Three regressions inherited by every PR rebased onto release/v3.8.51, caught and documented with the exact failing output.
The one that mattered most: src/shared/providers/webSessionCredentials.ts did not parse. The UC merge (#11513) inserted the uc: entry inside maxai.storageKeys and lost the array's closing ], plus the entry's }, leaving `ERROR: Expected "]" but found ":"` at line 351. That module is imported by the provider API routes, bulk-web-session, autoCombo's virtualFactory, keepaliveThreshold and dashboard components, so the break was live on the tip and flooded unrelated catalog tests with transform failures. That was my conflict resolution, not the contributor's code — thank you for catching it and for tracing it to the root commit rather than patching around the symptom.
Also fixed: the duplicate bin/cli/utils/volatileEnvPath.mjs entry in PACK_ARTIFACT_REQUIRED_PATHS (findMissingArtifactPaths reported it twice), and UC image models made prefix-addressable without letting them claim historical bare model ids belonging to other providers.
Reconciled on merge: #12394 landed the busy_timeout/probe work first, so src/lib/db/core.ts takes the tip's side. probeUtils.ts is the union of both rather than either side — this PR's message regex is wider (SQLite also reports "database table is locked", "database schema is locked" and "database is busy"), while #12394 added the driver code/errcode path that keeps a transient lock from being classified as corruption and renaming the database away. Taking either alone would have dropped the other half; this PR's own ENOENT test is what surfaced it.
Verified: 76/76 across uc-image, probe-9541-repro, web-session-contract, pack-artifact-policy, bulk-web-session-import and exclusive-connection-leases, and every changed .ts file parses.
Thanks @backryun.
GET /v1/providers/gemini-business/models returned nothing because gemini-business had no RegistryEntry: the listing route resolves the provider through getRegistryEntry and filters the unified catalog by owned_by, and open-sse/config/providers/index.ts only registered gemini and gemini-web.
Adds a registry entry mirroring gemini_webProvider — id gemini-business, alias gembiz, cookie auth — with the twelve ids from the executor's MODEL_CATEGORY_MAP. Each model is declared toolCalling: false, supportsReasoning: false, the same live-behaviour contract applied to gemini-web in #9356: the executor returns plain text, hard-wires the thinking mode and parses no tool calls.
Reconciled on merge: the only conflict was the reserved-prefix count assertion, which the tip had moved. Took the tip's text and measured the real value with this PR applied — 406 to 408, the gemini-business id plus its gembiz alias — rather than carrying the branch's number.
Validated in a combined worktree with all 25 PRs of this batch boarded together (typecheck:core clean, 443/443 node-runner plus 14/14 vitest, all static gates green), and re-verified standalone on the current tip after the other 24 landed: 33/33 across provider-node-reserved-prefix, gemini-business-model-registry-12107 and web-cookie-validation-fallback, with check:provider-consistency OK at 272 REGISTRY entries and 355 canonical providers.
Thanks @pacocartones.
The response de-obfuscation stripped the whole U+200B..U+200D range, so Persian/Kurdish half-spaces (U+200C), Arabic/Indic shaping and emoji ZWJ sequences (U+200D) were deleted from every assistant response — text, reasoning and tool-call arguments, streaming and non-streaming, every provider: ارائهدهنده came back as ارائهدهنده.
The request side only ever inserts a U+200D between two ASCII word characters, so the new stripObfuscationZeroWidth() removes a joiner only there, or at a string edge next to one so a word split across streaming deltas is still cleaned; U+200B and U+FEFF keep their unconditional removal. All seven copies of the old regex now go through the helper.
Validated in a combined worktree with all 25 PRs of this batch boarded together: typecheck:core clean, 443/443 node-runner tests plus 14/14 vitest across every test file the batch touches, and check-changelog-integrity, check:cycles (418 files), check:provider-consistency (272 REGISTRY entries, 355 canonical providers), check:docs-counts, check:docs-sync (42 locales) and check-file-size all green.
Thanks @pacocartones.
A thinking content part arriving with no signature — typical after a cross-provider hop where reasoning_content was converted into a thinking block — was stamped with DEFAULT_THINKING_CLAUDE_SIGNATURE. prepareClaudeRequest treats any non-empty signature on the latest assistant turn as genuine and preserves it verbatim, so the fabricated one reached Anthropic and the replay failed with "Invalid signature". A missing signature is now treated the same as an empty one, aligned with the stricter check claudeHelper.ts already used: the block is dropped rather than fabricated. Real signatures are still preserved verbatim and redacted_thinking is unchanged.
Validated in a combined worktree with all 25 PRs of this batch boarded together: typecheck:core clean, 443/443 node-runner tests plus 14/14 vitest across every test file the batch touches, and check-changelog-integrity, check:cycles (418 files), check:provider-consistency (272 REGISTRY entries, 355 canonical providers), check:docs-counts, check:docs-sync (42 locales) and check-file-size all green.
Thanks @pacocartones.
A proxy assigned to an opencode / opencode-go connection is pinned by the chat handler as the ambient proxy context before the executor runs. OpencodeExecutor only reads per-account proxies from providerSpecificData.accountProxies, so an API-key connection with none took the single-account fast path — which wrapped the dispatch in runWithDirectFetchContext(), and that direct sentinel makes patchedFetch bypass the ambient context and hit native fetch. The assigned proxy was discarded and the request egressed from the host IP, giving `403 This model is not available in your country` on geoblocked hosts. The fast path now applies the direct pin only when no ambient context exists.
Validated in a combined worktree with all 25 PRs of this batch boarded together: typecheck:core clean, 443/443 node-runner tests plus 14/14 vitest across every test file the batch touches, and check-changelog-integrity, check:cycles (418 files), check:provider-consistency (272 REGISTRY entries, 355 canonical providers), check:docs-counts, check:docs-sync (42 locales) and check-file-size all green.
Thanks @pacocartones.
POST /v1/images/generations through a combo returned a bare array instead of the OpenAI {created, data} payload: executeImageCombo() unwrapped one level too many, and the n used for cost calculation read the same double-nested shape, so it was always 0. The combo path now returns the handler payload unchanged, matching the direct-model path.
Second half: Codex image results emitted a data: URI in url whenever response_format was not b64_json, but OpenAI returns b64_json for the gpt-image-* family — clients that omit the field, Codex CLI's built-in image_gen among them, could decode neither shape. Codex now defaults to b64_json; an explicit response_format: "url" keeps its previous behaviour. Both land together because fixing one leaves Codex CLI failing at the other.
Validated in a combined worktree with all 25 PRs of this batch boarded together: typecheck:core clean, 443/443 node-runner tests plus 14/14 vitest across every test file the batch touches, and check-changelog-integrity, check:cycles (418 files), check:provider-consistency (272 REGISTRY entries, 355 canonical providers), check:docs-counts, check:docs-sync (42 locales) and check-file-size all green.
Thanks @pacocartones.
When every combo target is excluded because the request's max_tokens exceeds each target's known output limit, the terminal 400 now says so — requested max_tokens against the pool's highest known ceiling — instead of the unrelated "supports structured output for this request". Diagnostics (unmet, excluded[].reason, terminalReason) are unchanged; only the message for the output_tokens primary reason moves.
Validated in a combined worktree with all 25 PRs of this batch boarded together: typecheck:core clean, 443/443 node-runner tests plus 14/14 vitest across every test file the batch touches, and check-changelog-integrity, check:cycles (418 files), check:provider-consistency (272 REGISTRY entries, 355 canonical providers), check:docs-counts, check:docs-sync (42 locales) and check-file-size all green.
Thanks @pacocartones.
groq/compound and allam-2-7b were absent from the curated Groq registry, so the capability heuristic defaulted them to reasoning-capable and forwarded reasoning_effort verbatim — Groq answers HTTP 400. Declaring supportsReasoning: false makes applyThinkingBudget() strip reasoning_effort, output_config.effort and thinking, same class as #3258. The gpt-oss reasoning models keep the field.
Validated in a combined worktree with all 25 PRs of this batch boarded together: typecheck:core clean, 443/443 node-runner tests plus 14/14 vitest across every test file the batch touches, and check-changelog-integrity, check:cycles (418 files), check:provider-consistency (272 REGISTRY entries, 355 canonical providers), check:docs-counts, check:docs-sync (42 locales) and check-file-size all green.
Thanks @pacocartones.
Adds uncensored.com as two OpenAI-compatible providers mirroring UC's own surfaces: uc, the persona/subscription side over WebSocket with a durable Clerk credential minting a short-lived per-connect token (no API key, un-metered), as a full multimodal port — chat, tools, vision, doc-RAG, image, video, TTS; and uc-direct, the metered Developer API over REST with X-api-key. Same underlying models, two billing surfaces.
Reconciled on merge. 57 files conflicted; only seven carried UC content, the rest was drift from the older release line and took the tip's side.
- executors/index.ts: the tip has since refactored the executor map to lazy dynamic imports, so uc is registered in that shape. uc-direct needs no entry — it routes through the default OpenAI-compatible executor.
- imageGeneration.ts: the branch still carried the retired designerWeb import alongside ucImage; kept only the UC one.
- config/providers/index.ts, webSessionCredentials.ts and web-cookie.ts resolved additively against the MaxAI entries #11461 put on the tip an hour earlier.
- web-cookie.ts: the uc entry declared no serviceKinds, required since #11392, so provider validation would have thrown at load. Declared ["llm"]. uc-direct already declared it at the end of its own entry — an earlier pass of mine added a second one after id and TypeScript caught the duplicate (TS1117); the author's placement is what shipped.
Every count was measured against the merged tree rather than taken from the branch, and all three would have been wrong: reserved prefixes are 406, not 399; APIKEY_PROVIDERS is 237, not 234; providers are 355. PROVIDER_REFERENCE.md regenerated, the count updated across README/AGENTS.md/llm.txt and its 42 mirrors, package.json and 6 SVGs — every changed line in the protected surfaces is a digit substitution and nothing else, verified by masking digits and comparing the removed and added sets (90 lines each, identical). The executor-map golden snapshot went 134 -> 135.
The branch's file-size-baseline.json predates #12411's ratchet re-tightening and was discarded rather than merged; imageGeneration.ts (+12 for the uc-image format branch) was entered against the current baseline under a _rebaseline annotation, and no other cap moves.
Verified: typecheck:core clean, check:provider-consistency OK (271 REGISTRY entries, 355 canonical providers), check:docs-counts exit 0, check-file-size OK, check:cycles OK, 119/119 across the PR's test files, and 2/2 executor-map-golden.
Thanks @arminanton — two providers for two real billing surfaces, rather than one entry pretending to be both, is the right modelling.
MaxAI joins as a first-class signed provider: 13 chat models discovered live from /models/get_config plus 6 image models, routed through the standard /v1 endpoints with per-request X-Authorization signing, browserless onboarding, prompted tool-calling, vision input, image generation and document RAG.
Reconciled on merge — worth reading, because the branch forked 227 commits back and 77 files conflicted. Only five carried MaxAI content; the rest was drift from the older release line and took the tip's side, taking the diff from 113 files to 37 (then 93 as counted against the current base).
- executors/index.ts: the tip has since refactored the executor map to lazy dynamic imports, so MaxAI is registered in that shape rather than the branch's static import.
- imageRegistry.ts: kept only the maxai block. The branch still carried microsoft-designer-web, which #11754 retired.
- models/route.ts: the conflicting hunk was an unrelated Vertex/Anthropic URL change, not MaxAI — tip's side.
- volcengine agent-plan/coding-plan registries: git auto-merged both sides and produced a duplicated supportsVision key, which TypeScript rejects (TS1117). Removed.
One real integration break that only the combined state shows: the MaxAI entry declared no serviceKinds, which #11392 made required a few hours ago. Provider validation threw at load time and check:provider-consistency crashed outright. Declared ["llm"] — the image kinds derive from imageRegistry, per the convention in that PR's backfill.
Every count was measured rather than taken from the branch, and each would have been wrong: reserved prefixes are 402, not the 397 the branch computed from its stale 395 base; providers are 353, not 354. PROVIDER_REFERENCE.md regenerated, the count updated across README/AGENTS.md/llm.txt and its 42 mirrors, package.json and 6 SVGs — every changed line in those files is a digit substitution and nothing else, verified by masking digits and comparing the removed and added sets (90 lines, identical). The executor-map golden snapshot was regenerated: keyCount 133 -> 134.
The branch's file-size-baseline.json predates #12411's ratchet re-tightening, so it was discarded rather than merged — taking it would have silently undone that. The three files this PR grows (proxyFetch.ts +20 for the Windows/firefox_150 TLS profile, imageGeneration.ts +12, models/route.ts +48) were entered against the current baseline under one _rebaseline annotation; no other cap moves.
Verified: typecheck:core clean, check:provider-consistency OK (269 REGISTRY entries, 353 canonical providers), check:docs-counts exit 0, check-file-size OK, check:cycles OK, and 79/79 across the MaxAI suites plus 21/21 reserved-prefix and 2/2 executor-map-golden.
Thanks @arminanton — the provider work itself is thorough; it was the 227 commits of base that needed the attention.
On dashboard/memory?tab=engine the Embedding Model quick-select (and the rerank selector) built their lists from a keyword heuristic over the CHAT catalog (AI_MODELS) plus OpenRouter live discovery. Providers whose embedding models are not in that catalog never appeared — mistral, gemini, nvidia nim, groq, vercel-ai-gateway and others that serve embeddings on a standard OpenAI-compatible /embeddings endpoint — and typing such a model by hand failed at runtime with "Unknown embedding provider".
The fix is one generic mechanism rather than a list of per-provider patches: deriveEmbeddingProviderForChatProvider() turns any chat-registry entry with a /chat/completions base into an OpenAI-compatible /embeddings config, with curated EMBEDDING_PROVIDERS entries always winning; the embeddings service resolves a derived config for unknown-but-configured providers instead of rejecting them; deriveRerankProviderForChatProvider() does the same for Cohere-compatible /rerank; and both memory selectors fall back to a free-text provider/model input when no static catalog exists. No provider is special-cased by name, so adding one to the chat registry now makes it embedding- and rerank-capable here automatically.
Verified on the current release tip: merged clean, typecheck:core clean, check:cycles OK across 417 files, and 35/35 across the PR's five new suites (qdrant-quick-select-catalog, memory-provider-listings, rerank-provider-listings, embedding-generic-provider-fallback, rerank-generic-provider-fallback) plus the updated hard-session-lease-bypass-inventory and embeddings-handler.
Note: the base-red disclaimer in the description referenced #9985 against release/v3.8.50 — that window is closed and the current tip carries no open base-red, so nothing was inherited here.
Thanks @rqzbeh — deriving the capability instead of enumerating providers is the version of this that stays correct as the registry grows.
* fix(sse): map normalized xhigh to max for GLM-5.x+, DeepSeek-V4+, and provider aliases
* feat(sse): support native max reasoning effort and per-model clamping
* test(sse): add unit tests for Qwen 3.8, Claude 4.7+, GPT-5.6, and 2026 reasoning models
* fix(sse): align tests and file-size split for native max effort
Keep `max` as a first-class canonical tier. Split the new sanitizer
coverage out of base-executor-sanitize-effort.test.ts so the file stays
under testCap, and update discovery/catalog/vscode assertions to expect
native max instead of the old xhigh alias.
Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>
* fix(sse): keep combo effort lists and drop unused collectSSE helper
Combo vscode routes still advertise the 5-tier list. Canonical `max` is
preserved in discovery (#9160) and github model metadata. Remove the
unused collectSSE helper that failed the absolute ESLint gate.
Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>
---------
Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>
Co-authored-by: Chewji <Chewji9875@users.noreply.github.com>
* feat(compression): make proactive context-compression threshold a live setting
The proactive compression trigger ratio was a hardcoded COMPRESSION_THRESHOLD =
0.7 in chatCore. Operators could not move compression relative to a client's own
compaction point (e.g. Codex Desktop self-compacts at ~0.85 of its window, so
the 0.7 proxy threshold always preempts the client's compaction with the
proxy's lossier one — see #8932 for what that produced before 3.8.50).
New: key_value namespace 'compression', key 'proactiveConfig',
{"thresholdRatio": 0.7}. Clamped [0.1, 0.99], 30s TTL cache, ipFilter
persistence pattern (#6131), synchronous read stays in the hot path. Default
unchanged; missing/invalid rows fall back to 0.7.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
* test(compression): cover the live proactive-compression threshold (read, validity bounds, fallback, TTL)
Locks in getProactiveCompressionRatio() (src/lib/db/compression.ts), the
key_value-backed replacement for chatCore's hardcoded 0.7:
- shipped default 0.7 when no compression/proactiveConfig row exists
- 30s TTL cache: a fresh DB write stays invisible until the TTL lapses
(clock mocked via node:test mock timers, Date API — the module keeps
its cache private with no reset hook)
- valid override read from key_value, boundary values 0.1/0.99 included
- out-of-range ratios fall back to the DEFAULT (a validity window, not
clamping to the nearest bound — matching the shipped comment)
- broken JSON / non-numeric thresholdRatio: 0.7, without throwing
Guard verified by mutation: switching the window to clamping fails the
out-of-range case.
---------
Co-authored-by: root-cli (Hermes ops) <info@livewellwith.us>
Co-authored-by: Claude Opus 5 <noreply@anthropic.com>
Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>
Co-authored-by: Diego Rodrigues de Sa e Souza <diegosouza.pw@gmail.com>
* fix(kiro): do not permanently ban on 'User is not authorized to make this call'
* test(kiro): regression cover the 403 'User is not authorized' non-ban classification
---------
Co-authored-by: Deftera186 <Deftera186@users.noreply.github.com>
Co-authored-by: Diego Rodrigues de Sa e Souza <diegosouza.pw@gmail.com>
* feat(usage): devin-cli agentic quota + openrouter credits in Provider Limits
Two provider families with live quota APIs were missing from the Provider
Limits dashboard because their list entries were absent:
- devin-cli: new usage leaf querying the Codeium seat-management Connect API
(exa.seat_management_pb.SeatManagementService/GetUserStatus, protobuf over
POST with the raw `Basic <token>-<token>` auth header the CLI itself uses).
Surfaces the plan name plus daily/weekly agentic quota percentages with
reset timestamps from the GetUserStatus plan_status payload, via a minimal
hand-rolled protobuf encoder/reader (no proto dependency warranted for two
fixed messages).
- openrouter: the /key + /credits quota fetcher (#6842) was already wired
into the dispatcher but gated out of the bulk sync — add it to
USAGE_SUPPORTED_PROVIDERS and PROVIDER_LIMITS_APIKEY_PROVIDERS so key
limits and account credits actually surface.
* fix(build): externalize tiktoken so tiktoken_bg.wasm resolves at runtime
The vendored ChatGPT Web connector v4.0.7 (#12181) imports tiktoken
(get_encoding) at module level. tiktoken's node build reads
tiktoken_bg.wasm via a __dirname-relative fs.readFileSync during import;
when Next bundles the package the wasm asset is not traced into the server
chunk, and page-data collection for every route reaching the tokenizer
(e.g. /api/providers/[id]/chatgpt-web-codex-doctor) aborts with
"Missing tiktoken_bg.wasm" — breaking the whole standalone build.
Externalize it like the other runtime-resolved native/wasm packages
(sql.js, sqlite-vec, better-sqlite3): the require stays at runtime, where
node_modules/tiktoken/tiktoken_bg.wasm resolves normally.
* fix(openrouter): /credits balance survives a /key failure
OpenRouter is credit-based, not subscription-based: the authoritative
remaining-credits signal is GET /api/v1/credits (total_credits -
total_usage, the documented "get remaining credits" endpoint), while the
/key limit fields are optional per-key caps that most accounts never set.
fetchOpenrouterQuota previously treated /key as mandatory — any /key
failure (429 rate limit, transient error, unexpected shape) discarded the
whole payload and the Usage dashboard showed "OpenRouter (usage endpoint
unreachable)" even though /credits was reachable. Now:
- /key unavailable + /credits OK → credits-only quota (creditBalance =
total_credits - total_usage) instead of null
- /key 401/403 alone no longer means an invalid token; only a double
auth-rejection (both endpoints) does
- null is returned only when both endpoints fail, and the dashboard label
reflects that ("credits endpoint unreachable")
* fix(openrouter): render AI Credits as a USD credit count in Provider Limits
The Provider Limits card's dollar renderer only activates on
isCredits/creditCount rows (QuotaCardExpanded), but openrouter went through
parseGeneric — which drops `currency` and never sets those flags — so the
credits balance rendered as a meaningless "100% left" (the unlimited-credits
row is always 100%) instead of the actual credit count.
Route openrouter's `credits` quota through buildCreditsQuota() like the
DeepSeek/AgentRouter credits rows: label "AI Credits", dollar-formatted
balance. Free-tier request windows keep the generic percentage treatment.
* fix(usage): document DEVIN_SEAT_API_URL and split quota parsers
Keep fetchOpenrouterQuota and decodeProtoFields under the complexity
ratchets, and add the seat-management URL to the env/docs contract.
Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>
* test(usage): drop duplicated GLM quota-ordering test in provider-limits-ui
* test(usage): drop stale openrouter ACCEPTED_DIVERGENCE
OpenRouter is now in both USAGE_FETCHER_PROVIDERS and
USAGE_SUPPORTED_PROVIDERS, so the recorded aggregator divergence
is no longer real. Add the changelog fragment.
Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>
---------
Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>
Co-authored-by: Diego Rodrigues de Sa e Souza <diegosouza.pw@gmail.com>
* perf(compression): memory and OOM mitigations for large payload hashing and token estimation
* fix(compression): implement getMemoStats observability for result memo (#7847)
Adds the missing memo observability layer referenced by
tests/unit/compression/oom-memo-memory.test.ts and the monitoring API:
- resultMemo.ts: lifetime hit/miss counters + bounded time-ordered ring
buffer (10k entries, ~90KB) powering 1m/5m/15m/1h hit-rate windows;
getMemoStats() reports size/capacity/hits/misses/hitRate + windows.
- memoLookup() tags served results with stats.memoHit = true.
- clearMemoStore() also resets counters and the ring.
- compression/index.ts re-exports getMemoStats for the monitoring route.
- types.ts: optional memoHit field on CompressionStats.
- New GET /api/monitoring/compression route exposing the stats snapshot
(lightweight, no DB) for operators to track cache-hit efficiency.
* fix(compression): align memo contract with upstream #11727 — return caller object, reset lookup counter in clearMemoStore
* fix(compression): restore unwrapEventEnvelope in stream payload collector summaries
The OOM-mitigation commit accidentally replaced unwrapEventEnvelope(evt.data)
with asRecord(evt.data) in the summary builders and live push, breaking
translate-mode {event, data} envelope unwrapping (clientPayload type detection)
and failing 2 stream-payload-collector tests. Restored upstream semantics;
kept the jsonLength OOM optimization as the only delta in this file.
* refactor(compression): break down writeValue and writeEncodedString to pass complexity ratchets
Refactors jsonSha256 internal helpers (writeValue, writeEncodedString)
into small, single-responsibility sub-functions under the complexity
threshold (max cyclomatic 15, max cognitive 15). Preserves exact
JSON.stringify parity, circular reference guards on both arrays and
plain objects, and escape behavior (all 530 relevant tests pass).
* test(compression): make oom-memo heap assertion robust without expose-gc
The CI unit-test shard runner does not pass --expose-gc, so global.gc is
undefined and heapUsed can still momentarily hold GC-pending transients
(observed 53.4 MiB after a 3MiB body). Gate the retained-heap assertion
on forced collection being available (3 forced cycles for array buffers)
instead of skipping it silently, and keep it fully active when
--expose-gc is present.
* fix(compression): restore worker-pool offload path in runCompressionAsync
The OOM-mitigation refactor dropped the isCompressionWorkerEligible /
runCompressionInWorker dispatch at the top of runCompressionAsync, silently
removing the base's worker-thread offload for eligible large payloads.
Restore the block exactly as on release/v3.8.51, ahead of the result-memo
path, keeping the memoization and hashing improvements intact.
* docs(api): document GET /api/monitoring/compression and log route errors via pino
Add the new monitoring endpoint to docs/openapi.yaml following the
neighboring System entries, and replace the route's console.error with
the repo-standard pino logger.
* fix(skills): regenerate omni-resilience and add changelog fragment
Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>
---------
Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>
Co-authored-by: Andrian Balanescu <AndrianBalanescu@users.noreply.github.com>
Co-authored-by: Diego Rodrigues de Sa e Souza <diegosouza.pw@gmail.com>
* feat(combos): add universal handoff feature flag
Add a default-enabled runtime flag that lets operators disable universal context handoffs globally without changing existing combo configuration or requiring a restart.
* fix(i18n): seed the universal-handoff flag description key across locales
---------
Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>
* fix(sse): forward the upstream's real trailing usage in passthrough; estimate only at flush
#12151 injected estimated usage into the finish chunk and dropped the real
trailing usage block that genuine OpenAI upstreams send afterwards — metered
clients got estimates instead of real token counts. The estimate now leaves
via a canonical usage-only chunk at flush, only when the upstream stayed
silent; a real trailing block is forwarded verbatim and wins. The tool_calls
finish_reason normalization now materializes its own rewrite (it piggybacked
on the removed finish-time rewrite), and the dead collectSSE helper goes with
it (subsumes #12324).
* fix(i18n): seed the radarPage limits/training keys the #12320 UI already consumes
RadarCatalogTable.tsx references radarPage.colLimits / trainsOnPrompts /
trainsOnPromptsHelp but #12320 never added them to en.json, so the EN
fallback could not resolve the __MISSING__ markers across 42 locales.
Real translations for pt-BR and vi; the rest resolve via the EN fallback.
* fix(sse): carry the chat stream id into flush-time synthetic chunks
The estimated usage-only chunk emitted at flush used passthroughResponsesId,
which is only ever set on the Responses path — on the chat path the synthetic
chunk shipped id: null, breaking the string-id invariant pinned by
stream-numeric-ids.test.ts. Track the upstream chat-completion id in the
passthrough loop and reuse it (falling back to the Responses id, then a
generated one). Sibling sweep: 74 files importing utils/stream — 603/603.
The provider plugin manifest already exposed usage-fetch (40 providers, #11903); this publishes the second capability, usage-supported, so integrators can tell without reading TypeScript whether the server usage routes accept a provider. #11903closed#11722 after shipping only half of it and said so at the time — this is the follow-up it promised.
The two scopes genuinely differ and the docs now say how: usage-fetch resolves on id or alias (the dispatcher accepts both), usage-supported on id alone, because the runtime guard does a plain USAGE_SUPPORTED_PROVIDERS.includes(providerId) with no alias resolution. 42 providers carry both tags, 4 carry only usage-fetch (opencode, opencode-zen, openrouter, xai) and 3 only usage-supported (adobe-firefly, firefly, xiaomi-mimo-token-plan) — 7 measured differences, so neither tag implies the other. No list mutation, no new route, schemaVersion stays 1.
USAGE_SUPPORTED_PROVIDERS moved out of src/shared/constants/providers.ts into an import-free leaf at open-sse/services/usage/supportedProviders.ts, keeping the manifest's import graph light — the same move fetcherProviders.ts got in #11903, landed on the correct side of the workspace boundary.
Base note: the branch forked 46 commits before kilocode joined the list, so a wholesale take of its providers.ts would have silently dropped that id. Verified against the current release tip before merging — both sides hold the same 46 ids, nothing lost.
Verified on the current tip: typecheck:core clean, check:cycles OK across 417 files (the import-free-leaf claim holds), and 63/63 focused tests across provider-plugin-manifest, usage-fetcher-registration-coverage, adobe-firefly and agentrouter-quota-dashboard-rendering.
Thanks @maxmad64bis — and for finishing the half of #11722 that was left open rather than letting it sit.
* fix: resolve compression worker file using runtime anchors instead of import.meta.url
Replace workerUrl() function that used import.meta.url with resolveWorkerFile()
function that uses runtime anchors (process.cwd() and process.argv[1]) to locate
the worker file. This fixes webpack module resolution in Next.js standalone
bundles where import.meta.url is replaced with a stub pointing to build machine
path.
Also update Dockerfile to copy required worker-related scripts and adjust
npm install flags for better compatibility.
* fix(docker): restore base Dockerfile — keep npm ci --ignore-scripts supply-chain guard
Revert every Dockerfile change from this branch back to release/v3.8.51:
the branch dropped --ignore-scripts (reopening install-time script
execution for all transitive deps), swapped the reproducible npm ci for
npm install, invoked the nonexistent 'npm approve-scripts' command, and
broke the better-sqlite3 smoke test with a stray space in ':memory: '.
The worker-file fix does not need any Dockerfile change.
* test(compression): export runtime-anchor helpers and cover worker-file resolution
firstAncestorWith's doc already claimed 'exported for tests' without the
export; export it together with resolveWorkerFile and add unit coverage
for the runtime-anchor resolution: cwd anchor, dirname(argv[1]) anchor,
bounded walk-up (8-level cap boundary), prod-first .js-over-.ts ordering,
dev .ts fallback and the fail-open cwd fallback when nothing exists.
Fixtures live in mkdtemp sandboxes only — the repo tree is never touched.
---------
Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>
The weight table said stability accounts for "low latency stdDev / error rate". Grep errorRate in scoring.ts and you find it declared on ProviderCandidate and read nowhere — while combo.ts pulls 24 hours of usage history behind a ten-sample floor, falls back to real-time metrics, and hands every candidate an errorRate the scorer ignores. Two candidates, one failing 1% of calls and one failing 99%, scored identically at 0.459486.
This declares reliability as a sixteenth factor: 1 - failureRate, using the same formula, field precedence and rate-bounding speedRanking.ts already applies, so a corrupt reading means "nothing observed" rather than "fails every call". It ships at weight 0, leaving the ranking unchanged to the digit — the honest default, since which weight this deserves is a product call backed by traffic the author does not have. Two declared-but-silent factors already ship (cacheAffinity, resetWindowAffinity), so the pattern is not new. The stability row now describes what that factor actually computes: latency variance.
The rest is the mechanical 15 → 16 across nineteen documents and the forty-two llm.txt mirrors — sourced from check:docs-counts rather than a grep, the first real use of the gate #12316 extended.
Protected-surface note: this PR touches AGENTS.md, llm.txt and its 42 mirrors, and skills/omni-combos-routing/SKILL.md. Every changed line in those 45 files is a digit substitution and nothing else — masking all digits makes the removed and added lines identical, with no sentence added, removed or reworded. Reviewed and approved on that basis before merging.
Verified on the author's rebased head: check:docs-counts green (the gate that now enforces the count this PR moves), typecheck:core clean, and 71/71 focused tests across scoring-reliability-factor, combo-scoring-weights-schema-coverage, check-docs-counts-sync, lkgp-enabled-context, intelligent-routing-options and the combo-matrix auto integration suite.
Thanks @maxmad64bis — shipping the factor at weight 0 and saying plainly that the weight is someone else's call is the right way to land this.
* fix(guardrails): pass providerId to getResolvedModelCapabilities in checkComboVision (#12112)
Signed-off-by: Minxi Hou <houminxi@gmail.com>
* chore(quality): register combo-vision providerId test in the stryker tap set
---------
Signed-off-by: Minxi Hou <houminxi@gmail.com>
Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>
A passthrough stream could end with no usage even though the client asked for it via stream_options: {include_usage: true}, so providers that do not meter always showed 0 tokens. The fix estimates usage at the finish marker when the upstream stays silent (flagged estimated: true) and drops any duplicate trailing usage chunk so the client never sees two.
open-sse/utils/stream.ts:1982,1749 · open-sse/utils/usageTracking.ts:651,664
Six cases: the predicate (finish without usage but with content, trailing valid, empty response, tool-only) plus two SSE harness cases through createSSEStream passthrough.
Note on base: this branch forked 442 commits back and carried a base-red marker for #12109, which is now closed — the release tip has no open base-red issue. It merged cleanly against the current tip regardless.
Verified in a combined batch worktree: 174/174 focused tests across all 11 PRs of this batch (this PR's stream-passthrough-usage-estimation suite included), typecheck:core clean, check:cycles and check:docs-counts green.
Thanks @maxmad64bis.
With freeAccessPolicy: "strict" the read-only candidate listing silently dropped rows, so an operator could not tell "no free allowance left" from "the quota fetcher is broken" — in a listing whose own module header promises a candidate the routing path would skip "is never dropped". #9133 settled the same question for the resilience filter via a skip opt-out; the zero-cost guard never got one.
It gets it now: the guard is disabled for the inspector build exactly as the resilience filter already is, and every candidate carries freeAccessExclusion — null when satisfied, otherwise one of seven named reasons. The last three (exhausted, state-unknown, no-connection) are the point: they used to look identical because the row just disappeared. STRICT_ZERO_COST.md documents what each asks the operator to do.
Dispatch is untouched and a test pins that. The three existing guard suites were not modified — their 31 cases are the net and still pass. excludeTosAvoid still drops candidates without a reason; documented as a separate question rather than widened into here.
Verified in a combined batch worktree: 174/174 focused tests across all 11 PRs of this batch, typecheck:core clean, complexity 2706/3218, cognitive-complexity 1221/1437, check:cycles and check:docs-counts green.
Thanks @maxmad64bis — the reason table and the honest note about the order change (freshness before status) made this easy to review.
usage/fetcherProviders.ts exists, in its own words, "so the registration list can't drift from the dispatcher's switch statement". It drifted: #8006 added adobe-firefly and firefly to the dispatcher and to USAGE_SUPPORTED_PROVIDERS but not to this list, so the connection UI advertised usage support while the provider-plugin manifest, genericQuotaFetcher and the free-access quota cache all reported no fetcher — for two ids getUsageForProvider would happily serve.
Declaring them is what makes the balance actually get fetched (registerGenericQuotaFetchers wires a generic fetcher per declared id, and resolveFreeAccessState stops returning early), which the PR states plainly rather than burying as a side effect.
The test turns the docstring's prose invariant into enforcement: it reads the dispatcher's cases from source and compares both directions, and records each accepted difference against USAGE_SUPPORTED_PROVIDERS with a reason plus a staleness check, so the next drift can't hide among them. xiaomi-mimo-token-plan is left flagged as a real gap rather than widening the PR.
Verified in a combined batch worktree: 174/174 focused tests across all 11 PRs of this batch, typecheck:core clean, complexity 2706/3218, cognitive-complexity 1221/1437, check:cycles and check:docs-counts green.
Thanks @maxmad64bis.
* fix(usage): console-aware Token Plan guidance and subscription hint on bailian 401
The personal Token Plan is sold through two consoles with different portals,
gateway hosts and login tickets. Two operator-facing messages ignored the split:
- The quota guidance always said 'get the cookie at home.qwencloud.com', even
for connections served by the Alibaba Model Studio console — following it
verbatim produces a cookie the gateway rejects (console mismatch →
BailianGateway.Login.NotLogined). The guidance now derives the console from
the provider via resolveConsoleSite, matching what the fetcher will do with
the pasted cookie.
- Key validation mapped upstream 401 to a bare 'Invalid API key'. An expired
Token Plan subscription produces the exact same upstream 401 (observed live
2026-09-01: subscription ended 08-23, the working key started failing), so
the message now names the subscription as a cause worth checking.
* test(providers): align the remaining bailian 401/403 message pins to prefix match
search-provider-validation.test.ts pinned the exact 'Invalid API key' string for
the bailian validator; the message now also names an expired Token Plan
subscription. Same property asserted (401/403 => invalid), prefix match.
* feat(sse): add it/ru/zh caveman output instructions
* fix(sse): expose the dormant terse-prose translations through the catalog
* feat(sse): translate less-code to es/de/fr/it/ru/zh
* feat(sse): translate ponytail to es/de/fr/it/ru/zh
* feat(sse): translate i-have-adhd to es/de/fr/it/ru/zh
* test(sse): anchor wave-2 output-style translations to their own language
* feat(dashboard): offer every output-style language in the default-language selector
* feat(sse): let autoDetect pick the output-style instruction language
* docs(compression): consolidate the output-style tables and record full language parity
* fix(sse): trust finish_reason:length/max_tokens over the reasoning-ratio heuristic in response quality validation
A truncated response with empty content and reasoning_content present was
only rejected by validateResponseQuality() when reasoning consumed >=90%
of completion_tokens. A response truncated at a lower ratio (e.g. 63%)
passed through as "valid" even though the caller received no usable
content and finish_reason was explicitly "length" (or the alternate
"max_tokens" naming some providers use) -- an unambiguous truncation
signal the validator wasn't reading. Reproduced live against
nvidia/nemotron-3-super-120b-a12b: content:null, finish_reason:length,
reasoning_tokens 645/1024 (63%).
Trust finish_reason directly when it's reported, falling back to the
existing token-ratio heuristic only when it isn't. Does not affect the
deliberate-tiny-probe case (e.g. max_tokens:1 connectivity pings) --
those never produce reasoning_content, so the branch this change is in
doesn't run for them.
* docs(changelog): add fragment for #12262
---------
Co-authored-by: brick30llc-ctrl <admin@brick30.com>
getPersistedConnectionCooldownSkipReason() returned a skip for ANY connection
whose testStatus was `unavailable`, with no elapsed-cooldown check:
if (status === "unavailable") return `Skipping ...`;
That is the raw-label anti-pattern AGENTS.md warns about ("check whether code
is reading raw state instead of using getStatus()/canExecute()") — the
resilience layers are meant to recover lazily. The sibling helper directly
above it, getConnectionStatusQuotaCutoffReason(), does require
hasFutureRateLimitUntil() before treating `unavailable` as blocking.
Its stated justification — "Lazy recovery is unaffected: clearAccountError()
resets the status on first success" — does not hold on this path. This gate
runs BEFORE dispatch, so it prevents the very successful request that would
call clearAccountError(). And a row whose rateLimitedUntil is absent cannot be
rescued by the out-of-band recovery job either, because hasElapsedCooldown()
there requires a timestamp to be present.
Net effect reported in #12168: an entire combo pool answering
ALL_TARGETS_SKIPPED with recordedAttempts === 0 — zero upstream attempts, no
path back to healthy.
The original intent (do not burst into a connection AUTH just retired, before
the timestamp lands) is preserved, but bounded: the bare label is honoured only
while lastErrorAt is inside a grace window, mirroring ERROR_LABEL_GRACE_MS in
src/lib/quota/connectionRecovery.ts so the two never disagree about whether a
label is still meaningful. Past the window the request goes through, and one
real attempt either succeeds (clearing the status) or re-arms the cooldown with
a fresh timestamp.
Regression introduced by #11360, shipped in v3.8.50.
Two assertions in repro-combo-persisted-cooldown-preskip.test.ts encoded the
buggy behavior as intended ("skips an unavailable connection whose cooldown
already expired") and are realigned to the corrected contract, plus a case for
the orphan state (unavailable with no timestamps at all).
providerFailureThreshold / providerFailureWindowMs / providerCooldownMs shipped
in PROVIDER_PROFILES with no runtime consumer (2026-08-31 docs audit, P0.1).
Provider-level entries in providerCooldownTracker now honor them: the whole
provider only counts as cooling after providerFailureThreshold failures inside
providerFailureWindowMs, then cools for providerCooldownMs. Connection-level
entries keep the pre-existing exponential backoff, and the layer stays opt-in
(PROVIDER_COOLDOWN_ENABLED, default off) — default behavior is unchanged.
TDD: tests/unit/provider-cooldown-window-gate.test.ts written first (4 red on
the old behavior), then the wiring; legacy tracker suite aligned to the new
contract (23/23 green). Docs: AGENTS.md breaker section + RESILIENCE_GUIDE
opt-in layer subsection; executors soft-drift refresh (104 -> 106).
dispatchWithCooldownRetry arms a loop-safety timer (setTimeout, 10 minutes by
default) on every setTry iteration, so a combo that never produces a terminal
response still answers with a 504 instead of hanging. The only clearTimeout in
the whole file sat inside the `if (anySuccess)` branch — the comment said so
verbatim: "clear the safety timer on the happy path".
Every error exit therefore returned the response to the client while leaving a
600s timer pending, its closure retaining orderedTargets and the exhausted
provider/connection sets: all_targets_skipped, all_accounts_inactive, the
aggregated-status return, the final fallback, and the global-timeout branch.
The timer is also re-armed per setTry iteration with no clear in between.
Field evidence from the issue: two requests that failed quality validation
returned 502 to the client immediately, and "Combo loop safety timeout ...
force-terminating" was logged for both exactly 600 seconds later — the leaked
timers firing long after the requests were gone.
Fixed structurally rather than by sprinkling clearTimeout across the five
return sites: the handle is hoisted to function scope and released in a
finally, so a future `return` added to this function cannot silently
reintroduce the leak. The 504 backstop itself is unchanged.
Note the timer already called .unref(), so it never held the event loop open —
this is a memory-retention leak, not a hang.
Refresh the existing MIT-licensed miuuyy/codex-chatgpt-web vendor snapshot and its OmniRoute integration as one reviewable change.
Co-authored-by: backryun <backryun@daonlab.local>
* feat(usage): add Kilo Code balance and Kilo Pass quotas
* feat(usage): add Kilo Pass dashboard meter
* test(usage): cover Kilo Code quota integration
* docs(usage): document Kilo API endpoint override
* perf(sse): defer cloneLogPayload until after SSE collector cap check
Dropped SSE events no longer pay the structuredClone cost. The clone now
runs only for events that survive the maxEvents/maxBytes cap, eliminating
~9,800 wasted deep clones per streaming response (65-71% faster push).
Reducer snapshot isolation restored:
- OpenAI reducer stores first-chunk primitives instead of a chunk reference
- Responses reducer snapshots only needed fields, deep-cloning nested output/metadata
- getEvents() keeps defensive-copy semantics via cloneLogPayload
* chore: add changelog fragment for #12241
Local process execution failures (ENOENT spawn errors, binary missing, EPIPE, exit codes) were incorrectly treated as upstream provider failures, opening provider circuit breakers and cooling down valid connections. Added `isLocalExecutionError` guard to skip circuit breaker trips and connection disables when local host execution fails.
resolveWorkerPath() had two return branches: a process.cwd()-anchored
primary path and an import.meta.url-relative fallback. Turbopack's
dev-mode static worker-chunk detector partially resolves the
new URL(literal, import.meta.url) construct in the fallback branch
independent of which branch actually runs at runtime, producing an
inconsistent module graph node. turbo-tasks then panics on startup
with either 'inner_of_upper_lost_followers...' (aggregation_update.rs)
or 'there must be a path to a root...' (module_graph/mod.rs), and
Restart=on-failure just silently retries forever.
git bisect (443c96d28 good .. b7a0c5413 bad, 418 commits, 9 steps)
isolated this to 657d3a484 (#11732). Confirmed by isolation probe:
removing the Worker construct, or inlining a single non-branching
new Worker(new URL(literal, import.meta.url)) at the call site, both
avoid the panic; only the two-branch resolver does not.
The import.meta.url fallback was also silently dead in production:
the standalone bundle (webpack) freezes import.meta.url to the
build-machine path -- the same app-wide gotcha already documented on
GATE_DEP_REL in llmlingua/worker.ts's fail-open probe. Dropping that
branch fixes both problems with the same change: process.cwd() alone
is correct in every real runtime layout this app uses (dev via
run-next.mjs, and the production standalone bundle, where
outputFileTracingIncludes already copies the worker script preserving
its process.cwd()-relative path).
Verified live:
- Dev (Turbopack): npm run dev reaches '[Next] dev server listening on
...' cleanly; previously panicked and looped under Restart=on-failure.
- Standalone build: npm run build produced a clean webpack compile and
a working .build/next/standalone/open-sse/lib/deepseek-pow-worker.mjs
at the traced path; running the real solveDeepSeekPowAsync from cwd =
the standalone dir spawned the worker and returned the correct nonce
for a fabricated DeepSeekHashV1 challenge.
- node --test tests/unit/deepseek-pow-js-only.test.ts
tests/unit/deepseek-web.test.ts: 44/44 passing.
- eslint and tsc --noEmit: clean on the touched file (tsc's remaining
errors are pre-existing, in unrelated test files).
The leading system message was read as `typeof content === "string" ?
content : ""`, so a Chat-Completions content-part array — valid for
`system`, and what every prompt-caching client sends — collapsed the
whole system prompt into an empty `instructions`. Upstream accepted the
request and reported a normal prompt_tokens count, so the model answered
with no instructions at all and nothing in the response said so.
Mid-conversation system turns already handled the array shape (#7056);
only the first one did not. Reuses buildResponsesTextParts() and joins
the text parts, since `instructions` is a string rather than a part array.
Co-authored-by: Vadim Zhyvylo <zhyvylo@involve.software>
* fix(api): keep registry width and type on embedding models
* docs: changelog fragment for embedding registry fix
* test(api): cover embedding width and type merge
Exercises /v1/models rather than the registry in isolation: a synced
model colliding with an embeddingRegistry entry must keep the width the
registry states, and a synced model the registry names must be typed as
an embedding model.
Fails on catalog.ts before e7fbb62 (2 failures), passes after.
Refs #11759
Add a direct low-level mode for users who require explicit control over provider selection. score selects the highest configured weighted score directly while reusing the existing exploration rate.
Exact ties preserve configured candidate order. rules and all other strategies remain unchanged.
- streamReadiness: reset deadline on each received chunk (keepalive = alive)
with a hard maxTimeoutMs ceiling so truly-dead connections still fail fast.
Preserves operator's 20s/100s intent for dead pulls while allowing slow-but-alive
upstreams (reasoning warm-ups) to survive.
- chatCore + codexIdentity: auto-detect Claude Code CLI via user-agent/originator
headers and enable model echo for it. The response field now echoes
the originally-requested alias/combo (e.g. ) instead of the
resolved upstream id (e.g. ), so restores
cleanly without 'could not be restored' errors.
Refs: opensource-elearning/omniroute-fixes#1, diegosouzapw/OmniRoute#12185
Fase 2 da auditoria código×docs 2026-08-31: ~90 divergências corrigidas em README/llm.txt(+42 espelhos)/SVGs/AGENTS.md/25+ docs; correções semânticas (breaker 8/12/2 + DEGRADED, webhooks sem eventos fantasma, ROUTE_GUARD_TIERS completo, API_REFERENCE sem fantasmas, reasoning 200); gate check:docs-counts endurecido (versão em prosa, patterns anti-evasão, superfície +llm.txt/mcp-server/omni-mcp/tier-flow) +5 testes; fonte do gerador de agent-skills corrigida (107/32 → 110/33) e mesma família varrida do Copilot prompt, 39 locales, skills/README, CONTRIBUTING e 2 guias.
