The Devin CLI providers (devin-cli, devin-cli-agentic, devin-desktop; aliases
dv/dva) serve a catalog whose model ids EMBED the reasoning tier:
claude-opus-5-low, claude-opus-5-medium, … and gpt-5-6-sol-max/-low are
distinct upstream models (see registry/devin/catalog.ts).
applyClaudeEffortVariant stripped the trailing -{low,medium,high,xhigh,max}
from any id whose base is a known Claude model, regardless of provider. For
Devin lanes this dispatched a base id that does not exist upstream, e.g.
dva/claude-opus-5-low -> claude-opus-5 -> 400
'Model is not present in the current Devin catalog: claude-opus-5'
Only accidental double-suffixed ids (dva/claude-opus-5-max-low) survived,
because stripping the outer -low left the real claude-opus-5-max. Symmetrically,
the catalog synthesized -<level> variants on top of tier-embedded ids,
advertising phantom ids (dva/gpt-5-6-sol-max-low, dva/kimi-k3-*) that 400 when
called.
Three gates now treat Devin ids as literal:
- applyClaudeEffortVariant: early return for Devin providers (ids/aliases)
- appendClaudeEffortVariants: no -<level> variants for devin-prefixed ids
- appendSyncedEffortVariants: isSkippedEffortProvider now covers Devin
providers (they own their suffix mechanism — the tier IS the id)
Validated live on a self-hosted v3.8.51 deployment: dva/claude-opus-5-low,
dva/claude-5-fable-low and the whole tier-embedded catalog now dispatch; the
phantom variant ids disappear from /v1/models. Claude-lane stripping
(claude/cc, e.g. cc/claude-opus-5-high -> claude-opus-5 + reasoning_effort) is
unchanged and covered by existing + new characterization tests.
Co-authored-by: Neuron Mr White <whiteneuron@gmail.com>
Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>
* fix(usage) handle OpenRouter PAYG credit percentage
OpenRouter PAYG accounts without a per-key limit previously rendered
the credits row as 'total: 0, remainingPercentage: 100, unlimited: true',
treating /credits balance as unlimited even when a real credit pool was
present. Route the credit pool through the credits renderer with the
real denominator: total = totalCredits when positive, used = total -
creditBalance, remaining = creditBalance, remainingPercentage =
round(balance / total * 100), isCredits: true, unlimited: false. Per-key
limit still wins. A non-positive pool surfaces the row but never invents
a 100% bar.
Tests cover: explicit key limit, PAYG account credits without key limit,
key limit taking priority over account credits, and a balance without a
positive denominator.
* fix(usage) render OpenRouter PAYG quota as a metered percentage bar
The frontend parser was routing every OpenRouter 'credits' quota through
buildCreditsQuota(), which sets isCredits: true. QuotaCardExpanded
short-circuits on that flag and shows only the USD balance as a bare
number, so a real PAYG payload (used: 7.33, total: 10, remaining: 2.67,
remainingPercentage: 27) was rendered as '$2.67' instead of the '27% left
/ 7.33 / 10' bar the backend already computed.
Drop isCredits: true for any payload whose total is a positive finite
number - the row then goes through the normal normalizeQuotaEntry() path
with currency preserved as an extra. The balance-only fallback (total 0
or non-finite denominator, used by legacy /credits responses) still uses
buildCreditsQuota() so the row stays renderable, and never invents a
100% percentage.
The frontend test now asserts:
- PAYG positive denominator -> total: 10, remainingPercentage: 27,
currency: 'USD', isCredits !== true.
- Balance-only payload -> isCredits === true, creditCount === 2.67,
total: 0, no fabricated 100%.
- NaN denominator -> balance-only fallback.
- Non-credits keys -> unchanged normalizeQuotaEntry() path.
- Mixed payload -> normal quota row + PAYG row, both kept.
* docs(changelog): add OpenRouter PAYG fix fragment
* docs(changelog): remove self credit
* fix(gemini): preserve response-schema nullability across union flattening
cleanJSONSchemaForAntigravity flattens every union spelling of nullable before
the schema reaches Gemini: flattenTypeArrays turns ["string","null"] into
"string" and flattenAnyOfOneOf drops the {"type":"null"} branch. Correct for
tool parameters, wrong for response schemas — a model with nothing to say can
no longer answer null, so it returns the string "null" or fabricates a value,
and either reaches the client as schema-conformant data. Pydantic emits the
anyOf spelling for Optional[str], so the fabricating path is the common one.
A Phase 1b walk now records Gemini's sibling-key spelling, nullable: true, on
any node whose union carries null — before Phase 2 destroys the evidence. The
key is absent from GEMINI_UNSUPPORTED_SCHEMA_KEYS so it survives sanitizing,
and flattenAnyOfOneOf's Object.assign cannot clobber a key the surviving
branch lacks. Opt-in via { preserveNullable: true }, passed only by the
responseSchema call site; the three tool-parameter call sites keep the default.
Closes#12308
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
* docs(changelog): add fragment for #12308 gemini nullable schema fix
Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>
---------
Co-authored-by: Claude Opus 5 <noreply@anthropic.com>
Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>
* feat(sse): reserve the Antigravity account for the request's stream lifecycle
Re-land of the account-lease half of #10011 on the current release branch.
Its exact-model-scoping half had already shipped in #8050 and its quota half
lost to the tip's aggregate-family design (selectAntigravityQuotaWindowNames /
antigravityQuotaFamily.ts); none of that is reintroduced here. The lease is a
concurrency reservation only and never reads or writes quota state.
The Antigravity account selected for a request is reserved for the whole
streaming lifecycle of that request, so a concurrent retry — or the credential
handoff inside getProviderCredentialsWithQuotaPreflight — cannot re-pick an
account already committed to an in-flight upstream stream. The reservation is
scoped to (connection, callable upstream model) rather than the whole account,
so one account can still serve two different models at once; catalog ids that
resolve to the same upstream id (the gemini-3.7-flash tiers, all
gemini-3.7-flash-tiered) share one lease. When every eligible account is leased
for that model the request returns a structured 503 antigravity_pool_busy with
a bounded Retry-After instead of piling onto a busy account.
Opt-in behind ANTIGRAVITY_ACCOUNT_LEASE_ENABLED (runtime, default false). With
the flag off no reservation is taken, credentials carry no routing descriptor,
every release/hold is a no-op on an undefined lease id, and account selection
and dispatch behave exactly as before.
#10011's original test suite asserted family semantics for a lease that was
exact-model scoped and failed deterministically on its own head; the model ids
it used (gemini-3.5-flash / gemini-3-flash-agent) no longer exist in the
catalog. The contradiction is resolved in favour of one coherent semantic —
exact callable upstream model — and the tests assert it against the alias
tables as they are on this branch.
Co-authored-by: Ardem2025 <openclaw-auto@example.invalid>
* fix(sse): widen the Antigravity lease reservation result so auth.ts narrows it
The discriminated-union form of reserveAntigravityLeaseForSelection's return type
did not narrow under tsconfig.typecheck-api.json, so reading `reserved.lease`
after the `reserved.busy` early return raised TS2339 in the API Route Typecheck
gate. A single optional-property shape carries the same information and type-checks
everywhere.
Co-authored-by: Ardem2025 <openclaw-auto@example.invalid>
---------
Co-authored-by: Ardem2025 <openclaw-auto@example.invalid>
* fix(cache): fold the response output contract into the semantic cache signature
The signature hashed only {model, messages, temperature, top_p}, so two temp=0
requests with identical messages but different response_format shared a cache
key: the second was served the first's stored body under a 200, violating the
schema it asked for. tools/tool_choice had the same exposure.
generateSignature now takes an optional output contract — response_format,
text.format, tools, tool_choice, collected by outputContractOf() — and folds it
into the digest only when present, so plain-chat signatures (and every cache
entry already written for them) are unchanged. All three call sites pass it;
read/write symmetry is preserved because bodyForCacheWrite snapshots the same
body object the read path hashed (#cache-signature-asymmetry).
Closes#12307
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
* docs(changelog): add fragment for #12307 semantic-cache output-contract fix
Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>
* fix(cache): populate both constraint spellings in outputContractOf
The merge with #12734 left generateSignature reading the camelCase
constraints (toolChoice/responseFormat) with a snake_case fallback, but
outputContractOf only filled the snake_case keys, so the #12734
"signature is called with tool_choice/tools/response_format from body"
store tests failed on the merged branch. Set both spellings so either
caller shape reads the value it expects.
Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>
---------
Co-authored-by: Claude Opus 5 <noreply@anthropic.com>
Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>
Co-authored-by: amirrezakm <amirrezakm@users.noreply.github.com>
* feat(providers): update Openference free models and add Deyin to compatible agents
* docs(providers): regenerate PROVIDER_REFERENCE.md against the current tip
Post-merge regeneration so the diff only reflects the Openference free-model
addition, not stale eurouter/greenpt/count churn from an out-of-date local
generation.
Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>
* fix(providers): soften unconfirmed Openference free-forever claim
The Openference pricing page (openference.com/pricing) currently lists five
paid plans ($15-$120/mo) and no $0 tier in its structured pricing data, so
neither the old "3-day trial" note nor a "free forever" claim can be verified
against the source. Point readers to the pricing page instead of asserting a
specific duration.
Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>
---------
Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>
Co-authored-by: AnhLead <AnhLead@users.noreply.github.com>
* feat(routing): self-hosted unified OpenAI-compatible entry (RIC-738)
Divert /v1/chat/completions through the self-hosted provider adapters when
OMNIROUTE_SELF_HOSTED_PROVIDERS / OMNIROUTE_SELF_HOSTED_PROVIDERS_FILE is set:
one OpenAI-compatible contract in, auto-route to the selected provider
(x-omniroute-provider header, provider/model prefix, or first provider),
standard OpenAI error shape out. Optional OMNIROUTE_SELF_HOSTED_API_KEY guards
the entry (D5 reserved); unset = open loopback route. Upstream credentials stay
runtime-only and are stripped from echoed responses.
Brings in the provider-adapters baseline from sibling branch (RIC-737) that
this entry depends on. Includes 21 passing unit tests (provider selection,
model-prefix forwarding, header hygiene, auth, error normalization, SSE
passthrough, fall-through/misconfig), docs, env example, changelog fragment.
* feat(routing): deterministic routing strategies for self-hosted entry (RIC-740)
Add the M2 deterministic routing strategy engine (D3 可审计路由) to the
self-hosted unified entry: a declarative `strategy:` block expressing five
explainable, non-predictive policies — blacklist/whitelist hard filters,
cooldown circuit breaker, cost-priority, latency-aware ordering, and an
explicit fallback chain. The ordered candidate list is the fallback chain:
a failed primary (network or non-2xx) falls through to the next candidate and
each failure feeds the breaker. Every response carries an
x-omniroute-route-decision header answering "why this model / why not that
one". A pinned provider rejected by a hard filter returns 400 (never a silent
re-route); no eligible providers returns 503 with the full explainable
decision. No ML/predict dependency.
Covers the RIC-740 acceptance: 5 strategy types with unit tests + HTTP
fault-injection tests (primary down -> fallback works), config matching docs,
and no predict/ML deps. Adds docs, .env.example entries, and a changelog
fragment.
* refactor(routing): reduce complexity-ratchet violations in new self-hosted routing files
Extract cost/id validation, pin-blocked resolution, ordering, and env/file
source resolution into small helpers so routingStrategies.ts and
selfHostedEntry.ts stay under the complexity-ratchets cap. No behavior
change — the same 51 routing-strategies/self-hosted-entry/provider-adapters
tests pass unmodified.
Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>
* docs(routing): document the 5 self-hosted env vars in ENVIRONMENT.md
check:env-doc-sync failed because OMNIROUTE_SELF_HOSTED_PROVIDERS(_FILE),
OMNIROUTE_SELF_HOSTED_API_KEY and OMNIROUTE_SELF_HOSTED_STRATEGY(_FILE)
were present in .env.example but missing from
docs/reference/ENVIRONMENT.md. Add them under "6. Tool & Routing Policies".
Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>
---------
Co-authored-by: Ant Rich <ant@richants.com>
Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>
Co-authored-by: luyuehm <luyuehm@users.noreply.github.com>
* fix(sse): replay reasoning for Responses-API targets on plain turns and Anthropic clients
DeepSeek thinking mode requires the reasoning of every prior assistant turn
to be passed back once the request carries `tools`, including turns that
made no tool call. Since #10540 routed opencode-go/deepseek-v4-* to
`/responses`, the reasoning replay cache had two gaps on Responses-API
targets, and clients that drop `reasoning_content` hit intermittent
`400 The reasoning_text in the thinking mode must be passed back`.
1. Plain (non-tool-call) turns are keyed on a digest of the normalized
OpenAI transcript. Both capture sites used `translatedBody.messages` as
the history, which a Responses body (`input`) does not carry, so the
write-time digest never matched the read side. translateRequest now
reports the pivot transcript it digested via `onReasoningReplayHistory`,
and the streaming / non-streaming capture sites digest that transcript.
2. The Responses replay pass was gated on `sourceFormat === "openai"`, so
Anthropic Messages clients (Claude -> OpenAI -> Responses) got no replay
at all. The pass now runs on the OpenAI pivot for every source format,
right before the Responses conversion discards `messages`.
The reported transcript is a shallow snapshot of the digested fields only
and travels through a callback, not the body, so nothing new reaches the
upstream payload.
* docs(changelog): add fragment for #13031
* fix(sse): guard the Responses capture sites and skip plain-turn writes with no history
Review follow-ups for #13031:
- Add tests/unit/chatcore-reasoning-cache-write-guard-responses.test.ts:
runs the real handleChatCore against a mocked opencode-go/deepseek-v4-flash
Responses upstream (JSON and SSE), then asserts the next turn's upstream
body carries the replayed `reasoning` input item. Removing either capture
site fallback turns both cases red.
- Project the reported transcript down to the digested fields only
(tool_calls keep type/name/arguments, ids are dropped) and document that
`content` is shared by reference.
- Skip the plain-turn cache write when the history is empty: a real request
always has a prior user turn, so an empty history means the transcript
could not be recovered and a one-message digest can never match.
- Changelog wording: the pre-fix write digested only the assistant message.
* test(sse): select the /responses dispatch by URL in the Responses replay guard
Review follow-ups for #13031: the guard picks the upstream body by URL
(`/responses`) and asserts exactly one such dispatch per turn instead of
taking the last fetch, the streaming case asserts the same body shape as the
non-streaming one, and the `historyMessages` doc on
NonStreamingClientTranslateInput names the Responses-shaped fallback.
* docs(routing): name the replay-history hand-off without tripping the hook heuristic
The fabricated-docs gate treats any `onXxx` token in prose as a plugin hook
name and flagged `onReasoningReplayHistory` (a translateRequest option, not a
hook). Point at the option's home file instead.
* chore(quality): freeze chatCore.ts at 6159 for the Responses replay wiring
check:file-size in PR mode caps a frozen file at max(frozen, base). The rebase onto
the v3.8.51 tip (cde49c937) leaves chatCore.ts at 6159 lines against a 6146 ceiling:
the onReasoningReplayHistory callback on both Responses-capable translateRequest call
sites, reasoningReplayHistory on both non-streaming leg inputs, and the historyMessages
fallback at the streaming cache write. Record the growth with a justification key, as
#13033 did for the same file.
---------
Co-authored-by: jmche <jmche@users.noreply.github.com>
Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>
* fix(chatcore): block duplicate turn execution with 409 turn_in_progress
* test(sse): align turn-execution-guard 409 body expectation with buildErrorBody reason field
* fix(errors): preserve duplicate turn classification
* test(turn-execution-guard): assert ageMs range instead of exact 0
Comparing ageMs to an exact 0 was flaky under real scheduling
latency between the two synchronous calls in the test.
Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>
* docs(changelog): add fragment for turn execution guard fix (#12912)
Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>
* chore(quality): rebaseline chatCore.ts file-size for the turn-execution guard
The guard logic lives in the new open-sse/handlers/chatCore/turnExecutionGuard.ts
leaf; what grows chatCore.ts is the irreducible call-site wiring at the single
execution chokepoint (acquire, the 409 turn_in_progress early return, the
release/handoff bookkeeping and the try wrapper that scopes it).
Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>
* fix(sse): keep endpointPath outside the turn-guard try so the failure-usage closure can reach it
The try/finally that scopes the duplicate-turn guard block-scoped the
resolveChatCoreRequestFormat destructuring, but persistFailureUsage is defined
above the try and closes over endpointPath — every failure-usage write would
have thrown ReferenceError. Moved the destructuring above the guard (it is a
pure derivation from the request, so nothing else changes) and narrowed the
acquire result with an explicit === false, which the workspace tsconfig
(strict: false) needs to see the non-acquired arm's retryCount/ageMs.
check:open-sse-typecheck goes from 4 errors to 0; typecheck:core, eslint,
prettier and the PR's 4 turn-execution-guard tests stay green.
Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>
---------
Co-authored-by: Jihyun Son <jihyun.son@sk.com>
Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>
Co-authored-by: initguru <initguru@users.noreply.github.com>
* fix(copilot): fallback to copilot-chat on 403 identity denial for standard provider
* fix(copilot): document COPILOT_INTEGRATION_ID, extract identity fallback, add changelog
Adds the missing COPILOT_INTEGRATION_ID entry to .env.example (fixes
tests/unit/issue-7793-env-doc-sync-repro.test.ts), extracts the GitHub
Copilot 403 identity fallback out of open-sse/executors/base.ts into its
own module (open-sse/executors/copilotIdentityFallback.ts) to bring the
file back under the frozen file-size ratchet, and adds a changelog.d/fixes
fragment for the PR.
Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>
---------
Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>
Co-authored-by: tuandinh0801 <tuandinh0801@users.noreply.github.com>
* fix(streaming): per-provider fetch-start timeout cap override (#11526 follow-up)
Buffered gateways (opencode-go / command-code Console Go tiers) legitimately
buffer a whole reasoning generation before the first upstream byte, so their
streaming requests can exceed the default 110s headers-wait cap. #11526 capped
every streaming request at that ceiling, so these long generations died at
exactly 'Fetch timeout after 110000ms' (504) before any bytes arrived.
Add a per-provider fetchStartTimeoutCapMs registry knob (600s for opencode-go
and command-code) and project it into the executor's LegacyProvider so
resolveFetchStartTimeout caps only genuinely unbounded providers.
* docs(changelog): add fragment for fetch-start cap per-provider override
Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>
---------
Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>
Co-authored-by: alvinveroy <alvinveroy@users.noreply.github.com>
* feat(sse): add LLM Gateway DevPass quota tracking
Surface the LLM Gateway DevPass allowance (GET /v1/key) in OmniRoute's
quota telemetry, mirroring the OpenRouter API-key fetcher pattern.
- llmgatewayQuotaFetcher.ts: fetch + parse the DevPass /v1/key response
(decimal-string USD values), exposing two windows — monthly plan
credits and the 7-day premium-model window — with a 45s TTL cache.
Pay-as-you-go keys (devPlan "none") and 401/403 fail open (no quota).
- Register in chat.ts before registerGenericQuotaFetchers + register the
named windows for the dashboard cutoff modal.
- usage/llmgateway.ts leaf + usage.ts dispatch case so the Limits page
renders the monthly + weekly premium rows.
- Add "llmgateway" to USAGE_FETCHER_PROVIDERS, USAGE_SUPPORTED_PROVIDERS,
PROVIDER_LIMITS_APIKEY_PROVIDERS, and the dashboard label/order map.
- tests: 21 cases covering the parser, auth fail-open, cache TTL, window
exhaustion, preflight proceed/block, registration, and the usage leaf.
* docs(sse): add changelog fragment + codebase-doc entry for llmgateway quota
* refactor(sse): register llmgateway quota via quotaTrackersBatch
Move the LLM Gateway fetcher registration out of chat.ts (a frozen
file-size-baseline chokepoint) into quotaTrackersBatch.ts, the dedicated
side-effect module that exists precisely so new fetchers don't grow
chat.ts. The batch import runs at module load, before
registerGenericQuotaFetchers(), so the bespoke fetcher still wins over
the generic path. Fixes the file-size gate (chat.ts must not grow).
---------
Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>
resolveDirectHeadersTimeoutMs() (open-sse/utils/directResponseStartTimeout.ts)
now bounds only the pooled dispatcher attempt (attempt 0) with the flat
OMNIROUTE_DIRECT_HEADERS_TIMEOUT_MS watchdog. The fresh-socket retry (attempt
1) is by construction a brand-new socket with no zombie-socket risk (#10214's
rationale only applies to the pooled attempt), so when the caller already
attached its own deadline signal it now defers to a generous, configurable
backstop (OMNIROUTE_DIRECT_RESPONSE_RETRY_TIMEOUT_MS, default 600s) instead of
reusing the identical short flat window — fixing spurious 504s on healthy
slow-TTFB reasoning models that need well over 60s total for both attempts.
Regression test: tests/unit/proxyfetch-direct-response-start-flat-retry-budget-13703.test.ts
(RED against unmodified code: retry cut at 81ms against an 80ms flat budget
with ~1920ms of caller deadline unused; GREEN after the fix).
New env var documented in .env.example and docs/reference/ENVIRONMENT.md.
tlsProfileForProvider's return type in proxyFetch.ts is pulled into a named
alias so its signature stays on one line under prettier's canonical
formatting -- otherwise prettier's mandatory lint-staged reformat of this
frozen file grows it past the check:file-size baseline on every future touch.
base-red inherited: #14004 (docs env/docs contract, fixed separately in
#14022; chatHelpers file-size drift)
filterTargetsByRequestCompatibility ranked combo targets solely on the
chars/4 estimateTokens() heuristic. On a repetitive agent-session body the
estimate overstates real usage several-fold, so a manually-overridden
primary sized correctly for the real request got marked context-incompatible
and was reordered behind an unconfirmed catalog "emergency" member with a
large but unverified limit_context.
Fix: when the reorder branch promotes known-context-compatible targets,
split them by whether their pass came from an operator-set
model_context_override (trusted) or bare catalog metadata (advisory), and
also trust a near-boundary override rejection (required tokens within 5x
the override — covering the ~3.7x overestimate the issue measured) over a
catalog-only pass. An override target keeps or regains priority over an
unconfirmed catalog-only "known compatible" target; two override targets or
two catalog-only targets keep resolving purely on their own fit as before.
Regression test: tests/unit/combo-13870-chars4-overdrops-override-primary.test.ts
(RED before the fix — emergency member promoted to position 0 ahead of the
override primary; GREEN after).
⚠️ base-red inherited: #14004 — docs env/docs contract (fixed separately in
#14022), chatHelpers file-size drift. Not touched by this branch.
AgentRouter routes GLM models through the generic DefaultExecutor, which has
no GLM-specific handling. When the connection's OpenAI-compatible alternate
format is used, a Claude-style thinking:{type:"adaptive"} field survived
stripUnsupportedParams untouched and reached AgentRouter's upstream GLM
endpoint verbatim, which 400s (thinking.type "adaptive" is not supported by
glm models; must be one of enabled, disabled).
Add a mapThinkingType mechanism to paramSupport.ts's STRIP_RULES (in addition
to the existing drop/clamp mechanisms) and scope a rule to provider
"agentrouter" + model matching /glm-/i that remaps thinking.type from
"adaptive" to "enabled", preserving any other thinking fields (e.g.
budget_tokens) — mirroring the same mapping GlmExecutor already performs for
its own provider.
callerSupportsCcrRetrieve() matched the omniroute_ccr_retrieve tool by exact
string equality. MCP aggregators/gateways (Docker MCP Toolkit, Claude Code)
rename re-exposed tools with a namespace prefix (e.g.
mcp__docker__omniroute__omniroute_ccr_retrieve, or a dotted/slashed prefix),
so a fully MCP-capable caller reachable only under such a name was treated as
unable to retrieve at all -- skipping both the protocol-instruction injection
and, per the #7746 safety guarantee, the entire CCR compression engine for
that request.
Add matchesCcrRetrieveToolName(), a separator-bounded trailing-segment match
(__, ., /, :) that accepts a namespaced form of the tool name while still
rejecting a near-miss like omniroute_ccr_retrieve_v2 -- never a bare
substring/endsWith check.
* fix(resilience): treat a Cloudflare managed challenge as a fingerprint rejection, not a ban
A Cloudflare managed/JS challenge served in front of an upstream provider is the
same class of block as a Cloudflare 1010 — the edge refused the CLIENT's
signature — but it is a different product surface and carries none of the 1010
markers isCloudflareFingerprintRejection() looks for.
It therefore fell through the entire 403 ladder in classifyProviderError() to the
terminal default FORBIDDEN, which chatCore persists via writeTerminalStatus() as
testStatus=banned / isActive=false. That state never auto-recovers, so a single
challenge takes the whole provider offline until an operator reconnects in the
dashboard.
Observed on POST chatgpt.com/backend-api/codex/responses/input_tokens for a
healthy Codex OAuth account: the response carried cf-mitigated: challenge,
server: cloudflare and a ~12KB text/html interstitial with
window._cf_chl_opt = {... cType: 'managed', cZone: 'chatgpt.com' ...}. The same
connection refreshed its OAuth token successfully in the same second and served
normal /responses traffic seconds before and after, so the account was never
banned upstream.
Classify the interstitial as FINGERPRINT_REJECTION, reusing the existing
non-terminal precedent from #9929: authTerminalStatus already treats that type as
non-terminal, so the request falls through to the next combo target and the
account state stays untouched.
The markers are matched as full, distinctive Cloudflare-internal strings
(_cf_chl_opt, cdn-cgi/challenge-platform, the challenge-error-text span id
including its escaped-quote nested form) and never as the loose word
"challenge", so provider bodies discussing a challenge in prose are unaffected.
Tests cover the full interstitial, the gateway-nested error.message form, each
marker individually, prose false-positive guards, and regression guards proving
a genuine permission 403 and the ChatGPT Web Sentinel/Turnstile 403 (#8813) both
still classify as FORBIDDEN.
(cherry picked from commit 8204da668a)
* fix(translator): skip replayed web_search_call metadata in Responses-to-Chat
OmniRoute's web-search fallback emits a native web_search_call output item
alongside function_call/function_call_output. Responses clients keep that item
in conversation history and replay it in the next request's input. When the
follow-up turn routes to a Chat Completions target (Claude), the translator hit
its default unsupported-feature branch and returned a deterministic HTTP 400:
Unsupported Responses API feature: input item type 'web_search_call'
cannot be represented in Chat Completions
Skip the replayed metadata next to tool_search_call/tool_search_result. The
paired function_call_output still carries the search results, so no context is
lost and the sources are not duplicated into assistant history.
(cherry picked from commit 2b3e63a470)
* chore(changelog): credit the locked-fork landings of #13161 and #13304
Both PRs come from an organization fork (azox-ai) that refuses maintainer pushes
even with maintainerCanModify=true, so they cannot be re-synced in place and are
landed here by cherry-pick with the contributor's authorship preserved
(merge-gates §6). GitHub may not mark the PRs Merged, hence the explicit
credit in the fragments.
---------
Co-authored-by: anhth2 <anhth2@vng.com.vn>
An OpenCode Zen free-tier 403 ("free tier can only be used from within
OpenCode") reached the end of the executor loop unrecognized: nothing was
persisted about it, and the account that had just been refused was marked
successful, which clears the failure history driving its cooldown backoff. A
refusal was therefore improving the rotation health of the account it hit.
The refusal is now recognized by its own predicate, returned unchanged without
rotating (it is request-scoped, so every sibling account returns the same
verdict), and classified as a non-banning routing error, so the connection
records lastErrorType/lastError/errorCode and stays active.
The account-health reset is also reserved for HTTP successes at both call sites
in the loop, since the same reset ran on any status the loop did not handle in a
dedicated branch.
Co-authored-by: Max <maxmad64@gmail.com>
cleanResponse(text, strip = true) replaced every run of two or more
spaces with a single space. The non-streaming path (which tool mode
always uses, since it buffers the full completion before converting
<tool> text into tool_calls) runs cleanResponse with strip on, so any
code the model wrote through perplexity-web lost its indentation: every
nesting level came back as one space, making generated Python
unimportable. Tabs were untouched, which is what pointed at this
normalization step rather than the model.
Fold the space handling into CITATION_RE so the space before a removed
[n] marker goes with it ("text [1] more" still cleans to "text more"),
and drop MULTI_SPACE. Blank-line squashing and trim are unchanged.
Fixes#13968
Never alias window to the Node global without location. The wasm-bindgen
shim now installs a dedicated window with a Location-shaped object and
restores after WASM init/payload generation, so getLocationOrigin cannot
crash every route after TinyCMS is used once.
Co-authored-by: Cursor Agent <cursoragent@cursor.com>
Co-authored-by: Aref Alapour <aref-alapour@users.noreply.github.com>
Built-in providers that speak the OpenAI Chat wire format skipped
normalizeOpenAICompatibleTools(), which only ran for custom
openai-compatible-* connections. A client tool whose type is not
"function" (a named Claude server tool, a nameless hosted tool) was
forwarded verbatim, and agentrouter's GLM backend rejected the whole
request with 400 tools[0].type:type is illegal.
Extract the gate into shouldNormalizeFunctionToolsOnly(): custom
openai-compatible-* providers keep normalizing on every target, and a
conservative allowlist of built-in providers (agentrouter first)
normalizes on the OpenAI Chat target only. OpenAI itself stays off the
list, so its custom tools pass through untouched.
Closes#13789
* fix(providers): refresh uncloseai free-model roster after upstream rotation
hermes.ai.unturf.com rotated its lineup: /v1/models now serves exactly one
model (Lorbus/Qwen3.6-27B-int4-AutoRound, vllm, max_model_len 65536) while
every previously catalogued id (adamo1139/Hermes-3-Llama-3.1-8B-FP8-Dynamic,
qwen3.6:27b, gemma4:31b) returns 404 on /v1/chat/completions. Requests routed
through the static seed failed although the provider itself is healthy — a
live completion against the new id succeeds.
- registry seed: replace the three dead ids with the live one (+contextLength)
- FREE_MODEL_BUDGETS: 3 rows -> 1 (catalog totals 443 -> 441; counts synced in
README and free-tier-budget.svg)
- noauth authHint: verified-live-model pointer updated; PROVIDER_REFERENCE.md
regenerated
- regression test pins the live id and forbids the retired ids in both the
registry seed and the free catalog
Verified live on 2026-09-15 against https://hermes.ai.unturf.com/v1/models
and /v1/chat/completions.
* docs(providers): sync free-tier entry count after uncloseai merge
The uncloseai roster refresh (3 entries -> 1) dropped the live free-tier
catalog total from 491 to 489 once merged with the current release tip.
README.md and free-tier-budget.svg still quoted 491 after the merge;
update both to the real count so check:docs-counts-sync stays green.
Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>
---------
Co-authored-by: anon <anon@users.noreply.github.com>
Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>
Google validates every `functionDeclarations[].name` against one grammar and
rejects the WHOLE GenerateContentRequest when any single one is invalid, so six
`1c_*` tools in a 109-tool MCP catalog made every request 400, including
requests that would never call them (#13715).
`normalizeGeminiToolName` removed invalid characters, collapsed underscores and
stripped leading and trailing ones, none of which touches a leading digit. The
strip is also why a leading underscore was not a workaround: the client's
`_1c_probe` was normalized back to `1c_probe` and rejected for the same reason.
The prefix goes inside the normalizer rather than at the call site, because
that keeps the guarantee on the one value every path reads. `buildHashedGemini
ToolName` builds its name from the normalized string and inherits its first
character, so a fix applied later would hold for short names and fail silently
for the long ones a 109-tool catalog is full of. It also runs before the
collision check, so two names that newly collide are still separated by the
existing hashed path.
The reverse direction needs nothing: the sanitized name now differs from the
client's, so `buildChangedToolNameMap` carries the original and the response
translator restores the client's spelling on the model's functionCall.
Six cells: no declared name starts with a digit, a leading underscore reaches
Google letter-first, the reverse map returns the client's spelling, a
letter-first name is untouched, an over-long digit-first name keeps the
guarantee through the hash path, and two colliding digit-first names stay
distinct. Four mutations, all killed.
* fix(models): keep OpenRouter :batch variants out of chat routing
ModelSync imported OpenRouter's Batch-API-only variants into the chat
catalogue. A chat completion against one is rejected upstream with
404 This model is only available through the Batch API.
Use the /api/beta/batches endpoint instead.
which #13596 measured 91 times in 41 hours, third by volume, plus the
`model not found - locking mode` failover churn behind it.
OpenRouter's /models carries no endpoint metadata that separates a batch
variant from a chat one, so `classifyExplicitEndpoints` cannot decide it
and the rule belongs in `modelEndpointPolicy`, beside the OpenAI
image/video policy and for the same reason: the file exists so discovery,
import and catalog projection agree on one answer.
Matched as the exact `:batch` suffix, not "has a variant suffix" --
`:free`, `:nitro`, `:floor`, `:online`, `:extended` and `:thinking` are
routing hints on the same chat model, and excluding them would silently
shrink the routable catalogue. Applied unconditionally for this provider:
there is no "batch" endpoint name an upstream could declare next to a chat
one, and the already-stored rows carry the synthetic `["chat"]` default
that re-imported them in the first place.
Closes#13596
* docs(changelog): add fragment for the OpenRouter batch-variant fix
* fix(sse): release native Codex turn pin when pinned model is model-scoped unusable (#13564)
Long-running Codex sessions die with 400 NATIVE_CODEX_PINNED_MODEL_UNAVAILABLE whenever
the model pinned to the current turn becomes model-scoped unusable mid-session (per-model
quota lockout, connection cooldown, exhausted accounts). Claude Code has no equivalent
pin and already falls back to the next healthy combo model; Codex now matches.
Release the turn pin when all pinned provider+model targets are model-scoped unusable
and fall through to full combo routing, re-pinning to whichever model succeeds. Preserve
the pin on provider-wide outages (circuit breaker OPEN, provider cooldown) and when the
pinned target is still healthy.
Also prunes the stale ESLint suppression entry for combo.ts that this change orphaned
(createPinnedModelUnavailableResponse import dropped; pre-existing getBootstrapLatencyMs
remains the sole residual unused var).
* fix(api): resolve codex-settings apiKey via canonical resolver instead of 400 (#13563)
Applying Codex settings from /dashboard/cli-code/codex always failed with
400 "baseUrl, apiKey and model are required" when the dashboard sent an empty
apiKey (cloud mode with no management key selected) — baseUrl and model are
already Zod-gated, so that response could only ever fire on the empty key.
The codex-settings route had diverged from the sibling CLI tools (cline/forge/
openclaw/grok-build/jcode): an inline if(!apiKey) 400 guard plus a hand-rolled
getApiKeyById lookup, instead of the shared resolveApiKey(keyId, apiKey) helper
which resolves by keyId, falls back to the submitted apiKey, then to
sk_omniroute. This change makes codex-settings use the canonical resolver, so:
- empty apiKey + valid keyId -> the real DB key is written to auth.json
- empty apiKey + no keyId -> sk_omniroute default (config still applies)
- explicit apiKey -> written verbatim (unchanged)
* docs(changelog): add fragments for Codex turn-pin fallback and codex-settings apiKey resolution
Anthropic's Messages API rejects a tool whose `input_schema` carries a
composition keyword at the root with:
tools.N.custom.input_schema: input_schema does not support oneOf,
allOf, or anyOf at the top level
The refusal happens before inference, so a single MCP/agent tool carrying
a root-level union fails every request that ships the catalog, and combo
failover cannot recover from it.
Both conversion paths that build a Claude `input_schema` from a client
payload now flatten such a root union into a plain object schema:
object-compatible branches contribute their `properties` (root wins on a
name collision), the root is pinned to `type: "object"`, and only `allOf`
contributes `required` — `anyOf`/`oneOf` branch requirements are
alternatives and promoting them would refuse calls the original schema
accepts. Nested unions are untouched and clean schemas pass through
unchanged.
Closes#13552
* fix(compression): place output-style instruction in top-level system, not messages[0]
Anthropic-shaped bodies reject a synthetic system-role entry unshifted at
messages[0] (the claude passthrough forwards messages unchanged, and the
upstream API requires system content in the top-level system parameter).
Output styles and the caveman output mode now land their instruction in
the top-level system field when it exists (string or content-block array)
and only fall back to a trailing system message for OpenAI-shaped bodies,
which the claude system-role extraction hoists. Custom endpoint system
prompts skip the unshift whenever a top-level system field is present.
All compression options stay enabled; placement-only fix.
Fixes#12584
* test(compression): cover system-instruction placement branches 4 and 6
Adds direct unit cases for the placement ladder branches that had no
coverage, asserting where the instruction lands:
- branch 4: merge into a system message at index >= 1, leaving
messages[0] untouched (plus the skip-over-a-block-content system
message variant).
- branch 6: trailing append when the body has neither a `system` field
nor a string-content system message.
- block-array `system`: appends one block and is idempotent on a second
pass (first coverage of the array path of the marker check).
Also normalizes a malformed non-string/non-array top-level `system`
(e.g. `null`) to "" before the format branch in injectSystemPrompt and
injectCustomSystemPrompt. Previously such a body entered the `system`
branch, matched neither format, and silently dropped the prompt without
falling back to the messages path. Both new guards fail without this
change.
* chore(compression): add changelog fragment for top-level system placement
* fix(backend): stop redaction truncating the message after a path (#13144)
`findUnquotedPathEnd` may swallow the rest of a line when it cannot tell
where a path ends, so a Windows path with spaces cannot leak a
`Files\secret` suffix. Two things made that fire far wider than the
function documents at `:606`.
**1. The licence was granted on separator evidence alone.** Every API route
carries slashes, so an ordinary `/v1/x/y` in prose qualified as unequivocal
and truncated everything after it. The image-model 400 lost the one
sentence it exists to deliver:
built ...cannot be used on /v1/chat/completions. Use POST
/v1/images/generations instead.
delivered ...cannot be used on <path>
Now only a Windows path, a file URI, or a known POSIX filesystem root may
swallow the line. @diegosouzapw's `/zz` vs `/etc` probe on the issue is why
this is the condition and not the first-segment check I originally
proposed: both truncated identically, so the root was never the driver.
**2. The ambiguity branch ran before `resolvedExtensionEnd`.** A path whose
end is pinned exactly by a known extension was still treated as ambiguous
the moment any prose followed it, so the endpoint was discarded and the
line swallowed. A determinable extension leaves nothing to fail closed
about -- the whole path is still replaced, suffix included, and the tail
survives:
before Provider failed in <path>
after Provider failed in <path> with api_key='[REDACTED]'
Both halves are independently load-bearing: reverting (1) fails the two
route tests, reverting (2) fails the extension test.
Fail-closed is narrowed, not weakened. `/etc/shadow copy failed` has
nothing to anchor an endpoint on and still collapses to `reading <path>`.
Worth recording that the guard test for that cannot be killed by mutating
either mechanism alone -- the two are mutually redundant, so it takes
disabling both, which is also why this change cannot expose a suffix these
shapes did not already hide.
Test results against base:
chat-rejects-image-only-model red -> GREEN
dashboard-request-failed-redaction its delivered-log assertion now
passes; the test still fails on a
second, unrelated assertion (the
*internal* log is redacted where it
should stay raw) that base never
reached
tunnel-routes-error-sanitization unchanged, independent (a tripwire
asserting the shared sanitizer does
NOT cover a shape it now does)
836/840 pass across the sanitization, redaction and error suites; the
remaining failures are the two above plus mcp-public-error-boundaries,
which passes in isolation on base and on this branch and only flakes under
--test-concurrency=8.
* docs(changelog): fragment for the redaction truncation fix (#13144)
changelog.d/README says a PR adds exactly one fragment rather than editing
CHANGELOG.md, so the aggregation order stays deterministic and siblings
cannot conflict. This one was missing.
* test(redaction): narrow the headline case to the truncation it names
Rebased onto a base that has moved 40 commits; resolveEndpoint gained an
ignoreAmbiguity parameter and route-context callers in that window. The
merge keeps both: base's parameter, plus this PR's two changes (check the
resolved extension BEFORE the ambiguity branch, and drop
hasFilesystemEvidence from the swallow licence).
The headline assertion was that the whole message survives byte for byte.
On the current base the tail survives but the route itself still becomes
<path> in that particular message, because the quoted model slug earlier
in the line carries separators. That is a narrower, separate question from
the truncation this PR fixes, so the test now asserts the remediation
sentence survives and records the <path> substitution explicitly rather
than silently dropping the case.
* test(stryker): register the redaction-truncation test for mutation runs
errorPathRedaction.ts is mutation-tested, so a new covering test has to be
in tap.testFiles or the Stryker sandbox never runs it and its mutants
report as survived. Inserted in the alphabetical run beside the sibling
error-sensitive-redaction.test.ts, following #13036's precedent.
The list has a second, unsorted appended group; left that alone rather
than re-sorting a file this PR only needed one line in.
Requested in review by @diegosouzapw.
* fix(claude): forward client-negotiated thinking-binding-controls and thinking-display-updates betas
Gateways drops anthropic-beta tokens not on FORWARDABLE_CLIENT_BETAS, so
@ai-sdk/anthropic Fable 5.1 requests carrying thinking.block_binding were
rejected upstream with thinking.adaptive.block_binding: Extra inputs are
not permitted even when the client negotiated
thinking-binding-controls-2026-08-01 correctly. Same class for
thinking.display updates (thinking-display-updates-2026-08-18).
Regression guard: tests/unit/thinking-binding-controls-beta-forward.test.ts
* chore(changelog): fragment for #12989
* chore: trim comments
handleSonioxTranscription took no formData, built the job body from a fixed
three-key object, and reduced the transcript to { text }. Every client-supplied
parameter was therefore dropped in silence: the request returned 200, the flag
had no effect, and from the caller's side an unsupported parameter and a dropped
one looked identical. Diarization and Soniox's `context` were both unreachable,
and the per-token `speaker` attribution Soniox returns was discarded.
The handler now receives the form data (as the Deepgram one already did) and
maps what the caller asked for onto the job: diarization under the three
spellings callers reach for, `context` verbatim, and `language` as a Soniox
language hint. Keys are added only when requested, so a request carrying no
options produces byte-identical job bodies and the existing
audio-soniox-provider deep-equality assertions still hold.
Response shape stays `{ text }` by default. When diarization or
response_format=verbose_json is requested, the token stream is collapsed into
contiguous single-speaker runs and returned as OpenAI-style `segments` carrying
`speaker`, with `words` when word granularity is asked for.
Verified against a live Soniox account on a real two-party Persian phone call:
default response unchanged, 13 segments over 2 distinct speakers with turn
boundaries matching the dialogue, and `context` correcting a proper noun the
model otherwise gets wrong.
Closes#12947
Co-authored-by: Claude Opus 5 <noreply@anthropic.com>
Chat admission sampled host-wide /proc/pressure/memory, so a swapping
Docker host 503'd idle containers with resource_pressure. Prefer this
unit's cgroup memory.pressure and keep the host file as fallback.
* fix(command-code): floor tiny muse-spark output budgets so hidden reasoning cannot consume the whole budget
muse-spark models routed through command-code burn the entire output budget on
hidden server-side reasoning before emitting visible content. A small caller-set
max_tokens (e.g. 64) comes back as HTTP 200 with null content (out=64, reasoning=61).
Reuse the prefix-aware MUSE_SPARK_PATTERN to detect prefixed ids (meta/muse-spark-1.2-contributor,
cmd/meta/muse-...) and floor tiny budgets to 512 in both the /provider/v1 path and
the /alpha/generate fallback. No budget is synthesized when absent; large budgets untouched.
* docs(changelog): add changelog fragment for the command-code muse-spark budget floor (#12497)
* fix(combo): scope Claude model failures to the model, not the account
A priority combo whose steps are five models on one Claude OAuth connection
stops at step 1. hasPerModelQuota() returns false for the claude provider, so
markAccountUnavailable() records a model-specific 404 or 5xx against the
connection row, and getPersistedConnectionCooldownSkipReason() then skips every
sibling step before dispatch. Verified with a combo whose first step names a
model that cannot exist: the 404 lands and step 2 is never tried, while the
same combo works as soon as step 2 points at a different account.
A Claude OAuth connection multiplexes Fable 5, Opus 5/4.8/4.7/4.6, Sonnet and
Haiku behind one credential, which is the multiplexing hasPerModelQuota already
describes. Its quota is a separate question: a 429 on a Max subscription is
account-wide, and shouldMarkAccountExhaustedFrom429 pins that. So this adds
hasPerModelFailureScope() for the non-quota statuses and leaves 429 alone.
Combo exhaustion gets the same treatment for 404, which names one model the
account cannot serve rather than a bad connection. The empty-content 502 was
already exempt through isEmptyContentFailure.
Closes#12334
* fix(combo): extract Claude per-model failure scope to keep file-size caps
Move hasPerModelFailureScope into a leaf so frozen accountFallback.ts and
auth.ts stay at their file-size caps after #12334. Register
combo-claude-per-model-scope.test.ts in Stryker tap.testFiles so the
mutation-test-coverage gate sees the covering unit test.
Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>
---------
Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>
