explicit judgeModel is configured, the judge defaulted to panel[0]
before fan-out and was never reassigned. If panel[0] failed fan-out (timeout /
rate-limit / dropped straggler → it lands in `failures`, not `answers`), the
multi-answer synthesis path still dispatched the judge to that dead panel[0],
erroring the whole fusion request even though a quorum of other panel members
succeeded — exactly the failure fusion exists to tolerate.
Resolve the effective synthesis judge from a survivor when no explicit judge is
set: prefer panel[0] only when it survived, otherwise the first surviving
answer. An explicitly configured judge is still honored unchanged (operator
intent), and the answers.length===0 (503) and single-survivor branches keep
their existing semantics.
Co-authored-by: Chirag Singhal <chirag127@users.noreply.github.com>
LogsPage recomputed initialId from window.location on every render, but the
App Router syncs window.location only after the navigation commits. Closing
the detail modal re-rendered the page while the URL still carried ?id=X, so
initialSelectedId flipped null -> X and the child's one-shot deep-link effect
(guard still unarmed after the open-click render, where location was stale in
the other direction) reopened the modal. Only the second close worked.
Read the id once via lazy useState so the prop stays stable for the page's
lifetime; deep links still open the modal on mount.
Regression test reproduces the App Router ordering with a router.replace mock
that re-renders the page before committing the URL.
getExplicitModelOutputCap() (the clamp ceiling used by
resolveReasoningBufferedMaxTokens) only ever read the unvalidated synced
limit_output / registry / static-spec chain — it ignored the operator-settable
max_token capability override that getResolvedModelCapabilities() already
consulted. When a provider's synced catalog row reports a wrong limit_output
(e.g. ollama-cloud/deepseek-v4-flash: limit_output=1048576, same as
limit_context, while the real upstream cap is 65536), the reasoning-buffer
clamp trusted the bad number and inflated max_tokens 64000 -> 96000, which
upstream rejected with "exceeds model's maximum output tokens (65536)".
The override table (model_capability_overrides, "max_token" key,
/api/model-capability-overrides) is the existing, already-shipped remediation
path for exactly this class of bad catalog data, but reasoningTokenBuffer.ts
had no way to benefit from it. Extracted the override lookup into a shared
getMaxTokenCapabilityOverride() helper and made getExplicitModelOutputCap()
consult it first, so both read paths now agree.
TOTAL_MCP_TOOL_COUNT in open-sse/mcp-server/server.ts summed collection sizes
additively, double-counting tools registered in more than one collection.
The agent-skills trio (omniroute_agent_skills_list/get/coverage) is
intentionally defined in both MCP_TOOLS (schemas/tools.ts) and
agentSkillTools (tools/agentSkillTools.ts), inflating the reported count
from 96 unique tools to 99.
Replace the additive sum with countUniqueMcpTools() (new
open-sse/mcp-server/toolCount.ts), which unions all collection tool
names into a Set before counting, so any future overlap self-corrects
instead of double-counting.
Regression test: tests/unit/mcp-tool-count-dedup-6854.test.ts
Two related root causes in the stacked compression pipeline:
- #6479/#6491: a dispatched step whose engine legitimately finds nothing
eligible (session-dedup with no repeated blocks, ccr below its min-chars
threshold) returns `{ stats: null }`. `mergeStackStep()` silently dropped
that step from `engineBreakdown` with zero trace — no warning, no error.
Now records a `"<engine>: skipped (no eligible content)"` validation
warning for any null-stats step, covering every engine that follows this
convention (session-dedup, ccr, headroom, relevance, llm, llmlingua,
ionizer, readLifecycle), not just the two reported.
- #6480: `finalizeStackedResult` ran the aggregate `guardPipelineInflation`
check unconditionally, even when the loop-level `compressed` flag stayed
false (no step ever advanced `currentBody`). Since tokens are trivially
equal when nothing ran, the guard mislabeled a genuine no-op as
`fallbackApplied: true` with a misleading "reverted to original" warning.
Extracted the guard into `applyStackedInflationGuard()` in
`pipelineGuards.ts` (keeps `strategySelector.ts` under its frozen line
budget) and gated it on `compressed === true`.
Also fixes `compression-pipeline-inflation-guard.test.ts`'s wire test,
which passed a bare engine-id string to the pipeline; `normalizePipelineStep()`
only recognizes a fixed set of built-in string aliases and silently
downgrades any other string to `{ engine: "caveman" }`, so the test's
custom inflating engine was never actually exercised. Passing a step object
restores the test's original intent.
New regression tests: tests/unit/compression/repro-6479-6491-null-stats-silent-drop.test.ts,
tests/unit/compression/repro-6480-noop-guard-misfire.test.ts.
- preInitSqlJs() now memoizes an in-flight Promise (not just the resolved
adapter) per filePath, so concurrent BATCH/STARTUP/HealthCheck/
ProviderLimitsSync callers at boot share one full-file read+WASM decode
instead of each independently reloading the whole database — the
thundering-herd amplifier of the OOM condition #6632 already partly
fixed, left un-implemented by the reporter's own proposed fix (#6628).
- sqljsAdapter's run/get/all now unwrap a lone named-parameter object
(e.g. .all({ isActive: 1 }) for "WHERE is_active = @isActive", the same
call shape getProviderConnections() already uses against better-sqlite3)
before calling sql.js's stmt.bind(), expanding it to the @/:/$ sigil
variants sql.js's own named-bind path requires. Previously the object
was wrapped into an array and sql.js took the positional-bind path,
throwing "Wrong API use : tried to bind a value of an unknown type
([object Object])." whenever the sql.js WASM fallback driver was active
— exactly the error #6802 reported (misattributed to better-sqlite3).
Regression tests added to tests/unit/db-adapters/driverFactory.test.ts and
tests/unit/db-adapters/sqljsAdapter.test.ts, both proven RED against the
prior code and GREEN after the fix.
Enterprise-tier Claude accounts (default_raven_enterprise) don't get
five_hour/seven_day utilization windows from Anthropic's OAuth usage
endpoint — only an extra_usage credit-billing block. parseClaude()
only read data.quotas, so quotas stayed {} and the dashboard showed
"No quota data" even when extraUsage showed the account 100%
exhausted. parseClaude() now folds an enabled extraUsage block into a
credits-style quota row (mirroring parseCodex's bankedResetCredits
pattern), both when quotas is empty and when it's already populated.
No-auth providers (mimocode, opencode, ...) are always dispatched with a single
hardcoded connectionId ("noauth" — SYNTHETIC_NOAUTH_CONNECTION_ID in
src/sse/services/auth.ts). No provider_connections row ever has id="noauth", so
resolveProxyForConnection() in src/lib/db/settings.ts could never populate
connectionRecord for them, and its provider-level proxy lookup (Steps 6/8) only
runs when connectionRecord is present. A proxy assigned via Settings -> Providers
-> mimocode was therefore silently ignored, reproducing the reporter's "same
thing happen when i set the proxy directly in the provider menu" symptom.
Adds a best-effort fallback (src/lib/db/settings/noAuthProxyFallback.ts): when
connectionRecord could not be resolved, scan the known no-auth provider ids for a
configured provider-level proxy (registry first, then legacy) before falling
through to the global/direct steps.
Regression test: tests/unit/proxy-noauth-provider-6272.test.ts (RED on unfixed
code — resolved to level=direct/proxy=null; GREEN after the fix).
combo.ts's isContextOverflow400() guard required the literal word
'context' in the 400 error body before letting a combo fall through to
the next target. Kimi's exact wording ('Your request exceeded model
token limit: 262144 (requested: 308458)') never says 'context', so the
guard misclassified it as a body-specific error and halted the whole
combo instead of trying the next (larger-context) target.
accountFallback.ts's CONTEXT_OVERFLOW_PATTERNS already recognized this
wording one layer below (via checkFallbackError -> shouldFallback), so
the two independently-maintained classifiers disagreed and the
stricter one won. Export CONTEXT_OVERFLOW_PATTERNS from
accountFallback.ts and reuse it inside combo.ts's
isContextOverflow400() so both layers share a single source of truth.
Regression test: tests/unit/repro-6637-kimi-token-limit.test.ts
(RED on unfixed code -> GREEN after the fix). Existing #4519 guard
tests (tests/unit/combo-param-validation-fallback-4519.test.ts) still
pass, including the negative case that a genuinely body-specific 400
is NOT misclassified as overflow.
Collapse duplicate CI spend while keeping each gate's existence reason:
- quality.yml: TIA __RUN_ALL__ defers full unit to fast-unit 4-shard (#6781);
path filters via classify-pr-changes; docs-gates split; draft skip
- ci.yml: wire docs/i18n/code path filters; ESLint JSON artifact for quality-gate;
drop advisory typecheck:noimplicit; float actions/cache@v6
- TIA parity: memory/usage/combo/serial; **/*.test.mjs any depth; electron/bin
no longer force unit __RUN_ALL__
- check:complexity-ratchets: one ESLint walk, ruleId-isolated baselines + cache
- check:api-docs-refs + lib/apiRoutes: shared API route inventory
- husky pre-push: intentionally light (gates live in pre-commit); CLAUDE.md +
QUALITY_GATES.md docs synced
- collect-metrics / lint:json: path.resolve cache path; Windows-safe eslint bin
- env-doc allowlist for ESLINT_RESULTS_JSON / COMPLEXITY_ESLINT_REPORT
- release-green --full-ci expects check:api-docs-refs (not docs-symbols alone)
Tests: select-impacted, classify-pr-changes, api-routes lib, complexity-rule-count,
validate-release-green.
Reconciled after #6781 (fast-unit 2→4 shards) per maintainer request on #6716.
Co-authored-by: Diego Rodrigues de Sa e Souza <8016841+diegosouzapw@users.noreply.github.com>
* feat(providers): add Z.ai Web free web-cookie provider (#4056)
New zai-web web-session provider drives the free chat.z.ai consumer
chat UI via a pasted browser cookie, distinct from the existing
API-key zai/glm/glm-cn/glmt providers (api.z.ai). ZaiWebExecutor posts
to chat.z.ai/api/chat/completions with the cookie forwarded both as
Cookie and Authorization: Bearer <token>, and normalizes both z.ai's
internal delta_content/phase SSE envelope and a pass-through
OpenAI-shaped choices[].delta frame into standard chat-completion
chunks.
Registered in WEB_COOKIE_PROVIDERS, WEB_SESSION_CREDENTIAL_REQUIREMENTS,
the provider registry (GLM-4.6/4.5/4.5V models), the executor factory,
and tokenExtractionConfig.ts for in-app cookie capture.
* fix(providers): regenerate translate-path golden for zai-web + reduce cognitive complexity
* fix(providers): rename ZaiWebExecutor.buildHeaders to avoid incompatible BaseExecutor override
* chore(merge): re-sync with release/v3.8.47; move changelog bullet to changelog.d fragment (merge-storm proof)
* feat(usage): surface Antigravity weekly quota alongside the 5-hour window (#4017)
Antigravity enforces both a 5-hour and a weekly usage limit, but the agy/antigravity
quota widget only exposed the 5-hour window. The weekly limit isn't in the per-model
retrieveUserQuota response already fetched — it lives in a separate, undocumented
retrieveUserQuotaSummary RPC that groups models into families (Gemini Models, Claude
and GPT models) with one weekly bucket per family.
Adds a self-contained usage/antigravityWeeklyQuota.ts leaf: a cached, best-effort
fetch of that RPC + a pure parser that extracts the weekly-labeled bucket per group
(window inferred from bucketId/displayName text, matching the reverse-engineered
shape documented by third-party Antigravity clients) into gemini_weekly/
claude_gpt_weekly quota entries, merged into the existing quotas map the widget
already renders generically. A failed/unavailable RPC never affects the existing
per-model quotas.
Live VPS validation attempt (192.168.0.15, real antigravity account): both
retrieveUserQuota and retrieveUserQuotaSummary currently return 429
RESOURCE_EXHAUSTED for that account, so the live response shape could not be
captured directly. The parser was instead validated via TDD against the bucket
shape documented by CodexBar (steipete/CodexBar), a third-party Antigravity client
that reverse-engineered the same RPC, and is defensive against both response
envelopes it has observed (top-level groups[] and nested quotaSummary.groups[]).
* chore(merge): re-sync with release/v3.8.47 (restore CHANGELOG, keep own bullet)
* chore(merge): re-sync with release/v3.8.47; move changelog bullet to changelog.d fragment (merge-storm proof)
* feat(codex): echo requested effort-suffixed model id in Responses payloads (#3697)
Codex CLI compatibility shim: the Responses API response.created/
response.in_progress/response.completed payloads now carry a `model`
field (previously absent), and for Codex-CLI-originated requests it
echoes the client-requested effort-suffixed model id (e.g.
gpt-5.5-xhigh) instead of the bare upstream id (gpt-5.5), so the Codex
CLI status line/model button shows the active reasoning effort.
- openai-responses.ts translator threads the upstream model into the
Responses event objects (additive, omitted when unknown).
- New isCodexOriginatedHeaders() (codexIdentity.ts) reuses PR #3481's
originator/User-Agent detection, header-based so it still fires when
a combo routes codex/gpt-5.5-xhigh to a non-codex upstream.
- chatCore's existing opt-in #1311 echoModel pipeline now also fires
automatically for Codex clients on the Responses API, regardless of
the echoRequestedModelName setting.
- responseModelEcho.ts now also rewrites the nested response.model
field the Responses API uses (previously only top-level model).
- /v1/models keeps returning models: [] for Codex (unchanged, #3481).
Regression guard: tests/unit/codex-effort-model-echo-3697.test.ts.
Closes#3697
* chore(merge): re-sync with release/v3.8.47 (restore CHANGELOG, keep own bullet)
* chore(merge): re-sync with release/v3.8.47; move changelog bullet to changelog.d fragment (merge-storm proof)
* Add live model discovery for provider catalog
* Fix model discovery request headers
* fix(codex): sync live model limits with local catalog
* test(codex): split live model discovery coverage into dedicated route tests
* fix(codex): use chatgpt account id for live model sync
* Add GitHub-backed Codex model discovery fallback
* fix(providers): tighten oauth config tests and provider model display comments
* test: align client version expectations with release default
* fix(codex): keep discovery complexity within baseline
* fix: rebase live Codex model discovery onto release/v3.8.47, preserving kimi-web buildHeaders (#6308)
Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>
---------
Co-authored-by: Diego Rodrigues de Sa e Souza <diegosouza.pw@gmail.com>
Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>
* fix(i18n): backfill 194 missing pt-BR keys and add key-parity regression test (#6695)
* Merge branch 'release/v3.8.47' into fix/6695-i18n-drift
Resolve i18n key-parity and CHANGELOG-fragment conflicts:
- Convert the #6695 CHANGELOG.md bullet to a changelog.d/ fragment
(the fragment convention landed on release/v3.8.47 after this PR
branched, per changelog.d/README.md).
- Backfill 61 additional pt-BR keys that entered en.json on
release/v3.8.47 after this PR's original 194-key backfill, so the
PR's own key-parity regression test (tests/unit/i18n-pt-br.test.ts)
stays green against the moving release baseline.
* feat(dashboard): add search/filter input to the dashboard sidebar (#4013)
Adds a search box at the top of the expanded sidebar that filters nav
sections/groups/items client-side by label, so users don't have to
hunt through the growing nav tree. Reuses the existing common.search /
common.noResults i18n keys (no new locale edits needed) and the shared
Input icon="search" pattern. Matching sections auto-expand while
searching and the accordion/pin state is restored once the query is
cleared.
Filtering logic is extracted into a pure filterSidebarSectionsByQuery()
helper (src/shared/utils/sidebarSearch.ts) so it is trivially unit
testable independent of React/next-intl/next-navigation.
* fix(test): move Sidebar.search test to a runner-collected path (test-discovery gate)
* feat(routing): per-model web-search interception rule (#3384)
Adds a per-provider/per-model interceptSearch rule (src/lib/db/interceptionRules.ts,
key_value namespace interception_rules) that overrides the existing native
web-search bypass defaults (Codex/Gemini/Claude->Claude passthrough) in
webSearchFallback.ts. Wired at the existing prepareWebSearchFallbackBody() call
site in chatCore.ts. Resolution precedence: per-model rule > provider-level rule
> existing native-bypass defaults.
This lands Phase 1-2 of the plan (rule store + search interception). Web-fetch
interception and the dashboard UI toggle are tracked as follow-up phases.
* fix(db): register interceptionRules in localDb re-export layer (db-rules gate)
* fix(db): renumber interception_rules migration 119→120 (collision with model_capability_overrides)
OmniRoute's calculateCost() always estimated request cost from token
counts x static pricing, discarding xAI's exact provider-reported cost
when present. xAI's chat-completions usage object reports the precise
billed cost via cost_in_usd_ticks (docs.x.ai/developers/cost-tracking
and the API reference's usage schema: "TICKS_IN_USD_CENT: i64 =
100_000_000" => 1e10 ticks/USD, e.g. 37756000 ticks ~= $0.0038).
calculateCost()/computeCostFromPricing() now short-circuit to this
exact figure when present -- before any pricing DB lookup, so it also
works for models without a local pricing row -- and still fall back to
the token-based estimate when it is absent. The field is threaded
through both the streaming (extractUsage/normalizeUsage) and
non-streaming (extractUsageFromResponse) usage-extraction paths.
Corrected divisor vs upstream: the upstream PR used /1e12 (a 100x
under-report, e.g. reporting $0.00123 as the doc's $0.123 example);
this port uses the doc-verified /1e10 instead, confirmed against both
the cost-tracking guide and the API reference's usage-object schema.
Inspired-by: https://github.com/decolua/9router/pull/2453
Co-authored-by: ryanngit <74137224+ryanngit@users.noreply.github.com>
Auto-combo transparency + budget controls: the engine's budgetCap enforcement
always degraded to the globally cheapest candidate when every candidate
exceeded the cap - silently overspending instead of respecting the cap.
- engine.ts: budgetFallback "cheapest" (default, legacy) | "strict"
(BudgetExceededError when no candidate fits budgetCap)
- requestControls.ts: X-OmniRoute-Budget-Fallback header +
resolveRequestAutoControls() consolidating mode/budget/fallback parsing
- resolveAutoStrategy.ts / autoConfig.ts: thread combo-level
config.budgetFallback and catch BudgetExceededError into an HTTP 402
- chat.ts: switch to the consolidated resolveRequestAutoControls() helper
(net line reduction, stays under the frozen file-size baseline)
Regression guard: tests/unit/auto-combo-budget-fallback-3470.test.ts
QuotaCardGrid stacked every provider group vertically in a single
flex flex-col container, and each group's own card grid didn't go
multi-column until the md breakpoint. Provider groups now flow into a
2-column CSS multi-column layout on very wide (2xl) screens instead of
an unconditional vertical stack, and each group's card grid starts at
2 columns immediately, filling horizontal whitespace sooner on
narrower-but-not-mobile viewports.
Regression guard: tests/unit/quota-card-grid-horizontal-layout.test.ts
Add a manual per-model "Context Window Override" so an operator can correct
a provider's misreported context length (e.g. reports 1M when the real
limit is 128K) instead of the model getting silently dropped from combo
routing once the wrong value lands in the catalog.
Reuses the existing Feature-5004 model_context_overrides table
(source="manual") — already the priority-0 source getModelContextLimit()
(the function combo's context-window filter calls) reads ahead of the
models.dev/registry/static catalog — so no new resolver logic was needed,
only the missing write path:
- PUT /api/provider-models now accepts an optional contextWindowOverride
(number to set, null to clear), persisted via setModelContextOverride/
removeModelContextOverride.
- GET /api/provider-models surfaces the current override value + source
back on each custom-model row.
- CustomModelsSection.tsx: edit form gained a Context Window Override
input + a badge on the model row when an override is set.
Regression guard: tests/unit/provider-models-context-window-override-4125.test.ts
(manual override wins over a misreported catalog value, GET round-trip,
clearing via null, default-unchanged behavior).
* fix(kiro): route Amazon Q runtime by profileArn region for cross-region IdC
Enterprise AWS IAM Identity Center accounts whose IdC instance lives outside the two
Amazon Q Developer profile regions (us-east-1 / eu-central-1) - e.g. eu-north-1
(Stockholm), start URL https://d-XXXX.awsapps.com/start - showed no limits and returned
502 on every request.
Root cause: the backend used the IdC/OIDC token region (providerSpecificData.region, e.g.
eu-north-1) for every CodeWhisperer runtime call, hitting q.eu-north-1.amazonaws.com - a
host that does not exist as a Q Developer runtime endpoint. Per AWS docs ("Supported
Regions for the Q Developer console and Q Developer profile"), the Q Developer *profile*
(which produces the profileArn and hosts generateAssistantResponse / GetUsageLimits /
ListAvailableModels / ListAvailableProfiles) is only hosted in us-east-1 and eu-central-1,
regardless of the IdC region; "data is stored in the Region where you create the Amazon Q
Developer profile."
Fix (new open-sse/services/kiroRegion.ts) decouples the two regions:
- providerSpecificData.region stays the IdC/OIDC region, used ONLY for
oidc.{region}.amazonaws.com token mint/refresh.
- The runtime region is derived from the profileArn (resolveKiroRuntimeRegion):
profileArn region -> a valid stored profile region -> us-east-1. A stored IdC region
that is not a Q profile region (eu-north-1) is ignored for runtime.
- Profile discovery (discoverKiroProfileArnAcrossRegions) probes the Q profile regions
(EU IdC -> eu-central-1 first) with the cross-region SSO token instead of q.{idcRegion}.
Wired into: executors/kiro.ts (generateAssistantResponse targets the profile region),
services/usage/kiro.ts (getKiroUsage multi-region discovery + profileArn runtime region so
Limits resolves), services/kiroModels.ts (ListAvailableModels), and
src/lib/oauth/providers/kiro.ts (login-time postExchange profile discovery).
Adds tests/unit/kiro-idc-cross-region.test.ts (15 cases). All Kiro suites pass (60 tests).
* fix(kiro): probe the IdC region too during profileArn discovery (any IdC region)
Make profile discovery general for an IdC in ANY of the ~30 IdC-supported AWS regions
(us-west-2, ap-southeast-2, me-central-1, af-south-1, ...), not just eu-north-1.
buildKiroProfileDiscoveryRegions now probes the two documented Q Developer profile regions
FIRST (us-east-1 / eu-central-1, EU-first for EMEA IdC regions to cut latency), then appends
the IdC/stored region itself as a forward-compatible fallback: if AWS ever co-locates the
profile with the IdC or expands the profile-region list, a same-region probe still finds it.
Probing a region with no profile simply returns nothing and we fall through. The profileArn's
own region remains authoritative for every runtime call (resolveKiroRuntimeRegion), so a
newly-issued ARN in any region is honored automatically.
Adds ap-southeast-2 (APAC) cross-region coverage and updates the discovery-order tests.
---------
Co-authored-by: artickc <artur1992123@mail.ru>
* feat(proxy): implement latency-optimized proxy rotation strategy
Reconstructed onto release/v3.8.47 to drop unrelated main-drift (deps/electron/proxy
files belong to #6620, not this PR) and the direct CHANGELOG.md edit (fragments-first);
the author's env/docs/i18n deltas were re-applied cleanly onto the release tip.
Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>
* fix(proxy): add latency-rotation env var to .env.example
PROXY_LATENCY_WINDOW_HOURS was referenced in src/lib/db/proxies.ts and
documented in docs/reference/ENVIRONMENT.md, but missing from
.env.example, tripping the env/docs sync gate.
Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>
* chore(changelog): re-sync CHANGELOG.md to release tip
Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>
* refactor(proxy): extract latency-strategy helpers to keep frozen files under cap
Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>
* test(db-rules): expect 35 audited modules (proxyLatency joins INTENTIONALLY_INTERNAL)
Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>
---------
Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>
* fix(combo): clamp reasoning buffer to model output cap
* fix(routing): preserve near-cap reasoning max tokens
* fix(routing): getExplicitModelOutputCap falls through to registry cap on non-numeric synced limit_output
getExplicitModelOutputCap short-circuited to null whenever a synced
capability row existed, even if that row's limit_output was not a number
(models.dev commonly omits it). That silently disabled the reasoning-token
buffer clamp for any model with a synced row lacking an output limit.
Now only return the synced value when it IS a number; otherwise fall
through to registryModel.maxOutputTokens / spec.maxOutputTokens, matching
the ??-chain precedence already used by getResolvedModelCapabilities().
Adds a standalone regression test (proves the fallthrough returns the real
registry cap, not null) and hardens the #6274 fixture id so its no-output-cap
case does not prefix-match the real glm-5.2 static spec.
Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>
* chore(stryker): register ollama-quota covering tests (release drift from merge burst)
Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>
---------
Co-authored-by: Diego Rodrigues de Sa e Souza <diegosouza.pw@gmail.com>
Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>
* fix(lmarena): modernize Arena web provider + static Direct-chat catalog
Update the lmarena provider for arena.ai (product rebranded from LMArena):
- Route chat via arena.ai create-evaluation with Chrome TLS impersonation
(tls-client-node) and optional browser-minted recaptchaV3Token.
- Seed Text+Search (48) into the chat registry; seed Image (27) only into
IMAGE_PROVIDERS. Disable live HTML model discovery; resolve public names to
Arena UUIDs from the static TypeScript allowlist (no scrape JSON in-repo).
- Soft-exclude 404/502 model ids; slow/stop bulk test-all probes for this provider.
- Do not fold IMAGE_PROVIDERS/video specialty into the chat provider catalog when
a chat registry already exists (lmarena/openai/xai).
- Display name Arena (Free); keep wire id `lmarena` / alias `lma` for back-compat.
- Theme-aware provider icons: arena-light.svg / arena-dark.svg.
- Preserve split Supabase SSR cookie reconstruction for arena-auth-prod-v1.*.
* fix(providers): align provider-models-route test fixture + regen provider reference
Fold the topaz image-only catalog entry's apiFormat/supportedEndpoints
into the local-catalog test fixture (route now tags media-only
providers per the lmarena PR's staticModels.ts change), regenerate
PROVIDER_REFERENCE.md against the merged release providers.ts, and add
the changelog fragment for #6280.
Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>
* test: align web-cookie fallback suite — lmarena now has a registry entry (probe path)
Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>
---------
Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>
* feat: change default LIVE_WS_PORT from 20129 to 20132
Update the default WebSocket port for the live dashboard server from 20129 to 20132 across all configuration files, documentation, code comments, and tests. Also consolidate OMNIROUTE_DISABLE_LIVE_WS and OMNIROUTE_ENABLE_LIVE_WS into a single OMNIROUTE_ENABLE_LIVE_WS flag. Wire the live WebSocket server to start in-process via instrumentation-node.ts.
* feat: clarify NEXT_PUBLIC_LIVE_WS_PUBLIC_URL path usage and derive upgrade path from URL
Update .env.example and ENVIRONMENT.md to document that the pathname portion of NEXT_PUBLIC_LIVE_WS_PUBLIC_URL (e.g. /live-ws) is used as the WebSocket upgrade path by the dev proxy, handshake response, and client connection logic.
Extract deriveLiveWsPath() into shared/utils/wsPath.ts and wire it through:
- src/app/api/v1/ws/route.ts — handshake response path field
- src/hooks/useLiveDashboard.ts — build
* fix: use the standard URL API to safely parse and update the effectiveWsUrl
* build(docker): expose live WebSocket server port and configure CORS origins
Add LIVE_WS_PORT (20132), LIVE_WS_HOST (0.0.0.0), and LIVE_WS_ALLOWED_ORIGINS environment variables to all Docker Compose profiles and expose the WebSocket port mapping. Prevent infinite self-loop in standalone-server-ws.mjs by skipping proxy when the server itself is running on the LiveWS port.
* docs(env): fix comment formatting for HOST and HOSTNAME variables
---------
Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>
* feat(resilience): weekly-429 cooldown for fetcher-less providers (#3709)
Ollama Cloud free-tier accounts have a hard WEEKLY request cap. On cap the
upstream returns 429 "you (<account>) have reached your weekly usage
limit", but ollama-cloud is an apikey-category provider, so the existing
oauth-only shouldUseQuotaSignal gate in checkFallbackError skips the
subscription-quota-text classifier (Issue #2321) for its 429s -- the
account fell through to the generic exponential backoff (~1s, capped at
2min) and got retried every few minutes for the rest of the week (one
account took 285x429 in 48h).
Adds a new, ungated weekly-usage-limit text classifier that applies a 24h
QUOTA_EXHAUSTED cooldown regardless of provider category. Extracted the new
classifier -- together with the existing #2321 subscription-quota logic --
into a new open-sse/services/quotaTextCooldowns.ts module so the frozen
accountFallback.ts (file-size-baseline cap) didn't have to grow; net effect
shrinks accountFallback.ts by 20 lines.
This is Phase A of the plan (open-sse/services/accountFallback.ts:1038-1045
"weekly-429 cooldown"); Phase B (generic local request-counter preflight for
manual provider_plans dimensions) is a separate, larger follow-up per the
plan's own phasing.
* chore(6817): re-sync onto release tip; CHANGELOG entry → changelog.d fragment (fragments-first)
* fix(resilience): apikey-provider 429s honor explicit quota-exhausted text (#6638)
Ollama Cloud (and any other apikey-category provider) 429s skipped body-text
quota classification entirely; a genuine multi-day quota exhaustion was
misclassified as a plain rate_limit_exceeded with a few seconds of cooldown,
so combo routing retried the account immediately. shouldPreserveQuotaSignals()
now lets an explicit quota-exhausted signal (looksLikeQuotaExhausted) override
the apikey-category default, and parseDayGranularityResetMs() adds day-
granularity reset-hint parsing ("...reset in 3 days.") alongside the existing
Xh/Ym/Zs parsing.
Regression guard: tests/unit/issue-6638-ollama-quota.test.ts (RED before the
fix, GREEN after). Aligned two tests/unit/account-fallback-service.test.ts
cases that had codified the old buggy behavior for apikey-provider quota
text.
* chore(6731): re-sync onto release tip; CHANGELOG entry → changelog.d fragment (fragments-first)
* feat(xai): route xAI clients to Grok native /v1/responses endpoint
xAI ships a native /v1/responses endpoint (https://api.x.ai/v1/responses)
alongside /v1/chat/completions, but XaiExecutor extended BaseExecutor
without overriding buildUrl(), so every request always resolved to the
static chat-completions baseUrl regardless of target format — the last
genuinely-missing slice of decolua/9router#2439 (grok-build-0.1, the
reasoning-effort suffix routing, and bare grok-* routing were already
ported in prior cycles).
Add responsesBaseUrl to the xai registry entry and tag
grok-4.20-multi-agent-0309 (upstream's own Responses-only id) with
targetFormat: "openai-responses", mirroring the existing model-tag-driven
routing pattern already used by the gh executor (9router#102) and the
"openai" -pro heuristic in open-sse/executors/default.ts — the per-model
registry tag is the single source of truth that also drives chatCore's
body translation, so URL and body stay in lockstep. XaiExecutor.buildUrl
now checks getModelTargetFormat("xai", model) and resolves to the native
Responses endpoint only for tagged models, leaving every other grok-*
model on the existing chat-completions bridge.
TDD: tests/unit/executor-xai.test.ts adds a RED-then-GREEN case asserting
grok-4.20-multi-agent-0309 resolves to https://api.x.ai/v1/responses and
a control case asserting grok-4.3 still resolves to
https://api.x.ai/v1/chat/completions.
Co-authored-by: ryanngit <74137224+ryanngit@users.noreply.github.com>
Inspired-by: https://github.com/decolua/9router/pull/2439
* chore(6709): re-sync onto release tip; CHANGELOG → changelog.d fragment (fragments-first)
---------
Co-authored-by: ryanngit <74137224+ryanngit@users.noreply.github.com>
* feat(dashboard): request count log per provider, per date (#4009)
Some providers bill by request rather than by token, so operators need
a plain per-provider, per-date request count breakdown, not just token
aggregates. Adds a new getProviderDailyUsageRows() aggregation query
(src/lib/db/usageAnalytics.ts), a dedicated GET
/api/usage/requests-by-provider-date route (kept separate from the
frozen /api/usage/analytics route to respect the file-size baseline),
and a sortable, single-date-filterable table on Dashboard -> Analytics.
Closes#4009
* chore(6812): re-sync onto release tip; CHANGELOG → changelog.d fragment (fragments-first)
* feat(dashboard): add search to Playground model picker dropdown (#4086)
The shared ModelSelectModal (combo builder + CLI-code cards) already had
search, but the Playground's raw model <select> in StudioConfigPane stayed
a flat unsearchable list - unusable once a provider like OpenRouter
contributed 50+ models.
Adds a search input above the dropdown that filters options via
filterModelsByQuery() (Turkish-safe accent/case-insensitive match, reusing
matchesSearch()). The currently selected model always stays pinned in the
list even when it doesn't match the query, so typing never silently swaps
the active selection. Reuses the existing common.search i18n key already
translated in all 42 locales - no new key needed.
* chore(6811): re-sync onto release tip; CHANGELOG → changelog.d fragment (fragments-first)
* fix(translator): defer content_block_start until GLM streams the tool name (port from 9router#2077)
GLM 5.2 (and similar OpenAI-compatible upstreams) stream a tool call's id and
function.name across separate SSE delta chunks. The openai-to-claude streaming
translator emitted content_block_start immediately on the id-only chunk with an empty
name; the Claude SSE protocol cannot patch a block after emission, so the later
name-only chunk was dropped and Claude Code rejected the tool_use with an empty tool
name / "No such tool available:". Defer content_block_start until the name arrives
(start on args if they arrive first), and emit a start for any orphaned id-only tool
call at finish so content_block_stop is never orphaned.
Reported-by: itiwant (https://github.com/decolua/9router/issues/2077)
* chore(6730): re-sync onto release tip; CHANGELOG → changelog.d fragment (fragments-first)
* fix(translator): strip empty cloud_base_branch from Cursor Subagent tool call (port from 9router#2446)
The Responses->Chat tool-arg cleanup (stripEmptyOptionalToolArgs) only stripped
empty-string/empty-array optional args for Claude Code's Read tool. Cursor's local
Subagent tool call therefore passed through with the cloud-only field
cloud_base_branch: "", which Cursor rejects ("cloud_base_branch may only be specified
when environment equals cloud") before starting the subagent. Extend the cleanup to an
allowlist of Read + Subagent; arbitrary tools stay untouched.
Reported-by: like3213934360-lab (https://github.com/decolua/9router/issues/2446)
* chore(6729): re-sync onto release tip; CHANGELOG → changelog.d fragment (fragments-first)
* fix(antigravity): surface aborted Gemini tool calls off end_turn
Gemini/Antigravity aborts a turn with finishReason MALFORMED_FUNCTION_CALL
(or a sibling like UNEXPECTED_TOOL_CALL) instead of completing cleanly. Both
Claude-facing translators collapsed these to a clean end_turn, hiding the
aborted tool call as a successful completion:
- the OpenAI hub path (openai-to-claude.ts convertFinishReason default), and
- the DIRECT Gemini->Claude path (gemini-to-claude.ts), which is the one
Claude Code actually hits through an antigravity/Gemini-routed model.
Add isAbortFinishReason() to finishReason.ts and map these reasons to
tool_use on both paths; genuinely unknown reasons still fall back to end_turn.
Co-authored-by: anhdiepmmk <n08ni.dieppn@gmail.com>
Inspired-by: https://github.com/decolua/9router/pull/2462
* chore(6713): re-sync onto release tip; CHANGELOG → changelog.d fragment (fragments-first)
---------
Co-authored-by: anhdiepmmk <n08ni.dieppn@gmail.com>
* fix(volcengine): clamp Kimi max_tokens to Ark endpoint cap
VolcEngine Ark's Kimi coding-plan endpoint (ark.cn-beijing.volces.com)
enforces max_tokens <= 32768 server-side and returns 400 "integer above
maximum value, expected a value <= 32768" for anything over that ceiling.
OmniRoute's StripRule only supported dropping params outright, with no
numeric clamp mechanism, so a client sending a larger max_tokens (common
default, e.g. 65536) 400s outright against volcengine's kimi-k2-5-260127.
The 32768 cap is independently confirmed against two live-endpoint bug
reports hitting this exact Ark endpoint for both kimi-k2.5 and
kimi-k2.7-code (NousResearch/hermes-agent#51773, MoonshotAI/kimi-cli#1124),
not just upstream's own value — same cap upstream 9router#2460 uses.
StripRule gains two optional fields: `clampToModelMaxOutput` (clamp to the
model's own catalog maxOutputTokens ceiling, when set) and `maxOutputCap`
(a fixed endpoint-imposed ceiling); when both apply, the lower wins. The
new rule is scoped to the literal id `kimi-k2-5-260127` (OmniRoute's real
volcengine Kimi model, not upstream's `Kimi-K2.7-Code`), not a broad
/kimi/i regex, so it can never clamp an unrelated future Kimi listing
whose Ark cap may differ. glm-4-7-251222 (the other volcengine model) is
unaffected.
Inspired-by: https://github.com/decolua/9router/pull/2460
Co-authored-by: whale9820 <whale9820@users.noreply.github.com>
* chore(6712): re-sync onto release tip; CHANGELOG → changelog.d fragment (fragments-first)
---------
Co-authored-by: whale9820 <whale9820@users.noreply.github.com>
* fix(codex): surface capacity errors embedded in 200-OK SSE streams
Codex sometimes answers with HTTP 200 and a text/event-stream body whose
payload carries a transient error mid-stream (e.g. "Selected model is at
capacity...", server_is_overloaded, service_unavailable_error). Because the
outer HTTP status was 200, this looked like a successful response to every
caller — no retry, no circuit breaker, and no combo/account fallback ever
engaged, so a healthy account sat idle while the request silently failed or
truncated.
Add peekCodexSseTransientError() to open-sse/executors/codex.ts: it peeks the
first bytes of a text/event-stream Codex response, pattern-matches the known
transient-error signatures, and converts a match into a real 503 Response via
errorResponse() (Hard Rule #12 — sanitized, never raw upstream text). A 503 is
already a recognized provider-failure status in accountFallback.ts, so combo
routing and connection cooldown pick it up automatically. When no error
signature is found, the peeked prefix is prepended back onto the remaining
upstream body so the passthrough stays byte-identical to the unmodified
response.
Regression guard: tests/unit/codex-sse-capacity-fallback.test.ts — a
model-at-capacity payload and a server_is_overloaded/service_unavailable_error
payload both convert to 503; a normal single-chunk SSE stream and one split
across multiple network chunks both reassemble byte-for-byte unchanged.
Inspired-by: https://github.com/decolua/9router/pull/2452 (sub-bug #3 only —
OmniRoute already covers PR #2452's other two sub-bugs: service_tier "fast"
normalization and reasoning_effort "max" normalization).
Co-authored-by: ryanngit <74137224+ryanngit@users.noreply.github.com>
* chore(6710): re-sync onto release tip; CHANGELOG → changelog.d fragment (fragments-first)
---------
Co-authored-by: ryanngit <74137224+ryanngit@users.noreply.github.com>
open-sse/translator/request/claude-to-gemini.ts already guards against
sending thinkingConfig for gemma-4-* models (Gemma doesn't support it —
Vertex returns 400: "Thinking budget is not supported for this model"),
but the OpenAI-shape path (openai-to-gemini.ts) lacked the same guard, so
OpenAI-shape clients hitting a vertex gemma-4-* model still got a 400.
Mirrors the existing claude-to-gemini.ts guard: wrap the reasoning_effort
and Claude-shape thinking.budget_tokens branches with a model.startsWith
("gemma-4") check. Branch 3 (default includeThoughts for modern Gemini
models) already excludes non-"gemini" model ids and needed no change.
Inspired-by: https://github.com/decolua/9router/pull/2480
Co-authored-by: chy1211 <31048289+chy1211@users.noreply.github.com>
* fix(oauth): avoid bare-email dedup of Codex OAuth logins
When an incoming Codex OAuth connection has no verifiable workspace/account
id, do not merge it into an existing row on email match alone — that
silently overwrote the other account's token pair. Require a matching
chatgptUserId (a stable per-account JWT id) before merging; otherwise
insert a distinct connection row.
Co-authored-by: lucasjustinudin <34107354+lucasjustinudin@users.noreply.github.com>
Inspired-by: https://github.com/decolua/9router/pull/2477
* chore(6706): re-sync onto release tip; CHANGELOG → changelog.d fragment (fragments-first)
---------
Co-authored-by: lucasjustinudin <34107354+lucasjustinudin@users.noreply.github.com>
* fix(sse): unwrap bare {function:{…}} tools in openai→claude translation
Some OpenAI-shape clients send a tool as a bare `{ function: {...} }`
object, omitting the spec-required `type: "function"` parent wrapper.
The tools-mapping in openai-to-claude.ts (~line 366) only unwrapped
`tool.function` when `tool.type === "function"` was ALSO true, so a
bare-function tool fell through to `toolData = tool` (the wrapper
itself, with no `.name`), producing an empty `originalName` and
silently dropping the tool from the translated request — worse than
a 400, since the caller has no signal the tool never made it
upstream. Unwrap `tool.function` whenever present, independent of
the parent `type` field. Regression guard:
tests/unit/openai-to-claude-bare-tool.test.ts.
Co-authored-by: Samir Abis <me@samirabis.com>
Inspired-by: https://github.com/decolua/9router/pull/2473
* chore(6704): re-sync onto release tip; CHANGELOG → changelog.d fragment (fragments-first)
---------
Co-authored-by: Samir Abis <me@samirabis.com>
