Skip to content

Never stop coding. Free MIT AI gateway: one endpoint, 375 providers (150+ free), 1,400+ models Kimi, Claude, GPT, Gemini, GLM, DeepSeek, MiniMax. Works with Claude Code, Codex, Cursor, OpenCode, Cline & Copilot. Quota-aware auto-fallback, RTK+Caveman compression saves 15-95% tokens, MCP/A2A, Desktop/PWA. Built by 900+ contributors

3.4 MB
Raw Blame

Changelog

[Unreleased]

✨ New Features

  • feat(dashboard): adaptive context-budget dial on the compression settings panel — mode (off / floor / replace-autotrigger) and policy (reserve-output / percentage / absolute) persist via PUT /api/settings/compression contextBudget. Completes the dashboard half of #7005 (API + DB already shipped in #7183).
  • feat(sse): STRICT_ZERO_COST — opt-in, off-by-default freeAccessPolicy: "strict" setting that hard-verifies every auto-combo candidate against live quota state and per-connection economic safety before it can be dispatched, going beyond hidePaidModels's static catalog check. Adds curated hardStopGuaranteed metadata to FREE_MODEL_BUDGETS, a short-TTL quota cache reusing getUsageForProvider(), and a connection-safety guarantee: a candidate backed by multiple accounts has its allowedConnectionIds narrowed to exactly the connections independently verified SAFE, so dispatch can never use an unverified account. An excludeTosAvoid guard (default false) is available separately for contractual risk. See docs/routing/STRICT_ZERO_COST.md.

  • feat(services): show sanitized CLIProxyAPI account health from its authenticated management API without exposing credentials, file paths, or raw account metadata (#6342)
  • feat(credential-health): pace the credential health sweep per connection via provider_connections.healthCheckInterval (minutes, 0 = never), with CREDENTIAL_HEALTH_CHECK_INTERVAL as the global default (#8443)
  • behavior change: healthCheckInterval is a shared column — it paces both the OAuth token refresh and the credential health sweep, and 0 disables both. The connection editor defaults it to 60, so configured OAuth connections are now credential-checked at 60min instead of the previous ~10min (aligned with the probe-volume goal of #8443)
  • feat(providers): publish Poolside's Laguna Preview catalog statically — poolside/laguna-xs-2.1 and poolside/laguna-s-2.1 (262144 context, 32768 max completion, tools + reasoning, text-only), so the models are routable and visible before a key is configured instead of only after live discovery. Pins the catalog form of the XS id against the laguna-xs.2 variant carried by third-party listings. (#9085)
  • feat(modality-bridge): bridge Chat and Responses video parts through a strict trusted-loopback, quota-bounded FFmpeg broker; enforce HTTPS redirects/SSRF plus format, protocol, stream, pixel, frame, 50 MiB broker/remote, 36 MiB inline, and 120-second limits; propagate caller aborts; preserve the actual successful fallback model through cache/meta/headers; expose sampled latency and honest success telemetry; and ship the localized Video settings UI (#9760)
  • feat(radar): Persist local model display-name/enabled overrides and hide/restore tombstones, with authenticated catalog controls and feed safety precedence (#9830)
  • feat(radar): add curated-family combo suggestions, a guided combo page, and the read-only Radar MCP catalog tool (#9836)
  • feat(radar): add a signed live offers feed and supporter offers dashboard (#9912)
  • feat(radar): add signed Intel insights, supporter recognition, and local Radar CLI commands (#9923)
  • feat(radar): add a localized public news feed and dismissible dashboard launch banner, with the Radar announcement staged inactive for a separately authorized launch (#9926)
  • feat(admission): add lane-aware admission probes for combo/fusion/chaos fan-out (fail-open, queueing disabled), an env-wins OMNIROUTE_CHAT_VIRTUAL_LANES activation flag applied at boot, and adaptive-lane visibility in the omniroute_get_health MCP tool (related to #9654)
  • docs(mcp): complete the MCP server README tool reference so the schemas/ catalog is fully covered (agent-skills, oneproxy, web, tool-search, combo/routing, pricing and DB-health tools were previously only discoverable via omniroute_tool_search)
  • feat(cli): container-aware auto-config — setup-*, omniroute configure, omniroute config set and the CLI-tool config APIs now refuse to write into a containerised OmniRoute's ephemeral home (CLI exits 2, API returns 422 with containerEphemeralTarget) and point at the host-CLI or bind-mount setup instead; --allow-container-write / OMNIROUTE_ALLOW_CONTAINER_CONFIG_WRITE=true opt back in. Also fixes CLI_CONFIG_HOME so the Compose host profile's /host-home bind mounts are honoured instead of silently falling back to the container home. (#10057)
  • feat(dashboard): opt-in DASHBOARD_ALLOW_EMBED=vscode relaxes CSP frame-ancestors to 'self' vscode-webview: and drops X-Frame-Options for HTML pages only, so the dashboard renders inside the VS Code Simple Browser (OmniCopilot). Default posture unchanged — API routes stay unframable (#10273)
  • feat(resilience): warn when /healthz is served under event-loop lag ≥200ms so a slow 200 is visible as sick, not healthy (#10303)
  • feat(docker): add GET/HEAD /livez as a process-alive probe, distinct from /healthz readiness (#10316)
  • feat(providers): add Cloudflare AI Playground as a No Auth provider (cloudflare-playground, alias cfp) — free anonymous chat over the reverse-engineered cf_agent WebSocket protocol (PartySocket transport, no account/API key/cookies) with GLM 5.2, Kimi K2.7 Code, DeepSeek V4 Pro, gpt-oss-120B, Llama 3.3 70B, Qwen2.5 Coder 32B and 14 more curated models. The executor drives a headless Chromium via Playwright (the WS upgrade is TLS-fingerprint-gated), translates the cf_agent frame stream into OpenAI SSE, and surfaces upstream rate limits (3021) as HTTP 429. Fixes #10389
  • feat(providers): AI Horde accepts an optional registered API key and advertises only live image models that currently have workers (#10542)
  • fix(providers): AI Horde Check validates keys via /v2/find_user instead of the unauthenticated OpenAI models list (#10542)
  • feat(audio): proxy native ElevenLabs voices, text-to-speech, and speech-to-text HTTP routes through stored OmniRoute credentials, preserving query strings, multipart uploads, binary responses, and upstream errors (#10556).
  • feat(providers): complete Jina AI as one credential pool — dashboard jina-ai / jina-reader share a token, JINA_AI_API_KEY is a real fallback, Test probes GET https://api.jina.ai/v1/models (embeddings fallback hits jina-embeddings-v5-omni-small), embed/rerank logs keep connection_id, catalog adds jina-reranker-v3.5, Omni v5 multimodal {text}/{image}/{content} docs pass through intact, and OmniRoute proxies classify / segment / jina-search (s.jina.ai). Reader stays a separate r.jina.ai card with an explicit label. Gemini Embedding 2 (gemini/gemini-embedding-2, alias google/gemini-embedding-2) uses dashboard gemini keys (or GEMINI_API_KEY / GOOGLE_API_KEY only when none exist), forwards native multimodal parts, and maps N OpenAI input items to N :batchEmbedContents vectors instead of one aggregated :embedContent. (#10581)
  • feat(providers): accept response_format=ogg on /v1/audio/speech as an alias for the existing Opus/Ogg encoder (#10587)
  • Added Google AI Studio Gemini batch text-to-speech support through POST /v1/audio/speech.
  • feat(settings): add autoDisableBannedScope so permanent-ban auto-disable can target subscription/OAuth accounts only, leaving prepaid API keys in the routing pool (#10617)
  • feat(server): emit systemd sd_notify READY/WATCHDOG/STOPPING (generated unit becomes Type=notify with WatchdogSec=180) so a frozen server process is killed and restarted by systemd instead of lingering undetected
  • feat(providers): add the TabiToken NewAPI gateway (tabitoken) and teach the existing HCNSec entry (hcnsec) the three further protocols it actually serves. TabiToken leaves the NewAPI pricing endpoint public, so its catalog is read from the host rather than guessed: four Claude models, each reporting the Anthropic and OpenAI protocols. HCNSec shipped OpenAI-only; probing the host showed /v1/messages, /v1/responses and the Gemini /v1beta path all reach its token layer, so each is now declared as an alternate format — with its default format, base URL, auth scheme and regional catalog classification untouched. (#10668) — thanks @yawar-aquil
  • feat(sse): allow an alternate protocol to build its own upstream URL. AlternateFormat gained an optional urlBuilder, because the Gemini protocol carries the model inside the path ({base}/{model}:generateContent) and the existing chatPath/urlSuffix fields are constants that cannot express it. The route builder is extracted as buildGeminiGenerateContentUrl and shared with the native gemini provider so the two consumers cannot drift on the ?alt=sse streaming suffix. (#10668) — thanks @yawar-aquil
  • feat(call_logs): persist the per-call error family in call_logs.error_type and expose a failure breakdown (errorBreakdown) in the usage analytics endpoint, reusing the existing production classifier (#10670)
  • feat(proxy): the proxy-health sweep and GET /api/settings/proxies/egress now report an anonymous summary of egress-IP sharing — how many rotation groups share an egress IP and the largest number of accounts behind one IP — computed from persisted proxy_logs over a 24h window. No IPs and no account identities by default; PROXY_LOG_INCLUDE_IPS=true restores raw details. (#10677)
  • docs(guides): OmniRoute now serves VS Code's native Copilot Chat model picker through the OmniCopilot extension (Marketplace · Open VSX — Cursor, Windsurf, VSCodium, Theia…) — no Copilot subscription needed since VS Code 1.122. New docs/guides/VSCODE-COPILOT.md covers setup, how the picker collapses the dual-prefix catalog via GET /v1/models?prefix=alias, and the build-time DASHBOARD_ALLOW_EMBED=vscode flag that renders the dashboard in an editor tab (#10697)
  • feat(docker): DASHBOARD_ALLOW_EMBED is now a Docker build argument — docker build --build-arg DASHBOARD_ALLOW_EMBED=vscode produces an image whose dashboard renders inside the VS Code Simple Browser (OmniCopilot's dashboardOpen: "editor"). Previously the flag was only reachable from a source build: Docker silently drops a --build-arg with no matching ARG, so the operator got the default image and no error. Builder-stage only and empty by default — the runtime stages deliberately do not carry it, and the unframable default posture is unchanged (#10701)
  • feat(providers): new cursor-api provider (card "Cursor API", alias cua): connect a Cursor user API key (crsr_…) and route cursor-api/<model> through the existing Cursor agent executor (the key is exchanged for a 1h session token and cached), plus a /api/cursor-cli/* passthrough so the Cursor CLI itself runs through OmniRoute (CURSOR_API_ENDPOINT=http://<omniroute>/api/cursor-cli, CURSOR_API_KEY=<OmniRoute key>) with every RPC attributed and logged. The IDE cursor provider is unchanged. (#10729)
  • feat(api): GET /api/health now answers { status, timestamp } without a key. Until now the path had no route, so the management-auth boundary answered first with a 401 — indistinguishable from a wrong key or an unknown route, which left Docker HEALTHCHECKs and Kubernetes probes unable to tell "down" from "misconfigured". Kept deliberately minimal: version, uptime and memory stay behind the authenticated /api/monitoring/health (#PRNUM).
  • feat(routing): make Task-Aware Smart Routing's detection patterns operator-configurable via settings.taskRouting.patternOverrides (PUT /api/settings/task-routing) — the built-in patterns are English-only, so a non-English dashboard had no recourse short of turning detection off entirely; an override now replaces the pattern list for one task type without touching the rest (#10783)
  • feat(api): accept PATCH on /api/combos/[id], the verb the OpenAPI spec already documents (#10869)
  • feat(sse): add GLM-5.3 support (glm-5.3, glm-5.3-high, glm-5.3-low) across the z.ai first-party providers, mapping the upstream reasoning_effort request parameter to the existing 5.2 tier UX (#10896) — thanks @phuongddx
  • feat(home): add a live Recent Requests panel beside the home Provider Topology (polls GET /api/usage/call-logs?excludeTests=1 every ~3s, gated by the topology appearance toggle + page visibility). excludeTests is now an allowlist of real provider inference (/v1/% or /api/v1/%), applied before LIMIT, so connection-test/model-sync/management rows can never leak into the feed (#10897, extracted from #8450) — thanks @nguyenha935
  • feat(rankings): free provider rankings now expose a reliability field (raw testStatus/rateLimitedUntil per connection plus a healthy/degraded/down state, reusing the ProviderHealthState vocabulary of the provider health matrix) when the configured/available filters are active — derived from already-loaded data, without touching the ranking order (#10909)
  • feat(resilience): when an allowlisted provider (opencode family) answers 429 classified quota_exhausted or rate_limit_exceeded and its free-tier quota is bucketed by egress IP (#9611), every connection of that family sharing the IP is cooled down together before the rotation tries them — one guaranteed-failed upstream call per episode instead of N, on the combo path as well. For the allowlisted family a 429 now cools the connection instead of locking a single model. Exclusive allowlist, never terminal, best-effort when the egress IP is unknown (#10920).
  • feat(rankings): free provider rankings can now report what each provider actually served — reliability.usage (requests, successes, success rate over a window) behind the opt-in withUsage/usageRange query parameters, so a provider that answers every call with an error is no longer described as healthy (#10926)
  • feat(providers): add Logfare as a free OpenAI-compatible provider — dashboard card with a Free badge and request-logging disclosure (every prompt/completion is logged for research; opt out at logfare.ai/consent), live model discovery from https://logfare.ai/v1/models (20 models, 11 chat-capable: kimi-k3, deepseek-v4-pro, glm-5.2, gpt-5.6-luna, minimax-m3…), full chat/streaming through the existing OpenAI-compatible path, the real Logfare logo on the card, and a listing in the free-tiers guide. (#10987)
  • Run synchronous RTK and Caveman request compression in a bounded worker-thread pool, keeping large /v1/responses compression heaps outside the HTTP isolate while preserving strict fail-open behavior and per-engine telemetry.
  • feat(providers): let operators declare per-provider error rules through settings.providerErrorRules instead of patching the catalog — an operator-supplied rule for a provider is consulted before the built-in providerRuleRegistry, receives the raw error text, and has its declared scope/cooldown/reason actually honored end to end, for any provider (declaring the rule is the opt-in — no extra allowlist entry needed). Matches are plain case-insensitive substrings (never RegExp) and bounded to 50 rules to keep the hot path safe (#11104)
  • feat(combo): the shared per-request combo attempt budget is now operator-configurable via maxGlobalAttempts (combo config / comboDefaults cascade), instead of the hardcoded 30. Lower it to fail fast on a dead target pool, raise it for large combos; clamped to [1, 200] so an unbounded budget can never cause runaway background requests (#11134)
  • feat(api): /api/usage/om-usage gains a structured form — ?format=json returns the key's own usage as ApiKeyUsageLimitStatus + UsageSnapshot instead of text/plain. This is the surface a UI (the OmniCopilot panel) consumes to show a key holder their daily/weekly spend and quota reset. The route is self-service (the caller's own key, gated by allowUsageCommand), not the management surface; refusals come back as a discriminated { "allowed": false, "error": … } so a UI can tell "not allowed" apart from "allowed but nothing cached yet". The endpoint was previously undocumented in API_REFERENCE.md; it now has a section (#11190)
  • feat(api): /api/usage/om-usage?format=json now returns providers[] — every connection's quota snapshot, not just the single selected one — so a panel can render Codex / Claude / OpenCode side by side. The collector already gathered all of them; the single-pick provider field (kept) is a terminal presentation choice. Closes the per-connection gap from OmniCopilot #8 (#11192)
  • feat(providers): allow overriding the rate-limit queue wait timeout (maxWaitMs) per connection, alongside the existing rpm/tpm/tpd/minTime/maxConcurrent overrides — a single slow provider no longer has to lower the global wait budget for every other provider (#11251)
  • feat(dashboard): replace the hard Home → onboarding redirect with a dismissable first-run readiness card so returning users can stay on Home while new users still get a clear 4-step path (#11282)
  • feat(dashboard): lead Traffic Inspector with a purpose-first header that separates "what happened" from "how it happened", so beginners can read request outcomes without drowning in protocol detail (#11283)
  • feat(dashboard): add an Essentials sidebar preset that shows only the beginner core path (Home → Endpoints → API Keys → Providers → Health → Settings) while keeping Advanced tools reachable via Command Palette search (#11286)
  • feat(video bridge): harden the optional drill-down cache substrate with exact-path broker policy, canonical principal/session/media isolation, independent retained-byte quotas, cancellation-safe commits, rejection of excess or non-canonical Base64 padding and non-JPEG/truncated media, warning-sensitive full JPEG canonicalization that strips trailing polyglot bytes, server-derived dimensions, and auditable derivation metadata; production tenant binding and multi-resolution selection remain follow-up work (#11369)
  • feat(video): add an opt-in focused analysis mode that safely uses a normalized, 500-code-point latest-user hint for task-aware frame captions while preserving full-mode prompts, temporal-window isolation, and cache identity without storing raw task text (#11383).
  • feat(command-code): advertise low/medium/high/xhigh/max reasoning-effort suffixes for reasoning-capable models in the catalog and Combo Builder, with request-time resolution to reasoning_effort
  • feat(crof): advertise reasoning-effort tiers (none/low/medium/high/max) for live-discovered and seed models, so the catalog, Playground, and Combo Builder surface - aliases and requests resolve max upstream
  • feat(sse): add Cursor plan image generation via Agent CLI (IMAGE_PROVIDERS.cursor, format cursor-agent-image), reusing the chat Cursor OAuth connection
  • feat(routing): add the default-off DISABLE_CONTEXT_WINDOW_CHECKS feature flag to let operators bypass OmniRoute's local context-window and max-input-token check for direct single-model requests, leaving upstream limits, prompt compression, and output-token caps intact.
  • feat(catalog): surface runtime-learned reasoning_effort tiers in /v1/models capabilities.effort_tiers (learned set replaces synced metadata when present), map them to OpenCode ModelV2.variants in the OmniRoute plugin, and align dispatch -<tier> suffix validation to the effective (learned ?? synced) set — so the UI offers exactly the tiers the upstream accepts (e.g. {low, high, max} for oc/x-preview-f-free) and each advertised variant completes. Excludes codex/glm/kimi, which keep their own dedicated -{effort} suffix mechanism and never gain effort_tiers from this path (related to #7694, builds on #11232)
  • feat(usage): show Kimi Coding's fixed-order Code 5-hour/7-day quota windows plus Extra Usage status, balance, monthly spend/limit, and the official Additional Credits link on Dashboard → Quota cards.
  • feat(providers): copilot-m365-web now supports OpenAI tool calling — a router planning turn asks the substrate model (as a tool-selection assistant emitting CALL_TOOL: name({...}) / NO_TOOL_NEEDED text, which bypasses its plugin-registry refusal) and validated decisions surface as tool_calls with finish_reason: "tool_calls" in both stream and non-stream modes; also flattens the full message history (assistant tool_calls + compacted tool results) so multi-turn agent loops keep context, replies to SignalR type:6 keepalives, surfaces type:3 error frames instead of a silent empty stop, and suppresses writeAtCursor text from tool-progress frames
  • feat(api): add GET/POST /v1/multimodal-embeddings as an alias of /v1/embeddings so Jina-compatible clients do not receive HTTP 404 unknown_route — thanks @RaviTharuma
  • feat(opencode-go): expose Muse Spark 1.2 Contributor reasoning-effort aliases (minimal/low/medium/high/xhigh) in the Combo Builder
  • feat(providers): restore the operator-owned upstream timeout tier per connection via providerSpecificData.timeoutMs (preempts the maintainer-only model/provider registry tiers and the global FETCH_TIMEOUT_MS), and make the combo per-target timeout ceiling follow the selected connection
  • feat(cli): run omniroute serve --tray as a detached desktop process after server and tray readiness, with graphical login auto-start support.
  • feat(routing): add client-, provider-, and model-neutral exclusive managed session connection leases with API-key-bound generation fencing, durable SQLite ownership, explicit allowlist policy, and bounded 429 capacity retry semantics.

[3.8.52] — 2026-10-19

Living section — reconciled 2026-10-19 from all cycle commits (2f42a9ac19 → 0a1265ea81, 1,386 non-merge commits). Bullets carry the merged PR and its author; direct pushes are listed with their commit hash. Regenerated at each /generate-release phase.

📊 Release by the numbers

👥 People who contributed 318
📝 Commits in the cycle 1,386
🔀 Pull requests referenced 1,369
📋 Changelog entries 1,442
🙌 Contributors credited in entries 317
🤖 Automated dependency commits 17

Entries by type

Type Count
🐛 Fixes 1091
✨ Features 133
📚 Docs 55
🧪 Tests 51
🧹 Chore 45
⚙️ CI 30
⚡ Performance 21
♻️ Refactor 11
🏗️ Build 3
🔒 Security 1
📦 Dependencies 1

🏆 Top 25 contributors this cycle

By commits in 2f42a9ac19..0a1265ea81, author identities consolidated via .mailmap and the merged PR's GitHub login. Bots excluded.

# Contributor Commits
🥇 diegosouzapw 592
🥈 Dizzle (@maxmad64bis) 134
🥉 Paco Cartones (@pacocartones) 62
4 Bob.Hou (@HouMinXi) 49
5 WebPerson (@jonlwheat2-gif) 39
6 Lance Woodson (@woodsonl) 38
7 Fabio Junior Gruchovski (@fabioKoruja00) 19
8 Innokentiy Solntsev (@insoln) 19
9 Caio Figueiroa (@shipsfromrio) 18
10 小妍儿 ✨ (@xiaoyaner0201) 18
11 lorenzozane (@lorenzozanee) 16
12 Praveen K Palaniswamy (@yourspraveen) 15
13 Junior-HJ 14
14 skygunner 14
15 QuangBlue 12
16 Fouad Salkini (@fouadSalkini) 11
17 Jan Leon (@JxnLexn) 11
18 opensource-elearning 10
19 Tiago Rodrigues da Rosa (@tiagolabsit) 7
20 Aldo (@aldoeliacim) 6
21 Ronaldo Davi (@alltomatos) 6
22 Dmitry Galenko (@D4rk4) 6
23 Kizuno18 6
24 Jacob Stoner (@jacobsparts) 5
25 Stanislav (@LazyGatto) 5

✨ New Features

  • feat(providers): add the mainland BigModel.cn (Zhipu) chat-completions provider and correct Z.AI platform links (#12343) — thanks @jackjinke
  • feat(runtime): new REASONING_REPLAY_ENABLED feature flag (Settings → Feature flags, on by default) to stop Reasoning Replay from caching and re-injecting real reasoning text across turns; upstreams that require the reasoning field on replay turns still receive the internal placeholder when it is off (#12486 — thanks @jacobsparts).
  • feat(providers): add NoTrack (notrack.ai) as a free cookie-auth web provider — catalog id notrack-web (alias ntw), /api/dispatch SSE, session cookies uid/si_usr_id/si_ses_id. (#12534) — thanks @jeanbispo
  • feat(services): Added an LLMLingua embedded-service supervisor + HTTP sidecar wrapping the real @atjsh/llmlingua-2 prompt-compression package (ONNX TinyBERT model, downloaded from Hugging Face on first use). The compression engine dispatches /compress calls to the sidecar over HTTP, falling back to the existing in-process worker-thread backend when the sidecar is down. (#12967) — thanks @rqzbeh
  • feat(providers): Hugging Face connections support an optional Bill-To account sent as the X-HF-Bill-To header for cost attribution (#13016) — thanks @Abhijeet1520
  • feat(auto-combo): new nadir auto-router strategy (config.routerStrategy: "nadir") asks Nadir's decision API which model in the candidate pool the prompt needs and routes to it, leaving the connection choice to rules; opt-in, fail-open with a 30 s cooldown after failures, only the last user turn and the model ids leave the box (#13051) — thanks @doramirdor
  • feat(codex): Add explicit per-account paid-credit opt-in with an Extra credits On/Off button in the provider account list. Accounts with available OpenAI credits can continue after their subscription cutoff; credit availability is shown separately, while Spark limits and upstream cooldowns remain effective (#13337) — thanks @feci
  • feat(api-keys): Allow per-key service-tier overrides for OpenAI OAuth requests, preserving inherited defaults and enforcing explicit tiers across combos, retries and Responses transports. (#13554) — thanks @JxnLexn
  • feat(sse): opt-in per-model concurrency caps per connection — configure exact modelConcurrency ceilings inside rateLimitOverrides (dashboard editor or PATCH /api/providers/[id]), enforced as an atomic fourth gate alongside the global/provider/account semaphores with local queueing; unconfigured connections behave exactly as before. Also fixes rateLimitOverrides.executionMaxWaitMs: the schema accepted it but the DB allowlist rejected it, so saving it failed; a dashboard save now also preserves an API-set value (#13700) — thanks @Notaloop763
  • feat(sse): a request no longer dies when the upstream rejects thinking by naming the model instead of the offending field. OmniRoute now drops whichever reasoning field the body actually carries (reasoning_effort, reasoning, thinking, think) and retries once. Ollama rejects this way for Instruct-only models such as Qwen3-Coder, where the 400 reads "Qwen3-Coder:latest" does not support thinking. Neither existing detector could see anything to strip there, so the error went straight back to the client and every Claude Code request carrying a thinking block failed against a self-hosted non-thinking model. (#13868) — thanks @cryptiklemur
  • feat(providers): add Twinmind (twinmind / tm) unofficial web chat provider — Firebase refresh, grouped model discovery, SSE text_delta → OpenAI chunks, emulated tool calling for OpenCode/Claude Code (#14131) — thanks @artickc
  • feat(resilience): skip CLIProxy-backed combo targets before dispatch using management account health, ignoring stale per-model quota rejections (#14206) (#14301) — thanks @sprintberlin
  • feat(providers): WorkBuddy (www.workbuddy.ai, Tencent) is now a first-class provider with its own device-flow OAuth module, catalog entry and token refresh. It is a separate product from CodeBuddy CN (different host, account system and catalog, and a credential from one is not valid on the other), so it deliberately does not alias the CodeBuddy OAuth module — aliasing would point the state request at copilot.tencent.com and mint the wrong credential. The gateway is OpenAI-compatible but accepts only a single-turn request: it rejects the replayed history an agent client sends from turn two on, and reports each violation as a generic code that reads like a transport fault (11148 tool_call_sequence_broken, 11155 reasoning content from the previous turn must be passed back, 11128 Illegal API invocation from an unapproved channel, 11101 non-stream rejected). The repair therefore lands in core as two opt-in RegistryEntry flags — strictChatHistory and bodyStringReplacements — honoured by the existing transformRequest hook and inert for every provider that does not declare them, so no sidecar proxy is required. The catalog is left models: [] with passthroughModels: true and discovered from the authenticated GET /v3/config the official CLI itself uses, because WorkBuddy serves no OpenAI-shaped /models endpoint (/v1/models and /v2/models both 404). That payload is an authenticated GET /v3/config, which returns 22 chat models (observed live via the desktop app's own log); that list replaces the membership of the 26-entry catalogue the CLI ships in cli/product.json rather than extending it, so the image and video entries that exist only in the built-in file are kept out of a chat roster by their capability tags. (#14212) — thanks @zeeshanhaque21
  • feat(rerank): /v1/rerank expands named combos with ordered fallback from a local rerank node to hosted providers (#14286) — thanks @seanford
  • feat(providers): Added onomeo as an OpenAI-compatible API-key gateway (https://onomeo.com/v1), with live model discovery via passthroughModels. The free note states the current call-metered free tier (about 30 free models at 30 calls per 5 hours per account and 12 requests/min per key, inside a shared site-wide pool), that large models are paid from a USD balance with a $0.50/day beta cap, and the optional top-up from $5, and the hint says models are served by third-party upstreams and that onomeo is in public beta (#14297) — thanks @Owaid0723666
  • Share CLI capabilities, aliases, setup recipes and model flags across Node and server consumers, including 5dive; add persistent Gemini setup, private config writes, restricted child environments, signal-aware cleanup and registered native ACP lifecycle. (#14387)
  • feat(sse): Jina token balance in Provider Limits — jina-search, jina, jina-ai and jina-reader connections read the wallet total_balance from GET https://dash.jina.ai/api/v1/api_key/fe_user (60 s cache, fail-open) so rotation can skip exhausted keys (#14449) — thanks @HouMinXi
  • feat(cli): Add opt-in omniroute models test-add for local and remote contexts, with management-authenticated generation, streamed synthetic tool-call and continuation checks pinned to the selected connection. Register the custom model only after verified dispatch, unchanged configuration and transactional readback; return an informational validation receipt without executing real tools. (#14506)
  • feat(providers): add Claude Opus 5.5 (claude-opus-5-5) to the Anthropic API, Claude OAuth, Bedrock and Vertex catalogs — native 1M context, 128K output, five effort tiers (default medium), $4/$20 pricing, Fast Mode, Opus 5 fallback; disabled thinking, manual budgets and forced tool_choice are normalized before dispatch since Opus 5.5 rejects them. The cc/ Opus CLI default stays claude-opus-5, because claude/anthropic connections synced before the launch reject Opus 5.5 until their model list is refreshed (Refs #14612) (#14546). — thanks @ggiak / @HouMinXi / @PixmaNts
  • feat(plugin): filter the published catalog by capability — freeOnly, toolsOnly (on by default, hides models that cannot call tools; set toolsOnly: false for the full catalog), visionOnly (#14554) — thanks @maxmad64bis
  • feat(usage): Per-key weekly USD limits can use a calendar week instead of the earliest upstream account reset. OMNIROUTE_API_KEY_WEEKLY_WINDOW=calendar gives every key the local Monday 00:00 week in OMNIROUTE_API_KEY_WEEKLY_WINDOW_TIMEZONE (default: the process timezone); unset keeps the current behavior (closes #14601) (#14603) — thanks @xiaoyaner0201
  • feat(api-keys): new USAGE_LIMIT_IGNORE_UNPRICED feature flag (Settings → Feature flags, off by default). When on, per-key USD usage quotas count models that have no price as $0 instead of failing closed (#12341), so keys that use subscription or free models without pricing rows are no longer reported and blocked as over quota. Priced spend is still enforced, and dailyHasUnpricedUsage / weeklyHasUnpricedUsage still report the unpriced usage (#14799). — thanks @fouadSalkini
  • feat(providers): add Apmix (apmix.ai) as an OpenAI-compatible, subscription-based provider — one monthly plan with a weighted-token allowance, chat-completions + Responses API at https://api.apmix.ai/v1, live /v1/models discovery keyed to the plan's model union (seed catalog as offline fallback), and a usage fetcher (GET /v1/usage) surfacing the monthly allowance, self-imposed daily/weekly caps and top-up balance in Dashboard › Limits with quota-aware preflight (#14821). — thanks @aref-alapour
  • feat(providers): add BeatAPI as an OpenAI-compatible gateway (#14875) — thanks @erickkkyt
  • feat(api): add or remove models and combos on an API key atomically (#14880). — thanks @fouadSalkini
  • feat(dashboard): full-page API key access editor with tabs, replacing the permissions modal (#14885). — thanks @fouadSalkini
  • feat(dashboard): models and combos catalog with filters and browser-side health tests (#14887). — thanks @fouadSalkini
  • feat(codex): opt-in per-connection codexPromptCacheKeyScope: "thread" aligns the outbound prompt_cache_key with the converged thread id under session-mode fingerprint convergence, so it matches thread_id and stays scoped to the account instead of following a conversation across accounts on failover; the default (client) keeps forwarding the client's key unchanged (#14907) — thanks @gabrielalberton
  • feat(cli): setup-claude and the Claude profile auto-sync skip models of local-CLI providers (zcode, auggie, devin-cli-agentic, codex-app-server) whose binary is not detected on the host, print which providers were skipped and why, and take --include-local / --all to write them anyway (#14952) — thanks @shipsfromrio
  • feat(oauth): Claude Code connections can be added from a claude setup-token token. Dashboard → Claude Code → Add Connection now opens on a Setup Token tab (Browser Login is one click away and no longer starts on open); POST /api/oauth/claude/import-token verifies the token with one max_tokens: 1 request before storing a refresh-less OAuth connection named "Setup token", and re-pasting the same token updates that connection. The connect dialog also resets to its first step with an empty field on every open (#15093) — thanks @dpozimski
  • feat(providers): Added Y-API as an OpenAI-compatible API-key gateway (https://api.y-api.bestvirtualgoods.com/v1), with live model discovery via passthroughModels — its GET /v1/models requires a key, so no model list is seeded. Listed in AGGREGATOR_PROVIDER_IDS because its own catalog states that every model is served by the gateway from third-party upstream vendors rather than trained by it, and its free badge quotes only the models the publisher itself prices at 0 credit. (#15131) — thanks @jiweiyeah
  • feat(dashboard): per-account "Test message" action on each provider connection — a saved test model per connection (Edit connection) and a global test prompt in Settings → AI, backed by GET/PUT/POST /api/providers/{id}/test-message; the quota card footer actions also wrap responsively (#15138 — thanks @JxnLexn)
  • feat(proxy-subscription): generate the local proxy-core config beside the adopted file, opt-in per subscription (#15141) (#15564) — thanks @maxmad64bis
  • feat(proxy-subscription): verify the generated core configuration with the native binary before replacing the adopted file (#15142) (#15564) — thanks @maxmad64bis
  • feat(providers): Add GPT-6.1 Sol to the OpenAI and Codex catalogs with provider-specific limits, Standard pricing, Codex reasoning/Fast wiring, and a synchronized Codex 0.159.2 client identity (#15171) — thanks @xiaoyaner0201
  • feat(codex): add GPT-6.1 Sol support across execution and pricing pipelines — register gpt-6.1-sol and reasoning effort variants (ultra, max, xhigh, high, medium, low), bump Codex client version header to 0.159.2 to resolve upstream ChatGPT account tier-gating, surface upstream detail error diagnostics, retain Responses Lite parallel tool calls, and wire 2.5x Fast mode pricing. (#15192) — thanks @initguru
  • feat(providers): the Xiaomi MiMo Token Plan (mimotp) catalog now lists mimo-v2.6-pro and mimo-v2.6-flash (1M context, 128K output), ahead of Xiaomi retiring mimo-v2.5-pro / mimo-v2.5 on 2026-10-21 (#15219) — thanks @kang-heewon
  • feat(proxy-subscription): reload the local proxy core after a verified config replacement, paced at most once per 60 s per subscription (#15238) (#15564) — thanks @maxmad64bis
  • feat(plugin): filter the published catalog by provider (providersAllow, empty means full catalog) with an explicit empty-result guard (#15239) — thanks @maxmad64bis
  • feat(admin): Expose selector member set-aside state (motive, window, repeat count) in pool visibility (#15240). — thanks @maxmad64bis
  • feat(proxy-subscription): prune nodes the core check rejects, then retry verification before keeping the previous file (#15295) (#15564) — thanks @maxmad64bis
  • feat(proxies): accept operator-provided dated egress observations for pool members — an authenticated push route stores dated observed addresses per member behind a new opt-in flag, merged with the journal read for display and pool order with the freshest observation winning (#15314) — thanks @maxmad64bis
  • feat(proxy): Add OMNIROUTE_UPSTREAM_HTTP2_ENABLED=false to force HTTP/1.1 on Undici direct, HTTP(S)/SOCKS proxy, relay, and retry dispatchers without disabling certificate verification. HTTP/2 remains enabled by default (closes #15313) (#15318). — thanks @Kizuno18
  • feat(proxies): Log refusal-store writes and show transport proof in pool visibility (#15374) — thanks @maxmad64bis
  • perf(db): Serve repeated synced vision verdicts from a version-gated memo instead of re-reading the catalog (#15378) — thanks @maxmad64bis
  • feat(dashboard): flag a provider whose exits fail upstream together so the pool screen reads it as an upstream problem, and show upstream 4xx and 5xx as separate counts (#15386) — thanks @maxmad64bis
  • feat(combos): the /dashboard/combos auto-routing catalog now loads the live built-in auto/* combos from GET /api/combos/auto (kind badge and candidate count per card) with Control Center, Test and Duplicate actions; GET /api/combos/auto?id= materializes one auto combo and POST /api/combos/test can test an auto/* combo (up to 8 highest-weight candidates). auto/* combos with no live candidates are no longer listed in /api/combos/auto or advertised in /v1/models (#15397 — thanks @afonsoft).
  • feat(sse): Agentic pipeline combos can configure ordered executor backups for failed responses and transport errors while preserving successful streams (#15419) — thanks @potatosips.
  • feat(memory): opt-in claude-mem worker backend (pinned to 127.0.0.1), so /api/memory and A2A memory search read and write the same store the claude-mem Claude Code hooks fill; initMemoryBackends() now registers opt-in backends before configure() instead of failing silently on a non-SQLite primary/fallback (#15425) — thanks @engomarzaied-prog
  • feat(providers): ClinePass connections now show their 5-hour, weekly and monthly usage limits on the quota dashboard, read from Cline's /api/v1/users/me/plan/usage-limits, and reset-aware combo strategies can use those windows (closes #9740) (#15427) — thanks @PixmaNts
  • feat(i18n): the translation backend forwards the optional reasoning control when set (#15462) — thanks @maxmad64bis
  • feat(dashboard): token refresh failures can pause per account instead of blocking the whole provider (#15469) — thanks @maxmad64bis (with thanks to @xiaoyaner0201 for the scope proposal in #15444)
  • feat(plugin): trim the default view with per-provider fresh and showcase caps, tunable freshness window and new options (#15484) — thanks @maxmad64bis
  • feat(proxy-subscription): warn with a translated count when generation skips entries instead of reporting not generated or staying silent (#15564) — thanks @maxmad64bis
  • feat(providers): Perplexity API (Sonar, Agent) and Perplexity Search requests now send an X-Pplx-Integration: omniroute attribution header (#15573) — thanks @qirh
  • feat(dashboard): new Monitoring → Cooldown Manager page lists every connection that is out of routing (connection cooldowns, model lockouts, terminal states) and clears cooldowns per connection, for a selection, or for a whole provider, instead of opening each provider page. It also edits the stream-stall cooldown switch and the OAuth / API-key base cooldown and backoff steps. Backed by the new management-only GET/POST /api/resilience/cooldowns; terminal states (banned, expired, credits exhausted) are never cleared there. (#15630) — thanks @fouadSalkini
  • ZenMux image generation — route zenmux/ and zm/ models through the OpenAI Images API or Vertex AI :predict, normalize image output, and enforce Ming Design's text-only/model-selected dimensions contract. (#15670 — thanks @JeanRGW)
  • feat(embeddings): llama.cpp connections accept OmniRoute's canonical multimodal embedding items — image, audio (wav/mp3/flac) and video become llama-server content parts (image_url / input_audio / input_video) with inline data, one vector per item, for servers started with --embedding --mmproj (e.g. EmbeddingGemma 2). Remote sources go through the same SSRF-guarded fetch and 16 MiB request cap as Jina/Gemini (#15758) — thanks @hartmark
  • feat(providers): add the CodeBuddy International (codebuddy.ai) provider as codebuddy-intl (alias cbai), sharing the device-auth flow, executor, token refresh and quota fetcher with codebuddy-cn by realm (closes #15173) (#15799) — thanks @Junior-HJ
  • feat(combos): a successful model sync now flags combo steps pinned to models that are missing from the provider's synced catalog. The sync response lists them in staleComboRefs and an audit entry is written. Combos are never modified (#15800) — thanks @tiagovilasboas
  • feat(combos): opt-in COMBO_AUTO_PRUNE_STALE_STEPS feature flag (default off): a successful model sync removes the combo steps pinned to models missing from the provider's authoritative catalog, with one audit entry per removed step. Failed, degraded and free-only syncs never prune, and a combo is never emptied (#15925) — thanks @tiagovilasboas
  • feat(quality): the forgotten-sibling gate and test-impact selection now follow path-literal edges: tests that read a workflow, root config or doc as a file (new URL(…, import.meta.url), resolve/path.join, readFileSync) are mapped in a new deterministic artifacts section of the impact map, so changing .github/workflows/quality.yml flags and selects check-workflows.test.ts without ever forcing a full run (#16068) (#16129)
  • fix(quality): artifact edges require an existing tracked regular file and reject symbolic links, including linked parent directories. Import specifiers are excluded by occurrence, so a separate file read of the same JSON path still reaches the sibling advisory and additive test selection; deletion and skip/todo masking protections remain unchanged. (#16129)
  • fix(quality): scan quoted path spans without unbounded regular expressions, preserve occurrence-specific imports across whitespace, and retain parentheses inside joined path segments. Oversized or incomplete literals do not create partial artifact edges. (#16129)
  • feat(search): opt-in provider-reported search cost — a connection with providerSpecificData.trustReportedCost: true bills the usage.search_cost_usd its operator-configured endpoint returns (also on error responses) instead of the static costPerQuery, recorded once per request_id in request_cost_ledger; caller provider_options.baseUrl endpoints are never trusted. Per-key daily/weekly USD usage limits now include search spend (#16088) — thanks @thotypous
  • Set ONNXRUNTIME_NODE_INSTALL=skip explicitly for the shared dependency-install action and direct checkout install steps, include the action policy in the dependency cache identity, and document local shell usage. The legacy .npmrc key remains for compatibility; published-package consumer installs are outside this checkout policy. (#16246)
  • feat(video): MiniMax Hailuo text-to-video now works through a native create → poll → file-retrieve transport (previously listed but rejected with Unsupported video format) (#16508) — thanks @octo-patch
  • feat(providers): GitHub Copilot auto model (github/auto): Copilot picks the model, which Copilot Free accounts require. OmniRoute resolves the pick before routing, so a Claude pick uses the native /v1/messages path, and usage, cost and call logs are recorded under the model that served the request (#16679) — thanks @Primexz
  • feat(stt): Deepgram transcription supports opt-in speaker diarization and verbose_json speaker segments (#16685) — thanks @n1klaus
  • feat(sse): opt-in coalescing of Claude content_block_delta events (SSE_COALESCE_*_MS), SSE frame batching (SSE_FRAME_COALESCE_MS) and an interval-based stall watchdog (SSE_STALL_WATCHDOG=interval), all off by default (#16722) — thanks @WYSTAF / @wallerjenifernk530-hash
  • feat(image): Muse Image 1.0 text-to-image and reference-image edits on the Muse Code (Meta) provider, reusing the muse-code API key or device-flow credential (#16871) — thanks @dasepmoch
  • feat(settings): selective transfer of provider accounts (and the custom provider nodes they use) between installations — passphrase-sealed file, import preview and per-item add/skip/replace that never deletes anything (#16877) — thanks @minthanthtoo
  • feat(providers): Baseten opt-in thinking via chat_template_args.enable_thinking, per-model reasoning effort clamp and refreshed model seed (#16879) — thanks @elidrissi-amine
  • feat(providers): Qwen Cloud Token Plan text-to-speech (qct/qwen-audio-3.0-tts-plus, DashScope SpeechSynthesizer, region-aware) and per-model reasoning-effort enums so out-of-enum efforts are mapped on the first request (#16886) — thanks @alysson-souza
  • feat(providers): add DaoXE OpenAI-compatible gateway (daoxe) with live /v1/models discovery and passthrough models (#16889) — thanks @seven7763
  • feat(guardrails): per-type PII selection, custom masking patterns (ReDoS-gated) and detection counters for the opt-in PII masking (#16902) — thanks @lsilveirab39
  • feat(updater): Docker-aware dashboard update — image containers show the host docker compose pull + recreate command, and DOCKER_UPDATE_TRIGGER_FILE enables one-click updates through a host-watched request file inside DATA_DIR (no Docker socket) (#16907) — thanks @semihisikman
  • feat(cli-tools): applying Codex settings can also write an OmniRoute model catalog (model_catalog_json, opt-in checkbox, off by default) so Codex's /model picker lists routable provider-prefixed models; fails open with a warning when Codex can't be read (#16910) — thanks @kalfian
  • feat(providers): add NaN.Builders API-key provider (nan) — chat with live model discovery, embeddings, STT/TTS (Kokoro default voice), image generation, rerank and per-model token quota usage (#16912) — thanks @tasarren
  • feat(models): list Gemini 4 Argon on the Gemini API (API-key) provider and give it a default pricing row (intro paid tier $2 / $10 / $0.10 cached per 1M), so API-key usage is no longer billed at $0 (#16913) — thanks @WYSTAF
  • feat(usage): show the Codex Luna Reserve quota rows as display-only usage rows that never affect account selection (#16915) — thanks @raflyazf
  • feat(compression): opt-in antislop-ui and antislop-copy output styles that steer generated UI and copy away from generic AI defaults (purpose-tested visuals; no em dash, generic CTAs or invented facts; WCAG AA contrast, keyboard and mobile rules) (#16916) — thanks @royhanmh
  • feat(cli): omniroute setup-zed adds OmniRoute as an OpenAI-compatible provider in Zed (key stays in OMNIROUTE_API_KEY), plus a Zed card in the CLI catalog (#16919) — thanks @YoannDev90
  • feat(dashboard): per-connection decode speed (avg/last tok/s, TTFT, sample count per model) on quota cards (#16920) — thanks @raflyazf
  • feat(api-keys): Pin an API key to one model or combo with the new forcedModel permission: every chat request made with the key is routed there even when the client hard-codes or omits the model, while the key's model, combo and budget limits still apply. (#16921) — thanks @sandypratama2030-gif
  • feat(providers): Display each connection's last successful model-sync timestamp beside Auto-Sync and trigger an immediate per-connection model refresh when enabling it, with safe cache invalidation and user-visible sync-failure feedback (refs #16940) (#16947) — thanks @fabioKoruja00
  • feat(api): opt-in capability / min_context / kind / q discovery filters on GET /v1/models (output unchanged without them) and a new POST /v1/models/query deterministic model shortlist (#16945) — thanks @vianhanif
  • feat(search): forward allow-listed, Zod-validated Exa and Tavily request fields via provider_options.exa / provider_options.tavily, map time_range to both, add the Tavily finance topic and keep their extra result fields in provider_raw (#16950) — thanks @asuan-dev
  • feat(providers): "Ping now" button on Codex OAuth connection rows sends one small request right away so the account's quota window starts, without changing Auto-Ping. It skips when the session or weekly quota is exhausted, honours exclusive leases and the provider circuit breaker, and allows one ping per account at a time (#16954) — thanks @Chumr26
  • feat(search): Parallel web search and fetch provider (parallel-search, alias parallel): one API key serves /v1/search (web only, mode-based pricing for turbo/fast/basic/advanced) and /v1/web/fetch (Markdown only, capped at 100,000 characters; a page Parallel cannot extract returns 422) (#16960) — thanks @tommyyzhao
  • feat(providers): optional custom ID suffix for OpenAI-/Anthropic-compatible provider nodes (openai-compatible-<apiType>-<suffix> / anthropic-compatible-<suffix>), with 409 on duplicate ids instead of a 500 (#16961) — thanks @rafi-zadanly
  • feat(dashboard): Copyable connection-ID chip on every provider connection row, plus docs for per-request account pinning with the x-omniroute-connection header (strict pin, single-model requests, Claude Code ANTHROPIC_CUSTOM_HEADERS example) (#16963) — thanks @nhan14t1
  • feat(settings): scheduled wakeups for Codex OAuth accounts: a new section in the Codex Quota Auto-Ping card pings chosen accounts daily at a set time (timezone + weekdays) or every N minutes, so each account's quota window starts when you want. It skips when the circuit breaker is open, the account is cooling down or under an exclusive lease, or its quota is exhausted (#16965) — thanks @Chumr26
  • feat(providers): Add Demon Route as an OpenAI-compatible API-key gateway for open-weight models, with encrypted normal connection storage, live key-scoped model import / Auto-Sync, curated offline seeds, exact vendor/model ID preservation, Chat Completions and streaming, a 120-second minimum upstream wait, provider metadata and generated reference, a setup/troubleshooting guide, and provider-specific regression tests. Free-lane models require a confirmed deposit and are not advertised as an unconditional free plan (refs #16979) (#17001) — thanks @fabioKoruja00
  • feat(i18n): translate up to N locales per request with --locales-per-request in sync-ui-keys (#PRNUM) (#15562) — thanks @maxmad64bis
  • feat(dashboard): add Compact and Exact token display mode toggle to Usage Analytics toolbar (#13019) — thanks @ZaimMarzuki
  • feat(antigravity): dynamically synchronize live-discovered models outside the static catalog into quota normalization, and introduce family quota pools (Gemini Models, Claude & GPT Models) across 5-hour and weekly windows with dashboard group tracking (#13739) — thanks @TheDemonTuan
  • feat(providers): add ChatPlayground (chatplayground, alias cpl) web provider routing models across Azure, LMSYS, and Perplexity endpoints with dynamic model discovery from https://app.chatplayground.ai/api/models (#12690) — thanks @james-martinez
  • feat(providers): auto-minting Clerk session JWT authentication via https://clerk.chatplayground.ai/v1/client/sessions/{sid}/tokens using the __client cookie and sid (derived from the __session JWT cookie), with in-memory caching and expiry skew (#12690) — thanks @james-martinez
  • feat(providers): support multi-account rotation and failover natively via OmniRoute's Combo Router and account fallback engine (#12690) — thanks @james-martinez
  • feat(providers): OpenAI-compatible streaming translation for ChatPlayground via TransformStream, stripping internal CHAT_ID:[a-zA-Z0-9_-]+ tokens (#12690) — thanks @james-martinez
  • feat(usage): add ChatPlayground quota & limits monitoring via GET /api/user tracking daily query allowances (including 300 credits/day for Lifetime subscription accounts) and midnight UTC resets (#12690) — thanks @james-martinez
  • feat(dashboard): add clamp/full description display modes to the feature-flags cards. A persisted toggle in the Feature Flags grid toolbar (localStorage, hydration-safe) switches every card's description between clamp (2-line clamp + full text on hover, via a styled tooltip) and full (always expanded). (#12740) — thanks @jonlwheat2-gif
  • feat(network): new OMNIROUTE_LOCAL_PROVIDER_NODE_HOSTS lets operators list hostnames (typically a Docker/Compose service name such as reranker) that provider-node routing treats as local: eligible for /v1/rerank, /v1/audio/* and the local health check without the *_REMOTE_PROVIDER_NODES flags, and called directly instead of through HTTP(S)_PROXY or a pinned proxy — so a sidecar no longer needs a hard-coded container IP. Empty by default; IP literals and cloud-metadata names are rejected with a warning (#14635) (#14640) — thanks @xiaoyaner0201
  • feat(providers): DeepSeek-web tool-call parsing now recognizes native DSML invoke/parameter tokens (both the printable-ASCII and full-width tokenizer spellings) in addition to the existing <tool>{json}</tool> shapes, and preserves parallel tool calls and stable call IDs/indexes; custom OpenAI-compatible image/video generation providers now fail closed on a missing base URL instead of silently falling back to a built-in provider endpoint (#14103) — thanks @arashashrafii
  • feat(providers): Added SYNTX.ai (syntx / stx) as an unofficial/experimental JWT provider. Chat reuses chat_uuid across OpenAI multi-turn history, PUT /api/v1/user/settings for a 4k system_prompt.default, and keeps compact generate on tools:[]. Media covers image, video, speech, music, transcription, and upscale. Live /api/v1/llm/models discovery, Limits via subscription + rolling-window quotas, JWT validation, and a 10-minute stream-readiness window. (#14269) — thanks @artickc
  • feat(providers): add Token Market (tokenmarket) as an OpenAI-compatible multi-model API gateway with Bearer-key authentication and live model discovery from /v1/models. (#13191) — thanks @FitBBC
  • feat(providers): Added Unifically as an OpenAI-compatible API-key gateway (https://api.unifically.com/v1), with live model discovery via passthroughModels. Discovery is scoped to /v1/models?category=llm, because the same catalog also lists image, video and audio models that run through a separate async task API and are not chat-completion targets. (#14182) — thanks @unifically-dev
  • feat(auth): enable dual auth for Cline and Kilo Code providers (#11837) — thanks @raheemuddin786
  • feat(cli): add omniroute fcc-claude — FCC-style launcher with fallback chain (#11908)
  • feat(providers): add Astra to the curated GitHub Copilot catalog (#12759) — thanks @backryun
  • feat(search): implement /v1/alpha/search endpoint for Codex native web search (fixes #8674) (#13788) — thanks @amsjavan
  • feat(quality): version static gate profiles and detect script drift (#14401)
  • feat(quality): prepare non-mutating admission ruleset proposals (#14419)
  • feat(ci): bind workflow admission to exact candidate results (#14420)
  • feat(quality): common original-source coverage instrumentation (#14524)
  • feat(providers): add vendor-drift tooling for codex-chatgpt-web (#14194) (#14719) — thanks @alexdrgpy06
  • feat(dashboard): per-account quota display on provider page (#14803) — thanks @aref-alapour
  • feat(claude): add Claude Sonnet 5.5 and correct Sonnet 5 pricing (#15035) — thanks @HouMinXi
  • feat(resilience): expose live combo queue telemetry (#15092) — thanks @xiaoyaner0201
  • feat(providers): add TypeSafe System One client for Jev routing (#15278) — thanks @miguelenes
  • feat(dashboard): one-click show/hide toggles for sidebar sections (#15508) — thanks @for-the-zero
  • feat(images): support Gemini Antigravity image-to-image editing in /v1/images/edits (#15513) — thanks @Ardem2025
  • feat(identity): refresh pinned CLI versions from their release feeds (#15642) — thanks @HouMinXi
  • feat(sse): trace Anthropic thinking-signature failures without payload logging (#15685) — thanks @insoln
  • feat(dashboard): add Blacklist button for models with error test status (#15858) — thanks @kveita
  • feat(resilience): add typed provider availability states (#15918) — thanks @atescivitci-cmd
  • feat(sse): upload images for DeepSeek Web Flash (#16034) — thanks @ape1121
  • feat(server): headless mode — OMNIROUTE_HEADLESS=1 / serve --headless (rail 3.8.53) (#16079)
  • feat(dev): local candidate build/validate/promote loop with rollback (rail 3.8.53, #8084 local side) (#16180)
  • feat(grok-cli): allow a validated client version override (#16397) — thanks @TechNickAI
  • feat(api): support Grok subscription image generation (#15423) — thanks @QuangBlue
  • feat(api): add Firecrawl Map endpoint with account fallback (#15703) — thanks @lacijowo
  • feat(models): opt-in MODELS_DEV_EFFORT_TIERS_ENABLED (default off) uses the reasoning effort tiers models.dev publishes when neither an override nor the registry declares any, and treats a declared empty tier list as "no tiers" in /v1/models (#15944) — thanks @HouMinXi
  • feat(plugin): opt-in combosSource: "models" for the v2 plugin publishes the combo rows /v1/models already serves and skips the /api/combos request; the default "api" source and its every-member filtering are unchanged (#16167) — thanks @maxmad64bis

🐛 Bug Fixes

  • fix(sse): a 429 whose typed Google RetryInfo.retryDelay is longer than the request-queue budget (Gemini answers an exhausted daily quota with "86399s") no longer freezes the key's per-model rate limiter for the whole wait. #16528 already relays that window into the model lockout; the limiter still emptied its reservoir for the same window, so any call that reached the key before the window ended (after a resilience reset, a lockout cut short by the maxCooldownMs cap, or an alias the lockout did not name) sat out the full maxWaitMs and failed as a local 503. The wait now becomes a model lockout on that connection for the stated window (exact model on antigravity, plus the requested name when an alias resolved it) and the limiter stays unpaused. Prose and untyped hints and connection-wide limiters keep the pause, but a pause that outlasts a caller's remaining queue time now turns that call away at once instead of letting it wait the budget out, and the 429 header log no longer claims a pause it does not take (#16525) — thanks @shipsfromrio
  • fix(sse): a Claude conversation that moves to another account no longer fails on advisor results it cannot read. Anthropic encrypts advisor_redacted_result per organization and rejects one made elsewhere with Advisor tool result content could not be processed. OmniRoute now retries once on the same account with those results marked as an unavailable advisor error, then falls back as usual if that also fails. (#15103) — thanks @cryptiklemur
  • fix(anthropic): the dotted model ids that the first-party anthropic registry and /v1/models expose (claude-haiku-4.5, claude-sonnet-4.6, claude-opus-4.6, …) are now sent upstream in the dashed form the Messages API accepts (claude-haiku-4-5, …), including when an effort suffix such as -high is stripped. Before, they 404ed upstream, or were silently served by an older model through model-family fallback while the response still named the requested one. (#16425) — thanks @Louayzouaoui1
  • fix(db): gateway API keys are now stored encrypted at rest (the same enc:v1: envelope and STORAGE_ENCRYPTION_KEY as provider credentials) and authenticated by their SHA-256 key_hash only — the plaintext column is no longer matched. Existing rows are hashed and encrypted in place on first use; the admin reveal, key rotation and the settings export/import keep returning and restoring the raw key. (#16898)
  • fix(speech): audio and transcription upstream failures no longer echo the provider's raw response body (#15159). Two byte-identical-in-spirit upstreamErrorResponse helpers — open-sse/utils/audioResponse.ts (18 call sites: every TTS provider in audioSpeech.ts, awsPollyTts, fishAudioTts, geminiTts) and a duplicate definition inside open-sse/handlers/audioTranscription.ts (17 call sites, re-exported to openrouterTranscription.ts) — interpolated await res.text() straight into error.message, so any stack frame, absolute path or credential a provider chose to return reached the client. That is what the repo's own rule already forbids: open-sse/utils/upstreamErrorResponse.ts states that an opaque body must not be echoed "even sanitized fragments", and moderations.ts / ocr.ts already comply. Both helpers now build through the canonical errorResponse, reusing a JSON upstream's own message only after sanitizeErrorMessage and falling back to a fixed Upstream error (<status>) for opaque bodies. The duplicate is re-exported rather than redefined, so the next fix is a one-place change, and CORS headers are re-applied because the canonical builders do not emit them and these routes set CORS only on the OPTIONS preflight. The catalog did not report this; it is the largest remaining Hard Rule #12 surface. (#15388) — thanks @jonlwheat2-gif
  • fix(db): enforce autoBackupFrequency schedules and keepLastNBackups retention to prevent unbounded disk growth from automatic snapshots, with Windows file-lock retry backoff (#15550) (#15669) — thanks @exodus2512
  • fix(release): Drain base-red #15306 leftovers: align four unit tests with the intentional Gemini tool-result text folding (#16601) and single canonical User-Agent (#16445) contracts, and register 16 covering tests in stryker.conf.json tap.testFiles (#15306). (#16880)
  • fix(perf): pass the parse-model benchmark variant as a child-process argument so the internal A/B selector is no longer mistaken for an undocumented public environment variable by the docs contract gate; preserves the benchmark introduced in #16355 (thanks @opensource-elearning). (#16506)
  • test(authz): align the manage-scope bypass deny-list guard with #16366: /api/providers/{validate,import,bulk} are remote-reachable by design (spawn gated at the call site) and stay refused in the bypass list — the guard now asserts exactly that instead of a stale LOCAL_ONLY expectation. (#16881)
  • Send anthropic-beta: fast-mode-2026-02-01 when a Claude request body carries the top-level speed field, and keep the beta a client negotiated through the merge, so Claude Code's fast mode toggle no longer fails with a 400. (#16866) — thanks @cryptiklemur
  • fix(sse): forward the client-negotiated timing-2026-09-09 and inline-tools-2026-09-15 betas (Claude Code wire tokens, captured from [email protected]), so requests with message-level tool_addition blocks no longer get 400 ... tool_addition blocks require anthropic-beta from Anthropic (#15102) — thanks @cryptiklemur
  • fix(claude): Claude 5.x rejects cache_control nested inside tool_result.content[] ("cache_control may not be specified within tool_result.content") and caps total cache_control blocks at 4 per request. OpenAI role:"tool" messages whose content parts carry cache_control were translated verbatim, 400-ing most opus-5-5 traffic from such clients. prepareClaudeRequest now hoists nested tool_result.content cache_control markers onto the tool_result block itself and, when the target supports prompt caching, culls to the last 4 cache_control blocks when more are present. (#15647) — thanks @patrykkopycinski
  • fix(i18n): propagate the four CLI environment-variable contracts from #16441 and the in-memory quality tracker cold-start note from #16408 to every published documentation locale, preserving the original translations and verified source provenance (thanks @pacocartones and @maxmad64bis). (#16504) — thanks @skygunner
  • Synchronize the generated CLI key-policy skill with the existing model-access flags and guard its registry parity with a regression test.
  • fix(cli): omniroute skills skillssh install <url> now works. It posted { url }, which the install route has never accepted, so every install failed with a 400. The command accepts https://skills.sh/<owner>/<repo>/<skill> or <owner>/<repo>/<skill>, sends the { name, description, source, skillId } the route validates, rejects anything else before calling the API, and prints the server's error message when the install is refused. (#17033)
  • fix(cli): omniroute skills skillssh list printed the raw response object instead of the skills, because it read items while the route answers skills. It now lists the skills.sh results (id, name, source, installs) and prints the server's message when the request is refused. (#17035)
  • fix(cli): omniroute sync import --dry-run no longer replaces the live database. The command sent ?dryRun=true to POST /api/db-backups/import, a flag the route never read, so a "dry run" overwrote the operator's data. The route now honours dryRun (query string or multipart field): it runs the same integrity and table checks and returns the row counts and schema version a real import would load, without the pre-import backup and without touching the live database. The CLI goes further and never contacts the server in dry-run mode (an older server would still ignore the flag): it validates the backup locally and prints the summary. A real import now goes through the local CLI credential, which the route accepts next to the dashboard session and manage-scoped API keys; JSON files such as the output of sync bundle are refused before upload, and 401/403/400 end in a readable error instead of a bare status. (#16723)
  • fix(cli): omniroute sync pull works again. It sent POST /api/db-backups/exportAll, but that route only answers GET with a .tar.gz archive, so every run ended in Error: 405. The command now downloads the archive with GET and the local CLI credential (which the operator-only export check accepts), writes it to [outPath] or to the server's file name (owner-only permissions, never overwriting an existing file), and turns a 401/403/405 into a readable error instead of a bare status. The --source/--merge/--replace/--dry-run options were dropped: the route never supported them. (#16707)
  • fix(cli): the omniroute sync subcommands now match the routes the server actually serves. sync push sent {parts, dryRun, target} to POST /api/sync/cloud, which only accepts {action}, so it always failed with 400; it now runs one cloud sync ({action: "sync"}, never retried) and explains a disabled cloud sync or a missing manage-scoped credential. sync bundle sent the API key and an ignored ?parts= to GET /api/sync/bundle, which only accepts a sync token; it now takes --token (or OMNIROUTE_SYNC_TOKEN), fails before any request when none is given, and writes the {version, bundle} JSON to an owner-only file that it never overwrites. sync diff and sync resolve were removed: no server route computes a diff or lists conflicts, and both silently printed the cloud status instead. The ignored --target/--bundle/--dry-run (push), --include (bundle) and --from-cloud (initialize) options were dropped. Failures report the HTTP status, the server message and a hint, with the usual CLI exit codes. (#16721)
  • fix(providers): Route Command Code Claude models through the Anthropic Messages endpoint. (#15543) — thanks @lorenzozanee
  • fix(dashboard): the compression settings panel sends each save at once and rolls back only the fields of a failed save; the auto-trigger box commits once on Enter or blur and keeps rejected values as a draft, and the panel re-reads the settings after a save that threw so a value the server stored anyway is not lost (#15600) — thanks @woodsonl
  • fix(cursor): Bridge native Cursor tool execution requests to client-declared tools by JSON-schema shape when tool names diverge from the static allowlists, and expand the name allowlists for common harness names (Claude Code, OpenCode, Cline, Kilo, Codex, Cursor). The bridge remains strictly fail-closed: schemas that cannot express the event or carry unpopulated required fields retain typed rejections, and ambiguous multiple-candidate matches fail closed to null (#8168). (#15602) — thanks @Junior-HJ
  • fix(security): pruning stored backups now requires an operator credential, exactly like restoring one. DELETE /api/db-backups (cleanup — keep the latest N / drop older than D days) accepted any manage-scoped API key, so a machine credential handed to automation could delete every restore point but one; PATCH /api/db-backups (retention settings) is gated too, because the values it persists are what the automatic pruning applies after every backup — a keepLatest: 1 saved by a machine credential would have emptied the backup directory on the next snapshot. Both handlers now go through the same requireOperatorCredential() boundary as the restores: a signed-in dashboard session, the local CLI credential or an instance with login disabled pass; API keys of any scope and oma_ access tokens get 403 with the retention settings and the stored backups untouched; a missing credential stays 401. Listing (GET) and creating a manual backup (PUT) keep ordinary management auth (#16899 is the precedent). (#16958)
  • fix(resilience): delete the dead isProviderFailureCode export and correct the stale breaker-set pointers it was referenced from (#15159 B-04 + B-05). isProviderFailureCode in open-sse/services/accountFallback.ts had exactly one caller — its own unit test — so nothing in production used it, while the test pinned its policy: isProviderFailureCode(429) === true. That is the opposite of the whole-provider breaker policy, where 429 is per-connection cooldown / model-lockout scope and must never trip the provider breaker. A reader reaching for that helper would have found the 429-inclusive one. The single source of truth is PROVIDER_BREAKER_FAILURE_STATUSES in src/sse/handlers/chatPredicates.ts. The combo path's comments also still pointed at src/sse/handlers/chat.ts:206 for that set, which moved to chatPredicates.ts — a stale pointer in the very comment that had to be rewritten to drop the reference to the deleted helper, so both items ship together. (#15381) — thanks @jonlwheat2-gif
  • fix(resilience): a provider circuit breaker no longer stays DEGRADED forever when the provider stops receiving traffic. DEGRADED now returns to CLOSED after the longest OPEN hold for its last failure kind (cooldownByKind value, else resetTimeout, × maxBackoffMultiplier, at least one cooldown) with no new failure (lazily, on read; a restored row with no failure time decays at once), and a valid connection test closes a DEGRADED provider or connection breaker. A CLOSED partial count still recovers one success at a time, and OPEN and HALF_OPEN still recover only through a request probe. (#16594) — thanks @shipsfromrio
  • fix(ci): check:error-helper now trusts the call instead of the file (#15159 G-03). The gate used to skip an entire file as soon as it saw any import from a utils/error path — a file-scoped exemption applied to a call-scoped hazard — so one correct import { sanitizeErrorMessage } permanently excused every other sink in the file. That is how a live Hard Rule #12 leak shipped green: deepseek-web.ts imports the sanitizer for one call site while a second, file-local errorResponse builder forwarded raw upstream errBody.msg to the client. Trust is now resolved per line and per symbol: a line routing through a sanctioned builder/sanitizer is clean, a canonical builder counts only when the file actually imports it, and a file-local builder whose own body sanitizes is recognised too. Sanctioned builders are also recognised when called multi-line, and src/lib/api/errorResponse.ts (createErrorResponse, which sanitizes both exports and covers ~54 route handlers) now counts as a sanctioned module alongside open-sse/utils/error.ts. Two compliant shapes that the new rules over-reported are fixed as well: a raw error inside a message-first logger call (log.error("BATCHES", "sweep failed", { error: err.message })) is an audit row, not a leak. Removing the skip surfaced 21 pre-existing violations (8 provider-auth routes leaking error.message, the unsanitized toJsonErrorPayload passthrough the earlier E-15 PR missed, and executors forwarding a raw caught error); they are frozen in KNOWN_MISSING_ERROR_HELPER with per-class justifications and self-clean via assertNoStale, so the gate now blocks new leaks while the backlog is worked down. (#15376) — thanks @jonlwheat2-gif
  • fix(security): restoring a backup now requires an operator credential, exactly like exporting one. POST /api/db-backups/import (uploaded .sqlite, dryRun=true included), POST /api/db-backups (restore a stored backup by id) and POST /api/settings/import-json (legacy JSON restore) accepted any manage-scoped API key, so a machine credential handed to automation could replace the live database or the whole settings/connections/API-key set irreversibly. The three routes now go through the same requireOperatorCredential() boundary as the exports: a signed-in dashboard session, the local CLI credential or an instance with login disabled pass; API keys of any scope and oma_ access tokens get 403 with the live data untouched; a missing credential stays 401. Listing, creating, retention and cleanup of backups keep ordinary management auth. omniroute sync import keeps working on the server host (it already sends the local CLI credential) and its 401/403 hint now names the operator rule instead of suggesting a manage key. (#16899)
  • fix(inspector): removed the Traffic Inspector "TLS intercept" toggle, which never did anything. The HTTP_PROXY listener only CONNECT-tunnels HTTPS and nothing read the flag, yet the docs said turning it on would decrypt HTTPS bodies. POST /api/tools/traffic-inspector/capture-modes/tls-intercept, the tlsIntercept field of GET /capture-modes, the matching CLI command and the INSPECTOR_TLS_INTERCEPT env var are gone. The docs now say plainly that HTTP_PROXY mode records HTTPS as tunnel metadata only, and point to AgentBridge, Custom Hosts or TPROXY decrypt for body capture. (#16692)
  • fix(db): restoring a settings JSON backup (POST /api/settings/import-json) no longer loses API key and provider connection state. The importer rewrote every key from only 8 columns, so revocation, expiry, ban, connection/combo allow-lists, scopes, IP allow-lists, rate limits and spend limits fell back to their defaults on restore; it now maps every column the backup carries from the live table schema, updates existing rows only with the columns the backup has, and only ever makes a key's lifecycle stricter (a key revoked, banned, expired or deactivated since the backup stays that way, and its stored secret is never replaced). Connection credentials now go through the same at-rest encryption as normal writes, and max_concurrent, group, proxy toggles and other later columns are restored instead of reset (#16693) — thanks @xiaoyaner0201
  • fix(transcription): the Kie.ai async transcription failure path no longer returns the raw upstream response body to the client (#15159 E-02). handleKieAudioTranscription built its createTask error body with a hand-rolled Response.json({ error: { message: err.message } }) and no sanitizer — and the caught error is the upstream body, because kieExecutor.createTask throws new Error(await res.text()). An upstream that returned a stack frame or a credential-shaped string therefore reached the browser verbatim. The branch now uses the canonical errorResponse builder, which the same file already imported and which the Kie poll path 40 lines below already used. The upstream status is still preserved, and the CORS headers are merged back on because that route only sets them on the OPTIONS preflight, never on the POST response. (#15379) — thanks @jonlwheat2-gif
  • fix(security): /api/oauth/kiro/auto-import can no longer be opened to remote callers through the LOCAL_ONLY manage-scope bypass list. The route reads host-local Kiro credential files (kiro-cli data, ~/.aws/sso/cache), so it now sits in the same never-bypassable set as the spawn-capable routes: PATCH /api/settings rejects it (and any broader prefix such as /api/oauth/kiro/) with BYPASS_PREFIX_NOT_ALLOWED, the runtime snapshot drops it from a stored bypass list, and the request-time check refuses it for a manage-scoped key from a non-loopback address. (#16691)
  • fix(ws): the live dashboard WebSocket sidecar now accepts a Bearer API key only when it holds a management scope (manage/admin); inference-only keys are refused with the existing UNAUTHORIZED close while dashboard sessions and the env master key are unchanged. Its internal event endpoint (POST /__omniroute_event) also stops trusting a loopback peer that arrives with proxy-hop headers (x-forwarded-*, x-real-ip, forwarded, via), so a same-host reverse proxy can no longer relay fabricated dashboard events; the in-process publisher is unaffected. (#16955) — thanks @AndreAlmeidaDC
  • fix(mcp): one shared internal hop for the MCP server, read lazily (#15159 M-06). The server→server call from an MCP tool to OmniRoute's own API existed as three copies — server.ts:199 (omniRouteFetch, lazy env read), tools/advancedTools.ts:37 and tools/pickFastestModel.ts:17 (both private apiFetch, env read frozen at module load) — and had already drifted three ways. All five consumers now import one hop from the new open-sse/mcp-server/internalFetch.ts, the module M-01's prescribed split already names. (#15468) — thanks @jonlwheat2-gif
  • fix(ci): Require both admission gates in the release merge queue and run applicable CI lanes for verified same-repository Mergify candidates, while rejecting ordinary drafts, forged queue identities and missing or skipped required results. (#16957)
  • fix(db): migration 207_remove_opencode_free_provider no longer aborts the migration run with no such column: api_key on a provider_connections table that predates the credential columns — the runner adds the missing nullable columns first (base-red #15306, regression from #16310). (#17006)
  • fix(models): preserve upstream model namespaces when resolving static tool and context capabilities, so AI Horde models cannot inherit DeepSeek capabilities from a matching leaf name. Provider-qualified public spec lookups and the vision-only leaf fallback remain supported. (#16939)
  • fix(dashboard): the provider Import button no longer reports "All models already imported" when model discovery actually failed (#15159 B-03). /api/providers/<id>/models?refresh=true flags a degraded response with source + warning — local_catalog when remote discovery failed, cache or github_catalog with a warning — and the server already defines the condition as isDegradedDiscovery. The dashboard read data.warning but pushed it into the import modal's log lines while the headline status stayed allModelsAlreadyImported (or noModelsFound), so an operator whose provider had silently fallen back to a local catalog concluded there was nothing to import and the missing models stayed missing. The warning is now promoted to the headline status, with the success-flavoured line demoted to a log entry. The decision lives in a pure classifyModelImport() helper that reuses the server's own isDegradedDiscovery rather than a second local copy, so the two cannot drift — and intentional: true (reka / voyage-ai / t3-web, whose local catalog is the intended and only source, #5460 / #5465) is still correctly treated as healthy. (#15384) — thanks @jonlwheat2-gif
  • fix(db): complete the at-rest protection of nested connection credentials started in #16863: the shared providerSpecificData secret allowlist now also covers the extraApiKeys rotation pool (every entry enveloped), refresh-flow sub-tokens (copilotToken, dcaToken, clientSecret, recaptchaV3Token) and web-session cookies/tokens; rows already on disk in plaintext are swept once on the module's first read, mergeConnectionProviderSpecificData patches are enveloped too, the compliance audit redaction derives from the same list, and the providers API masks extraApiKeys like apiKey (masked placeholders echoed on update resolve back to the stored keys, never overwrite them) and strips every other declared field. (#16966) — thanks @jackyfan01
  • fix(rate-limit): withRateLimit no longer starts the scheduled work of a caller that was aborted while it waited for its provider slot. With limiter capacity free, Bottleneck dispatches the job before abandonQueuedJob() can pull it from the queue, so the work ran with an already-aborted signal; work that only listens for a later abort event then held the limiter slot and Bottleneck's ref'd expiration timer for executionMaxWaitMs (10 min by default). The dispatched job now rejects without calling the work, as it already did after a queue-wait timeout, and tests/unit/rate-limit-manager.test.ts exits right after its last test instead of about 10 minutes later. (#16566) — thanks @shipsfromrio
  • dashboard: the connection rate-limit badge now shows "Protected" when the auto-enable safety net (autoEnableApiKeyProviders / RATE_LIMIT_AUTO_ENABLE) already covers an active API-key connection, matching what the limiter actually does (#15790)
  • fix(sse): a Claude refusal and Bedrock content_filtered or guardrail_intervened stops now reach OpenAI-format clients as content_filter, and model_context_window_exceeded as length, instead of a clean stop, in the Claude stream translator and the Bedrock executor. (#16453) — thanks @pacocartones
  • fix(api): /api/v1/registered-keys (list, issue, fetch, revoke) is now a management-only surface like /api/keys: it takes a dashboard session, the local CLI token, or an API key with manage/admin scope, and refuses inference-only API keys with 403. Before, the /api/v1/ prefix routed these handlers through the inference-key check, so any active API key — whatever its scope — could list every registered provider key and revoke them all. Response shapes on the allowed path are unchanged; docs/openapi.yaml and docs/reference/API_REFERENCE.md now document the auth tier. (#16878)
  • fix(db): authoritative live catalogs no longer veto registry models that a provider explicitly marks as dispatch-discoverable (registryDispatchUnion + per-model targetFormat); z.ai opts in first, because its Anthropic-compat /models surface omits the coding-plan glm-5.3-flash family and a fresh authoritative snapshot then rejected dispatchable models with model_not_in_catalog. Unioned rows are listed alongside discovered ones and stamped catalogOrigin: "registry". On an account not entitled to the model, the first dispatch fails upstream and model lockout re-gates that connection. (#15599) — thanks @woodsonl
  • CI: add a narrowly scoped, owner-reviewed Relay integration queue that preserves original commits and requires both aggregate gates without enabling paid batching; prevent integration candidates from falling through to the ordinary squash queue and document activation/readback requirements. (#16716)
  • fix(relay): the relay's per-(token, client) throttle now keys on the client address the authz pipeline resolves at the trust boundary — the connection's peer, or the client a loopback / private-network / Cloudflare / OMNIROUTE_TRUSTED_PROXIES proxy reported — forwarded to the route as a stamped header the pipeline strips from client input. Before, it keyed on X-Forwarded-For / X-Real-IP as sent, so a direct caller could reset the bucket by rotating the header (the per-token limit still applied). Clients behind the same reverse proxy keep separate buckets; when no stamp exists every caller of a token shares one bucket and the forwarding headers remain log-only. (#16975)
  • fix(relay): minting a relay token (POST /api/relay/tokens) now requires a signed-in dashboard session or the local CLI, so an API key — even a manage-scoped one — can no longer create new relay credentials; listing, editing and revoking keep accepting a manage key. A malformed request body now answers 400 Invalid JSON body, and an unexpected failure answers a sanitized 500 instead of echoing the raw exception message. (#17032)
  • fix(authz): relay token management (GET/POST /api/relay/tokens, GET/PATCH/DELETE /api/relay/tokens/{id}) now requires a dashboard session or a manage-scoped API key even when the dashboard runs with login disabled — relay tokens are reusable inference credentials, so minting, listing, editing and revoking them joins the always-protected tier alongside the other credential-issuing routes. Automation that already sends a manage key is unaffected. (#16976)
  • release: two more release-branch reds drained — the A6 keyless catalog integration cases stay keyless when CI exports INITIAL_PASSWORD, and the OpenCode usage happy-path test passes a connection id now that quota reads resolve the connection proxy (#16344). (#17059)
  • release: three more unit regressions drained from the release branch — the two Antigravity quota tests follow the live Gemini 3.8 tier that #12608 switched to, and the free-tier budget card SVG is regenerated for the catalog after #16310 and #12608 (478 models, 12 providers). (#17054)
  • release: four unit regressions drained from the release branch — the MCP server bundle imports again on plain Node (modelFallbackAuthorization and keyConnectionScope now load apiKeyPolicy lazily, and apiAuth loads next/headers lazily, so the bundle no longer carries a Next-only import nor the auth init cycle), the route-guard locale proof follows the current paragraph order, the parseModel cache-key test asserts the anthropic/claude-opus-4.5 contract introduced by #16811, and the family-fallback effort test opts in to the cross-version hop that #16810 made consent-gated. (#17007)
  • fix(i18n): propagate the pre-dispatch model-cooldown recovery contract from #16360 / #16319 to all published resilience-guide mirrors, including bounded waits, transient retry hints and the non-retryable lock exclusions (thanks @skygunner). (#16504) — thanks @pacocartones / @maxmad64bis
  • fix(api): preserve the Responses API message phase field in non-streaming output. (#16076) — thanks @jacobsparts
  • fix(sse): accumulate fragmented tool-call names and resolve qualified, dotted and unique bare declarations in both Responses streaming paths. Buffer arguments until a declared identity is ready, preserve that identity and call type through completion, and reset pending state when a call ID is replaced. Explicit flat functions keep precedence over namespaced custom tools; ambiguous or unknown names wait for closure and retain their wire spelling, with the existing MCP continuation fallback preserved. Declaration context retains repeated-letter fragments; complete-name and progressive resends remain idempotent. Before publication, a fragment that extends another declared name takes precedence over an ambiguous suffix resend. Without declarations, suffix resends retain the deterministic deduplication policy. Request-ledger aliases also retain their original namespace and custom type when long wire names were hash-truncated; literal flat declarations and ambiguous aliases remain flat. (#16106) — thanks @ape1121
  • docs(routing): correct the routing-strategy count and lists across the agent guides and the Auto-Combo reference (#15159 DOC-01). The code exposes 20 public strategies in ROUTING_STRATEGY_VALUES, but AGENTS.md claimed 19 with quota-weighted missing from its list, open-sse/services/AGENTS.md claimed 17 with four strategies missing (p2c, quota-weighted, cache-optimized, pipeline), docs/routing/AUTO-COMBO.md claimed 19 in two places and was missing the quota-weighted table row entirely — despite AGENTS.md calling that table "the full strategy table" — and the hand-maintained Bosnian mirror carried the same 19. quota-share stays excluded, since it is deliberately internal (INTERNAL_ROUTING_STRATEGY_VALUES). A new test asserts all four documents against ROUTING_STRATEGY_VALUES directly, so the prose cannot drift from the code again. (#15382) — thanks @jonlwheat2-gif
  • fix(security): machine credentials can no longer take over the operator identity or weaken the login defenses through the settings routes. POST /api/settings/require-login is operator-only once a password exists (dashboard session, local CLI, or an instance with login disabled) — API keys of any scope and oma_ access tokens get 403 — and a dashboard session must send currentPassword to replace the password; the local CLI keeps the recovery path and the first-run bootstrap window is unchanged. PATCH /api/settings refuses machine credentials on every security-policy key (requireLogin, newPassword, oidc*, corsOrigins, bruteForceProtection, the LOCAL_ONLY bypass list) even when they know the password, oma_ tokens need the admin scope for any /api/settings mutation, and corsOrigins / bruteForceProtection now take the password re-auth and are dropped by the JSON import — the Security tab asks for the current password before changing them. Refusals are audited as settings.update_failed with reason: OPERATOR_REQUIRED. (#16897)
  • fix(skills): skillRegistry.resolveVersion now honors the >=, <= and == constraints. It read only the first character of the constraint as the operator, so ">=1.0.0" was parsed as > against "=1.0.0" and never matched any version. (#16204) — thanks @Louayzouaoui1
  • fix(skills): a skill that outlives the executor timeout is now recorded as timeout instead of error, so execution history and analytics can tell timeouts from handler failures. The race timer is also cleared once the handler settles; it used to stay pending for the full timeout (30 s by default) after every execution. (#16216) — thanks @Louayzouaoui1
  • fix(skills): POST /api/skills/skillssh/install now validates source (a plain GitHub owner/repo), skillId (letters, digits, _, -) and version (X.Y.Z, the registry's own contract) before building the GitHub raw URL, answering 400 with the usual { message, details } body instead of fetching whatever path the values spelled out. fetchSkillMd() applies the same check at the fetch boundary, so direct callers cannot pivot the request with .., ?, #, %2f or extra segments either. (#16959) — thanks @AndreAlmeidaDC
  • fix(authz): the Headroom, Tailscale and cloudflared status routes now run their host binary probes (which, --version, tailscale status --json) only for loopback/LAN callers. These routes stay reachable from a remote dashboard, which now gets the same response built from persisted state and flagged probed: false; local responses are unchanged. (#16872)
  • fix(quality): allowlist the check-workflows.test.ts assert reduction (73->67) from #14497's fail-closed refactor - the graceful-fallback asserts were replaced by fail-closed assert.throws, verified legitimate, not test masking. (#15575) — thanks @Junior-HJ
  • fix(ci): check-api-typecheck, check-dashboard-typecheck and check-open-sse-typecheck now run tsc through runBuildTool() instead of spawning npx.cmd without a shell, which Node refuses on Windows (EINVAL), so the three gates crashed there before checking anything. ensure-native-sqlite runs npm.cmd rebuild through the shell on Windows instead of always reporting a failed rebuild. runBuildTool() now returns the tool's stdout. (#16581) — thanks @shipsfromrio
  • fix(tests): Compression worker-pool and RTK TOML test assertions now cross-platform — worker specifier compares URL pathnames via pathToFileURL, and the 0600 permission check is POSIX-only. Both failures were Windows-only test bugs, not runtime regressions. (#15572) — thanks @Junior-HJ
  • fix(ci): seven scripts/ entry points (check-mutation-test-coverage, check-vitest-exclusions, check-new-key-coverage, select-impacted-tests, mutation-radiography, feature-triage, sync-changelog-i18n) now reach main() when run on Windows. Their `file://${process.argv[1]}` guard never matches a Windows path, so each one exited 0 without doing anything, and a local check-mutation-test-coverage --strict passed where CI failed. They now use the pathToFileURL(process.argv[1] || "").href guard most scripts already use, and tests/unit/build/scripts-direct-run-guard.test.ts keeps the old form out of scripts/. (#16580) — thanks @shipsfromrio
  • Prevent externally supplied Video Bridge audioTranscript metadata from claiming server-owned provenance; preserve client transcript fusion and log-safe redaction while reserving audio-bridge labels for the server STT adapter. (#16061)
  • Connect Video Bridge's existing Audio Bridge STT orchestration to actual video guardrail processing. A video part must explicitly set audioTranscription: true and the operator must enable modalityBridgeVideoAudioTranscriptionEnabled; both branches reuse admitted bytes and the attempt deadline. STT failures retain the visual description, and server-generated cues receive a log-safe redaction shadow. Whole-video result caching is disabled while the operator STT flag is enabled to prevent reuse across consent/model changes; the bounded, principal-scoped STT cache remains available. (#15780)
  • fix(ci): keep recurring baseline headroom reports in workflow summaries and retained artifacts without recreating a permanent issue; preserve measurements and quality gates (#12149). (#16111)
  • fix(video): Keep transcript redaction shadows out of diagnostic logs and redact video transcript fields in pre-dispatch rejection snapshots while preserving live request and response payloads (#12150, #12430). (#15984)
  • fix(guardrails): protect raw video transcript fields in retained request logs when description fails or the video limit leaves a part untouched, while preserving the live provider payload (#12430). (#16128)
  • fix(i18n): reject invalid UI message syntax in CI and repair five rich-text translations without exposing markup (#12560). (#15912)
  • fix(i18n): ICU-escape 381 locale message values containing raw <tag> / {...} literals so next-intl no longer throws INVALID_MESSAGE and falls back to the raw key; adds a check:ui-icu script that compiles every catalog with IntlMessageFormat (not yet wired into CI as a blocking gate — pending owner coordination) (#12564 — thanks @chirag127)
  • fix(build): externalize tiktoken so tiktoken_bg.wasm resolves at runtime under Turbopack (#12564 — thanks @chirag127)
  • fix(providers): refresh the shared Antigravity/AGY Google Cloud Code catalog to Gemini 3.8 Flash tiers and reject stale upstream discovery entries (#12608) — thanks @backryun
  • fix(routing): reject individually hidden chat models before upstream dispatch, including aliases, compatible-provider prefixes, combo targets, and connection defaults; preserve API-key policy and catalog prefix settings (#12666). (#16110)
  • docs(docker): explain Gemini regional errors, connection-scoped proxy checks and the prerequisites and rollback for host-level IPv6 tuning (#12762). (#15914)
  • fix(poixe): allow billing aliases such as qwen3-32b:free after syncing a catalog that lists only bare model IDs (#12765). (#15916)
  • Fix setup-codex profiles referencing an undefined OmniRoute provider on fresh configurations, preserving existing provider/auth settings and keeping supplied API keys out of generated files (#12782). (#15960)
  • Reuse direct-auth provider alias pools in combo/auto discovery so Antigravity CLI and OpenCode accounts remain visible to scoring, quota, tags and affinity while preserving account restrictions. (#12844) (#15927)
  • fix(cli): Individual provider tests use the server probe when no local recipe exists, report skipped offline tests without marking credentials unhealthy, and include connections without expiration records in provider status. Refs #12849. (#15957)
  • fix(dashboard): Organize sidebar navigation with compact links, collapsible subgroups, clearer active-page styling, accessible keyboard controls, and a mobile close button. (#12869) — thanks @yugui923.
  • fix(codex): Explain models withheld by the effective HTTP client version during discovery, preserving account inventory authority and refreshed catalog diagnostics (#12893) — thanks @guanbear
  • fix(providers): Meta AI (muse-spark-web) now fetches a fresh WS access token via the browser pool instead of reusing the stale page-embedded token, and preserves the nested [1,1,5] proto envelope when patching the conversation id, fixing WS timeouts after a few messages (#12914) — thanks @wray-lee
  • fix(claude): drop assistant thinking / redacted_thinking blocks that carry no signature from the history before it is replayed to the genuine Anthropic upstream (Claude passthrough), so reasoning relayed from an OpenAI-compatible leg no longer triggers Invalid signature in thinking block on the next turn — signed blocks are forwarded untouched (#12917) — thanks @atescivitci-cmd
  • Preserve valid caller correlation IDs on Responses requests, accepting X-Request-Id after X-Correlation-Id, and return the same ID used by call logs on streaming, early keepalive and JSON responses. (#15924)
  • fix(ops): add OMNIROUTE_EGRESS_IP_LOCK_PROVIDERS to override the hardcoded egress-IP-bucketed lockout provider set (opencode/opencode-go/opencode-cli) — disable it entirely (none/off/false/0) or replace it with an explicit comma-separated list (#13003) — thanks @alvinveroy
  • fix(opencode): send the CLI identity headers and x-opencode-session on background calls (model-catalog discovery, quota fetcher) that previously bypassed buildHeaders() and went out with a bare Bun fetch user agent — the session id is a stable per-connection fingerprint seeded by any of the workspace keys the providerSpecificData validator accepts (openCodeGoWorkspaceId/opencodeGoWorkspaceId/workspaceId), falling back to the connection id, so background traffic groups under one upstream identity instead of a fresh anonymous UUID per call (#13005) — thanks @alvinveroy
  • Validate every aggregate provider-count claim across the main documentation and provider skills, and refresh stale references to the measured catalog total (#13205).
  • fix(codex): Preserve discovered native reasoning levels and defaults through synchronization, dashboard/catalog variants, and request dispatch; keep Max unchanged and protect exact upstream model IDs from suffix parsing (#13224) — thanks @TPOHH
  • fix(embeddings): LAN/CGNAT OpenAI-compatible embeddings nodes with a stored API key now send Authorization: Bearer on the outbound request, matching dashboard Check. Keyless LAN nodes stay no-auth (#6925) (#13234)
  • fix(security): four independent security-audit fixes. The DNS-rebinding guard (GHSA-cmhj-wh2f-9cgx) now also covers the provider dispatch path: for operator-supplied base URLs (providerSpecificData.baseUrl, including openai-compatible-* nodes) the connection is validated at connect time against the same guard mode as the URL check (dispatchGuarded() in open-sse/executors/dispatchPin.ts), closing the classic TOCTOU window on the highest-risk override path; it rides on the ambient patched fetch as a dispatcher and is skipped whenever an outbound proxy, TLS impersonation or the direct sentinel applies, so configured proxies and request logging are never bypassed. The rate limiter no longer fails fully open when Redis errors — it falls back to the in-memory limiter instead of allowing every request unconditionally. Boot warns loudly in production when STORAGE_ENCRYPTION_KEY is missing and keeps serving: an empty key is the documented encryption-disabled contract, so production must not exit. Persisted JWT_SECRET/API_KEY_SECRET are encrypted at rest. And the MCP scope-enforcement gap is closed: withScopeEnforcement() now forces scope checks for any caller resolved from a real HTTP Authorization header that does not already hold manage/admin scope, even when OMNIROUTE_MCP_ENFORCE_SCOPES is off by default — so a key granted only the narrow mcp:connect bypass scope can no longer invoke arbitrary MCP tools once remote access is enabled (#13330) — thanks @JJAbrams-eng
  • fix(deepseek-web): handle native DeepSeek Web DSML tool calls (<calls> / <invoke> / <parameter>, including root-level invokes) so they parse into tool calls instead of leaking as plain text (#13345) — thanks @espokaos-ops
  • fix(routing): make complexity-aware auto combo routing always-on — the complexityAwareRouting opt-in flag was orphaned by migration 103 (stripped as a legacy key on every save), so the feature was permanently unreachable and a short prompt (e.g. "olá") could tie-score into an expensive model instead of a cheap one (#13386, #13387) — thanks @ClaudioASantana
  • fix(models): Keep concurrent catalog timeouts in one shared build window and yield during API-key filtering so large authenticated catalogs do not stall unrelated requests. (#16115)
  • fix(network): skip Chrome TLS impersonation for Groq (api.groq.com); Cloudflare 1010s that JA3 while native undici reaches the API (#13445) (#13225)
  • fix(cli): attach compatible provider-node credentials through the management API so keys add preserves the node endpoint and never falls back to an unbound SQLite credential (#13452). (#15922)
  • Add dashboard controls for hiding auto and no-thinking catalog entries, with saved settings applied to model discovery without changing routing permissions (#13562). (#15986)
  • fix(combo): hop Astra quality-empty 200 to sibling effort alias and Gemini 3.8 Flash empty-content to medium then low (#13603) — thanks @HouMinXi
  • Exclude Qoder PAT/CLI accounts from caller-tool requests, preserve HTTP account selection, and document the PAT runtime and OAuth configuration limits. (#13697) (#15919)
  • sse: clearProviderFailure() now also clears the network-error and per-connection failure dedup windows, so a fresh network blip right after a provider breaker reset is counted instead of swallowed (#13887, #13899 (#16077) — thanks @onexer88-blip).
  • fix(providers): use genuine OpenCode CLI tool name (read) instead of _noop for free-tier placeholder to resolve 403 FreeTierError (#13935) (#16858) — thanks @themirzaalibaig
  • Accurate log-export counts: streamed exports now finish with the actual emitted row count, including after concurrent deletion or hydration failure. Preflight counts are explicitly estimates; the dashboard reads the final count without parsing the full download. Proxy-log pagination is documented as a live, best-effort export. Fixes #13999. (#15933)
  • Backport the upstream v6.1.7 conservative 500,000-character ChatGPT Web reasoning preflight limit for Medium/High and Pro Extra High, including multipart stages and final parts. Instant, Pro mode, and Luna keep their existing limits. This partial backport retains the v4.0.7 vendor pin; the full refresh and live acceptance remain pending. Refs #14194. (#16145)
  • fix(providers): isolate tool-only model catalogs so TinyFish/web-fetch providers stay out of MCP discovery and the LLM list, and one unavailable catalog no longer aborts aggregate listing (#14286) — thanks @seanford
  • Imported and synced provider models now expose an editable maximum output-token limit that persists through model sync and is enforced by request output budgeting. Clearing the value restores the discovered default. The provider model list also preserves discovered vision support and shows its badge only when explicitly supported. (#15963)
  • Protect semantic-cache embedding credentials when the configured provider or endpoint changes, validate outbound embedding URLs and refuse redirects. Mask saved embedding/Redis secrets in settings responses, encrypt them with the configured storage encryption key, and prevent approximate cache reuse when its text projection omits request context or tool/output contracts. (#15921)
  • fix(security): chatgpt-web-codex CDP proxy now fails closed without a token, auto-provisions a shared secret in both compose files and the executor sends X-Omni-Cdp-Token on connectOverCDP (#14486) (#15825)
  • fix(security): pin Google service-account token exchange and reject redirects; warn once when CLI salt persistence falls back, preserving existing tokens (#14487, items 3–4). (#15913)
  • fix(api): expose per-member supportsVision on /v1/combos model steps so the member blocking a combo's multimodal capability can be identified (#14587) (#15823)
  • fix(usage): covered normalized-token image generation/edit paths now record successful priced calls in usage_history against the API key and serving connection; each physical image attempt keeps its own priced usage row even when completion timestamps and token counts match; genuinely unpriced/no-usage calls remain in attributed call_logs only to avoid the fail-closed unpriced-usage guard; pricing-store failures preserve reported token evidence, Codex dedicated Images usage keeps its existing response envelope, and covered combo/edit call_logs rows carry the API key and connection. Direct SYNTX edits, Antigravity-native edits, and native provider usage that is not normalized into the wrapper remain outside this change (refs #14618) (#14619) — thanks @xiaoyaner0201
  • CLIProxyAPI recovery: Return the request body actually used after reasoning retries, retry a known unsupported field once after a 400, and reuse observed effort limits on later requests. Preserve private tool aliases and namespace identities through body copies and response restoration without sending them upstream. Refs #14629. (#16137)
  • Fix CommandCode CLI reasoning retries to update nested parameters while retaining the complete request envelope, and recover primary endpoint 400 errors for rejected request fields using the shared bounded recovery helper. (#16136)
  • tests: Require explicit recovery ownership for new executor overrides, using an AST guard and a fixed legacy identity snapshot. Document the override contract; linked behavioral tests remain responsible for proving payload and retry correctness. Refs #14629. (#16140)
  • GLM OpenAI recovery: Keep the returned request body aligned with reasoning retries, retry a known unsupported field once after a 400, and reuse learned effort limits on later OpenAI requests. Unknown first attempts and the Anthropic transport retain their existing behavior. Refs #14629. (#16135)
  • Fix 9router recovery for named unsupported request fields on HTTP 400, preserve the body actually sent after retries, and reuse learned reasoning-effort limits for subsequent requests to the same forwarded model. (#16133)
  • fix(cursor): finish a Cursor stream that hits the 5-minute safety timeout after partial output with finish_reason "length" + [DONE] instead of dropping the response mid-answer (#14727) (#15835)
  • fix(claude-web): forward text file attachments to claude.ai instead of hardcoding empty attachments (#14742) (#15720)
  • fix(providers): QwenCloud Token Plan quota now parses the new Monthly Quota window instead of going blank (#14763) (#15832)
  • fix(providers): GLM no longer rejects registry models (e.g. glm-4.6v) with "not available in the active live catalog" when a connection's synced snapshot predates them (#14779) (#15833)
  • fix(proxy-health): skip sweep probes already proven by recent single-provider production successes (a recent failure still gets its live probe), opt-in via PROXY_HEALTH_PASSIVE_SKIP (default off: sweep probes every proxy as before) (#14805) — thanks @maxmad64bis
  • fix(sse): opencode park-and-resume replay no longer serves proxy-set-aside accounts by default; the #14750 "serve anyway" fallback is now opt-in via OPENCODE_PARK_REPLAY_SERVE_SETASIDE (#14851) (#15834)
  • fix(providers): ensureThinkingBudget no longer injects a fabricated 4096 max_tokens floor when the model has no registered maxOutputTokens (#14888) (#15821)
  • fix(docker): the runner-cli image now installs @qoder-ai/[email protected] (exact pin, like the other bundled CLIs), so Qoder PAT (pt-*) logins no longer fail with spawn qodercli ENOENT / cli_not_found inside the container (#9292) (#14901) — thanks @Ch3fUlrich
  • docs(docker): document the EACCES: permission denied fix for /app/data on self-host (chown to the node UID 1000, or an opt-in named volume with its empty-volume/backup caveats); docker-compose.selfhost.yml keeps the ./data bind mount as the active default (#14910 — thanks @matheusgcdj)
  • fix(providers): venice-web now mints a Clerk session JWT from the __client cookie and calls the real outerface.venice.ai inference endpoint (#14922) (#15846)
  • fix(sse): keep Cursor's prompt cache across agentic turns by sending a session-stable conversation_id (derived from the client session or conversation fingerprint) instead of a random one per request, and report Cursor's metered usage including cache reads/writes. (#14973) — thanks @QuangBlue
  • fix(providers): Codex GPT Image models now use the dedicated Images generation/edit routes with the requested image model forwarded, while GPT-5.6 image catalog entries retain the hosted Responses tool path; this removes chat-tool-decline failures and carries upstream usage for later priced image accounting (refs #14617) (#14976) — thanks @xiaoyaner0201
  • fix(resilience): the proxy health sweep no longer wakes idle edge-relay proxies (deno/vercel/cloudflare) every 10 min; they are probed on a separate 6h cadence (PROXY_HEALTH_RELAY_INTERVAL_MS), cutting Deno "Memory Time" burn (#14984) (#15837)
  • fix(images): route POST /v1/images/generations to image models synced from user-defined OpenAI-compatible providers (#14989) (#15838)
  • fix(usage): add GET /api/usage/by-run?run_id= — per-run/session cost lookup joining the existing call_logs.session_tag and correlation_id columns to request_cost_ledger, no schema change (#PENDING) (#14992) — thanks @Ch3fUlrich
  • ci(quality): the release-green sweep now reports pricing data untouched for 90 days, so stale tier pricing shows up in the report instead of drifting silently (#15003) — thanks @maxmad64bis
  • fix(cursor): cursor-api connections now discover Cursor's live AvailableModels catalog (using the exchanged session token) instead of persisting the static registry as their authoritative catalog, so newly released models such as claude-opus-5-5-* are routable; a Cursor AI Model Not Found reply without a usage-reset hint is reported as a 404 model-not-found error instead of a 429 rate limit (#15024). — thanks @mdc2122
  • fix(context): calibrate the chars/4 guard estimate with provider-reported prompt_tokens (#15054) — thanks @skygunner
  • fix(providers): Freebuff now shows an official-client-only risk notice (free mode is suspended when used via proxies) and drops the "automated harvester" wording (#15058) (#15842)
  • Auto-Combo ToS policy: Separate settings state from rendering to keep the new control within the function-size limit, and use the canonical provider term in its Simplified Chinese text. Refs #15059. (#16132)
  • fix(auto): excludeTosAvoid now defaults to on, so providers whose catalog verdict is tos: avoid (e.g. Antigravity) no longer receive auto traffic on fresh installs; this changes the auto candidate pool on existing installs that never set it — opt back out with PATCH /api/settings {"excludeTosAvoid": false} (#15059) (#15839)
  • Apply the provider's ToS verdict to auto-routing candidates missing from the per-model budget catalog, and clarify first-run setup, account requirements, promotional credits, and aggregate free-tier estimates (#15059). (#15979)
  • Expose the global ToS-avoid opt-in on the Combos dashboard, keep unconfirmed settings disabled, preserve the last confirmed value when saving fails, and refresh the auto catalog after saving.
  • Explain the default auto exclusion before connecting affected providers in the provider detail and onboarding pages, and align first-run documentation with account requirements.
  • fix(providers): adding a second web-cookie provider account under an already-used connection name no longer overwrites the first account's session (#15070) (#15844)
  • fix(browser): gemini-web now applies the configured provider proxy to its Playwright context, and headed browsers (zai-web) start a private Xvfb when the host has no display (Docker -web image now installs xvfb), with a clear 503 instead of a bare 502 when none can be started (#15076, #15300) (#15851)
  • fix(quota): Keep Codex quota auto-ping usage reads and warm-up requests on the selected connection proxy without enabling direct fallback (#15084) — thanks @xiaoyaner0201
  • fix(capabilities): Apply explicit model-compat vision overrides consistently to model catalog and combo capability projections while preserving provider scope and Custom Models precedence (#15086) — thanks @xiaoyaner0201
  • fix(proxy): Count no-auth account proxy references during deletion and keep unresolved or inactive required proxy bindings from silently dispatching direct (#15087) — thanks @xiaoyaner0201
  • fix(codex): hand off live HTTP Responses SSE after complete non-empty text/reasoning/tool delta events instead of buffering until the combo target deadline; preserve lifecycle-only capacity-error detection (#15094) — thanks @xiaoyaner0201
  • fix(dashboard): the provider test playground conversation is now persisted in localStorage per provider and API key, so it survives switching between the Test and Logs tabs and reopening the slide-over (#15097). Messages are reloaded when the selected key changes, validated on read, written only after a turn completes, and storage errors no longer break the panel. (#15114) — thanks @Yashxp1
  • fix(cursor): large tool results no longer reach the model cut or empty — MCP text is fitted under Cursor's 40000-byte inline limit and held reads under its 100000-character limit, on a line boundary with a note naming where to continue; held reads set range_applied and a total_lines that covers the lines before the slice, so ranged reads stop failing with "Offset N is beyond file length" (#15117). — thanks @QuangBlue
  • fix(idempotency): the idempotencyWindowMs setting (Settings → Cache) now actually sets the replay window — it was saved and shown by /api/cache but the store always used the 5s default (#15127) — thanks @MumuTW
  • fix(idempotency): the replay key now includes the calling API key, so two API keys sending the same Idempotency-Key, model and body no longer share one cached response; retries from the same key still replay (#15128) — thanks @MumuTW
  • fix(oauth): replace both token expiry fields on reauthentication and clear stale OAuth errors, preventing an immediate background refresh from disabling a freshly signed-in account (#15135 — thanks @JxnLexn)
  • fix(cline): expose the free-bucket cline-free/* model ids in the Cline catalog and stop labelling paid vendor ids (DeepSeek V4 Flash, Step 3.7 Flash, MiniMax M3) as "(Free)", which caused 402 Insufficient balance at zero credits (#15136) (#15848)
  • fix(dashboard): the Engine Combos pipeline editor and the sidebar settings page give each drag-and-drop list an SSR-stable id, so their drag handles hydrate without React's attribute-mismatch error and their aria-describedby points at the drag instructions dnd-kit renders (#15137) — thanks @woodsonl
  • fix(proxy): Steer the local core selector off every set-aside pool member, whatever the refusal motive, instead of quota refusals only (#15140) — thanks @maxmad64bis
  • fix(cli): omniroute status resolves its version from the package location instead of cwd (no more "unknown" on global installs), and omniroute cache status reports the real HTTP status instead of "not available" (#15148) (#15850)
  • fix(combo): preserve terminal upstream error codes such as invalid_encrypted_content in combo errors so Responses clients can recover from rejected reasoning replay. Mixed failure classes retain their aggregate status and generic code. (#15154) — thanks @shipsfromrio
  • fix(cache): API keys configured with cache_default_mode=bypass now skip semantic-cache writes as well as reads, including streaming responses. The persisted database semanticCacheEnabled toggle now gates the request path alongside the general settings toggle. (#15155) — thanks @shipsfromrio
  • fix(providers): sanitize DuckDuckGo upstream error types and anti-abuse override codes before returning them from the executor, and enforce the same error boundary in the DuckDuckGo and Copilot M365 local response builders without changing status codes or retry behavior (#15159). (#16125)
  • fix(dashboard): list every provider, model, account and API key in the Logs tab filters, not only the values in the loaded page, so a key with no row in view (or no traffic yet) can still be selected (#15160) — thanks @QuangBlue
  • fix(providers): add gpt-6-sol and gpt-6-luna to the OpenAI public API registry with the correct 1,050,000-token context window — they were missing and fell back to the 128k defaultContextLength, causing combos that included either model to advertise a capped-8x context window (#15023, #15164 — thanks @shipsfromrio)
  • fix(combo): a native Codex turn pinned to a model is no longer terminated with 400 NATIVE_CODEX_PINNED_MODEL_UNAVAILABLE when the model only has a short transient lockout (e.g. a 5s server_error set by another client); the router waits it out and continues on the pinned model (#15168) — thanks @QuangBlue
  • fix(dashboard): model import no longer reports "All models already imported" as a success when provider discovery fell back to the local catalog; the dialog now shows a distinct warning phase with the fallback message when importWarning is present and no new models were found (fixes #15069) (#15169) — thanks @shipsfromrio
  • fix(cli): omniroute stop accepts --port, falls back to the port after a stale PID file, and no longer prints "Server stopped." when lsof is missing (#15177) (#15716)
  • Align three nightly test contracts with the integrated Docker, correlation, and vision behavior: exercise the native rebuild shell boundaries, inspect real correlation sender calls while allowing structured error options, and require lossless vision reroute to block when no usable target exists. This changes tests only; it does not certify a Docker build or all nightly jobs. (#16251)
  • fix(docker): a root-built image started with OMNIROUTE_BASE_PATH now prefixes the webpack/Turbopack chunk loader, the client-reference moduleLoading.prefix and compiled CSS url() assets, so the dashboard hydrates behind a non-stripping subpath proxy instead of 404ing every lazy chunk at /_next/ (#15181, refs #9124)
  • fix(cli): omniroute serve --daemon and serve --no-recovery now spawn the server child with process.execPath instead of a bare node, so they start under launchd/cron/service managers whose PATH does not include the Node install (#15183, refs #9156)
  • fix(ci): docker-publish no longer leaks stale digests from self-hosted runners into version manifests (#15184) (#15836)
  • fix(sse): an isolated 5xx after a stream already relayed output no longer records a model-only lockout that benched the model for every other client; three such failures in a row still lock it (#15186) — thanks @QuangBlue
  • fix(codex): stop synthesizing a session/thread/turn identity envelope for Codex requests whose client never presented one. A request without Codex identity headers (OpenAI-compatible SDKs, non-Codex harnesses) previously received a fabricated x-codex-turn-metadata including a sandbox:"none" claim that no real Codex client emits; the ChatGPT backend rejects tool-bearing turns on the newest models with This request was blocked by our safety systems. Reason: Potentially unintended activity. Session/full fingerprint convergence now applies only when the client already carries a Codex identity — an identity header or a body session (prompt_cache_key / session_id / conversation_id) — and the fabricated envelope fields are gone. Device-level installation identity is unchanged, and a client-supplied x-codex-turn-metadata is still merged rather than replaced. (#15187) — thanks @Neuron-Mr-White
  • fix(i18n): Localize cache and access-token settings, feedback, scopes, and dates across all 67 UI catalogs, preserving technical identifiers and API values (#15189) — thanks @xiaoyaner0201
  • fix(db): a partial cavemanOutputMode settings write keeps the stored values of the fields it leaves out, so sending only autoClarity or only enabled leaves the other two as they were (#15206) — thanks @woodsonl
  • fix(ci): scripts/dev/new-worktree.sh checks the pre-commit hook where git actually runs it: an absolute core.hooksPath as-is, a relative one against the new worktree, and the main repository's .git/hooks when core.hooksPath is unset, so a worktree whose hooks are active no longer fails the closing check (#15209) — thanks @woodsonl
  • fix(routing): unknown models now carry a provenance flag on pricing and latency bootstrap fallbacks, so guesses are traceable instead of silent (#15224) — thanks @maxmad64bis
  • fix(db): a migration file that hits "duplicate column name" now gets its missing columns instead of being marked applied with none of them added, which left routes failing with "no such column" (#15236) — thanks @maxmad64bis
  • fix(dashboard): proxy health now reports headers-wait abandonments and client aborts as separate counters instead of transport failures (#15248) — thanks @maxmad64bis
  • fix(usage): Ollama Cloud usage cookie now accepts a bare value, name=value or a full Cookie header, and non-sign-in redirects are no longer reported as "authentication expired" (#15256) (#15854)
  • fix(cli): omniroute serve now scopes its port preflight to TCP listeners, so a client socket left on the port (for example a long-lived gateway connection in CLOSE_WAIT after its peer exited) can no longer make it report a free port as busy and crash-loop the service on restart — a real conflict is still detected (#15211) (#15259) — thanks @aldoeliacim
  • fix(kiro): Preserve activity from native signature-only reasoning frames so the stream content watchdog can wait for the answer without exposing signatures. Metadata-only stalls, the active-stream deadline, and client cancellation remain enforced (#15260). (#16121)
  • fix(cursor): Composer turns no longer end on a KV save that arrives while a tool call is still in flight (after an internal get_mcp_tools call, or while a client tool call is streaming), so the client tool call is returned instead of a one-line narration with finish_reason: "stop" (#15261) — thanks @QuangBlue
  • fix(dashboard): content-heavy modals scroll instead of clipping: the Provider Quota "USD Cost" modal body now scrolls inside its max-h-[88vh] panel (it was overflow-hidden without a bounded body, so rows past the fold were unreachable), the Cloudflare/Deno/Vercel relay modals cap at 90vh and scroll on short viewports, and the Orchestration history drawer no longer overflows 375px screens (#15262) — thanks @HDBR
  • fix(combo): return an actionable 400 context_length_exceeded instead of the opaque "Maximum combo retry limit reached" 503 when the attempt budget is spent mostly on context-window rejections (#15289) (#15856)
  • Return an actionable context-length error when round-robin or nested combo attempts are exhausted predominantly by context overflow, preserving diagnostics and the attempt limit (#15289). (#15934)
  • fix(ci): check:file-size PR mode judges test files against their base line counts too, so an innocent PR passes when a test file already drifted past its ceiling on the base (#15298) — thanks @woodsonl
  • fix(sse): clamp Xiaomi MiMo V2.5/V2.6 reasoning_effort on OpenCode gateways so xhigh/max/ultra become high and minimal becomes low, instead of opencode-go/mimo-v2.6-pro and mimo-v2.6-flash answering HTTP 400 "Invalid request parameters" (#15299) — thanks @fidelix
  • dashboard: declare the Add API Key bulk state before the open-reset effect so react-hooks/immutability no longer reports setMode/setBulkText accessed before declaration (#15306). (#16842)
  • types: narrow IntegerParam / Copilot /auto failures with ok === false and type Transformer.cancel so api-typecheck stops regressing nine TS2339s (base-red #15306).
  • test(auto-combo): seed provider connections in the auto/coding:fast and auto/coding:reliable weight-composition tests, since #15979 excludes the ToS-avoid keyless opencode pool by default and an empty pool returns default weights before the tier overlay (#15306) (#16151) — thanks @shipsfromrio
  • Correct the SQLite driver regression test to verify Bun stays on bun:sqlite while preserving Node's network-exclusion and report-restoration assertions. (#15306) (#16149)
  • Translate Burmese and Filipino Codex paid-credit controls and quota statuses, including the full quota-filtering exception, without changing the translation-ratio policy. (#16148)
  • fix(tests): match exact Codex discovery hosts and preserve token boundaries when masking skill-manifest comments, with adversarial URL and dangling-link regression coverage. (#15306). (#16141)
  • test: give auto/<family> materialization 90s so the ~50-candidate GLM cartesian pool can finish inside vitest (base-red #15306). (#16690)
  • i18n: use the canonical zh-TW glossary term 提供者 (not 供應商) on the four ToS-avoid auto-combo strings so check-glossary-consistency passes (#15306). (#16827)
  • fix(api): the dashboard model test works again with REQUIRE_API_KEY=true when the caller authenticates with a manage-scoped key, the CLI token or an oma_ token instead of a dashboard cookie. The test calls the /v1/chat/completions and /v1/responses handlers in-process. Since #16403 those handlers check authentication themselves, so the anonymous internal request got 401 Authentication required. Both internal requests now present the per-process self-loop bearer, which is never sent over the wire. They also send x-omniroute-no-memory: true, so the probe gets no injected memory or skills (#15306) (#16589) — thanks @shipsfromrio
  • Include five existing regression suites in the mutation-test selection so strict coverage checks account for their assertions. (#16130)
  • tests: Align the Notion executor's exact catalog expectation with the 20 fallback models added in #16113, preserving the original 17 IDs and the checks that internal codenames stay out of the public catalog. (#16138)
  • tests: Provider icon regression coverage prepares the real icon catalog during test collection, verifies the lazy loader's placeholder-to-fallback transition, and unmounts React roots between cases. This avoids cold dependency transformation consuming individual test deadlines while preserving URL, prototype-guard, and asset-provenance assertions. (#16134)
  • refactor(quality): extract balanced JSON boundary scanning and prose retry-delay parsing into small helpers, retaining the existing error-module exports. Extract Responses turn-grouping input data into a fixture while retaining every assertion in the original suite, and move the independent VS Code service-tier test into a dedicated file unchanged. Restores the four release file-size budgets without increasing ceilings, removing coverage, or changing error sanitization. (#16681)
  • fix(tests): allow only the retention HTTP fixture's allocated loopback host and port through MockAgent, so the retention matrix reaches the real routes while other loopback ports and external hosts remain blocked (#15306). (#16122)
  • Image-route regression guard: Require the error-helper allowlist declaration before checking that sanitized image routes are not frozen, so an empty or unrelated source file cannot silently satisfy the negative assertions. Refs #15306. (#16139)
  • test(api, oauth): two unit tests now follow behavior that already landed. The messages/count_tokens provider-count case expects claude-opus-4-6, the dashed id that #16425 sends to the Anthropic Messages API, instead of claude-opus-4.6. The Antigravity OAuth enrichment case pins the IDE version cache to ANTIGRAVITY_IDE_FALLBACK_VERSION and builds its User-Agent pattern from that constant. It was still matching 2.1.1 after #16512 moved the fallback to 2.5.5, so loadCodeAssist failed its mock and projectId came back empty. Clears two v3.8.52 base-reds (#15306) (#16584) — thanks @shipsfromrio
  • fix(proxies): Share the set-aside store across duplicated server module copies, so a connection reset or 429 recorded through one copy keeps that egress set aside for every reader (#15316) — thanks @maxmad64bis
  • fix(translator): Give chronologically reused tool calls unique paired Gemini wire IDs after resolving thought signatures with the original IDs. Reject overlapping duplicate calls locally instead of guessing result pairing; native Responses passthrough and client history stay unchanged (closes #15312) (#15319). — thanks @Kizuno18
  • fix(api): the standalone server (Docker dev/run-standalone.mjs) now owns SIGTERM/SIGINT, so graceful-shutdown cleanup (spend flush, call-log flush, DB checkpoint) runs before exit instead of being preempted by Next's own handler calling process.exit(143) (#15331) — thanks @QuangBlue
  • fix(api): npm run dev / npm start no longer cut claimed WebSocket upgrades (Responses WS, /v1/ws bridge): Next's lazily self-attached upgrade listener now sits on a relay and only receives upgrades nobody else claimed (HMR) (#15332) — thanks @QuangBlue
  • fix(resilience): a pooled relay attempt that exceeds RELAY_FETCH_TIMEOUT_MS no longer escapes as an unhandledRejection and kills the server (exit code 7) — the process crash guard now classifies RELAY_TIMEOUT/relay_timeout as a recoverable upstream timeout, like the direct-fetch DIRECT_RESPONSE_START_TIMEOUT case (#PR) (#15340) — thanks @Theadd
  • fix(compression): the compression panel and the compression settings tab show a Retry when GET /api/settings/compression fails, and keep their controls hidden until a load succeeds, so a failed load can no longer save defaults over the stored output styles, context budget or engines (#15346) — thanks @woodsonl
  • fix(api): GET /api/combos/auto now resolves candidate capabilities once per request instead of once per candidate for every listed variant, so listing auto combos no longer holds the event loop on large pools (#15353) — thanks @maxmad64bis
  • perf(db): drop two conversation turn node indexes no query reads (#15354) — thanks @maxmad64bis
  • fix(auto-combo): a provider whose quota fetcher returns nothing readable (failed fetch, missing credentials, message-only or malformed snapshot) no longer scores as a full quota in auto-combo; it ranks strictly below any real reading without being blocked or evicted, and unlimited plans now report unlimited: true so they still score as full (#15358) — thanks @shannonlowder
  • fix(api): log the auto-combo variants GET /api/combos/auto skips instead of failing silently (#15362) — thanks @maxmad64bis
  • fix(cursor): expose Grok 4.7 context, reasoning and fast variants from the Cursor catalog and preserve their parameters upstream (#15366). (#15911)
  • fix(sse): name the egress proxy on every opencode rotation log line (#15375) — thanks @maxmad64bis
  • fix(proxies): a member refused for the request region (403/451 with a region signal) is set aside briefly instead of retried first by every following request (#15383) — thanks @maxmad64bis
  • fix(providers): the freetheai gateway now points at freetheai.org — api.freetheai.xyz is retired (answers HTTP 530), signup moved off Discord, and model ids are namespaced fta/<group>/<model> and must be sent in full (#15385) — thanks Vibhek Soni (FreeTheAI) (#15520) — thanks @jonlwheat2-gif
  • fix(providers): Gemini Web connection test no longer reports valid for any pasted value — it now requires a __Secure-1PSID cookie, a signed-in session (SNlM0e) on 200, and rejects redirects to non-Google hosts (#15387) (#15853)
  • fix(api): log admission rejections on the chat and responses routes to the request journal with their status and reason (#15390) — thanks @maxmad64bis
  • fix(plugin): stop requesting the virtual-entries route and publish the gateway /v1/models entries as served (#15392) — thanks @maxmad64bis (diagnosis in #14889 — thanks @bilalsafdardev)
  • fix(shared): share request context stores across duplicated modules (#15395) — thanks @maxmad64bis
  • fix(rankings): free provider ranking now follows the documented free-model catalog (#15396) — thanks @maxmad64bis
  • fix(i18n): keep quoted placeholders working in Maltese and Korean messages (#15399) — thanks @maxmad64bis
  • fix(auth): answer a paused no-auth provider with a retryable cooldown (#15402) — thanks @maxmad64bis
  • fix(grok-cli): Grok Build discovery prefers the grok-cli registry context window over the 256k that Grok Build /v1/models reports for every model, and never stamps a default; the reconciler drops stale auto:discovery pins when discovery reports no window, so grok-4.5/4.6/4.7 advertise their real 500k window (#15412) — thanks @QuangBlue
  • fix(kiro): Kiro model discovery no longer stamps a 200k context window when tokenLimits.maxInputTokens is missing, so the reconciler cannot pin a fake auto:discovery override over the registry / models.dev window (#15413) — thanks @QuangBlue
  • fix(combo): skip the redundant second LKGP write when a combo success already pins the combo-level key (#15432) — thanks @maxmad64bis
  • fix(sse): keep thinking-only streams alive past the content-stall budget (#15433) — thanks @maxmad64bis
  • fix(sse): record the received Trae upstream status on proxy failures so health stats count them as upstream instead of transport failures (#15434) — thanks @maxmad64bis
  • fix(auth): serve catalogued free-access models only on credits-exhausted OpenRouter connections (#15436) — thanks @maxmad64bis
  • fix(auth): limit the no-auth free-tier pause to the refused model so sibling models stay eligible (#15443) — thanks @maxmad64bis
  • fix(deepseek-web): native single-pipe |DSML| and full-width DSML tool calls are now bound to the request nonce and validated against the tool schema; an unbound or invalid envelope is repaired once in a fresh session and, if still unusable, answered with a 502 instead of leaking the raw envelope as a 200 text reply. The double-pipe dialect and the canonical <tool> envelope keep the existing #9343 policy (#15448) — thanks @arashashrafii
  • fix(sse): a dead Codex refresh token no longer trips the provider circuit breaker, so healthy accounts keep serving while the dead one asks for re-authentication (#15449) — thanks @maxmad64bis (with thanks to @xiaoyaner0201 for the diagnosis and repro in #15444)
  • fix(cli): the server bootstrap (scripts/build/bootstrap-env.mjs), npm run env:sync and the desktop app now resolve the default DATA_DIR exactly like the runtime and the CLI: an existing ~/.omniroute wins over $XDG_CONFIG_HOME/omniroute. With XDG_CONFIG_HOME exported, the first start persisted JWT_SECRET / STORAGE_ENCRYPTION_KEY / API_KEY_SECRET to ~/.config/omniroute/server.env while the database and the CLI used ~/.omniroute, so a provider credential saved on one side could not be decrypted on the other. (#15450) — thanks @grapeslush
  • fix(open-sse): replay a 403 or 451 title refusal carrying insufficient_quota once in the other tool shape, so a wrong shape no longer surfaces as exhausted credits (#15451) — thanks @maxmad64bis
  • fix(opencode): count silent streamed replies that would have tripped the 15 s first-byte window while the guard stays off (#15455) — thanks @maxmad64bis
  • fix(cc-bridge): preserve OpenAI tool history on Claude Code compatible targets (#15457) — thanks @skygunner
  • fix(combo): keep declared priority order ahead of session stickiness (#15458) — thanks @skygunner
  • fix(combo): persist and clear per-combo universal handoff config (#15459) — thanks @skygunner
  • fix(combo): the auto combo pool check types the custom-model row so typecheck:noimplicit:core passes (#15461 — thanks @HouMinXi)
  • fix(sse): record the received Cursor upstream status on proxy failures so health stats count them as upstream instead of transport failures (#15463) — thanks @maxmad64bis
  • fix(cli): the --tray launcher no longer kills a still-booting worker after a fixed 60s (autostart on Windows never converged while systray2 installed); default readiness deadline is now 240s and configurable via OMNIROUTE_TRAY_READY_TIMEOUT_MS (#15464) (#15855)
  • fix(cli): the serve port guard ignores a listener on another address when the server binds a specific host (#15472 — thanks @HouMinXi)
  • fix(providers): a 400 or 429 reading "endpoint is unavailable" now pauses only that model for 5 minutes (#15474) — thanks @maxmad64bis
  • fix(opencode): only a 403 or 451 naming the refused model clears borrowed tools; a 429 leaves them alone (#15475) — thanks @maxmad64bis
  • fix(api): log wrapper admission rejections to the request journal with their status and reason (#15482) — thanks @maxmad64bis
  • fix(noauth): paused free-tier models now show a pause note on the account card instead of unexplained failures (#15483) — thanks @maxmad64bis
  • fix(db): Preserve explicitly saved provider-connection priorities instead of silently renumbering them on edit; keep the internal move-to-top operation and return its persisted rank (#15485) — thanks @piyush97
  • fix(sse): a native Responses/Codex turn whose pinned model is only temporarily unavailable (quota, rate limit, overload, open circuit) now gets a retryable 429/503 with error.code: "model_cooldown" and a Retry-After header instead of the terminal 400 "start a new turn", so opaque continuation state and pending tool calls survive the wait; an account-level block keeps the pin for those turns, while a plain turn with a healthy alternate still auto-resumes there as before (#15486) — thanks @JxnLexn
  • fix(providers): sync missing account catalogs before model tests (#15487) — thanks @JxnLexn
  • fix(sse): map a Codex/Responses response.incomplete (max_output_tokens, content_filter) to finish_reason length/content_filter instead of a clean stop that was then rejected as a 502 "Provider returned empty content" (#15489) (#15867)
  • fix(build): repair partially traced external package dirs missing package.json (e.g. zod) in the standalone bundle so Windows portable MCP endpoints stop returning 500 (#15493) (#15715)
  • fix(combo): a local per-target timeout records a model lockout without tripping the provider breaker (#15494 — thanks @HouMinXi)
  • fix(usage): price separate reasoning token buckets instead of a negative cost adjustment (#15496) (#15717)
  • fix(chatgpt-web): __Host- cookies from imported storage state or a Cookie header are now injected host-only (no leading-dot domain), so the ChatGPT Web browser session is no longer rejected (#15500) (#15860)
  • fix(timeout): configure the streaming fetch-start response-header cap while preserving its 110-second default. (#15542) — thanks @lorenzozanee
  • fix(sse): a finished Claude stream with end_turn or stop_sequence and no content block stays a 200 (#15505 — thanks @HouMinXi)
  • fix(sse): PII transform in warn/off mode now passes Responses text deltas through byte-identical (no buffering/re-chunking); redact snapshot sanitization guarded by test (#15507) (#15845)
  • fix(sse): keep Responses API tool-call ids/arguments intact and stop replaying output_item.done in the PII stream transform (#15507) (#15824)
  • fix(sse): a replayed Responses event is dropped when sequence_number is a whole-number string (#15510 — thanks @HouMinXi)
  • fix(providers): Muse Code OAuth requests use the configured Responses API endpoint instead of Chat Completions. (#15576) — thanks @lorenzozanee
  • fix(api): honor user pricing overrides (PATCH /api/pricing) in GET /v1/models catalog pricing (#15528) (#15831)
  • fix(reasoning): preserve explicit thinking opt-outs—including OpenRouter reasoning.enabled: false—through preprocessing and OpenAI/Responses/Claude translation, apply native off switches to automatically detected compatible backends, and restore native-thinking-off defaults for Claude Code classifiers that omit reasoning controls, preserving real permission verdicts and explicit overrides (#15533) — thanks @D4rk4
  • fix(video): retry transient 429/503 on async video job status polls with Retry-After-aware backoff instead of aborting (#15536) (#15822)
  • fix(auto-combo): RulesStrategy now honours the routing weights (mode pack / category overlay) instead of always using DEFAULT_WEIGHTS on virtual auto/* routes (#15546) (#15813)
  • fix(auto): auto/subscription now fails closed when its pool is empty instead of widening to every active provider's full catalog (#15546, complements #15813) (#15852)
  • fix(auto): auto/thrifty keeps its subscription-ladder rung order after scoring instead of letting the score re-sort escalate to metered models first (#15546, complements #15813) (#15852)
  • fix(auto): unpriced models no longer default to a flat $1/M in auto-combo scoring; they take the priced-pool median cost (neutral), and namespaced ids retry the bare model-name pricing lookup (#15546, complements #15813) (#15852)
  • fix(sse): Register batch quota trackers after imports resolve, so a cold chat route no longer dies with an empty HTTP 500 ((0 , e.Zd) is not a function) (#15548 — thanks @wedreamer).
  • fix(qoder): Qoder PAT validation accepts any exit-0 model catalog instead of requiring the literal MODEL header, and queued/busy (10605) qodercli responses are classified as a retryable 503 instead of a long-cooldown 502 (#15559) — thanks @christianmahardhika
  • fix(providers): honor the Custom Models vision override when a combo step uses the custom node alias prefix (aegy/<id>) (#15563) (#15806)
  • Disable cloud sync by default for new installations, block uploads while disabled, and require explicit credential-upload acknowledgment in the Endpoints enable dialog. Show the remote destination and distinguish cloud sync from local Storage backups. (#15565) — thanks @adivekar-utexas
  • fix(providers): Antigravity Claude Opus 5.5 / Sonnet 5.5 work again. Antigravity serves each reasoning tier as its own upstream id (claude-opus-5-5-high, …) with no bare claude-opus-5-5, but the Claude effort-variant normalizer stripped the tier and dispatched the bare id, so every request failed with 429 Resource has been exhausted on all accounts. Tiered Claude 5.x ids on antigravity / ag / agy are now dispatched verbatim, and the catalog no longer synthesizes effort variants on top of an Antigravity Claude 5.x base id. (#15566 — thanks @edosulai)
  • fix(compression): the caveman context page shows one Auto-Clarity control, and its Auto-Clarity, language and pack saves keep changes saved elsewhere after the page loaded; turning a language pack off keeps English enabled, and the two translation keys only the removed control used are deleted from every UI catalog (#15571) — thanks @woodsonl
  • fix(compression): the Compression Hub, the combos manager and the RTK page show a Retry when GET /api/settings/compression fails, and keep the controls that depend on the loaded settings off screen until a GET succeeds, so a failed load can no longer render defaults as the stored settings (#15583) — thanks @woodsonl
  • fix(api): GET /api/models now reports available consistently with the live-catalog gate used at dispatch, so models an authoritative provider catalog omits are no longer shown as available (and listed ones are no longer hidden) (#15585) (#15862)
  • fix(compression): the omniglyph context page shows a Retry when GET /api/settings/compression fails, and keeps its enable toggle and profile select hidden until a load succeeds, so a failed load can no longer save default engines over the stored row (#15587) — thanks @woodsonl
  • fix(resilience): settle acquired provider probes once under their generation lease, normalize provider aliases, and advance priority and round-robin combos immediately on local circuit refusals while preserving upstream 503 accounting (#15588). (#15594) — thanks @TheViniAlmeida
  • fix(dashboard): the caveman page's Output Mode card previews the exact block requests carry; the injector's composition moves into the exported buildOutputStylesInstruction() so preview and injection cannot drift, and the dead contextCaveman.preview keys are removed from all 67 UI catalogs (#15589) — thanks @woodsonl
  • fix(ci): the Fast Quality Gates loop runs the import-cycles gate in ratchet mode (--ratchet, ceiling quality-baseline.json → metrics.cycles) like ci.yml already does, so PRs against the release branch stop failing on the frozen 14-SCC baseline that #15281 turned into a hard exit (#15590) — thanks @woodsonl
  • fix(quality): new-code ratchet gates (complexity, dead-code) diff against the GitHub merge ref's first parent — the live base tip — instead of the stale pull_request.base.sha, so base commits landing after a PR is opened are no longer counted as the PR's new code (#15590) — thanks @woodsonl
  • fix(dev): the dev server runner (run-next.mjs) retries its listen with bounded backoff on EADDRINUSE, ending the integration-suite port race where a leftover socket claims the harness port during the ~15-18s prepare() window and kills the boot (#15590) — thanks @woodsonl
  • fix(compression): legacy cavemanOutputMode injection now matches the old caveman text in Hungarian (it no longer falls back to English) and drops the extra space before the shared boundary sentence in Japanese and Chinese, including terse-cjk and multi-style selections. Prompt-cache prefixes for existing ja/zh selections change once on upgrade. (#15591) — thanks @woodsonl
  • fix(dashboard): surface failed saves and loads in the compression exclusions panel (save-failed alert, load-failed alert that locks the editor so Save cannot store an empty list over the stored exclusions, 15-second request timeouts, loading-aware count), and make the RTK context page queue its config saves one at a time with per-edit rollback, shape-checked replies with one read-back inside a shared 15-second budget, number fields that commit once and clamp to the schema bounds, an unload guard with commit-on-unmount, a failure toast after the page closed, and a load-failure alert that blocks saves (#15592) — thanks @woodsonl
  • fix(compression): the Compression Hub's two settings controls stay enabled while a save is in flight, so saves overlapped and a failed one reverted the whole screen to its own starting snapshot, wiping a newer save the server had already stored; saves now go out one at a time through a queue shared by every mounted Hub, a failed save rolls back only its own fields, its error clears when a later save stores the field that failed or at the next edit, a queued save that a later one replaces is skipped, a stalled PUT fails after 15 seconds instead of holding the queue, the page asks before unloading while a save is pending, and a Hub mounted behind the queue shows what the server stored (#15593) — thanks @woodsonl
  • fix(dashboard): Compression engine pages now save only the fields edited on that page. A save re-reads the stored settings and applies the edits onto the fresh stored copy, so saving an engine page no longer reverts thresholds and tool strategies the Compression Settings tab saved after the page loaded. An emptied text field clears its stored key, ultra.enabled survives engine-page saves, and the Ultra page saves on a default install where the model path was never set. (#15595) (#15612) — thanks @woodsonl
  • fix(compression): CCR retrievalRampFactor keeps fractional values across save and reload — the read normalizer clamps to [1, 100] without flooring, so a saved 1.5 no longer disables the retrieval ramp (#15603) — thanks @woodsonl
  • fix(compression): the Aggressive and Ultra engine pages' Preview now runs with the form values the operator set: /api/compression/preview threads the sent config.aggressive/config.ultra into single-engine and pipeline previews instead of silently using built-in defaults, and the engine config page strips empty-string form values so Ultra's default modelPath: "" no longer fails preview and save validation (#15604) — thanks @woodsonl
  • fix(compression): the Lite, Aggressive and Ultra engine pages no longer show a "Preserve system prompt" checkbox. The per-engine field never took effect: every apply site reads the flag only from the global compression settings, the settings normalizers drop the per-engine copy, and the Lite page never sent it, so unchecking the box saved silently and snapped back on reload. The global "Preserve system prompt" control stays the real switch, stored combo step configs that still carry the key keep validating, and the orphaned locale label went away with it (#15605) — thanks @woodsonl
  • fix(compression): current engine settings rows (aggressive, ultra, headroom) now win over legacy aggressiveConfig/ultraConfig/headroomConfig rows in getCompressionSettings(). The unordered namespace read let the longer legacy key, read second via the (namespace, key) index, shadow every save on the Aggressive, Ultra and Headroom engine pages: saves returned 200 while GET and live compression kept the legacy values. A legacy row with no current counterpart still applies, and a parsed non-object legacy value still resets the engine to defaults. (#15607) — thanks @woodsonl
  • fix(compression): the legacy default-combo fallback now respects the x-omniroute-compression header and the lossy request policy: a header-chosen plan (named combo, engine:<id>, default, off) always wins, and without an opt-in the default combo runs only its safe steps, so an operator-edited default combo can no longer run rtk or other lossy engines on requests that never asked for them (#15609) — thanks @woodsonl
  • fix(dashboard): compression engine config pages treat an emptied number field as "not set" instead of saving 0, which silently changed compression for fields that accept zero (aggressive.minSavingsThreshold, ultra.compressionRate, ultra.minScoreThreshold) and 400'd saves for fields with a floor (maxTokensPerMessage, minRows, minBlockChars, minChars, retrievalRampFactor). Emptied fields now save as unset, show the schema default as a placeholder, keep an unparseable entry from unsetting the value, and reject overflow input visibly; lite's cap still clears when emptied. (#15612) — thanks @woodsonl
  • fix(compression): the compression engines map now merges by engine id on PUT /api/settings/compression instead of storing the map whole, and the Omniglyph page and context settings panel send only the engine they changed — so a toggle on one page no longer wipes what the other page (or a failed settings GET) never loaded. (#15613) — thanks @woodsonl
  • fix(docs): Correct two setup notes that contradicted the code. .env.example and the docs called OMNIROUTE_ALLOW_PRIVATE_PROVIDER_URLS=true required for local providers, but OMNIROUTE_ALLOW_LOCAL_PROVIDER_URLS already allows local and LAN URLs by default, and the private flag turns off the cloud-metadata block on provider validation and model discovery. The bootstrap INITIAL_PASSWORD warning said an unset variable fell back to CHANGEME; it now reports the CHANGEME placeholder, an unset variable (the onboarding wizard asks you to create the password), and a whitespace-only value separately (#15618) — thanks @woodsonl
  • fix(stream): passthrough keeps the real trailing usage-only summary (cache reads) after the finish-chunk usage instead of dropping it (#15619) (#15826)
  • fix(kiro): fall back from the branded kiro.dev gateway to CodeWhisperer on 400 "Improperly formed request" / REQUEST_BODY_INVALID for Builder ID and amazon-q (#15621) (#15828)
  • fix(sse): route an explicit provider/<id> to the named provider when its active customModels or synced catalog has the id, instead of freezing it as an exact id cataloged by another provider (sensenova misrouted to xkiro) (#15622) (#15864)
  • fix(providers): Keep Claude accounts routable when active critical usage warnings still report remaining quota; report the actual reset of the blocking account and model windows. (#15624). — thanks @feci / @HouMinXi
  • fix(sse): a stream content stall no longer cools down the account that served it. The stall watchdog fails that one request (usually a long reasoning turn with no output yet), but the 60s cooldown also took the account out of routing, so on a pool with one usable Codex account every request in that minute failed over or returned 503. The error stays recorded on the connection, and operators can opt back in with resilienceSettings.streamStallCooldown.enabled. Codex GPT-6.x models and the xhigh/max effort tiers now also get the high-reasoning readiness budget. (#15628) — thanks @fouadSalkini
  • perf(usage): cache slow GET /api/usage/analytics responses for 60 s per query and share in-flight computations, so repeated dashboard visits (API Manager, Costs, Budget, Analytics) no longer block the event loop — and every other request, including /v1 traffic — for several seconds on large usage_history tables. Tunable with OMNIROUTE_ANALYTICS_CACHE_TTL_MS (0 disables) and OMNIROUTE_ANALYTICS_CACHE_MIN_COMPUTE_MS (#15629) — thanks @HDBR
  • fix(executors): stop forwarding the caller's User-Agent (e.g. Python-urllib) to generic API-key providers such as Groq, so the provider/env/connection User-Agent applies and Cloudflare no longer returns 403 (#15632) (#15863)
  • fix(sse): a 404 model_not_found on one model now locks only that model for every provider instead of cooling the whole connection (#15633) (#15723)
  • fix(cursor): bare Connect not_found stream errors now classify as 404 model-not-found so the unavailable model is locked (#15671) (#15723)
  • Copilot models rejected with model_not_supported are temporarily omitted from the public catalog when no eligible account advertises them. Repeated requests retain the specific unsupported-model diagnostic; explicit clear or expiry restores listing without modifying synced inventories or combo definitions. (#16147)
  • fix(sse): lock a Copilot model for the connection after a direct 400 model_not_supported so it is not retried on every request (#15634) (#15849)
  • fix(usage): scale Claude extra-usage amounts by decimal_places so SGD 60.00 no longer shows as 6,000.00 (#15635) (#15718)
  • fix(conversations): parallel agents on one API key no longer re-insert their whole history as a new conversation on most requests. Reconnect candidates were the 20 most recently seen conversations in a bucket every agent on the key shared, so with more than 20 active agents a session's own conversation fell out of the window and its next request copied hundreds of turn nodes again (about 1.3M rows, ~1 GB a day, on one deployment). Candidates are now the bucket's conversations that already hold some of the request's turns, and when the client sends a session id (x-claude-code-session-id, Codex session_id, opencode x-opencode-session, or Claude's metadata.user_id session) the bucket is that session; the old bucket is still searched so existing conversations keep continuing. Migration 206 recreates the (conversation_id, content_hash) turn-node index (dropped by migration 201) that the new candidate lookup probes, so it stays a fast indexed search instead of a ~0.3–0.4 s row-by-row scan per request on busy installs. (#15637) — thanks @fouadSalkini
  • fix(usage): bound the in-flight pending-request map by retained bytes, not entry count alone — the 5000-entry cap bounded COUNT but not MEMORY (each entry retains up to four preview-truncated payloads, landing a worst-case burst near 250 MB). A 256 MB byte ceiling now evicts stale-marked-then-oldest entries after every insert, instead of waiting for the 5-minute age sweep; the UPDATE paths are measured too, so the ceiling cannot no-op. Pending ids also gained a monotonic suffix: the previous 24-bit random suffix collided ~52% of the time across a 5000-request burst. (#15640) — thanks @alvinveroy
  • fix(api): Preserve token limits, modalities, and reasoning tiers for custom-provider combos saved with public model prefixes and internal node IDs (#15668) — thanks @ekinnee
  • fix(auto-combo): auto/claude-opus|sonnet|haiku now pin the candidate pool to that Claude family instead of scoring every connected model (#15675) (#15813)
  • fix(compression): use Unicode-aware word boundaries in Russian rule packs and language detector so Cyrillic rules actually match (#15677) (#15719)
  • fix(compression): Unicode-aware word boundaries in the de/es/fr/it/pt-BR/hu rule packs and a working user overlay dir $DATA_DIR/compression/rules (#15677) (#15812)
  • fix(api): GET /api/usage/provider-limits no longer serves cached quota entries for deleted provider connections — all three connection-delete paths now purge the connection's providerLimitsCache entry (#15699) — thanks @trespassmk
  • fix(sse): the stream content-stall watchdog no longer cancels reasoning turns that are still thinking. Claude thinking deltas with no visible text and Responses reasoning items now restart its budget, so long xhigh turns stop failing after the readiness timeout and clients no longer fall back to non-streaming retries that reverse proxies cut off (Cloudflare 524) (#15700) — thanks @fouadSalkini
  • fix(sse): keep the word json in Responses input when response_format is json_object and only the system prompt mentions JSON, so Codex no longer answers 400 (#15702) — thanks @QuangBlue
  • fix(claude): forward the client-negotiated message-threads-2026-08-12 beta so Claude Code's top-level thread field is accepted upstream (#15705) (#15808)
  • fix(sse): a failed deduplicated request with no joiners no longer raises an unhandled rejection that exits the server with code 7 — fixed via #15750, regression test added (#15706) (#15817)
  • fix(memory): rebuild memory_fts bound to memories.memory_id via migration 203 so legacy NULL-id rows no longer leave orphan FTS docs or hide new memories (#15707) (#15810)
  • fix(i18n): sync-ui-keys --locale=<code> writes the locale catalog again instead of failing with ReferenceError: localePath is not defined (#15713).
  • fix(cli): provision STORAGE_ENCRYPTION_KEY in getDefaultDataDir() instead of hardcoded ~/.omniroute when DATA_DIR is unset (#15730) (#15807)
  • fix(open-sse): a replayed shape refusal reuses the verdict read once instead of reading the refused body again (#15737) — thanks @maxmad64bis
  • fix(sse): a request refused for quota on its own tools is retried once with the observed tool names appended, and the retry answer is served when it passes (#15738) — thanks @maxmad64bis
  • fix(models): After catalog state changes, a timed-out /v1/models rebuild no longer returns the previous model list as a successful response. (#15739) — thanks @lorenzozanee
  • fix(claude): stop .env.example pinning an outdated CLAUDE_USER_AGENT and ignore a stale claude-cli/<older> env UA below the pinned Claude Code version, so Opus 5.5 is no longer rejected (#15740) (#15808)
  • fix(sse): record rejected chat requests in the request journal with status and reason (#15741) — thanks @maxmad64bis
  • fix(sse): hold enabled per-connection maxConcurrent overrides until response bodies finish, fail, or are cancelled, so receiving HTTP headers cannot admit extra streaming requests; preserve legacy caps, clear stale limits when credentials change, and prevent an unobserved deduplication rejection from terminating the server when admission fails (#15750) — thanks @D4rk4
  • fix(speech): Xiaomi MiMo TTS now sends audio.format as the mp3/wav enum instead of an IANA media type, so POST /v1/audio/speech no longer fails upstream with Param Incorrect (#15752) — thanks @LazyGatto
  • fix(providers): Xiaomi MiMo quota now reads the real Token Plan usage from the MiMo console API (plan label, used/limit Credits, currentPeriodEnd reset) through a new per-connection console-cookie field, and the self-tracked fallback budget becomes configurable (providerSpecificData.monthlyTokenLimit / XIAOMI_MIMO_MONTHLY_TOKEN_LIMIT, default 4.1B Lite Credits) — no more permanent Exhausted after a plan upgrade (#15753) (#15759) — thanks @LazyGatto
  • fix(sse): Mistral reasoning models (e.g. labs-leanstral-1-5) that stream delta.content as typed part arrays ([{"type":"thinking","thinking":[…]}]) no longer leak the array into response.output_text.delta or Chat-Completions content — text parts become the content string and thinking parts reasoning_content, so strict Responses clients stop failing with "malformed response.output_text.delta delta" (#15757) — thanks @hartmark
  • fix(embeddings): with two or more connections of the same local embedding provider (llama.cpp, LM Studio, Ollama, Lemonade), a request is no longer sent to a connection whose configured default model names a different model — local servers ignore the request's model, so the wrong connection silently returned vectors from another model and dimension (#15761) — thanks @hartmark
  • fix(providers): keep upstream reasoning.supported_efforts when syncing named OpenAI-style providers so /v1/models exposes effort_tiers (#15762) (#15815)
  • fix(api): count_tokens no longer returns a provider input_tokens of 0 for non-empty input; falls back to the local estimate (#15763) (#15809)
  • fix(translator): omit the "(empty response)" text block next to tool_use in non-streaming OpenAI-to-Claude responses (#15764) (#15805)
  • fix(authz): gate the Devin CLI-spawn fallback on peer locality so a leaked JWT via tunnel cannot spawn processes (#15766) — thanks @jonlwheat2-gif
  • fix(mcp): sanitize the internal hop's error at the throw so upstream details never reach the client (#15767) — thanks @jonlwheat2-gif
  • fix(mcp): make the internal hop's env key a stdio-only fallback (#15768) — thanks @jonlwheat2-gif
  • fix(ci): restrict Docker runner-alert schedules to the upstream repository and enforce the scheduled-job guard across all Docker jobs while preserving manual dispatch and publish conditions (#15782).
  • fix(fetch): Avoid replaying inference POSTs after ambiguous response-start timeouts, preserve caller cancellation, and give unsafe methods the existing final request budget (#15787) — thanks @D4rk4
  • Keep sibling models eligible after a 413 explicitly naming the failed model or a 504 attributed to its model endpoint; generic connection failures retain their cooldown and cascade protection (#15788). (#15939)
  • fix(imageRegistry): split image-registry data from flag-gated logic so the client bundle does not leak server-only registry code (#15791) — thanks @Junior-HJ
  • fix(executors): make registerLazyExecutor HMR-safe so hot-reload no longer double-registers executors (#15792) — thanks @Junior-HJ
  • Restore direct compression/Auto-Clarity contracts and consolidate the POSIX capability check without removing CLI lifecycle cases or weakening the release test policy. (#15793)
  • fix(error): unwrap Tencent/CodeBuddy {Response.Error}, {data.Response.Error}, and {code,msg} envelopes in parseUpstreamError so CodeBuddy 4xx surfaces the real upstream message and code instead of a generic status (#15797) — thanks @Junior-HJ
  • fix(dashboard): Model Overrides picker search ranks results by relevance (exact provider, then provider prefix, then model-id prefix, then substring) and widens the filtered view to 300 rows, so a short query like bai surfaces the intended provider's models instead of hiding them behind baichuan/baidu neighbors (#15801) — thanks @woodsonl
  • fix(cache): include structured tool calls and outputs in both exact cache signatures while preserving plain-message keys (#15816). (#15909)
  • fix(claude): claude-haiku-5-5 requests keep thinking.type: "adaptive" instead of being downgraded to "enabled", which Anthropic rejected with HTTP 400 (#15841) — thanks @aldoeliacim
  • fix(muse-code): validate device authorization URLs and expiry, bound requests, and sanitize polling responses while preserving renewable DCA credentials (#15869). (#15920) — thanks @caniko
  • fix(tests): align Muse Code's non-JSON polling regression with its fixed-error contract and reject exposure of arbitrary upstream response text (#15869). (#16116)
  • fix(embeddings): llama.cpp connections discovered through model sync (the synced local-endpoint route) keep accepting canonical image/audio/video embedding items — the route rebuilt the provider config and dropped the multimodal capabilities added in #15758 (#15881) — thanks @hartmark
  • fix(quota): quota readings use the assigned connection egress path (#15885) — thanks @maxmad64bis
  • fix(monitoring): a silent event-loop freeze now leaves a trace — the server logs one line at resume with the stall length, memory and in-flight counts, and reports the count plus longest stall on the monitoring health payload (#15888) — thanks @maxmad64bis (diagnosis from #13891 — thanks @morpheus9393)
  • Report whether connection restrictions come from an API key, quota scope, combo pin, or routing allowlist, and log distinct reasons when synthetic no-auth fallback is unavailable without changing credential selection or access rules. (#15959)
  • Preserve Codex web-search history during Responses compaction and declare its replay tool without allowing new tool calls (#15901). (#15926)
  • fix(resilience): the COOLDOWN_RETRY "cooldown elapsed" log now reports the retry number (retry N/M) instead of an attempt index that could read 4/3 (#15905) — thanks @wakqasahmed
  • fix(dispatch): rebuild the shared cloud dispatcher pool after an unreachable error instead of reusing its stale sockets (#15908) — thanks @maxmad64bis
  • fix(resilience): Keep structured content-policy refusals (cyber_policy / content_policy_violation) request-scoped: they now surface as a 400 with their public error code instead of a synthetic 502, and no longer cool the account, trip the provider breaker or trigger a same-target stream retry. Stopping the whole combo on such a refusal (instead of advancing to the next target) is opt-in via the new CONTENT_POLICY_REFUSAL_COMBO_STOP_ENABLED flag, default off (#15917) — thanks @insoln.
  • fix(auth): Preserve the validated API key owner and management scope on file and batch list requests sent with a bare x-api-key, including requests carrying a dashboard session; reject unknown keys instead of falling back to session visibility (#15931).
  • Cline model discovery now includes the current free-model bucket alongside the full paid catalog, preserving each upstream model ID and retaining the last complete cache when either catalog is unavailable. ClinePass remains subscription-only; catalog discovery does not change upstream API-key access restrictions. (#15940)
  • fix(images): the Codex hosted image_generation path now forwards background (e.g. "transparent") to the tool and downscales inline reference images above 400 KB to a 1024px long edge, which the backend otherwise rejects with 503 (#15941) — thanks @HDBR
  • fix(providers): 9router requests now reach the reactive reasoning_effort 400 clamp-and-retry, so an upstream that names its accepted enum is retried once with the clamped value instead of surfacing the raw 400 (#PENDING, part of #14629) (#15946) — thanks @pacocartones
  • KIE image generation now selects Flux Kontext Pro/Max through explicit catalog metadata and maps legacy Z-Image picker IDs to the documented z-image upstream ID. Existing Flux Kontext requests retain their dedicated API contract; the canonical Z-Image model is also selectable. (#15947)
  • Generated Agent Skill API examples now use the canonical local base URL unless a caller explicitly supplies another host, preventing ambient environment variables from changing committed documentation. Explicit configured-host examples remain supported, and regression tests cover both behaviors, including isolation and restoration of NEXT_PUBLIC_BASE_URL. (#15948 — thanks @BenjaminAronsson)
  • Added OrcaRouter's verified base token prices for all seven fixed catalog models and corrected the Gemini Flash rate. The adaptive auto route remains unpriced; context-tier and time-of-day adjustments remain outside these estimates. (#15954)
  • Generated REST CLI commands now require and encode URL path parameters, including routes without explicit OpenAPI parameter declarations. Path-level parameters and operation overrides are preserved, and repeated placeholders are all replaced. (#15955).
  • Standalone images now retain the source CLI's tsx loader, esbuild compiler and installed platform binaries, preventing the missing-loader failure before CLI commands can start. (#15956)
  • Keep full translation catalogs out of the global error boundary's client bundles by generating small localized error catalogs before production builds. All supported languages and the English fallback remain available. (#15958).
  • Release stream-controller callbacks after completion, disconnection, errors, and aborts so retained response controllers do not keep completed request bodies alive through callback closures. (#15961)
  • Recognize Gemini thought signatures and OpenAI-compatible reasoning deltas as stream progress, preserving content-stall timeouts for lifecycle-only streams and the separate empty-output guard. (#15962)
  • fix(ci): the route-guard membership gate now walks each route's runtime import graph (G-09) and fails when a route that is neither local-only nor frozen reaches a NEW spawn-capable module; the 38 spawn modules already reachable when the walk landed are frozen per module (KNOWN_TRANSITIVE_SPAWN_SINKS) and fixed-argv plumbing is listed in TRANSITIVE_SPAWN_INFRA_MODULES, both stale-enforced. Spawn detection now keys on runtime child_process/worker_threads imports with comments ignored, so prose like -> spawn(), RegExp#exec, regex literals and exec(sql) method signatures no longer count (#15965 — thanks @jonlwheat2-gif)
  • Stop sending the Claude Code identity beta by default for the plain Anthropic API-key provider, while preserving OAuth/Claude Code headers and explicit compatible-provider overrides. (#15968)
  • Reject conflicting node identity before binding a new compatible-provider credential through a bare-type fallback, and report misplaced Cloudflare account IDs instead of discarding them. (#15987)
  • Restore long namespaced Responses tool identities from the current request's explicit function-call history, preserving current tool declarations and avoiding ambiguous or flat-name remapping. (#15974)
  • fix(providers): Google AI Studio keys can serve native image generation through the new gemini image provider (gemini/gemini-3-pro-image-preview, gemini/gemini-3.1-flash-image-preview), with a timed upstream call and n served by bounded single-candidate requests (#15976) — thanks @LuisAlejandro
  • fix(monitoring): warn once when the stall sampler reports an unusable reading, reusing the existing sampler fallback (#15978) — thanks @maxmad64bis
  • fix(sse): first-byte watchdog suite holds its own loop so it reports 3 pass instead of 3 cancelled (#15980) — thanks @maxmad64bis
  • fix(sse): count refused requests on the web executor as upstream 4xx/5xx instead of transport failures (#15981) — thanks @maxmad64bis
  • Keep translated README migration counts synchronized with the SQL migration files and check all configured README mirrors in the documentation gate. (#15982)
  • fix(ci): isolate nightly Schemathesis runtime state and block fuzzing when the server is unavailable or its port is already occupied (#15988) (#15998)
  • fix(sse): in-request Codex 429 / Antigravity BYOP 422 account rotation now stays inside the request's connection allowlist (API key allowed_connections, combo-step allowedConnectionIds, quota pool) instead of picking any account of the provider (#15990) — thanks @fouadSalkini
  • fix(db): keep token_ledger when resetting usage data (#16005) — thanks @pacocartones
  • fix(playground): Preserve selected provider/model and concurrent Config edits, send image-generation prompts in the required prompt field, and render image results with safe download links instead of raw JSON (#16006) — thanks @D4rk4
  • fix(sse): Preserve structured Responses refusals through translation and combo validation, and retain received text when a matching terminal message snapshot is empty. Refusals remain refusals instead of empty-response failures (#16007). — thanks @insoln
  • fix(profile): format XP numbers with the UI locale (#16008) — thanks @pacocartones
  • fix(profile): expose the badge detail panel as a modal dialog (#16009) — thanks @pacocartones
  • Retain bounded, content-free upstream and translated response shape counts for malformed non-streaming failures, independently of truncated payload artifacts (#16011). — thanks @insoln
  • Keep explicit Codex model-overload cooldowns on the exact model while preserving transient duration, Retry-After, quota/auth protections and generic connection-failure handling (#16014). — thanks @insoln
  • Detect recent event-loop lag peaks instead of a lifetime average, and offer opt-in, content-free timing of synchronous better-sqlite3 adapter operations (#16015). — thanks @insoln
  • fix(images): Route unannotated synced image models using the configured provider node endpoint type, preserving explicit model overrides and connection affinity (#16017) — thanks @D4rk4
  • fix(providers): Keep Codex app-server turns open during retryable errors and settle disconnected turns without hanging or starting work on a closed socket (#16029). (#16045)
  • fix(db): publish durable pre-migration snapshots when hard links are denied, while preserving complete backups and backup listing, restore, and retention (#16035). (#16044)
  • fix(db): avoid synchronous database-wide statistics and integrity scans when inference reads semantic-cache or modality-combo settings; retain administrative database diagnostics (#16037) — thanks @insoln
  • fix(codex): registered Codex -ultra aliases (gpt-6.1-sol-ultra, gpt-6-astra-ultra, …) now resolve to their registry base like -max/-xhigh, so they are no longer hidden from a live-authoritative catalog that only reports the base (#16038) — thanks @alltomatos
  • fix(quota): replace quadratic USD cost history matching with an indexed nearest-unused lookup, yield between calculation batches so HTTP requests can progress, and skip detailed cost loading for aggregate-only keys (#16041) — thanks @insoln
  • fix(bedrock): Converse prompt-cache usage is now read from the fields Bedrock returns (cacheReadInputTokens and cacheWriteInputTokens), so cache usage no longer reports 0, and prompt_tokens includes the cached tokens, as the cost calculator already assumes (#16043) — thanks @pacocartones
  • fix(translator): a null stop is dropped instead of being sent as stop_sequences: [null] to Claude and stopSequences: [null] to Gemini, which both reject it (#16047) — thanks @pacocartones
  • test: pin the generic headerless 429 fallback so it keeps rotating accounts instead of locking a model (#16051) — thanks @maxmad64bis
  • fix(providers): give MiMo Token Plan model tests a 60-second minimum timeout (#16213) — thanks @lorenzozanee
  • Tests: preserve the first-visit locale refresh ordering regression, checking that cookie, local storage, language and text direction are persisted before refreshing server content; isolate mock and React-root cleanup between cases (#16059).
  • fix(usage): normalize every decimal comma (not just the first) in the SYNTX token-balance and percent-left parsers, closing CodeQL #1107 (#16063)
  • fix(proxy-subscription): record throttled selector repeats on the subscription row so the list shows why nothing moved (#16067) — thanks @maxmad64bis
  • fix(streaming): preserve valid empty native completions through JSON, SSE, combos and deduplicated requests for official OpenAI, Anthropic, Codex, OpenRouter and documented Azure endpoints, while retaining unknown-origin, incomplete-stream and upstream-error guards (#16072). (#16112)
  • fix(dashboard): the resilience settings page surfaces the server's validation detail (offending field and reason) instead of the generic "Invalid request" toast when a PATCH is rejected, reusing the shared readFetchErrorMessage helper introduced for compatible-node prefix errors (#16085) — thanks @farshadasadpour
  • fix(providers): Kiro GPT-5.6 Sol/Terra/Luna use the 1M-token context window Kiro rolled out on 2026-09-14, so usage derived from contextUsagePercentage is no longer under-reported ~3.7x (#16086) — thanks @thotypous
  • fix(sse): silent Responses stall shadow miss is counted without a debug line (#16102) — thanks @maxmad64bis
  • fix(subscriptions): Refreshing a disabled subscription no longer deletes its missing nodes (#16104) — thanks @maxmad64bis
  • fix(providers): model discovery now sends the connection-level customHeaders, so a key that needs a header on every request (e.g. anthropic-workspace-id for an Anthropic key not scoped to a workspace) can import models (generic, claude and anthropic-compatible-* discovery paths), not only chat (#16108) — thanks @shipsfromrio
  • fix(sse): preserve mid-conversation system instructions for Claude Sonnet 5.5 and Haiku 5.5 instead of appending them to the top-level system prompt on each turn; retain legacy model handling and existing beta-header selection (#16127). (#16131)
  • fix(providers): the Anthropic API-key and anthropic-compatible connection tests no longer report a key as valid when Anthropic answers 400 "credit balance is too low" or "not scoped to a workspace" (every real request would fail the same way), and the probe now sends the connection-level customHeaders like the chat path does (#16146) — thanks @shipsfromrio
  • fix(i18n): keep rich-text tags across locales and Maltese CLI paths (#16162) — thanks @maxmad64bis
  • fix(proxy): attach a manually added proxy to an existing subscription from the proxy update route, and detach it again with a null subscription (#16166) — thanks @maxmad64bis
  • fix(pack): Include the global-error catalog generator imported by the published isolated-build entrypoint without allowing unrelated i18n scripts (#16171).
  • fix(resilience): Background-job history write and pruning errors are logged without escaping the scheduler or retrying completed handlers (#16181) — thanks @W25X80
  • Memory expiration: Serialize Date expiration values on UPSERT as ISO strings, matching inserts and preserving the existing record, owner, metadata and access telemetry. (#16182) (#16219)
  • Restore SQLite memory search when its optional token budget is omitted, including searches through the memory manager and keyless A2A memory recall. Respect the requested result count while preserving ranked order and the existing token budget. An internal SQLite limit of zero returns no hits; negative, fractional, and nonfinite limits are rejected. Fixes #16183, #16184, and #16185. (#16215)
  • A2A discovery: Align the legacy Agent Card’s static names, descriptions, and examples with the English v1 card, and remove the outdated fixed skill count while preserving legacy protocol fields and dynamic fleet entries. Fixes #16188. (#16226)
  • Preserve built-in and valid custom prompt-injection rules when another custom pattern has invalid regex syntax. Emit one bounded configuration warning per evaluation without logging pattern or request text, while preserving disabled modes, thresholds, and HTTP rejection behavior. Fixes #16189. (#16221)
  • fix(sse): record unknown previous response rejections in the request journal (#16193) — thanks @maxmad64bis
  • fix(sse): count refused requests on the grok-web executor as upstream 4xx/5xx instead of transport failures (#16194) — thanks @maxmad64bis
  • fix(gamification): Ignore malformed badge criteria consistently during evaluation and hidden-badge prerequisite checks, preserving valid awards, owner isolation, and the earned snapshot (#16195). (#16239)
  • Classify Cloud Agent statuses such as pending approval and waiting_for_plan_approval as awaiting_approval, preserving terminal and running precedence without adding provider plan-approval capabilities (#16196). (#16229)
  • Fix custom webhook delivery results and delivery history to retain the final HTTP status when 5xx retries are exhausted. Terminal network failures continue to use status 0. Fixes #16197. (#16230)
  • Gamification: Reject non-finite token-transfer amounts before ledger access and sanitize unexpected transfer errors through the canonical error boundary, preserving fractional amounts, idempotency and balance checks. Fixes #16198. (#16235)
  • Restrict Cloud Agent provider checks and lookups to registered IDs, rejecting inherited object properties while preserving the four agent singletons and their listing order. (#16232)
  • Fix gamification events with no positive XP reward creating zero-score leaderboard entries or refreshing existing score timestamps. Preserve positive awards, recognition-only Radar badges, and historical ranking rows. (#16238)
  • Preserve Slack, Discord, and Telegram payloads for unknown programmatic webhook events with a generic description. Ignore inherited description and color keys, retain the original event in custom deliveries and audit records, and keep successful wildcard deliveries from accumulating failures. Fixes #16201. (#16233)
  • Gamification: Keep XP-to-next-level finite for non-finite totals and consistent with the level calculation's safe-integer cap, while preserving finite negative XP debt and normal level boundaries. Fixes #16202. (#16231)
  • Stop advertising compatible-provider nodes on GET /v1beta/models when they do not uniquely own a non-reserved routable prefix (missing, reserved or losing-collision prefixes), so the Gemini catalog no longer lists models under a name that cannot be routed (#16207). (#16218)
  • fix(models): /v1beta/models names custom-provider models under the operator-configured provider-node prefix instead of the raw node UUID, matching the public identity /v1/models has published since #8327. (#16316) — thanks @skygunner
  • fix(models): /v1beta/models now enforces the same requireAuthForModels gate as /v1/models, so the Gemini catalog no longer lists models to unauthenticated callers on installs that require catalog auth. (#16316) — thanks @skygunner
  • fix(proxies): set aside a slow egress only with cross-egress serve evidence (#16212) — thanks @maxmad64bis
  • Endpoint: Add the cloud upload destination, credential warning and consent messages in all 67 UI locales so the real translator resolves them without missing-message errors. Preserve the full upload disclosure and local-backup distinction. Fixes #16220. (#16241)
  • Family fallback now skips hidden models and active connection-scoped model lockouts while preserving eligible sibling order (#16223, fixes #16227; thanks @turbolego). (#16810)
  • Resolve Antigravity Gemini 3.7 Flash high/medium/low requests against the fresh catalog of the selected connection before global aliases, preserving tiered-only accounts and rechecking the same connection after a 404 discovery sync. Explicit overrides, other model families, stale-catalog fallback and conservative concurrency leases retain their existing behavior. Fixes #16237. (#16254)
  • fix(dashboard): reveal streamed pages while the tab is hidden instead of staying on "Loading OmniRoute..." (#16245) (#16930)
  • fix(routing): count the abandonment of the first ordered combo target as a fallback, so a failover that abandons target #1 reports a non-zero count in the documented X-OmniRoute-Fallback-Attempts response header and in totalFallbacks instead of staying at 0 (#16250) — thanks @M3stryX
  • fix(images): the OpenAI-compatible image deadline now covers the response-body read, so an upstream that stalls after sending headers ends with 504 instead of hanging; opt-in IMAGE_GENERATION_LONG_JOBS adds the per-connection image timeout, caller cancellation (499) and SSE keepalive heartbeats for long Playground image jobs (#16260) — thanks @D4rk4
  • fix(docs): compile the getting-started, provider and comparison guides linked from the documentation index, use canonical featured-link slugs, and preserve required page titles and sensitive-page exclusions. Adapted from PrinceThummar011’s PR #16338. (#16261) (#16764) — thanks @PrinceThummar011
  • fix(sse): a finished text turn whose final event arrives without its closing blank line no longer triggers a wasted continuation re-request (#16267) — thanks @maxmad64bis
  • fix(conversation-reconnect): elect the reconnect point with the longest run of turns already on record, so a repeated tool result followed by new content appends to the same conversation instead of copying its whole history (#16268) — thanks @maxmad64bis
  • fix(app-log): clock-prefixed lines now file under the real component and file messages drop terminal color codes (#16270) — thanks @maxmad64bis
  • fix(calls): streamed replies now record whether they carried text or a tool call (#16271) — thanks @maxmad64bis
  • fix(resilience): PATCH /api/resilience accepts and returns the stream recovery block, so stored values can be changed without direct database edits (#16272) — thanks @maxmad64bis
  • fix(settings): accept the five stored retention durations missing from the database settings schema so the dashboard retention field takes effect (#16273) — thanks @maxmad64bis
  • fix(dashboard): the connection quota card now shows the payload-level quota message next to parsed rows instead of hiding it whenever a row exists, so a self-tracked fallback (e.g. xiaomi-mimo with an expired console cookie) can no longer look like a live plan reading; the MiMo hint also gains the same cookie capture instructions Qwen prints. (#16274) — thanks @LazyGatto
  • fix(token-refresh): skip the retry loop when there is nothing to refresh (#16275) — thanks @maxmad64bis
  • fix(usage): a usage write blocked by a locked database is retried once after a short delay instead of being dropped; if the server stops during the wait the row is lost as before, on a best-effort basis (#16276) — thanks @maxmad64bis (with thanks to @W25X80 for surfacing the SQLite write failures in #16176)
  • fix(security): Kiro token refresh validates the stored AWS region before building the OIDC endpoint, so a tampered region can no longer send the refresh credentials off AWS (#16280) — thanks @nitsuah
  • fix(sse): DeepSeek connections on the Claude wire format no longer receive type: "custom" tool definitions, which the endpoint rejected with HTTP 400 (#16281) — thanks @0xH3li0 / @kabeza
  • fix(sse): Codex tool schemas no longer forward pattern values with Unicode property escapes (\p{...}), which the Codex backend rejected with HTTP 400 (#16282) — thanks @minhtuancn
  • fix(sse): Qoder's 10605 queue-throttle envelope now returns 503 upstream_busy with Retry-After instead of 403 authentication_error, so healthy accounts are not parked on the auth cooldown (#16283) — thanks @chenli971014-sketch
  • fix(sse): unwrap ```json fences from non-streaming responses when the client requested JSON mode (response_format / text.format), and tell prompt-instructed providers not to fence JSON (#16284) — thanks @47vigen
  • fix(audio): AssemblyAI transcriptions no longer fail with 401: the API key is sent as a raw Authorization header (no Bearer scheme), and the language field now maps to AssemblyAI's language_code (#16285) — thanks @igorgomes3
  • fix(codex): detect the fake-200 "servers are currently overloaded" output-text stream and convert it to a 503 so account fallback engages (#16287) — thanks @dungartoriaaa
  • fix(codex): retry once on the same account without undecryptable reasoning ciphertext instead of failing the turn with invalid_encrypted_content (#16288) — thanks @ryanngit
  • fix(sse): the Qoder usage view re-exchanges a PAT after its cached job token is rejected, and job-token TTLs are read in either unit and capped at 24h (#16289) — thanks @AmirSys-dev
  • fix(sse): xhigh is no longer force-rewritten to max for max-tier model families once the provider+model is known not to accept max, fixing repeated 400s on custom DeepSeek V4 hosts (#16290) — thanks @66666-design
  • fix(call-logs): unclassified call-log errors now fall back to the HTTP status: bare 403s are stored as forbidden and retired-model 410s as model_not_found instead of unknown (#16291) — thanks @maxmad64bis
  • fix(perplexity-web): publish a received 400+ status on the active proxy sink instead of counting it as a transport failure (#16292) — thanks @maxmad64bis
  • fix(sse): server_tool_use blocks with non-Anthropic ids (e.g. z.ai call_…) and their paired results are dropped before a request reaches Anthropic, so a mixed combo no longer 400s on every following turn (#16293) — thanks @f-liva
  • fix(sse): cached thought signatures are scoped by model family, so switching between Claude and Gemini on one Antigravity connection no longer fails with "Corrupted thought signature" (#16294) — thanks @louisphamdev
  • fix(sse): Gemini tool schemas drop errorMessage and cache_control keywords and expand shorthand property entries like "metadata": "object", fixing HTTP 400s (#16295) — thanks @louisphamdev / @WellArtDev
  • fix(sse): requests to Gemini 3.x through the direct Gemini and Vertex paths no longer end on a model turn, which Gemini rejected with HTTP 400 (#16296) — thanks @ArchdukeViel
  • fix(sse): OpenAI↔Claude translation now preserves allowed_tools, parallel_tool_calls: false and tool-level strict (#16297) — thanks @Android-Login-Analysis
  • perf(usage): repeated usage-history views share one cached response for up to a minute instead of recomputing every five seconds, so the screen spaces its refreshes as designed and serving traffic no longer stalls (#16300) — thanks @maxmad64bis
  • fix(logs): keeping the console log view open no longer slows other requests (#16301) — thanks @maxmad64bis
  • fix(call-log): Request log rotation now honors the dashboard retention setting, falling back to the configured variable or 7 days (#16302) — thanks @maxmad64bis
  • fix(models): Retired models reporting 410 lock for 24 hours on status alone (#16303) — thanks @maxmad64bis
  • fix(sse): a translated stream ending without a closing blank line now forwards the final usage event instead of replacing it with an estimate, so local quota counters report the provider counts (#16304) — thanks @maxmad64bis
  • fix(sse): new opt-in flag COOLDOWN_RETRY_SKIP_REFUSALS_ENABLED (default off): when every account is cooling down and the earliest cooldown was a 403 or 410, the request returns the refusal at once instead of waiting the cooldown out and retrying, and a renewing agentrouter connection-scoped quota 403 is persisted as 429 so it stays retryable. With the flag off only a 401 skips the wait, as before (#16305) — thanks @maxmad64bis
  • fix(providers): remove only the keyless "OpenCode Free" provider (opencode, alias oc), preserving authenticated Zen/Go models and request contracts and the OpenCode client/plugin integrations. Migration 207 preserves keyed legacy connections by moving them to Zen, removes keyless configuration, and retains historical usage and call logs. See docs/reference/REMOVED_PROVIDERS.md (#16310).
  • fix(resilience): a combo whose every target is skipped by a short transient model cooldown no longer answers a hint-less 503 ALL_TARGETS_SKIPPED in ~100ms (#16319). The zero-attempt path now waits the lock out (quota/auth/not-found locks excluded) and re-dispatches; when no wait is allowed it answers 429 with the earliest transient lock expiry instead. (#16360) — thanks @skygunner
  • Build: documented the expected production-build warnings in docs/ops/BUILD_WARNINGS.md, stopped the login route from logging [SECURITY] FATAL: JWT_SECRET is not set during next build, and widened the Fumadocs MDX build-cache warning filter to cover the dist/eval-*.js bundle (#16320). (#16324) — thanks @opensource-elearning
  • fix(i18n): mount the /home route's message provider so the home page and its sponsor banners stop rendering MISSING_MESSAGE in production builds (#16321). (#16323) — thanks @opensource-elearning
  • fix(build): check:pack-artifact runs npm through the bundled npm-cli.js instead of trusting npm_execpath, so bun run no longer dumps the Bun ELF binary into the validator's error output (#16321). (#16323) — thanks @opensource-elearning
  • Fix CLI key-policy commands to use the authenticated management API, support explicit unrestricted model access, preserve unrelated limits, and omit credentials from policy output. Refs #16329.
  • perf(combos): list endpoint shares alias tables and resolved windows across combos in one request (#16331) — thanks @maxmad64bis
  • fix(health): Shrink the provider health matrix payload by omitting idle models by default, with includeIdleModels=true to restore the full list (#16333) — thanks @maxmad64bis
  • perf(usage-analytics): resolve each provider and model price grid once per GET /api/usage/analytics request instead of once per aggregated row, cutting about a third of the uncached response time on large usage_history tables; amounts are unchanged (#16334) — thanks @maxmad64bis
  • fix(api): Accept partial database settings sections with a 200 response (#16335) — thanks @maxmad64bis
  • fix(favicon): redirect to the default icon with an absolute URL instead of failing (#16336) — thanks @maxmad64bis
  • fix(proxy-subscriptions): keep the last switch trace per selector so multi-selector subscriptions show each selector's member, time and cause (#16339) — thanks @maxmad64bis
  • fix(usage): keep request detail files with the redirected test database (#16340) — thanks @maxmad64bis
  • fix(quota): remaining console quota readings use the assigned connection egress path (#16344) — thanks @maxmad64bis
  • fix(resilience): a stream-lifecycle error the pipeline already accounted for — the empty upstream flush stamped empty_response with the __omniroutePendingRequestCleared marker (#16347) — no longer escapes the combo retry wait as a process-fatal uncaughtException (exit code 7); the crash guard swallows it with a warn log, the same treatment RELAY_TIMEOUT received in #15339. (#16359) — thanks @skygunner
  • fix(translator): Claude Code WebSearch routed to Gemini/Antigravity now enables Google Search grounding and returns server_tool_use + web_search_tool_result blocks instead of plain text (#16351) — thanks @wray-lee
  • fix(checks): repo quality checks now run in a throwaway data directory instead of the operator database (#16353) — thanks @maxmad64bis
  • fix(call-logs): Hold the artifact ceiling even when files are already gone (#16354) — thanks @maxmad64bis
  • perf(cache): opt-in OMNIROUTE_CACHE_METRICS_MIN_COMPUTE_MS serves the usage-history part of the cache page from the shared response cache (history may then lag up to the analytics TTL); unset keeps every read live (#16361) — thanks @maxmad64bis
  • fix(resilience): saving resilience settings now stores only the values set in the request, so documented variables keep taking effect until their own value is changed (#16362) — thanks @maxmad64bis (diagnosis in #4165 — thanks @KooshaPari)
  • fix(db): pre-migration snapshots are now pruned by the configured retention outside the migration window, even with automatic backups off (#16363) — thanks @maxmad64bis
  • fix(sse): a Codex encrypted-reasoning replay rejection that arrives on a 200-OK SSE stream is no longer masked as a generic 502 STREAM_EARLY_EOF (#16364) — the readiness gate surfaces the same stable 400 invalid_encrypted_content the non-streaming path returns, and the deterministic rejection no longer buys a same-connection retry of a payload that will fail identically. (#16375) — thanks @skygunner
  • fix(sse): an LMArena response at status 400 or above now sets that sink's upstream status instead of looking like a transport failure (#16367) — thanks @maxmad64bis
  • fix(notion-web): a Notion web response at status 400 or above now sets that sink's upstream status instead of looking like a transport failure (#16371) — thanks @maxmad64bis
  • fix(logs): keep writing structured logs to the current file after rotation (#16373) — thanks @maxmad64bis
  • perf(db): opening or saving storage settings no longer scans the whole database; the stored integrity result is shown with its date and a manual check runs it on demand (#16374) — thanks @maxmad64bis
  • fix(sse): skip the emergency fallback redirect when its target is missing from the live catalog (#16376) — thanks @maxmad64bis
  • fix(stream-recovery): a stream cut after engagement with nothing rendered is now re-sent once on the same stream instead of surfacing a 502 (#16377) — thanks @maxmad64bis (diagnosis in #11902 — thanks @Xxx91n)
  • fix(api): job-preset video generations now write one request-log row and one usage row each, on success and on failure, so a paid generation can be matched to a provider bill and a provider that keeps failing is visible. The video route binds the request-scoped API key, so both rows are attributed to the calling key instead of landing with api_key_id NULL, and the usage writer falls back to that context as the request-log writer already did. The OpenAI video logger's endpoint/durationMs field names, which the writers ignore, are corrected to path/duration (#16378) — thanks @LuisAlejandro
  • fix(proxy): Prefer a selector member no sibling selector serves when switching (#16379) — thanks @maxmad64bis
  • fix(resilience): a 401 whose body attributes the failure to the model (Model <id> is not supported — a delisted model on a healthy key) now locks only that model for the not-found cooldown instead of pushing the connection toward a terminal state, so auto-combo candidate filters stop re-selecting the dead member while sibling models keep flowing (#16380); mirrors the #15633 404 model_not_found treatment. (#16388) — thanks @skygunner
  • fix(translator): filterToOpenAIFormat now rebuilds each known-type content part from a per-type field allowlist instead of forwarding unknown fields, so client metadata riding on a text/image/audio part (e.g. a compaction message_id) no longer reaches strict OpenAI-compatible validators — Cerebras 400 property … is unsupported / Mistral 422 extra_forbidden — and no longer burns every strict target in a combo (#16384). cache_control keeps its #2069 preservation contract; the Moonshot video_url extension stays opt-in. (#16395) — thanks @skygunner
  • fix(logs): Bound the CLI logs route to a tail read so large log files no longer stall served requests (#16386) — thanks @maxmad64bis
  • fix(opencode): Park requests with every egress set aside for quota until the earliest release instead of returning 429 (#16389) — thanks @maxmad64bis
  • fix(call-logs, proxy-logs): a call log row or proxy log batch blocked by a locked database is retried once after a short delay instead of being dropped; retries retain the original database handle and cleared proxy batches stay cleared (#16390) — thanks @maxmad64bis
  • fix(sse): keep the events read before a cut stream in its failure record (#16391) — thanks @maxmad64bis
  • fix(api): Skip usage-counter reads and monthly cost-ledger aggregation for API-key quota dimensions without configured limits. Keep configured quotas enforced and dashboard usage totals unchanged. (#16392) — thanks @woodsonl
  • fix(sse): /v1/images/generations and /edits now return n images for providers whose upstream returns one image per call (Black Forest Labs, Stability AI, Leonardo, Ideogram, Haiper, Hyperbolic, Magnific/Freepik, NVIDIA NIM, Cloudflare Workers AI), using bounded, all-or-nothing calls with seed offsets; BFL no longer forwards an undocumented num_images (#16399) — thanks @yourspraveen
  • fix(memory): with PII_RESPONSE_SANITIZATION on, memory now extracts facts from the sanitized response text instead of the raw model output, on both streaming and non-streaming paths (#16401) — thanks @yourspraveen
  • fix(security): close authentication bypass via uppercase path variants. The authz proxy matcher's custom regex groups (e.g. /:v1seg([vV]1)/:path*) do not reliably match uppercase aliases in production, so requests like POST /V1/chat/completions skipped the proxy and reached the dispatch layer with no API key (the route handlers performed no authentication of their own). The critical inference handlers (chat/completions, completions, messages, responses, responses/[...path]) now enforce enforceClientApiRouteAuth directly as defense-in-depth. (#16403) — thanks @foxy1402
  • Keep Anthropic 404 errors naming the requested model scoped to that model, preserving sibling models on the same connection.
  • Preserve the explicit Anthropic provider for dotted Opus 4.5 requests and normalize its effort variants to dashed wire IDs while retaining the bare legacy alias. (#16811)
  • fix(routing): require explicit opt-in for internal fallback across model versions, preserve same-version effort recovery and report the dispatched model after fallback; keep API-key model restrictions effective (Refs #16406). (#16810) — thanks @turbolego
  • fix(sse): combo hops back to OmniRoute's own listener skip public admission pressure again in production builds. The self-hop token (#13593) was a module-level value, so the copy of selfHop.ts that stamps outgoing loopback requests (the global fetch patch, installed from the instrumentation bundle) and the copy that verifies them (the chat route bundles) each generated a different random token and the bypass never matched. The token now lives on globalThis, like the self-loop secret since #14740. (#16410) — thanks @yourspraveen
  • Merge both guardrail disable headers with key, body, and metadata selections instead of dropping the legacy header; prompt-injection block mode remains mandatory (#16412, #16413). (#16805)
  • fix(monitoring): dashboard polling reuses the health snapshot for 30 seconds to reduce repeated event-loop aggregation. (#16426) — thanks @lorenzozanee
  • fix(auth): keep inactive API-key groups restrictive, retain their denies, and accept grants only from active groups (#16416). (#16808)
  • fix(codex): close client-facing stream immediately upon forwarding authoritative Responses terminal events on HTTP transport and cover all terminal types on WebSocket transport, preventing streaming hangs after tool calls (#16427) — thanks @Junior-HJ
  • Preserve Arena ELO last-sync freshness across restarts by reading the persisted source timestamp in authenticated sync status.
  • fix(openai): send explicit non-reasoning mode for GPT-6 Luna function tools on Chat Completions, preserving Responses reasoning (#16440).
  • fix(resilience): a Gemini 429 whose retry window lives only in the structured google.rpc.RetryInfo detail (error.details[].retryDelay) is now parsed by parseUpstreamError and relayed on the rebuilt error response's Retry-After header, so the model lockout honors the real upstream window (e.g. 1104s) instead of the 3s base cooldown while the per-connection limiter pauses for the full delay (#16442). (#16528) — thanks @skygunner
  • fix(memory): reindex all pending memory vectors in bounded sequential batches without overlapping workers, and retry loading the static Potion embedding model after a transient failure instead of caching a rejected promise until restart (refs #16444); write cached Potion downloads atomically and discard corrupt tokenizer/weights files for safe retries (#16888) — thanks @fabioKoruja00
  • fix(executors): Normalize Grok Build's forced native web-search choice when search is the only offered tool, avoiding ModelToolChoice rejection while preserving mixed/custom tool choices (#16447) — thanks @DavyMassoneto
  • fix(sse): Preserve forced search tools through Claude-to-Cursor translation and replay the completed server-owned search loop as Anthropic SSE, so streaming clients receive the final answer (#16448) — thanks @DavyMassoneto
  • fix(chat): the two local pre-dispatch 429s, TOKEN_LIMIT_EXCEEDED and GEMINI_TPM_EXHAUSTED, now carry X-OmniRoute-Provider and X-OmniRoute-Model like every routed response, so a client can tell which provider and model was throttled; status, body and error codes are unchanged (#16456) — thanks @pacocartones
  • fix(resilience): Claude's 429 "…monthly spend limit" (extra-usage cap reached) is now classified as a long-window quota and parks the connection for the subscription cooldown instead of being retried on every request (#16462) — thanks @tiagolabsit
  • fix(dashboard): the Utilization page now plots the tightest quota window of each bucket (one point per provider or connection) instead of an arbitrary window, so an exhausted weekly window is no longer hidden behind a half-full session window (#16463) — thanks @tiagolabsit
  • fix(claude): the "block extra usage" policy now also parks a Claude connection when extra usage is enabled and a plan window (5h session or weekly) is exhausted, for accounts whose usage API never reports extraUsage.queued; it stays blocked until the latest exhausted window resets (#16464) — thanks @tiagolabsit
  • fix(i18n): the pt-BR Claude connection pill no longer reads "Bloquear Extra ON" when extra usage is allowed; the label now names the state ("Uso extra") like the other locales (#16465) — thanks @tiagolabsit
  • fix(translator): Gemini/Antigravity tool schemas no longer carry empty-string enum values (which Gemini rejects with enum[N]: cannot be empty); the value is dropped, and the enum with it when it was the only one, so one such MCP tool no longer fails every Gemini leg of a combo (#16466) — thanks @tiagolabsit
  • fix(claude): the server-side-fallback-2026-06-01 and server-side-fallback-2026-07-01 betas that Claude Code sends with its fallbacks body parameter are now forwarded, so the request no longer fails with 400 fallbacks: Extra inputs are not permitted (#16467) — thanks @tiagolabsit
  • fix(sse): flatten root oneOf/anyOf/allOf tool schemas into one object schema and keep tool parameters whose names match JSON-Schema keywords when sanitizing Kiro tool schemas (#16469) — thanks @ryanngit
  • fix(settings): Stop offering weighted as an account-level fallback strategy: it had no account implementation and silently ran fill-first, and choosing it as the combo default could rewrite the global account strategy. Legacy saved weighted settings are still accepted and normalised to fill-first; combo-level weighted is unchanged (closes #16459) (#16473) — thanks @xiaoyaner0201
  • fix(dashboard): Distribute Proxies now asks for confirmation before replacing every account assignment in the group, a provider's account list gains a bulk "Remove account proxies" action that clears individual assignments (accounts fall back to provider/global proxy settings), and account proxy badges are clickable to open that account's proxy settings (#16474) — thanks @JxnLexn
  • fix(claude): preserve signed Claude thinking across OpenAI/Responses tool turns — the exact signed blocks are kept server-side per tool call and restored before the owning tool_use (same provider + model only; placeholder fallback and signature recovery kept) (#16475) — thanks @ryanngit
  • fix(usage): Claude OAuth quota polling keeps the last known quota on 429, honors Retry-After and coalesces concurrent polls (#16480) — thanks @ryanngit
  • fix(auth): accept a valid x-api-key even when an unrelated Authorization bearer is present (e.g. Claude Code with a claude.ai session); identity and API-key policy bind to the credential that validated (#16483) — thanks @omar-nahhas
  • fix(logging): request capture now masks credential query parameters (e.g. ?key=) in logged provider URLs and the live pending-request detail, masks *-secret/*-password headers, and writes call-log artifacts and log archives with 0600/0700 permissions (#16484) — thanks @ryanngit / @omar-nahhas
  • fix(dashboard): offer every canonical Codex Fast Tier model in the settings checklist while preserving saved selections and custom models (#16486). (#16694)
  • fix(sse): normalize tool-call IDs longer than 64 characters before sending to OpenAI, keeping assistant calls and tool results linked (#16490) — thanks @mwpn
  • fix(combos): resolve client-namespaced combo names (<namespace>/<combo>) to the combo when the namespace is not a known provider; explicit provider/model pins still reach the provider (#16493) — thanks @Jordannst
  • fix(translator): split inline <think> reasoning into Claude thinking blocks on OpenAI→Claude streams for textual-reasoning routes (DeepSeek-R1 mirrors, QwQ, K3, MiniMax-M3), instead of leaking the raw markup into the text block (#16494) — thanks @sirwhy
  • fix(kiro): retry once with a corrective instruction when Kiro emits a malformed tool_call wrapper before any output, instead of failing the turn; gated by the KIRO_TOOL_CALL_REPAIR_ENABLED flag (on by default) (#16495) — thanks @SemonCat
  • fix(kiro): route Kiro API-key connections to the Amazon Q regional endpoint and auto-detect the key's region at import (#16497) — thanks @ryanngit
  • fix(docker): the Docker/standalone launcher (scripts/dev/run-standalone.mjs) now applies the #5172 RAM-scaled V8 heap default when OMNIROUTE_MEMORY_MB is unset or invalid, instead of a fixed 512 MB; an explicit OMNIROUTE_MEMORY_MB and an existing NODE_OPTIONS heap flag still take precedence (#16498) — thanks @omnitrixAG
  • fix(translator): Claude clients now get refusal when an OpenAI-format upstream stops on content_filter or a Gemini upstream blocks on safety, recitation or a blocklist, instead of a clean end_turn; a truncated non-streaming completion now gets max_tokens instead of the raw length, as the streaming path already did. (#16501) — thanks @pacocartones
  • fix(i18n): document the opt-in CodeBuddy system-prompt length cap, default identity-only replacement, and replacement-reason logging in all 66 environment-reference translations. Extend the mirror contract tests to prevent missing variable rows and stale source/target hashes (follow-up to #16510; thanks @ntdatt812 and @Azzt17). (#16504) — thanks @pacocartones / @maxmad64bis / @skygunner
  • docs: translate the remote token privilege ceiling and stored-settings/legacy-import local-only bypass protections in all 66 locale mirrors. Refresh only the corresponding translation records after validating each source and target preimage, and add regressions requiring localized policy literals and matching content hashes. (#16504) — thanks @pacocartones / @maxmad64bis / @skygunner
  • Translate the Vision Bridge API-key scope, authenticated self-loop enforcement, and no-permitted-model fallback into all 66 documentation locales, with content-bound translation state and regression coverage (#16504). — thanks @pacocartones / @maxmad64bis / @skygunner
  • fix(sse): a completed stream whose client disconnects now hands the inner reader to the bounded terminal drain, so the SSE heartbeat interval is disposed without further downstream reads and delayed usage from a completed tool handoff is no longer dropped (follow-up to #16269, fixes #16258) (#16505) — thanks @xiaoyaner0201
  • fix(translator): a non-streaming Responses upstream that ends incomplete on max_output_tokens or content_filter now reaches a Chat Completions client with finish_reason length or content_filter instead of a clean stop, as the streaming path already did (#15867). (#16507) — thanks @pacocartones
  • fix(sse): CodeBuddy (CN and International) keeps long system prompts instead of swapping them for a neutral one; CODEBUDDY_SYSTEM_PROMPT_MAX_LEN opts back into a length cap, each replacement logs which rule fired, and a kept prompt that trips the content filter is retried once with the neutral prompt (#16510) — thanks @ntdatt812 / @Azzt17
  • fix(translator): normalize tool-call arguments to valid JSON before upstream dispatch — missing/null/empty arguments become "{}" and malformed replayed fragments no longer trigger strict-upstream 400s (#16511) — thanks @dungartoriaaa / @vianhanif
  • fix(antigravity): raise the IDE fingerprint fallback to 2.5.5 and resolve the live version before quota fetches so newer-model quota buckets are returned (#16512) — thanks @FerryAr
  • fix(executors): strip per-message fields an upstream 400 names as unsupported (Groq reasoning_content, pydantic extra_forbidden) and retry once; Groq requests now drop per-message reasoning_content pre-send (#16513) — thanks @rusak47
  • fix(server): requests that offer a cleartext Upgrade: h2c (the JDK HttpClient used by JetBrains IDEs / Air, among others) are now answered in HTTP/1.1 instead of having the connection dropped with "HTTP/1.1 header parser received no bytes"; WebSocket upgrades and TLS connections are untouched (#16514) — thanks @tiagolabsit
  • fix(responses): a non-streaming /v1/responses client served by a Chat Completions upstream now gets status: "incomplete" with incomplete_details when the upstream stops on length or content_filter, instead of completed, as the streaming path already did (#14806). (#16517) — thanks @pacocartones
  • fix(sse): Ollama-compatible /api/chat stream now ends with a single done:true carrying tool calls, no longer drops an unterminated final line, and keeps tool calls whose arguments are malformed JSON (#16519) — thanks @ntdat812
  • fix(sse): /v1/responses no longer drops the last upstream event when the stream ends without a trailing blank line (#16520) — thanks @ntdat812
  • fix(sse): record partial token usage (provider-reported, or estimated from the streamed output and flagged as estimated) for streams the client aborts or that fail mid-answer, instead of zero; status 499/success semantics unchanged (#16521) — thanks @brimob-sowax / @ntdat812 / @ntdatt812
  • fix(memory): the static memory embedder (potion-base-8M) reads its vocab from tokenizer.json instead of requesting a vocab.json that the model repo does not ship, so it can load from an empty cache dir (refs #16444) (#16522) — thanks @developerjillur
  • fix(cli-tools): settings writers (codex, kilo, cline, letta, smelt, pi, crush, openclaw, droid, jcode, Continue, Hermes, Oh My Pi) refuse to overwrite a config file that exists but cannot be read or parsed (409 naming the file) instead of silently replacing it; Oh My Pi also backs up models.yml/config.yml before every write and can assign its modelRoles to OmniRoute models without touching other providers' roles (#16523) — thanks @ntdat812 / @krzemienski
  • fix(sse): drop duplicate finish chunks in passthrough streams and fail over on empty completions carrying an upstream native_finish_reason error (#16524) — thanks @askidmobile
  • fix(tests): preserve proxy environment variables during setup and restoration on Windows, preventing six false TLS proxy-context failures (#16526). (#16697)
  • fix(backup): use relative tar archive paths for Windows Git Bash compatibility and clean partial exports after failures or timeouts (#16527). (#16696)
  • fix(sse): honor Google RPC ErrorInfo quotaResetDelay/quotaResetTimeStamp as an upstream retry hint and accept compound durations (151h58m18s) in retry-delay parsing (#16529) — thanks @fasilu
  • fix(providers): Groq catalog no longer advertises decommissioned models (llama-4-scout, llama-3.3-70b-versatile, qwen3-32b); the retired Llama ids stay reasoning-denied for keys that can still reach them (#16530) — thanks @andikadevs
  • fix(opencode): alias over-length tool names (>64 chars) for OpenCode Zen/Go and restore them in responses (#16531) — thanks @MakdumIbrohim
  • fix(providers): surface the Antigravity "verify your account" link as a Verify button on the connection row, cleared after the next success (#16532) — thanks @sasazemzulin058-debug
  • fix(translator): coerce non-object tool-schema shapes (array properties, primitive property values, shorthand string sub-schemas, tuple items) for Gemini/Antigravity and Codex instead of failing every account with a 400 (#16537) — thanks @AnandkumarMall / @coozgan / @bayhnf
  • fix(translator): id-less tool results are paired with their pending call instead of dropped (Chat → Claude and Responses → Chat), orphan outputs from stateless Responses continuations are kept as labelled context, and an empty tool_calls: [] no longer bypasses OpenAI-format normalization (#16539) — thanks @ntdatt812 / @claytontavaresdan / @vianhanif
  • fix(sse): generated images (Gemini inlineData → delta.images) are no longer dropped from OpenAI-format streams (#16540) — thanks @ntdatt812
  • fix(opencode): route the Grok family (grok-4.5 + effort aliases, new grok-4.6 on Go; grok-4.6 and grok-build-0.1 on Zen) to the Responses endpoint instead of /chat/completions (#16541) — thanks @thienpvt / @dqphong0302
  • fix(api): report a provider with no connected account as provider_not_configured instead of invalid_api_key (single-model 401) or model_not_found (combo 404); HTTP status and type are unchanged, so clients that keyed re-auth on invalid_api_key no longer do for this case (#16542) — thanks @haumanto
  • fix(claude): drop the Claude Code diagnostics telemetry field that Anthropic rejects with 400 "Extra inputs are not permitted" (#16543) — thanks @arcslash
  • fix(translator): a non-streaming Claude response now maps refusal, model_context_window_exceeded, stop_sequence and pause_turn the same way the streaming path does, so a Gemini client sees a refusal as SAFETY and a context-window stop as MAX_TOKENS instead of a clean STOP. (#16544) — thanks @pacocartones
  • fix(domain): combo health and combo forecast skip quota windows the upstream did not report a fraction for (fractionReported: false, e.g. Antigravity's chat_* windows) instead of reading their stored placeholder as 0% remaining, which showed healthy providers at 0% and raised target_low_quota (#16546) — thanks @AndresDiagoM
  • fix(dashboard): report complete combo health target metrics instead of dropping unreported windows (#16549) — thanks @jacobsparts.
  • fix(resilience): Unpinned combo targets now find model lockouts recorded under the concrete connection that served a previous attempt, so short transient locks are waited out and redispatched instead of immediately returning ALL_TARGETS_SKIPPED (#16551). (#16582) — thanks @skygunner
  • fix(monitoring): reuse bounded call-log statistics snapshots across health matrix polls and cache light database diagnoses separately from full integrity scans, preserving live connection state and reset/repair invalidation. (#16552) (#16770)
  • fix(catalog): a combo's max_output_tokens now honours a member's max_output_tokens capability override, as the per-model entry already did (#16556) — thanks @pacocartones
  • fix(logging): retain bounded, sanitized upstream error bodies in call-log details even when detailed pipeline logging is disabled (#16558). (#16758)
  • fix(api): /v1/images/generations now bills the images the provider actually returned instead of max(n, returned), so a provider that delivers fewer images than requested is no longer charged for the missing ones (follow-up to #16398) (#16565) — thanks @yourspraveen
  • fix(antigravity): keep a client-declared google_search callable alongside other function tools in Antigravity (Cloud Code) requests instead of promoting it to native grounding, which the v1internal endpoint rejects with HTTP 400; a sole search tool still gets native grounding. Plain Gemini (AI Studio / Gemini CLI) behavior is unchanged. (#16567) — thanks @Neuron-Mr-White
  • fix(pricing): Antigravity and Antigravity CLI connections price every model they serve (Gemini 4 Argon, Gemini 3.8 Flash, Claude 5.5, Flash-Lite, and the existing Pro/Claude 4.6/GPT-OSS rows that were only under the legacy ag key), so a single call no longer fail-closes a per-key USD quota (#16574) — thanks @xiaoyaner0201
  • Correct the context and output limits for twelve verified OpenCode Go models so imported models no longer fall back to a 200K context window. Preserve the narrower GPT-6 Luna and Space Bunny input limits, operator overrides, synced metadata, and documented request formats.
  • fix(providers): Codex gpt-image-2.5-sunburst is registered as a Codex image model, so reference-image edits reach the dedicated Codex Images edits route instead of returning 400, and Sunburst appears in the image catalog next to Flare (#16583) (#16585) — thanks @xiaoyaner0201
  • fix(providers): expiry-first ranks accounts on the reset of their weekly (or monthly) window instead of the nearest 5h session reset, so subscription pools spend the account whose week ends first (#16586) — thanks @feci
  • fix(usage): apply cached-input pricing to translated streaming usage in the spend ledger and shared-quota USD recorder, preserving nested cache-write and reasoning details — fixes #16590 (#16591) — thanks @xiaoyaner0201
  • fix(ci): scan process.env references with Node so env-doc-sync works on Windows (#16592) — thanks @shipsfromrio.
  • Authenticate internal Chaos, Evals, and Issue Agent chat calls when API keys are required, preserving explicitly selected keys and keeping synthetic tasks out of user memory (#16593). (#16804)
  • fix(providers): OpenCode zen free-tier sessions are minted with the client's descending Identifier encoding so x-opencode-session stays inside the accepted counter window instead of 403 FreeTierError (#16598) — thanks @MRQC-st. Refs #16328.
  • fix(oauth): recover the credential-health cache after a successful persisted manual connection test without clearing independent quota cooldowns, and add the default-off CREDENTIAL_GATE_AUTH_DIAGNOSES_ONLY flag so only authentication diagnoses (not timeouts, quota or outages) block routing when enabled (#16599) — thanks @KiaroSama
  • fix(providers): Preserve Claude text following tool results in Gemini function responses, avoiding mixed turns rejected by Vertex Gemini Flash (#16443 (#16601) — thanks @christianjgilman). — thanks @lorenzozanee
  • fix(translator): route gemini-4-* (including vertex/gemini-4-argon) through the thinkingLevel path instead of sending a numeric thinkingBudget, and support unnamespaced canonical spec lookup for provider-qualified model IDs (#16603) (#16665) — thanks @SkiiOS
  • fix(streaming): drain upstream finalization within the disconnect grace period when clients close after a terminal event, preserving usage and call-log completion (#16604). (#16766)
  • fix(translator): an unclosed <invoke> / <tool_call> tag (or a held DSML prefix) in the middle of an OpenAI-compatible stream was swallowed when the stream ended, so a reply that merely mentioned such a tag lost everything from that point to the end. The held tail is now released as text at finish (#16606) — thanks @pacocartones.
  • fix(devin-cli): a client disconnecting mid-stream no longer crashes the server. The executor kept calling controller.enqueue()/close() from the child's close handler after the response body was cancelled, which throws ERR_INVALID_STATE inside an event listener (an uncaught exception). The stream now defines cancel(), which stops the child and silences further writes (#16607) — thanks @pacocartones.
  • fix(rerank): a null entry in documents made the NVIDIA and DeepInfra request builders throw a TypeError (they read doc.text; the Voyage builder already used doc?.text), which the injection guard turned into a 500 after the provider credentials and quota had been consumed. A null entry now becomes an empty passage, as in the Voyage branch (#16609) — thanks @pacocartones.
  • fix(modelDiscovery): upstream OpenAI-compatible /models rows that advertise snake_case context_window, max_context_window_tokens, context_window_tokens, or a boolean vision flag now map onto inputTokenLimit / supportsVision (#16610) — thanks @cmbergamo.
  • fix(context): Layer-3 purify_history used to drop the lone user turn when it kept only the newest tool tail, so vLLM/Qwen rejected the request with HTTP 400 "No user query found in messages." The newest user message now survives purification (#16613) — thanks @pompousman.
  • fix(sse): some Cloudflare Workers AI models stream a digit as a JSON number ("delta": {"content": 7}). Only string content counted as text in the passthrough, so those chunks were dropped and the digits vanished from the reply («ABC123» came out as «ABC»). A finite numeric delta.content is now turned into its string before the chunk is forwarded, recorded or counted (#16614) — thanks @pacocartones. Refs #16554.
  • fix(chat): a null (or non-object) entry in tools[], or a null entry in input[], made generateSessionId and the OpenAI→Claude tool conversion throw a TypeError (Cannot read properties of null (reading 'name')) and the request ended as a 500. They are now answered with a 400 and a clear message (tools: Expected array of objects, input: Expected array of objects or strings), in the same place and style as the messages entry guard (#12643, #13755). Plain string items in input stay valid (#16615) — thanks @pacocartones.
  • fix(translator): keep Claude document (PDF) blocks when translating to OpenAI-hub providers such as Antigravity, Gemini and OpenAI-compatible targets (#16617) — thanks @Shurelia
  • fix(antigravity): normalize Oh My Pi harness markers in system instructions (Claude and Gemini branches) and drop the used_claude* telemetry keys to avoid false 429 quota errors (#16618) — thanks @ersaasyhar
  • fix(i18n): 24 UI strings of the Spanish locale were still English (the quota-share time windows, the paid Codex credits row, the Codex cooldown label, the request timeline labels); they are now translated with the wording of the neighbouring Spanish keys (#16620) — thanks @pacocartones.
  • fix(oauth): parse numeric-epoch expiresAt (seconds or ms, number or string) on the request-time refresh path, so the proactive refresh fires, epoch-seconds values no longer refresh on every request, and a still-valid rotated DB token is reused (#16621) — thanks @elielsousa-pathbit
  • fix(mitm): revert the hosts entries, or flag Repair, when the AgentBridge/MITM child process exits unexpectedly after a successful start (#16622) — thanks @khanghack222
  • fix(codex-images): allow up to 16 GPT Image edit references on the dedicated Codex Images route while retaining the existing eight-reference bound for the hosted Responses image tool (#16623) — thanks @Bl0ck154
  • fix(providers): Google PSE validation now rejects an invalid API key or Search Engine ID (cx) (HTTP 400) instead of reporting it as valid; quota-exhausted 429 stays valid (#16625) — thanks @Jordannst
  • fix(sse): keep output_config.format (structured output) for the official Anthropic API-key provider in Claude passthrough (#16626) — thanks @raflyazf
  • fix(dashboard): reset the Add API key modal between opens so a saved or cancelled credential no longer reappears in the next add (#16628) — thanks @hamedniroomand / @semihisikman
  • fix(usage): Cost estimates now apply long-context pricing tiers (OpenAI >272K input tokens, xAI Grok 4.6 ≥200K) and use corrected GPT-5.4 and Claude Haiku 4.5 cache rates (#16629) — thanks @Primexz
  • fix(i18n): 61 UI strings of the Russian locale were still English (the quota-share time windows, the Codex cooldown and paid credits, the proxy subscription core errors, the per-model concurrency caps, the request timeline labels); they are now translated with the wording of the neighbouring ru keys (#16630) — thanks @pacocartones.
  • fix(i18n): 61 UI strings of the Japanese locale were still English (the quota-share time windows, the Codex cooldown and paid credits, the proxy subscription core errors, the per-model concurrency caps, the request timeline labels); they are now translated with the wording of the neighbouring ja keys (#16633) — thanks @pacocartones.
  • Keep the video retention HTTP fixture isolated from raw-socket egress during application imports and resource cleanup, including mixed IPC/TCP socket options, in addition to the installed Undici dispatcher guard (#16634).
  • fix(i18n): 61 UI strings of the Korean locale were still English (the quota-share time windows, the Codex cooldown and paid credits, the proxy subscription core errors, the per-model concurrency caps, the request timeline labels); they are now translated with the wording and glossary of the neighbouring ko keys (#16636) — thanks @pacocartones.
  • sse: Cloudflare Workers AI no longer rejects tool-call-only assistant turns with HTTP 400 — missing or null content is now sent as an empty string (#16637) — thanks @zasasamar2129
  • dashboard: make Quota Tracker row hide/show independent per account instead of per provider, keeping legacy provider-level preferences as a fallback (#16638) — thanks @ShiDai567
  • fix(i18n): UI strings of the Traditional Chinese locale were still English (the quota-share time windows, the Codex cooldown and paid credits, the proxy subscription core errors, the per-model concurrency caps, the request timeline labels); they are now translated with the wording of the neighbouring zh-TW keys (#16639) — thanks @pacocartones.
  • fix(usage): preserve OpenRouter's provider-reported usage.cost for OpenAI-format clients and use it as the exact cost in usage accounting (#16640) — thanks @AlexanderMakarov
  • fix(i18n): UI strings of the Vietnamese locale were still English (the quota-share time windows, the Codex cooldown and paid credits, the proxy subscription core errors, the per-model concurrency caps, the request timeline labels); they are now translated with the wording of the neighbouring vi keys (#16641) — thanks @pacocartones.
  • sse: Gemini translators keep functionResponse parts last when merging a tool result with a following user turn, fixing the Gemini 4 400 "Requests ending with a model turn are not supported" (#16642) — thanks @daniadasaya
  • fix(mitm): AgentBridge cert download, status, trust and untrust now use the active certificate. Root-CA installs get ca.crt instead of the legacy server.crt, with a filename for each mode and a not-found message that matches the mode (#16643)
  • sse: move bare property-map tool parameters under properties so Gemini no longer 400s with "Unknown name … Cannot find field" (#16644) — thanks @Deepleakk
  • fix(responses): map provider-injected tool-call prefixes (functions.exec) back to the declared tool name instead of forwarding an undeclared name the Responses client rejects (#16645) — thanks @vianhanif
  • sse: drop the reasoning alias once reasoning_content carries the replay, fixing GLM/Kimi 400 "duplicate field reasoning_content" (#16646) — thanks @likongkong666
  • fix(i18n): UI strings of the Turkish locale were still English (the quota-share time windows, the Codex cooldown and paid credits, the proxy subscription core errors, the per-model concurrency caps, the request timeline labels); they are now translated with the wording of the neighbouring tr keys (#16647) — thanks @pacocartones.
  • fix(sse): a combo falling back from a native Claude OAuth step no longer hands the next provider a body already rewritten for Claude Code (tool names renamed to Read, billing/sentinel in system[], metadata, output_config, cache_control), so OpenCode clients stop getting tool calls for tools they do not have (#16648) — thanks @AndresDiagoM
  • fix(i18n): UI strings of the Polish locale were still English (the quota-share time windows, the Codex cooldown and paid credits, the proxy subscription core errors, the per-model concurrency caps, the request timeline labels); they are now translated with the wording of the neighbouring pl keys (#16649) — thanks @pacocartones.
  • fix(i18n): UI strings of the Dutch locale were still English (the quota-share time windows, the Codex cooldown and paid credits, the proxy subscription core errors, the per-model concurrency caps, the request timeline labels); they are now translated with the wording of the neighbouring nl keys (#16650) — thanks @pacocartones.
  • fix(i18n): UI strings of the Indonesian locale were still English (the quota-share time windows, the Codex cooldown and paid credits, the proxy subscription core errors, the per-model concurrency caps, the request timeline labels); they are now translated with the wording of the neighbouring id keys (#16651) — thanks @pacocartones.
  • fix(i18n): UI strings of the Ukrainian locale were still English (the quota-share time windows, the Codex cooldown and paid credits, the proxy subscription core errors, the per-model concurrency caps, the request timeline labels); they are now translated with the wording of the neighbouring uk-UA keys (#16656) — thanks @pacocartones.
  • fix(i18n): UI strings of the Thai locale were still English (the quota-share time windows, the Codex cooldown and paid credits, the proxy subscription core errors, the per-model concurrency caps, the request timeline labels); they are now translated with the wording of the neighbouring th keys (#16657) — thanks @pacocartones.
  • fix(i18n): UI strings of the Hindi locale were still English (the quota-share time windows, the Codex cooldown and paid credits, the proxy subscription core errors, the per-model concurrency caps, the request timeline labels); they are now translated with the wording of the neighbouring hi keys (#16658) — thanks @pacocartones.
  • fix(i18n): UI strings of the Arabic locale were still English (the quota-share time windows, the Codex cooldown and paid credits, the proxy subscription core errors, the per-model concurrency caps, the request timeline labels); they are now translated with the wording of the neighbouring ar keys (#16659) — thanks @pacocartones.
  • fix(api): the list routes of the dashboard and the management API (/api/batches, /api/files, /api/logs/detail, /api/cache/entries, /api/webhooks, /api/v1/management/proxies and /api/v1/management/proxies/assignments) parsed ?limit=, ?offset= and ?page= with a bare Number/Number.parseInt; a value such as abc reached the SQLite LIMIT ? bind as NaN (datatype mismatch) and the request ended as a 500 or threw out of the handler. They now answer 400 <name> must be an integer through one helper, as /v1/batches and /v1/files already did for limit; out-of-range integers are still clamped as before (#16661) — thanks @pacocartones.
  • fix(chat): a JSON body that is valid JSON but not an object (null, a string, a number or true) made handleChat throw (Cannot use 'in' operator to search for 'messages' in x, Cannot convert undefined or null to object) and the request ended as a 500 on /v1/chat/completions, /v1/messages and /v1/completions. It is now answered with a 400 Request body must be a JSON object, the message /v1/responses already used (#16668) — thanks @pacocartones.
  • fix(api): a null JSON body made POST /api/logs/detail, PATCH and PUT /api/settings and DELETE /api/providers throw (Cannot read properties of null) and the request ended as a 500, and POST /api/logs/detail answered a string, a number or true body with a 200 that switched detailed logs off; they now answer 400 for a body that is not an object (POST /api/logs/detail also for malformed JSON, which was an unhandled rejection). Input validation, not a crash fix: other requests and the process are unaffected either way. (#16671) — thanks @pacocartones.
  • fix(providers): reassemble fragmented tool-call arguments in the Devin Desktop executor (#16674) — thanks @wusaki0723
  • fix(resilience): park GLM/Z.AI "insufficient balance" 429s (code 1113, 余额不足) and new-api insufficient_user_quota / "out of funds" 400s as credits_exhausted, so the pool rotates instead of retrying a key with no credit (#16675) — thanks @IEatCodeDaily / @brimob-sowax
  • fix(security): exporting stored credentials now needs the instance operator. GET /api/settings/export-json, GET /api/db-backups/export and GET /api/db-backups/exportAll accept a signed-in dashboard session, the local CLI, or an instance with login disabled; API keys of any scope (including manage) and oma_ access tokens get 403. The check is the same one that guards sync token issuance and admin CLI token minting (requireOperatorCredential() in src/lib/api/operatorCredential.ts) (#16680).
  • fix(usage): redeeming a Codex or Grok Build reset credit now releases the account's local quota cooldown and model lockouts, so it re-enters rotation immediately (#16682) — thanks @JunDev76
  • fix(dashboard): connection reorder (▲/▼ on tied priorities and "Reorder by availability") now writes valid 1-based priorities and reports failures instead of silently saving the wrong order (#16684) — thanks @StevanusPangau
  • fix(providers): align xiaomi-mimo-token-plan discovery-class metadata — declaring modelsUrl moves getDiscoveryClass from static-only to openai-compat and exposes the URL on the registry entry. Live /v1/models refresh already ships via #15308 (#16686) — thanks @AbdullahPesteli
  • Correct the model-refresh warning for providers that intentionally use a local catalog: explain that discovery is not configured instead of reporting an unavailable upstream API.
  • fix(translator): Claude → Gemini/Antigravity pairs each tool_result with its own tool_use when an id is reused across turns, ending a persistent 400 INVALID_ARGUMENT in long sessions (#16698) — thanks @WYSTAF
  • fix(translator): keep Claude turns whose only block is container_upload or another unrecognized type instead of silently dropping them — forwarded verbatim to Claude targets, a type-only notice on OpenAI-format targets (#16699) — thanks @semihisikman
  • fix(codex): fall back to the next combo model when Codex streams "model unavailable / not offered on this account" in a 200 SSE — surfaced as a model-scoped 400 (combo advances, model locked on that connection) instead of a failed 200 turn (#16700) — thanks @mrnim94
  • fix(sse): Gemini responses without totalTokenCount now report total_tokens as prompt + completion instead of 0 (non-streaming, SSE accumulator and streaming paths) (#16702) — thanks @WYSTAF
  • fix(providers): Inception Labs keys are now validated with a 1-token chat probe. Its /v1/models is public and returns 200 for any key, so the setup dialog and connection Test used to accept any string. mercury-2.5 is also added to the model seed. (#16703) — thanks @defyma
  • fix(translator): function tools keep their source-format strict semantics on every Responses route — an omitted or null Chat/Claude strict now reaches Responses as false instead of being strict-normalized, Responses → Chat no longer leaks strict: null, and Claude → Chat carries explicit boolean strict (#16704) — thanks @juniorgaudencio00-code / @semihisikman
  • fix(resilience): Gemini 429s that report a daily (PerDay) quota violation now lock the model until the daily reset (still bounded by the model-lockout cap) instead of re-trying the exhausted key every ~20s (#16705) — thanks @scottchristian
  • fix(translator): content-filtered turns reach Claude clients as stop_reason "refusal" (stream and non-stream, including Gemini SAFETY/RECITATION/BLOCKLIST/PROHIBITED_CONTENT), and non-stream "length" maps to "max_tokens" (#16708) — thanks @huohua-dev
  • fix(usage): keep reasoning_tokens when OpenAI/Responses upstreams are translated to Claude format, and read nested output_tokens_details in usage accounting (#16710) — thanks @raflyazf / @rifuki
  • fix(translator): OpenAI→Claude now honors an explicit max_tokens / max_completion_tokens cap even when tools are present, so 1-token prompt-cache warm requests no longer generate 32k output tokens; small explicit caps (e.g. 4096) with tools are now hard limits (#16711) — thanks @kvyb
  • fix(mitm): /api/settings/mitm (cert download and certExists) and the AgentBridge repair step now use the active certificate, so repair no longer leaves the root CA trusted. "Regenerate cert" (the cert/regenerate endpoint, the dashboard action and settings/mitm) never changes the cert model any more: legacy installs get a new leaf, and root-CA installs get a rotated CA. The old CA is untrusted first, rotation is refused while the bridge runs, and the response says the new CA must be trusted (#16712)
  • fix(translator): OpenAI→Claude streaming no longer merges two tool calls that reuse an upstream index, keeps tool calls that arrive without an id or with arguments before the name, and ends a stream that stops without a finish reason with an explicit error event instead of leaving the client hanging; Read accepts path/filePath, Claude tool_result errors stay visible to OpenAI-format models, and opencode raises tiny max_output_tokens budgets (Muse Spark Responses ≥ 8000, others ≥ 16) (#16713) — thanks @andyvandaric / @Reinazhard / @semihisikman
  • fix(translator): keep the developer role for OpenAI targets that accept it (openai, azure-openai, azure) instead of always lowering it to system, so the per-model preserveOpenAIDeveloperRole toggle takes effect; custom openai-compatible-* nodes and github keep system unless opted in (#16715) — thanks @semihisikman
  • fix(copilot): route not-yet-synced Responses-only GPT models (gpt-6-*, gpt-5.5+) to /responses instead of /chat/completions (#16717) — thanks @amirsalaar
  • Preserve a retryable response with Retry-After when a combo peer is temporarily unavailable at the credential gate and another target returns a payment failure (#16718). This also preserves retry hints for active pinned connections with persisted cooldowns, excluding terminal and inactive accounts. (#16776)
  • fix(pricing): price parenthesized Codex effort ids (e.g. gpt-6-luna(max)) at the tier/base rate instead of $0 (#16719) — thanks @raflyazf
  • fix(translator): expose Claude Code deferred MCP tools (tool_addition) to non-Anthropic providers instead of a 400 or a silently dropped tool (#16720) — thanks @semihisikman
  • fix(mitm): bin/antigravity-bridge.mjs now finds its certificate the way the AgentBridge server does. It used to read only server.crt/server.key from CERT_DIR or $HOME/.omniroute/mitm, so it ignored DATA_DIR and exited with "Certificate files not found" on root-CA installs (ca.crt/ca.key, the default model). It now reads <DATA_DIR>/mitm (with CERT_DIR still an explicit override), keeps the legacy leaf where one is active, signs a per-host leaf from the root CA otherwise (respecting MITM_ROOT_CA_ENABLED), and exits at startup with a mode-specific message when the files are missing (#16724).
  • fix(github): strip unsupported temperature for Copilot GPT-5.4+ and GPT-6 models (gpt-5.5, gpt-5.6-*, gpt-6-astra), which otherwise returned HTTP 400 (#16725) — thanks @m00nie
  • fix(grok-cli): keep hosted web_search/x_search tools and the function tool_choice forces when capping Grok Build requests at its 200-tool limit (#16728) — thanks @ryanngit
  • fix(translator): Gemini/Antigravity schemas keep literal unions typed (0|1|2|null → bounded integer, string unions keep every value) and reject runaway $ref expansion with a clean 400 that skips account cooldown (#16729) — thanks @amirho1
  • fix(usage): Ollama Cloud quota now reads the documented /api/balance with the API key — monthly credit allowance, legacy session/weekly windows and purchased USD credits — keeping the settings-cookie scraper as fallback. Incomplete monthly balances remain unknown, API balances survive past reset timestamps, and positive purchased credits prevent included-only quota blocks without clearing real upstream 429 cooldowns. Confirmed purchased-credit readings reuse the existing quota cache (#16732) — thanks @x-adam
  • fix(proxy): relay proxy test waits up to 30s so free-plan Vercel / Deno / Cloudflare relays are not reported dead during a cold start (#16733) — thanks @erikyao0125
  • fix(codebuddy): normalize empty finish_reason / function_call shells in CodeBuddy SSE and drop empty tools: [] (#16735) — thanks @asuan-dev
  • fix(models): combo context/output limits in /v1/models ignore seats whose provider has no active connection (#16738) — thanks @amirsalaar
  • fix(kiro): preserve the upstream output-token limit in live model discovery (#16742) — thanks @ryanngit
  • fix(api): treat empty streaming finish_reason strings as null so clients do not see a blank reason (#16744) — thanks @alltomatos.
  • fix(usage): stop bounded log/quota previews from retaining whole request bodies — truncateLogString (open-sse/utils/requestLogger.ts) and truncatePreviewString (src/lib/usage/usageHistory/helpers.ts) built their result from String#slice(), which V8 can represent as a SlicedString whose parent is the entire multi-megabyte source. Any retained 64 KB / 1200-char preview therefore kept the full request body alive. Both now flatten the result (Buffer round-trip), matching the structuredClone guard the completed-details path already uses. A cut that would keep a trailing high surrogate is shortened by one code unit first, so the round-trip does not replace a split pair with U+FFFD. (#16752) — thanks @LazyGatto
  • Fixed: Non-sync Obsidian MCP tools reject callers without a resolved API-key identity before accessing vault credentials; authenticated global fallback and sync-token behavior remain supported. (#16757) (#16778)
  • fix(errors): mask house access tokens in public error text (#16759) — thanks @maxmad64bis
  • fix(mcp): dynamic skill tools now run as the authenticated key principal instead of the client-chosen session id (#16760) — thanks @maxmad64bis
  • Avoid warning about volatile package configuration when the CLI only loads unchanged generated defaults. Continue warning for effective customized values and unknown keys without exposing their contents.
  • fix(tls): cap non-streaming TLS responses at 10 MB instead of buffering without limit (#16769) — thanks @maxmad64bis
  • fix(usage): route usage readings through the assigned connection proxy (#16773) — thanks @maxmad64bis
  • fix(plugin): freeOnly now publishes the free-tier models actually served (alias and canonical keys share the provider free status and budget), keeps single-use credits and retired models visible only without the preset, and labels uncapped budgets (uncapped) instead of an empty suffix (#16775) — thanks @maxmad64bis
  • fix(plugin): publish preset-passing entries and member-passing combos (#16779) — thanks @maxmad64bis
  • Correct free-model classification for Bytez paid 70B/72B tiers and unverified Nous Hermes-4 entries; retain Bytez Mistral-7B without an unsupported token allowance and refresh the documented catalog counts. Closes #16781. (#16791)
  • fix(providers): keep the operator-disable marker authoritative even when isActive is true (#16784) — thanks @jacobsparts.
  • fix(plugin): plugin warnings now reach the host log, each refresh logs a stage summary, and the freeOnly warning counts unmatched free-tier lines (#16785) — thanks @maxmad64bis
  • fix(streaming): mask credentials in streaming responses under the same opt-in (#16786) — thanks @maxmad64bis
  • Preserve explicit vision capabilities for named combo members without registry or synced metadata, keeping the advertised vision and derived modalities available for confirmed vision combos while leaving unknown members conservative.
  • fix(arena-elo-sync): serve the persisted dataset date on the authenticated sync status after a restart (#16793) — thanks @maxmad64bis
  • fix(vision): Keep image payloads raw for confirmed all-vision combos with compatible-node steps; reconcile a node's public routing prefix and provider-owned slash-shaped IDs without treating foreign nodes or explicit non-vision verdicts as vision-capable. Covers OpenAI/Anthropic-compatible nodes and nested/mixed combo regressions; self-loop authentication remains governed by #14740 (refs #16801) (#16937) — thanks @fabioKoruja00
  • fix(vision): Behavior note — a compatible-node combo step is now judged only from that node's own evidence (its custom-model override, model-compat override or synced model row); a compatible node without any such row is treated as non-vision, so more of those steps take the Vision Bridge describe path instead of receiving raw images. Add a vision override or sync the node's models to keep images raw (refs #16801) (#16937) — thanks @fabioKoruja00
  • fix(auto-combo): stop per-candidate registry scans and DB reads on the auto/* request path (#16814) (#16933)
  • fix(copilot): skip unnecessary refresh of healthy Copilot sub-tokens, while still probing GitHub /user so a verified 401 demotes the access token (#16818) — thanks @gsmarcil.
  • fix(sse): Guard against uncaught upstream BodyTimeoutError responses without an error code, preventing avoidable server restarts (#16819) — thanks @fabioKoruja00
  • fix(dashboard): the Codex auth.json import dialog now accepts the top-level credential layout ({ access_token, refresh_token, id_token, … }) that POST /api/oauth/codex/import already imports, instead of rejecting it with "The file is not a valid Codex auth.json" (#16821) — thanks @suhanrain
  • fix(sse): Buffered Antigravity/Gemini SSE responses now preserve upstream cache-read tokens as prompt_tokens_details.cached_tokens for non-streaming clients (#16829) — thanks @xiaoyaner0201
  • fix(db): keep the SQLite credential store, its WAL sidecars and every DB backup owner-only (0600 files, 0700 backup dirs), repairing world-readable files from older installs (#16840) — thanks @elvis460
  • fix(usage): show Cloudflare Workers AI daily neuron usage (GraphQL Analytics, vs the 10k/day free allocation) in Provider Limits (#16841) — thanks @askidmobile
  • fix(command-code): fail the CLI-fallback request on an embedded server error before any content so combo fallback triggers (#16852) — thanks @jefripunza
  • fix(providers): Restore Ollama Cloud live model discovery via the documented /api/tags endpoint and normalize tag names as routable model IDs (#16856) — thanks @fabioKoruja00
  • fix(resilience): Scope ModelScope's invalid_model 401/403 denials to the rejected model, preserving healthy sibling routes and combo failover (#16859) — thanks @fabioKoruja00
  • fix(providers): Correct OpenCode Go context and output limits for both DeepSeek Flash aliases and MiniMax M2.5, replacing stale imported placeholders with upstream-verified limits while honoring manual overrides; cover all confirmed Go model windows with regression tests (#16860) — thanks @fabioKoruja00
  • fix(providers): kimi-web and deepseek-web connection tests now time out after 15s instead of hanging on a stalled upstream (#16862) — thanks @ntdat812
  • fix(security): Encrypt provider management keys and Qwen/Alibaba/Volcengine/Xiaomi console cookies and tokens plus Ollama Cloud usage cookies and nested OAuth, API-key and session-cookie metadata inside providerSpecificData at rest; keep lazy/eager reads, legacy migrations and cookie-session deduplication working; prefer actual browser cookies over shared host metadata during session deduplication to avoid merging separate accounts; redact dashboard-sensitive fields from compliance audit details (#16863) — thanks @fabioKoruja00
  • fix(providers): register DeepSeek V4 Flash Vision (exp) on OpenCode Go and Command Code with explicit vision support (#16865) — thanks @Spoon94
  • fix(sse): strip malformed regex pattern constraints from tool schemas before OpenRouter dispatch (#16867) — thanks @nitsuah
  • fix(providers): Reject Chat Completions and direct Responses image inputs to Cloudflare Playground's text-only bridge with a clear 400 before launching a browser, and report missing Playwright/Chromium runtimes as actionable host errors instead of generic upstream failures (#16868); sanitize unexpected browser startup errors before returning HTTP responses — thanks @fabioKoruja00
  • fix(dashboard): the console log stops dragging the reader back to the newest line while they read older entries; it follows the tail only while scrolled to the bottom (#16891) — thanks @ntdatt812
  • fix(codex): quota auto-ping only counts a ping as sent after a completed Responses event; failed, incomplete or unfinished 200 streams now enter the failure cooldown (#16892) — thanks @anndev-69
  • fix(dashboard): single source for the provider-limit quota colour thresholds; drop the dead helpers that used a contradictory 70/30 policy (#16893) — thanks @ntdatt812
  • fix(video): list the Vertex Veo 3.1 / 3.1 Fast GA model ids (veo-3.1-generate-001, veo-3.1-fast-generate-001) in the video catalog (#16894) — thanks @coozgan
  • fix(compression): skip output-style prompts (less-code / caveman) on requests that carry images, audio or files, so they no longer degrade vision answers (#16906) — thanks @raflyazf
  • fix(providers): DeepSeek API Flash models (deepseek-flash, deepseek-v4-flash) now advertise vision support, so image requests are no longer filtered away from them (#16908) — thanks @GimingRao
  • fix(translator): honor the terminal response.done status on Responses→Chat streams — failed/cancelled now surface as upstream errors (combo fallback fires), incomplete keeps its length/content_filter finish reason, and its usage is extracted (#16909) — thanks @amirsalaar
  • fix(resilience): concurrency-limit 429s ("too many concurrent requests") now fail over to the next connection with a short fixed cooldown instead of escalating a healthy connection's backoff (#16911) — thanks @PIKACHUIM
  • fix(usage): report Command Code cache-write tokens (inputTokenDetails.cacheWriteTokens) as cache creation in usage and cost (#16914) — thanks @Spoon94
  • fix(translator): qualify a namespaced Responses tool_choice (and namespaced allowed_tools entries) to the flattened Chat tool name instead of the bare leaf (#16917) — thanks @JessenReinhart
  • fix(combo): an unknown pre-dispatch skip reason is logged and the request keeps routing instead of failing (#16925) — thanks @maxmad64bis
  • fix(zcode): the ZCode executor now runs through the native generateWorkspaceText model call, keeping the caller's tools, image inputs and usage, and bounding maxOutputTokens into the model's native range on every turn. Image input on a stock ZCode worker needs the strictly opt-in helper scripts/ops/zcode-image-runtime.mjs, which is never run automatically, only writes a SHA-pinned patched copy to a separate directory and never edits the installed runtime (see docs/ops/ZCODE_IMAGES.md) (#16926) — thanks @megamen32
  • fix(sse): opt-in STREAM_LATE_OUTPUT_PROBE_ENABLED (default off): a stream whose visible output starts after the first 64 forwarded chunks is reported as having emitted output (one chunk in 16 keeps being probed), so a lone 5xx after that point no longer records a model-only lockout that benched the model for every other client; three such failures in a row still lock it, and firstOutputMs is back-filled to the first output chunk of the sampling gap (#16927) — thanks @maxmad64bis
  • fix(providers): Gemini TTS (*-tts), transcription (*-transcribe) and gemini-2.5-flash-image requests no longer carry a thinkingConfig, which these models reject with 400 Thinking is not enabled for this model on every path, including reasoning_effort: "none" (fixes #16928) (#16931) — thanks @yourspraveen
  • fix(quota): bespoke quota providers now refetch usage after an upstream 429 instead of serving a cached snapshot for up to a minute (#16932) — thanks @maxmad64bis
  • fix(docs): Clarify that the Ollama Cloud balance URL belongs to the upstream service, not an OmniRoute API route, preserving the strict documentation-accuracy check after #16732 (#16936) — thanks @diegosouzapw
  • fix(providers): Count imported Web Cookie sessions (authType: cookie) on the matching provider cards and section totals, while keeping cookie credentials excluded from unrelated API-key and OAuth cards (refs #16943) (#16951) — thanks @fabioKoruja00
  • fix(translator): keep a client-sent service_tier through the Anthropic → OpenAI pivot for OpenAI, Azure OpenAI and Codex targets, so the billed tier and the served tier no longer diverge (#16946) — thanks @raflyazf
  • fix(resilience): Permit longer per-process quota-fetch spacing (up to 60s) and make its additional jitter configurable with OMNIROUTE_QUOTA_FETCH_JITTER_MS, keeping the existing defaults and bounded validation. This reduces synchronized bursts; it cannot guarantee prevention of upstream token revocation (refs #16948) (#16952) — thanks @fabioKoruja00
  • test(cursor): lock the KV blob ack request_metadata echo with round-trip tests (#16949) — thanks @mannnrachman
  • fix(sse): the two stream-deadline paths (withDeadlineSignal, createStreamDeadlineSignal) merge the client and deadline signals with the self-cleaning mergeAbortSignals instead of AbortSignal.any, whose composites Node pins in memory while they carry an abort listener — a completed streaming request no longer stays reachable with its wire body; a deadline expiry still emits the route error frame (#16974) — thanks @oooohhoo
  • fix(kimi): Kimi Coding effort-suffixed ids (kimi-coding/kimi-k3-high, kmc/kimi-for-coding-low) are dispatched as the bare wire model (k3) with the suffix as reasoning_effort instead of failing with upstream 401 "model id does not exist"; scoped to kimi-coding/kimi-coding-apikey, Cursor's genuine kimi-k3-high ids are untouched (#16985) — thanks @castocolina
  • fix(api): POST /api/chaos/run and POST /api/skills/collect/chaos answered 500 when the request body was not JSON; both now answer 400 with Invalid JSON body (#16991) — thanks @pacocartones
  • fix(api): POST /api/skills/marketplace/install and POST /api/skills/skillssh/install answered 500 when the request body was not JSON; both now answer 400 with Invalid JSON body (#16992) — thanks @pacocartones
  • fix(quota): opt-in QUOTA_WINDOW_BOUNDED_PARK lets the per-window quota policy honour the bounded exhaustion park, so a passed park deadline or an observation older than the park cap stops excluding an account until a fresh quota fetch; off by default (#17000) — thanks @copystring
  • fix(api): POST /api/combos/duplicate no longer returns raw error.message in the 500 body (#17002) — thanks @trespassmk
  • fix(claude): API-key Claude requests that ask for thinking.display: "summarized" (e.g. @ai-sdk/anthropic / OpenCode) now get readable thinking text: OmniRoute no longer adds the redact-thinking-2026-02-12 beta on its own for those requests, which made Anthropic return empty, signature-only thinking blocks (#16407). OMNIROUTE_REDACT_THINKING=false disables the OmniRoute-added redact beta entirely; client-negotiated redact betas and CC-compatible requestDefaults.redactThinking opt-ins are unchanged. Claude OAuth requests keep the beta by default (it is part of the Claude Code OAuth fingerprint); the new default-off CLAUDE_OAUTH_SUMMARIZED_THINKING_UNREDACTED feature flag extends the behavior to OAuth. (#17003) (#17040) — thanks @huynle-dev
  • fix(qoder): timeout or abort now kills the whole qodercli process group (no orphaned ~780 MB CLI processes), and an opt-in process-wide FIFO cap (QODER_CLI_MAX_CONCURRENCY, unset = no cap; QODER_CLI_QUEUE_TIMEOUT_MS → 429 qodercli_busy + Retry-After) lets a small host bound how many qoder/* runs spawn at once; an experimental opt-in persistent ACP transport (QODER_CLI_TRANSPORT=acp) is also available (#17004) — thanks @mad-gooze
  • fix(gamification): POST /api/gamification/rotate, POST /api/gamification/servers and POST /api/gamification/federation/score called request.json() outside a try, so a body that is not valid JSON threw an unhandled SyntaxError instead of answering 400; they now answer 400 Invalid JSON body, as the other routes do. (#16730) — thanks @pacocartones
  • fix(rate-limit): idle limiter eviction now emits one aggregated line per watchdog pass instead of one line per limiter, with per-key detail on the debug channel (#PRNUM) (#15560) — thanks @maxmad64bis
  • fix(sse): key resource-pressure runtime singleton on globalThis to survive duplicated bundle copies (#PRNUM) (#16144) — thanks @trespassmk
  • test: harden messages-route memory profiler against orphaned heap snapshots (#PRNUM) (#16144) — thanks @trespassmk
  • fix(usage): Antigravity family weekly buckets gemini_weekly and claude_gpt_weekly are kept by the provider-limits sanitizer, so /api/usage/provider-limits and the quota card can show the weekly window next to the 5-hour window instead of dropping those keys as unknown model ids (#15359) — thanks @edosulai
  • fix(antigravity): diagnose upstream failures from a bounded allowlist projection instead of the raw Google error body — call logs keep the validation classification (status, provider code/status/reason, category, field, schema keyword) while the persisted response body and its headers stay free of provider text that can echo request content (the client-facing envelope keeps surfacing the real upstream message, #13591) (#13311) — thanks @iamzidar
  • fix(routing): a connection's defaultModel (resolveBareModelToConnectionDefault, added in #4825) now only replaces a placeholder bare model (auto or empty); a concrete bare name such as claude-haiku-4-5 is no longer silently swapped for another provider's default. BARE_MODEL_DEFAULT_OVERRIDE_LEGACY=true restores the old behaviour. (#16550) — thanks @shipsfromrio
  • fix(combo): do not hop a request-scoped streaming refusal (invalid request / context overflow) onto a family-sibling effort alias of the same model; the combo advances to the next target instead (regression from #13603) (#15903)
  • Fix the bundled MCP server deadlocking on startup (antigravity executor ↔ executeAttempt async init cycle from #13739), keep the media and provider-detail client bundles free of the DB-backed feature-flag module (#15423 regression) while still honouring the GROK_SUBSCRIPTION_IMAGES_ENABLED dashboard override, and restore the canonical 提供者 rendering for "provider" in the new zh-CN/zh-TW strings. (#15902)
  • fix(db): the first better-sqlite3 open no longer blocks for seconds on hosts with slow reverse DNS — the addon's musl probe calls process.report.getReport(), whose network section reverse-resolves every open TCP handle; the driver now constructs it with process.report.excludeNetwork on and restores the caller's value (~6.5s → ~0.6s first capability lookup under vitest, #15106) (#16092)
  • fix(i18n): zh-CN codexKeyServiceMode uses the glossary term 提供者 instead of 提供商, and the route-namespace map is regenerated to carry the new namespace (#13554) (#16092)
  • fix(translator): coerce the pending Responses tool-call id to a string before queueing it, restoring the API route typecheck gate that #16539 broke (openai-responses.ts TS2345) (#16624)
  • dashboard: declare rateLimitAutoProtected on ConnectionRowConnection so the auto-protected rate-limit badge from #b16bf50f type-checks (base-red #15306). (#16496)
  • Preserve whether cache-creation tokens are included in prompt totals across Chat, Responses, and Claude translations, preventing negative round-trip input counts and downstream double counting while retaining legacy unmarked usage behavior. (#16100) — thanks @Kizuno18
  • Strip a trailing text-only assistant turn before official Claude OAuth dispatch. Claude returns 400 This model does not support assistant message prefill for that shape; the shared strip only covered Mistral.
  • fix(codebuddy-intl): send the X-Conversation-* header family on CodeBuddy International chat requests to pass upstream channel verification and avoid HTTP 400 code 11128 (#16922) — thanks @Junior-HJ
  • fix(codex): account selection honors each connection's synced model inventory, so models available only to one subscription are not sent to other accounts during selection or fallback; reasoning-effort aliases (gpt-6-sol-high, -max, -ultra, ...) are matched through their base model id (#15133) — thanks @JxnLexn
  • fix(codex): treat the live (or cached) Codex model inventory as authoritative — static registry entries now only enrich ids the account already lists, or serve as an explicit offline fallback; a failed refresh prefers the account's last cached catalog over the GitHub manifest (#15132). User-visible change: the static effort-suffix aliases (-low/-medium/-high/-xhigh/-max/-ultra) and the retired codex-auto-review id (#14903) are no longer listed by /v1/models or the dashboard model picker unless the account's inventory advertises them. (#15132) — thanks @JxnLexn
  • fix(api): persist comboAutoPromoteEnabled through PATCH /api/settings by adding it to updateSettingsSchema (Zod previously stripped the key with HTTP 200, so the sticky combo promote flag could not be turned off via API once enabled; same class of bug as #13562 / #13800) (#15237) — thanks @hussmanne
  • fix(combos): Gemini combo probes send reasoning_effort: none so thinking does not eat the health-check budget; truncated finish_reason: length responses are no longer rewritten as empty-content 502s
  • Keep duplicated auto-combo snapshot weights at a total of 100 for large or highly skewed candidate pools, preserving every candidate with a positive fractional weight when integer weights exceed the budget. (#16942)
  • fix(sse): honor an explicit combo target timeout when it is longer than the upstream request timeout (#16309) — thanks @jacobsparts
  • fix(compression): terminate idle compression worker threads on eviction. finish() scheduled remove(slot, false), so the idle timer only dropped the pool's reference and the worker_threads thread, its MessagePort and its private heap (~100 MB each) survived for the whole process lifetime — every successful chat request on the published 3.8.50 npm package leaked one more thread, invisible to heapUsed, until the host hit its memory cap. Ported from release/v3.8.51 as an npm hotfix on top of 3.8.50 (#13091, #14933, #11804) (#14988) — thanks @anhtahaylove / @webzidev
  • Docs: sync release counts (370 providers, 202 migrations); regenerate PROVIDER_REFERENCE.md. (#15929)
  • Docs: sync release counts (364 providers, 200 migrations); regenerate PROVIDER_REFERENCE.md. (#15873)
  • Docs: sync release counts (367 providers, 200 migrations); regenerate PROVIDER_REFERENCE.md. (#15879)
  • Docs: sync release counts — 360 providers (onomeo #14297) and 200 migrations; regenerate PROVIDER_REFERENCE.md. (#15866)
  • Docs: sync the provider count to 363 after the d7 provider merges (Unifically #14182, BeatAPI #14875, ChatPlayground #12690); regenerate PROVIDER_REFERENCE.md. (#15870)
  • Keep CursorExecutor.buildHeaders synchronous (and expose getCursorAgentCliVersionSync) so the executor honors the BaseExecutor contract again; the async override from the CLI-version refresh broke the types and would hand a Promise to callers that do not await. (#15731)
  • fix(dashboard): The feature-flag cards for OUTBOUND_SSRF_GUARD_ENABLED, OMNIROUTE_ALLOW_PRIVATE_PROVIDER_URLS, OMNIROUTE_ALLOW_LOCAL_PROVIDER_URLS and RERANK_REMOTE_PROVIDER_NODES still described the pre-correction guard semantics (local URLs blocked by default, metadata hosts never routed to). The dashboard renders the raw English description from featureFlagDefinitions.ts, so the four descriptions now carry the same corrected semantics as docs/reference/FEATURE_FLAGS.md (PR #15616), pinned by unit tests (#15617) — thanks @woodsonl
  • fix(gemini): send Gemini 3.8 thinkingLevel instead of a numeric thinkingBudget, and omit includeThoughts unless the client asked, so hidden thoughts stop eating maxOutputTokens
  • Preserve the active application log during legacy-log archival when its configured path is directly inside DATA_DIR/logs or nested more deeply; unrelated legacy siblings are still archived and removed. (#16816)
  • fix(server): enable TCP keepalive on main server connections to detect half-open sockets and avoid hanging streaming responses during silent network drops. (#16396) — thanks @Junior-HJ
  • fix(api): Prioritize chat and responses endpoints in detectTestKind and classifyModelSupportedEndpoints so multimodal chat models are not probed as audio transcription models or classified as specialty media surfaces. Previously, if a custom model had audio-transcriptions or media endpoints among its supportedEndpoints, detectTestKind checked audio transcription first and dispatched a test probe to /v1/audio/transcriptions, which rejected non-transcription providers with Invalid transcription model: <model>. Use format: provider/model. Chat capability now takes precedence during test detection and endpoint classification. (#15082) — thanks @lepkov
  • fix(sse): Fix Native Codex long-session OOM paths: release Antigravity abort bindings when SSE streams finish, measure stream backpressure in bytes for every provider, shed new requests on runaway process RSS (OMNIROUTE_RSS_PRESSURE_MB), and isolate heuristic ultra compression in a worker bounded by a heap cap that now applies to all worker modes (OMNI_COMPRESSION_WORKER_MAX_OLD_MB), so a pathological large /v1/responses history fails open instead of driving the gateway into host OOM. (#15277) — thanks @mdigitalbh81
  • fix(auth): the OIDC callback no longer signs anyone in when oidcAllowedSubjects is empty, and the settings route rejects any update that would leave OIDC enabled with an empty allowlist, not only the one that enables it (#15049) — thanks @HouMinXi
  • fix(opencode): the synthesized OpenCode User-Agent now auto-refreshes from the npm registry (opencode-ai/latest, 6h cache, pinned fallback offline) instead of rotting on a hardcoded pin. (#14199) — thanks @DangTAnh
  • fix(providers): OpenCode Go keeps a third-party coding agent's own User-Agent (e.g. my-coding-agent/1.0) instead of rewriting it to opencode/<version>, as Go's client requirements ask; generic SDK / HTTP-library UAs are still replaced and the Zen free-tier policy is unchanged (#15311) (#15453) — thanks @yourspraveen
  • fix(opencode-plugin): models served through generic openai-compatible-chat-* adapters (e.g. ih/glm-5.3) are no longer labelled with an unrelated provider in the OpenCode picker — the bare-model-id enrichment fallback now keeps only the model name unless the matched entry's provider owns the row, so a foreign provider's label, free-tier budget and pricing are not carried over (v1 and v2 plugins) (#14966) (#15465) — thanks @yourspraveen
  • OpenCode plugin v2: toolsOnly defaults to false again, so the full catalog is published; set toolsOnly: true to hide models without tool calling (reverts the default introduced by #14554). (#15811)
  • Unmount the provider audio-badge test's React roots and await teardown before removing the DOM, preventing scheduler work from accessing a disposed jsdom window during parallel UI tests. (#16815)
  • Fix POST /api/providers throwing a ReferenceError (HTTP 500) when allowNoCredential is true: the route called providerAllowsOptionalApiKey without importing it. (#15721)
  • fix(authz): stop loopback-locking /api/providers/{validate,import,bulk} — the devin spawn is call-site gated. A reverse-proxied OmniRoute dashboard (docker + nginx) could never validate ANY provider key: POST /api/providers/validate returned 403 LOCAL_ONLY for every request arriving through the proxy hop, and the Add API Key "Check" button rendered that as a bare red "Invalid" badge — the object-shaped {code:"LOCAL_ONLY",message} error body was silently swallowed by the modal (it only rendered string errors). The blanket LOCAL_ONLY pattern came from #15242 (audit #15159 S-01) to cover the only spawn reachable from these routes — the devin CLI-key fallback validator (webProvidersB.ts, fixed binary + argv). #15766 later established the precise pattern for exactly this class (gate the spawn at its call site on the trusted peer-locality header, the /api/providers/[id]/models cursor precedent) and threaded allowLocalSpawn into the test/create/test-batch routes, but the blunt regex stayed and kept every remote dashboard locked out. Now: the pattern is dropped; validate/bulk/import thread allowLocalSpawn: getRequestPeerLocality(request) !== "remote" into validateProviderApiKey (#15766 shape) so a remote peer still cannot reach the devin CLI-spawn fallback; and validationErrorMessage() renders {code,message} envelopes so a pipeline rejection can never masquerade as a key rejection again. Discovered live on a docker/nginx v3.8.52 deployment: a TokenRouter key that returns 200 OK from api.tokenrouter.com/v1/models showed as "Invalid" through the proxy and validated fine once the gate moved. (#16366) — thanks @woodsonl
  • ui: cancel deferred canvas initialization fits on teardown and invalidate the disposed React Flow generation, preventing callbacks into unmounted graphs while preserving active and replacement graph fitting. (#16689)
  • tests: unmount CompressionCockpit React roots before removing their DOM containers, allowing lifecycle cleanup to run before the jsdom environment is disposed. (#16689)
  • ui: invalidate the previous graph immediately on key replacement, including the interval before the replacement's deferred initializer arrives. (#16689)
  • Dispose Qdrant save-notification timers when leaving the settings card, keep each saved notice visible for its own interval, and ignore save completions after unmount; unmount the UI test fixtures before discarding their DOM. (#16848)
  • fix(quality): prevent cancelled, failed, or incomplete complexity measurements from appearing as passing release-green metrics; retain advisory classification for complete, measured ratchet drift without changing baselines or CI workflows. (#16533)
  • Release-branch hygiene: document OMNIROUTE_API_KEY_COOLDOWN_MS, OMNIROUTE_ESTIMATOR_CALIBRATION and GROK_SUBSCRIPTION_IMAGES_ENABLED in .env.example + ENVIRONMENT.md, allowlist the Next-internal NEXT_MANUAL_SIG_HANDLE, move the orphan Cloudflare tunnel-error test into tests/unit/ so it actually runs, and register 10 covering tests in stryker.conf.json. (#15802)
  • Release-branch hygiene: the hard-lease bypass inventory now lists the Firecrawl Map route (#15703) and the priority-0 re-read in providers.ts (#15485); the #11759 catalog test pins CATALOG_BUILD_TIMEOUT_MS so a slow cold build on a loaded host no longer returns 503 mid-assertion. (#15827)
  • Release-branch hygiene: add the Y-API entry (#15131) to the provider translate-path golden snapshot. (#15847)
  • Release-branch hygiene: re-measure the reserved provider-prefix count (418) after ChatPlayground (#12690) and BeatAPI (#14875) both landed, and write the ChatPlayground upstream endpoints in its provider doc as full upstream URLs so check:fabricated-docs no longer reads them as local routes. (#15871)
  • Release-branch hygiene: reset the connection test modal state in its open handler instead of the loading effect, clearing the react-hooks/set-state-in-effect lint error introduced with #15138. (#15872)
  • Release-branch hygiene: classify the hard-lease bypass sites added by #13788 (alpha/search), #15761 (local embeddings allowlist) and #15138 (per-account test message). (#15874)
  • fix(providers): the Codex provider page and Test All list the reasoning-tier rows (<model>-<tier>) the account advertises again, as /v1/models already did. They had disappeared when Codex switched to the account's live inventory (#15132), which skipped the tier step added by #13224. Six unit tests on the release tip that still expected behavior from before recently merged PRs (#15754, #14976, #15736, #15378, #15830, #15164) were updated to match. (#15884)
  • Include a nested error object in post-keepalive Responses failures for LiteLLM-compatible clients, preserving status and retry metadata and displaying bounded, sanitized upstream details. (#16099) — thanks @Kizuno18
  • Restore readable ocx1: compaction summaries before routed translation and native Codex replay, validating canonical base64 and UTF-8 while preserving opaque native state and rejecting unsupported translated state. (#16101) — thanks @Kizuno18
  • fix(api): /v1/models no longer lists another provider's models to a restricted API key — an aggregator row whose root carries a vendor namespace (cline's deepseek/deepseek-v4-flash, openrouter/free) no longer matches an allowlist entry meant for that vendor, and a blocked id is no longer re-admitted through its root (#15409) (#15452) — thanks @yourspraveen
  • fix(translator): sanitize $ref and $-prefixed keys in Gemini/Antigravity function_response.response payloads. Google Cloud Code and Gemini endpoints parse $ref inside function_response.response as unresolvable OpenAPI references and fail with HTTP 400 INVALID_ARGUMENT. Keys are now renamed to _ref during translation, preventing session-bricking 400 loops when tools output JSON Schemas or OpenAPI specs. (#16153) — thanks @aliepratama
  • fix(providers): clamp SenseNova DeepSeek V4 Flash reasoning_effort xhigh/max to high (upstream lists xhigh then 400s it)
  • providers: add StepFun step-5-preview (1M context, vision + reasoning, verified against the live api.stepfun.ai catalog) and a modelsUrl for live model discovery; document the China vs international platform split (api.stepfun.com vs api.stepfun.ai keys are not interchangeable). (#15904) — thanks @hongkongkiwi
  • fix(providers): Vertex AI project-scoped execution now routes requests with region eu to the canonical EU residency endpoint aiplatform.eu.rep.googleapis.com instead of the global endpoint, allowing projects with constraints/gcp.restrictEndpointUsage organization policies to complete requests successfully (#15178) — thanks @fabioluissilva
  • Render the complete Video Bridge promotion security-fixture instruction on two lines so the adversarial text fits inside the frame; verify actual FFmpeg glyph widths before accepting the corpus. (#16509)
  • Preserve short scene events selected by experimental segment-aware video sampling through the final caption budget, and separate its corrected result-cache identity from the previous oversampled policy. (#16518)
  • Release-branch hygiene: freeze the file-size ceilings grown by #15766/#15757 and register two covering tests in stryker.conf.json after the d7 board-1 merge wave. (#15829)
  • Release-branch hygiene: freeze the file-size ceilings grown by #15132, #15756 and #14103 after the d7 board-2/3 merge waves. (#15843)
  • i18n: translate the two sidebar section show/hide labels added by #15508 into all 66 UI locales. (#15861)
  • fix(typecheck): remove duplicate KeyHealth export (#15769)
  • fix(ci): enforce exact-SHA release-green evidence and verdicts (#14376)
  • fix(quality): reject mismatched and contradictory acceptance evidence (#14380)
  • fix(build): require typed executable MITM artifacts (#14385)
  • fix(quality): supervise gate processes and retain failure evidence (#14388)
  • fix(ci): pin lockfile installer and align acceptance runtime (#14395)
  • fix(ci): isolate pending builds by candidate ref (#14398)
  • fix(ci): recognize Actions output path in environment contract (#14400)
  • fix: align queue skills and merge contracts with exact-candidate gates (#14432)
  • fix: fail deployment when the target cannot be reached (#14488)
  • fix(ci): fail closed on incomplete workflow audits (#14497)
  • fix(video): limit snapshot cue redaction to video carriers (#14507)
  • fix(providers): declare claude-opus-5 thinking efforts in claude registry (#14548) — thanks @HouMinXi
  • perf(dashboard): cut page payloads, keep /v1/models cache warm, add per-account quota panel (#14835) — thanks @aref-alapour
  • fix(sse): refuse traversal in the Codex /v1/responses subpath (#15090) — thanks @HouMinXi
  • fix(sse): scan Devin and PromptQL tool markup in linear time (#15099) — thanks @HouMinXi
  • fix(security): require admin for API-key routes and reserve login slots during password checks (#15146)
  • fix(cache): fold generation params into semantic cache signature (#15149) (#15156) — thanks @Laksopan23
  • fix(electron): resync electron/package-lock.json with its package.json (#15157)
  • fix(models): accept "responses" as a supported endpoint (#15158) — thanks @HouMinXi
  • fix(release): let an Electron dispatch build the ref it is dispatched on again (#15165)
  • fix(poe): add a text part to image-only user turns sent to Poe (#15179) — thanks @bbmy85552
  • fix(sse): make MAX_TOOLS_LIMIT configurable via env var (#15193) — thanks @70nyIT
  • fix(dashboard): recognize Codex Pro Standard, Extra and Max plan variants (#15161) (#15197) — thanks @alltomatos
  • fix(deps): bump transitive undici, brace-expansion and fast-uri to their patched releases (#15210)
  • fix(api): synthesize string response.id for /v1/responses frames (#15202) (#15213) — thanks @yugui923
  • fix(backend): sanitize messages in createErrorResponse / createErrorResponseFromUnknown (#15159 E-13) (#15215) — thanks @jonlwheat2-gif
  • fix(backend): sanitize upstream error text in deepseek-web errorResponse (#15159 E-09) (#15217) — thanks @jonlwheat2-gif
  • fix(backend): sanitize fetch error in blackbox-web errorResponse (#15159 E-01) (#15221) — thanks @jonlwheat2-gif
  • fix(backend): sanitize caught error in playground simulate-route (#15159 E-14) (#15222) — thanks @jonlwheat2-gif
  • fix(backend): sanitize upstream error text in image upscale handlers (#15159 E-15) (#15226) — thanks @jonlwheat2-gif
  • fix(ci): register 19 covering unit tests in stryker.conf.json tap.testFiles (#15227) — thanks @jonlwheat2-gif
  • fix(translator): pair unpaired Responses function_call items with synthesized outputs (#15216) (#15231) — thanks @skygunner
  • fix(backend): sanitize MCP advanced tools errors + G-11 gate for tool-result shapes (#15159 M-05/G-11) (#15233) — thanks @jonlwheat2-gif
  • fix(authz): loopback-gate the two provider routes that spawn child processes (#15159 S-01) (#15242) — thanks @jonlwheat2-gif
  • fix(plugin): redact credentials before the debug JSONL is written (#15159 O-05) (#15244) — thanks @jonlwheat2-gif
  • fix(build): alias Turbopack hashed externals to bare names in standone bundles (#15255) — thanks @for-the-zero
  • fix(combo): keep the caller's connection pin through the safety-net redirect (#15280) — thanks @HouMinXi
  • fix(ci): make check-cycles honest — real graph walk + cycles ratchet (#15159 G-01/G-02) (#15281) — thanks @jonlwheat2-gif
  • fix(ci): bundle-size ratchet no longer passes when it cannot measure (#15159 G-04) (#15282) — thanks @jonlwheat2-gif
  • fix(ci): prune 6 stale eslint suppressions that made npm run lint exit 2 (#15159 G-05) (#15283) — thanks @jonlwheat2-gif
  • fix(db): a cookie connection's name must not overwrite another account's session (#15159 B-01) (#15290) — thanks @jonlwheat2-gif
  • fix(dashboard): widen the sidebar search corpus beyond the visible label (#15159 B-02) (#15291) — thanks @jonlwheat2-gif
  • fix(cli): expose built-in auto combos to Claude Code discovery (#15301) — thanks @VihaanR
  • fix(providers): live model discovery for xiaomi-mimo / xiaomi-mimo-token-plan (#15308) — thanks @LazyGatto
  • fix(sse): reject premature upstream EOF in Responses translation (#15310) — thanks @fidelix
  • fix(security): never put an escape-folded ACP version probe on a shell command line (GHSA-jw7m-33xr-hmhj) (#15324)
  • fix(deps): next 16.3.6 (next/og RCE), axios 1.20, grpc-js 1.14.5, dompurify 3.4.16 and patched npm-bundled copies in the image (#15325)
  • fix(cli): spawn npm without DEP0190 on win32 (#15328) — thanks @Yi-111-a
  • fix(opencode): declare targetFormat openai-responses for opencode-go gpt-6-luna (#15333) — thanks @c4lyp5o
  • fix(sse): preserve Copilot discovered endpoint routing (#15337) — thanks @zebrajaeger
  • fix(quality): clear two base-reds on release/v3.8.52 (gate-manifest alias + stryker registration) (#15342)
  • fix(quality): release-green no longer runs pull_request-only ci.yml steps (#15344)
  • fix(dashboard): cancel the compression tab's saved-badge timer on unmount (#15350)
  • fix(mcp): read cost totals from the analytics summary (#15355) — thanks @HouMinXi
  • fix(ci): register the cycles ratchet script in the gate manifest (#15356) — thanks @HouMinXi
  • fix: prevent proxied stream crashes and normalize Groq reasoning (#15357) — thanks @kalpakprod
  • fix(providers): do not treat a Claude weekly warning as exhaustion (#15361) — thanks @HouMinXi
  • fix(providers): show a No Auth provider's real requirement on its banner (#15363) — thanks @HouMinXi
  • fix(backend): keep TLS fingerprinting when a request carries Next.js metadata (#15364) — thanks @HouMinXi
  • fix(dashboard): show the message inside a Cloudflare tunnel error object (#15365) — thanks @HouMinXi
  • fix(responses): stop marking native collaboration calls as plaintext (#15368) — thanks @HouMinXi
  • fix(dashboard): read OpenCode model limits from the catalog (#15406) — thanks @HouMinXi
  • fix(auth): revoke one access token by prefix and stop rewriting last use (#15407) — thanks @HouMinXi
  • fix(providers): pace Codex usage fetches that bypass the quota gate (#15408) — thanks @HouMinXi
  • fix(cli): send --allow-no-credential to the server (#15416) — thanks @HouMinXi
  • fix(auth): let a cache-scoped key reach the cache routes (#15430) — thanks @HouMinXi
  • fix(devin): mark legacy CLI transport as text-only (#15431) — thanks @mohammadraufzahed
  • fix(devin): surface CLI auth and include auto-routing (#15446) — thanks @mohammadraufzahed
  • fix(providers): update the pinned Copilot CLI version (#15476) — thanks @lorenzozanee
  • fix(providers): declare Claude target format for GHE Copilot models (#15478) — thanks @lorenzozanee
  • fix(api): route provider-model handlers through management auth (#15479) — thanks @lorenzozanee
  • fix(combo): include monthly quota in reset-aware scoring (#15480) — thanks @lorenzozanee
  • fix(routing): allow same-model failover after disabling a turn-bound account (#15481) — thanks @JxnLexn
  • fix(providers): omit thinkingBudget 0 on Flash-Lite models (#15491) (#15492) — thanks @claw-io
  • fix(pii): stop redacting code syntax as IPv6 (#15498) — thanks @branben
  • fix(executors): route GHE Copilot GPT-6 models to Responses (#15499) — thanks @lorenzozanee
  • fix(providers): load the ChatGPT web validator only when it is tested (#15516) — thanks @HouMinXi
  • fix(combo): skip hedging when the request body is large (#15535) — thanks @HouMinXi
  • fix(memory): bound synchronous lexical query work (#15549) — thanks @seanford
  • fix(sse): stabilize standalone compression worker resolution (#12797) (#15557) — thanks @insoln
  • fix(translator): resolve the Claude thinking output cap with the routed provider (#10139) (#15561) — thanks @Juriseagle
  • fix: honor Codex quota filter opt-out (#15574) — thanks @KiaroSama
  • fix(providers): append new models from GitHub catalog in enrichCodexModelsFromGithubCatalog (#15525) (#15578) — thanks @alltomatos
  • fix(sse): scope type-only context errors in combo streams (#15597) — thanks @insoln
  • fix(sse): accept first-party Claude empty end_turn in combo quality gate (#15598) — thanks @insoln
  • fix(db): insert call_logs row before writing its artifact (#15608) — thanks @insoln
  • fix(chatgpt-web): restore browser sessions and native Playground (#15615) — thanks @ElioBonfim
  • fix(antigravity): strip stream field from Google request envelope (#15582) (#15631) — thanks @alltomatos
  • fix(quota): register the tracker batch after its binding is ready (#15643) — thanks @HouMinXi
  • fix(deps): bump transitive deps to patched releases (Dependabot batch) (#15662)
  • fix(codex): refresh discovery client version (#15663) — thanks @zitus91
  • fix(combo): stop memory-pressure guard 503s from poisoning connection health (#15674) — thanks @HouMinXi
  • fix(skills): repoint deleted omniroute-* skill links to omni-* targets (#15676) — thanks @Poid-ZA
  • fix(sse): skip Trae upstream status on a 2xx with no body (#15686)
  • fix(security): never authorize an API key from the Redis auth cache alone (GHSA-66vh-35g3-78qv) (#15687)
  • fix(security): require the relayed OAuth callback to echo the active attempt's state (GHSA-3fxv-j4h8-9mgg) (#15688)
  • fix(models-dev): carry reasoning effort tiers through sync and discovery (#15701) — thanks @HouMinXi
  • fix: use Unicode-aware word boundaries for non-ASCII language detection (#15708) — thanks @dajiaohuang
  • fix(codex): honor the ultra alias set for parenthesized suffixes (#15710) — thanks @HouMinXi
  • fix(grok-web): use x1: prefix for x-statsig-id fallback (#15711) — thanks @fomathpro-arch
  • fix(changelog): start the providers-route fragment with a bullet (#15722)
  • fix(cursor): derive imported token TTL from JWT exp instead of hardcoded 24h (#15724) — thanks @patrykkopycinski
  • fix(cursor): fail over on empty-turn completions instead of fabricating clean 200s (#15725) — thanks @patrykkopycinski
  • fix(opencode-plugin-v2): restore the redis optional-peer entries dropped from the lockfile (#15727)
  • fix(antigravity): accept Gemini Flash tier ids on tiered-only accounts (#15659) (#15735)
  • fix(gamification): use canonical getRank for rank badge criterion (#15742) — thanks @SoloDrex52
  • fix(combo): hop fill-first on 403 balance without deactivating (#15745) — thanks @RaviTharuma
  • fix(opencode-plugin-v2): use advertised combo name as model id (#15755) — thanks @ebouther
  • fix(sse): trim oversized tool enums on opencode dispatch (#15756) — thanks @timootten
  • fix(mcp): stop echoing the caller identity into scope denials (#15159 S-04) (#15771) — thanks @jonlwheat2-gif
  • fix(mcp): sanitize the JSON-RPC transport error envelope (#15159 S-05) (#15773) — thanks @jonlwheat2-gif
  • fix(quality): exclude private root manifests from dependency review (#15777)
  • fix(docs): resolve generated skill curl examples to the configured loopback port (#15778) — thanks @ggdayup
  • fix(ci): verify native runtimes before accepting dependency installs (#15781)
  • fix(resilience): preserve account cooldown in weighted availability (#15784) — thanks @insoln
  • fix(combo): preserve protected-stop diagnostics and actual terminal status (#15786) — thanks @insoln
  • fix(quality): align static cycle scan with the frozen CI ratchet (#15798)
  • fix(backend): restore six fixes reverted by the #13717 squash (#13445, #13463, #13572, #13560, #13398, #13626) (#14062) (#15830) — thanks @HouMinXi
  • fix(db): stop alibaba drained-model writes from dropping the model catalog cache (#15857) — thanks @woodsonl
  • fix(a2a): resolve agent-card and MCP version from package.json (#15868) — thanks @atescivitci-cmd
  • fix(dashboard): rtl-safe mobile sidebar drawer and immediate dir flip on locale switch (#15878) — thanks @waleedkkk
  • fix(security): bind the MITM proxy to loopback and stop /api/providers/client from returning credentials (GHSA-qxg2-rm3h-4cxp) (#15886) — thanks @CodeSavvy77
  • fix(deps): next 16.3.8 (#15887)
  • fix(oauth): preserve active cooldown diagnostics on token refresh (#15891) — thanks @insoln
  • fix(sse): preserve and classify buffered Responses failures (#15893) — thanks @insoln
  • fix(sse): classify invalid_request as request-scoped (#15895) — thanks @insoln
  • fix(sse): retain safe transport cause codes in failure diagnostics (#15897) — thanks @insoln
  • fix(models): expose chat-scoped visibility to legacy dashboard clients (#15900) — thanks @tikhomirov
  • fix(security): reject encrypted-envelope credentials at provider create/update validation (#15932) — thanks @hongkongkiwi
  • fix(chatgpt-web): reuse persistent CDP sessions and current app mentions (#15935) — thanks @ElioBonfim
  • fix(security): sanitize toJsonErrorPayload before it reaches a client response (#15949) — thanks @jonlwheat2-gif
  • fix(security): sanitize the image routes' proxy-error catch and unfreeze both sites (#15950) — thanks @jonlwheat2-gif
  • fix(security): route /v1/search's internal-exception branch through buildErrorBody (#15951) — thanks @jonlwheat2-gif
  • fix(security): sanitize the Kie speech createTask failure body (#15952) — thanks @jonlwheat2-gif
  • fix(ci): give check-dead-code a real exit taxonomy and make it runnable on Windows (#15964) — thanks @jonlwheat2-gif
  • fix(providers): keep vendor-owned fal ids and encode per-app duration/resolution (#15975) — thanks @LuisAlejandro
  • fix(guardrails): reroute nested tool-output images to vision models (#15977) — thanks @LuisAlejandro
  • fix(dashboard): clear the two TS2345 dashboard-typecheck regressions on the release tip (#16000)
  • fix(catalog): stop appending a second effort tier to a Codex alias (#16018) — thanks @HouMinXi
  • fix(opencode): declare real 1M context for mimo v2.5/v2.6 -free models (#16021) — thanks @Tushar49
  • fix(codex): resolve an effort-suffixed alias against the synced base model (#16028) — thanks @HouMinXi
  • fix(api): stop the supervisor when the dashboard asks to shut down (#16033) — thanks @HouMinXi
  • fix(sse): strip every percent sign when parsing SYNTX quota left (#16040)
  • fix(sse): a partial OpenCode account sync no longer rejects the other keyless free models (#15540) (#16042)
  • fix(mcp): lazy-load API key database lookup (#16046)
  • fix(combo): scope quota-weighted scoring to the requested model's quota windows (#16054) — thanks @HouMinXi
  • fix(quality): point the provider-asset provenance manifest at the squash commit that landed notrack-web (#16055)
  • fix(i18n): give the /home route a provider layout — it rendered raw keys (#16056) — thanks @jonlwheat2-gif
  • fix(security): refuse non-loopback MITM peers and hard-gate /api/providers/client (#16060) — thanks @CodeSavvy77
  • fix(catalog): yield while filtering skipped models (#9147) (#16064)
  • fix(combo): drop malformed custom models from the auto candidate pool (#16069)
  • fix(docs): parse HTML text in migration count checks (#16089)
  • fix(bedrock): merge consecutive same-role messages (#16105) — thanks @lorenzozanee
  • fix(release): drain four v3.8.52 base-reds left by #15891, #15927 and #15968 (#16107)
  • fix(skills): repoint remaining omniroute-cli skill links (#15665) (#16109) — thanks @afonsoft
  • fix(notion): update fallback catalog with live models and internal codenames (#16113) — thanks @imnotscare
  • fix(providers): SambaNova free catalog matches its own table (#16117) — thanks @pacocartones
  • fix(changelog): turn the codebuddy-intl fragment into a bullet line (#16118) — thanks @Junior-HJ
  • fix(memory): keep Obsidian memory files inside the vault (#16168) — thanks @pacocartones
  • security(rail): local-only route sweep, public-creds structural guard, LTS security baseline and support window (rail 3.8.57) (#16209)
  • perf(rail): reproducible LTS performance baseline script and preliminary devbox report (rail 3.8.56) (#16211)
  • fix(sse): preserve input image in body.image for Gemini Antigravity image edits (#16222) — thanks @zeydsama
  • fix(sse): ignore retry hints outside the valid date range (#16234) — thanks @GGsXPosiedon
  • fix(dashboard): prevent Codex account row collapse (#16248) — thanks @jktan0504
  • fix(resilience): apply absolute RSS threshold to background sampler (#16262) — thanks @sadSanta-07
  • fix(perf): escape backslashes in LTS baseline markdown cells (#16263)
  • fix(models): preserve hidden Codex models for inference (#16299) — thanks @JxnLexn
  • fix(sse): correct 'transfered' typo in context handoff prompt (#16322) — thanks @brashdaniel-droid
  • fix(antigravity): expose cachedContentTokenCount in OpenAI usage details (#16337) — thanks @ElissonNadson
  • perf(sse): bound parseModel and canonical-provider caches (#16341) — thanks @opensource-elearning
  • perf(build): enable incremental TypeScript for all typecheck configs (#16342) — thanks @opensource-elearning
  • perf(sse): defer two conditional chatCore imports to first use (#16346) — thanks @opensource-elearning
  • perf(sse): route parseModel hot-path logging through level-gated logger (#16355) — thanks @opensource-elearning
  • fix(github): auto vision header, Responses body for uncatalogued codex ids, GHE targetFormat override (#16357) — thanks @DuqueSouza
  • fix(combo): opt-in outputLimitAdvisory demotes instead of drops output-limited targets (#16387) — thanks @octrow
  • fix(models): preserve account catalogs across disable and reactivation (#16404) — thanks @JxnLexn
  • fix: repair guarded transport and release CI contracts (#16420) — thanks @KiaroSama
  • fix(agnes): fold the intl max catalog id back onto the free model (#16423) — thanks @HouMinXi
  • fix: preserve account quota preferences (#16424) — thanks @KiaroSama
  • fix(providers): add uncloseai gateway entry + allow noauth providers in API + add bun-pack (#16430) — thanks @opensource-elearning
  • fix(sidebar): accessible accordion header toggle via native button (UX-1) (#16434) — thanks @KrzysiekSko
  • fix(sse): derive the auto-combo noauth pool from emulated tool calling (#16436) — thanks @opensource-elearning
  • fix(security): drop global TLS-verification override from the inspector HTTP proxy snippets (#16468) — thanks @123abab123789
  • fix(security): require a dashboard session to issue sync tokens and expire/revoke them with their issuer (#16478) — thanks @gsmarcil
  • fix(security): harden the LOCAL_ONLY manage-scope bypass deny-list and settings import (#16479) — thanks @123abab123789
  • fix(security): apply API-key connection and model scope to count_tokens and Vision Bridge (#16481) — thanks @TheCrabe
  • fix(api): stop manage-scoped keys from minting admin CLI tokens (#16489)
  • fix(auth): re-auth the OIDC identity keys and keep session state out of import-json (#16491) — thanks @123abab123789
  • fix(guardrails): apply API-key routing policy to the audio and video bridges (#16492)
  • fix(api): apply API-key connection scope to direct /v1 media and embedding routes (#16499) — thanks @TheCrabe
  • fix(antigravity): update CLI fallback to 1.3.4, add Argon/Claude 5.5 models, and preserve weekly quota keys (#16559) — thanks @artickc
  • fix(vision): add gemini-4 to VISION_MODEL_ID_FRAGMENTS (#16568) — thanks @Neuron-Mr-White
  • fix(sse): align reactive compaction ceiling to input cap (#16627) — thanks @Junior-HJ
  • fix(sse): enable reactive compaction for native Codex passthrough (#16632) — thanks @Junior-HJ
  • fix(video): preserve fallback transcripts as safe text content (#16635)
  • fix(cli): não devolver chaves em texto puro a credenciais de máquina em /api/cli-tools/keys (#16654) — thanks @AndreAlmeidaDC
  • fix(db): import JSON preserva estado das chaves e cifra credenciais (GHSA-p42h) (#16655) — thanks @xiaoyaner0201
  • fix(sse): keep adaptive thinking effort at the level the client chose (#16673) — thanks @HouMinXi
  • fix(opencode): serve the gated free tier from /inference/v1 and give placeholder tools non-empty schemas (#16826) — thanks @velizarofff
  • fix(sse): send Arena create-evaluation cx telemetry and fix its cookie probe (#16895) — thanks @themirzaalibaig
  • fix(i18n): use the canonical 提供者 for provider in the zh Zed tool description (#16964)
  • fix(oauth): honor connectionId on email-less re-auth instead of duplicating the connection (#16986) — thanks @castocolina
  • fix(health): don't condemn refresh-less connections whose known expiry is still future (#16987) — thanks @castocolina
  • fix(combos): hiding a base model also hides its generated effort variants (#16988) — thanks @castocolina
  • fix(sse): compress Anthropic tool_result blocks in safe pipeline (#16999) — thanks @DarknessKiller
  • fix(auth): keep a provider wildcard inside the provider it names (#15415) — thanks @HouMinXi
  • fix(sse): always emit is_error on Anthropic tool_result blocks (#15754) — thanks @timootten
  • fix(sse): preserve assistant text from Responses snapshots (#15772) — thanks @insoln
  • fix(sse): a normal empty stop is a valid empty turn, not a provider failure (#16072) (#16091) — thanks @skygunner
  • fix(sse): explicitly cancel inner reader on terminal disconnect to prevent heartbeat leaks (#16269) — thanks @sadSanta-07
  • fix(sse): forward client app-attribution headers to upstream (OpenRouter Inkling 403) (#16445) — thanks @woodsonl
  • fix(catalog): retain explicit vision evidence for named combos (#16789) (#16796)
  • fix(resilience): show when a blocked provider serves again on the connection card (#16082) — thanks @maxmad64bis
  • fix(proxy-subscriptions): count early selector skips (no control URL / unmapped key, in memory, reset on restart) and show them in the Subscriptions tab only when they concern a configured selector (#16164) — thanks @maxmad64bis
  • fix(auto): the auto/* request path shares one lazy pool preparation (unknown ids never prepare), with candidate members, order and weights unchanged; the opt-in AUTO_REQUEST_RESOLVED_CAPABILITIES flag (default off) prepares the pool with resolved capabilities and quality scores, which also feed the snapshot weights (#16419) — thanks @maxmad64bis
  • fix(credential-health): opt-in retest of connections the gateway switched off by itself (OAuth refresh retry budget, stream expiry, missing Antigravity project, ...). With the CREDENTIAL_HEALTH_RETEST_AUTO_DISABLED feature flag on (default off), such a switch-off is marked autoDisabledAt and the credential-health sweep retests it at most once per 6 h each and 5 per sweep; a passing test switches it back on. Operator-disabled connections, accounts switched off by the permanent-ban auto-disable (banDisabledAt), banned or deactivated accounts and inactive rows with no marker are never retested (#16190) (#16588) — thanks @shipsfromrio
  • fix(sse): a finished /v1/messages request no longer stays in memory through the provider circuit breaker — the chat handler handed the process-wide breaker two callbacks created per request (onStateChange, isFailure), and because a closure shares its function's V8 context the breaker kept the last request's whole scope alive, body included (a 100k-token body survived forced GC, JON-562); both callbacks now live at module scope. (#17103)
  • fix(logs): in a source checkout (dev server, tests) the call-log artifact worker starts with TSX_DISABLE_CACHE=1: tsx lists its whole disk cache on a thread's first transform, and on a long-lived dev box (585k entries) the worker's first write waited 5-7 s on that listing instead of ~0.4 s. Compiled workers (npm, standalone, Docker) are unchanged. (#17133)
  • release: the catalog visibility settings suite (#13562) waits for the settings round trip with a realistic window — its switches stay disabled until the settings load, Testing Library's default 1 s wait timed out inside the full Vitest UI suite on a loaded runner (1,203 ms in #17130's CI), and with every response delayed 1.5 s the previous suite failed 7 of 8 cases; the assertions are unchanged. (#17139)
  • fix(muse): normalizeMuseBaseUrl() stripped trailing slashes with /\/+$/, which backtracks quadratically on a long run of slashes (about 5 s for a 50k-slash base URL). It now uses the linear stripTrailingSlashes() helper; two unit tests that picked a mocked upstream by URL substring now compare the parsed host instead (CodeQL #1155, #1153, #1148). (#17058)
  • fix(codex): Recognize x-opencode-session-id, x-session-affinity, and x-session-id as inbound client session aliases for fingerprint convergence and turn-state provenance, preserving native session-header precedence and existing fingerprint-mode boundaries. (#17096) — thanks @aldoeliacim
  • fix(docs): the /docs pages that contain a markdown link render again instead of falling back to the error boundary (#11834 destructured the awaited fumadocs-ui/mdx module namespace instead of its default export, so defaultMdxComponents.a was undefined and every DocsLink threw React error #130 during SSR; it also pushed a literal default key into the components map, producing [object Object] hrefs). Pages without links were never affected. (#17079) — thanks @Theadd
  • fix(i18n): the docs translation pipeline formats each mirror with the repo's Prettier options (prettier.config.mjs + .editorconfig, printWidth 100) instead of Prettier's defaults, so the commit-time prettier --write pass no longer re-wraps code fences and leaves the mirror matching its recorded target_hash. (#17129)
  • fix(compression): Preserve assistant and instruction context during live-zone compression so current tool observations stay complete through the session-dedup and lite pipeline, while completed tool history remains eligible and cached prefixes stay unchanged. (#17097) — thanks @aldoeliacim
  • fix(authz): management credentials now take a real credential to obtain when login is disabled. With requireLogin=false, an anonymous caller could create an API key with the manage or admin scope, grant that scope to an existing key, regenerate or reveal a management key, or mint a CLI access token at any scope — and each of those credentials passes the always-protected tier (credential and log exports, shutdown). Those operations now need a dashboard session or the local CLI in every login mode, and a manage-scoped key keeps its write ceiling for CLI tokens with login off, as it already had with login on. Inference-only keys are unaffected: they are still created, edited, regenerated and revealed exactly as before. (#17064)
  • release: four more release-branch reds drained, all tests left behind by contracts that merged on 2026-10-18 — the ConnectionRow suites mock useLocale (#16947), the Docker heap anchor follows build.yml's 10240 MB webpack budget (#16938), and the call-log readers in upstream-error-call-log-16558 and reasoning-routing-pipeline wait for the detail artifact to settle now that rows are inserted before it (#15608). (#17060)
  • release: the quota-fetch throttle suite (#6911) measures the throttle again instead of process start-up — since #15885 the first quota fetch of a process lazily imports the proxy transport and opens SQLite to look up the connection's proxy between the throttle and the upstream call, which bunched the first two DeepSeek fetches together on loaded runners; the suite now answers the proxy lookup without the database and pays those imports once before timing. (#17102)
  • release: the provider translate-path golden records WorkBuddy's gateway-compliance headers (User-Agent: WorkBuddy/5.5.6, the X-Agent-* / X-IDE-* set, X-Domain, X-Requested-With) that #17046 introduced on purpose without regenerating the snapshot; no other provider entry changed. (#17102)
  • release: the WorkBuddy registry test pins the four gateway-compliance body rewrites #17046 placed ahead of the generic Codex/OpenAI ones (Anthropic billing header dropped, Claude Code and Codex CLI identity sentences reworded); the order is part of the contract. (#17102)
  • test(sse): the specificity detector's <5ms for 20 messages check measures the fastest of ten warm runs instead of one cold call, which a busy CI runner stretched to 5-18 ms in 4 of 32 PR Vitest jobs; the 5 ms bound is unchanged. (#17159)
  • fix(sync): sync token management now stays authenticated when login is disabled. With requireLogin=false, an anonymous caller could issue a sync token and use it to download the config bundle (decrypted provider credentials and API keys), even though /api/settings/export-json already refused the same caller; the token also kept working after login was turned back on. Issuing, listing and revoking sync tokens now need a dashboard session or the local CLI in every login mode (/api/sync/tokens joins the always-protected tier), and tokens issued that way by earlier releases (issuedBy: "open") no longer open the bundle — issue a new one from the dashboard or the CLI. (#17053)
  • release: the quota-fetch throttle suite (#6911) keeps 100 ms of slack for scheduling noise — it measures spacing at the mocked fetch, and on a loaded runner a 30–50 ms hiccup on the first call alone shrank a 200 ms spacing to 168 ms; it now uses a 400 ms interval with a 300 ms floor (an unthrottled pair still lands 1–2 ms apart) and a 150 ms cache-hit ceiling. (#17108)
  • fix(providers): codebuddy-cn thinking models can inline replayed reasoning_content as a <thought> block across turns, so it survives a gateway that drops the field — opt-in via the default-off CODEBUDDY_CN_INLINE_REASONING flag (#13632) (#14475)
  • Fix the dashboard on phones: Material Symbols icons now honour their Tailwind size classes (the icon font's unlayered font-size: 24px overrode every utility), provider screens, filters, toolbars and model cards fit and fill the viewport, page/section titles stay on one line with header actions below, toggle thumbs are centred and spinners rotate around their own centre. Adds shared responsive primitives (ActionBar, FilterPills, ReorderControl, ProviderCardGrid, AutoGrid, CapabilityBadge, compact Select). (#15532) — thanks @HDBR
  • docs(quality): QUALITY_GATES.md and RELEASE_CHECKLIST.md (and their 66 locale mirrors) no longer call bare check:cycles advisory: it exits 1 whenever any cycle exists (fail-closed since #15281), and the release checklist now runs check:cycles:ratchet, the frozen-ceiling form CI uses; the check-cycles.mjs header comments say the same. The quality-rail membership guard also rejects a plain (non-ratcheted) cycles entry in the fast-gates gates=(...) array (#15682 — thanks @woodsonl)
  • fix(usage): credit-metered providers (Kiro) are billed by the credits the upstream reports (meteringEvent, converted with KIRO_CREDIT_PRICE_USD, default $0.02/credit) instead of by per-token pricing rows that describe no real tariff and charged the same request twice. The exact provider-reported cost is persisted per request in usage_history.provider_cost_usd (any provider, xAI and OpenRouter included) with provider_credits as provenance (migration 212), and per-key USD quotas, usage stats, provider window costs, the costs dashboard, retention rollups and the metered budget all read it instead of re-pricing those rows' tokens; KIRO_REQUIRE_CREDITS=true fails a Kiro response that reports no credits (#16087) — thanks @thotypous
  • Skip family fallback candidates locked under their provider wire alias while preserving the selected account and catalog visibility checks. Refs #16227; complements #16223 and #16810. (#17113)
  • fix(plugin): the v1 plugin no longer fetches /api/combos/auto on each catalog refresh; the auto/* entries come from /v1/models as served by the gateway (#16228) — thanks @maxmad64bis
  • fix(sse): opt-in ANTHROPIC_COMPATIBLE_MID_CONVERSATION_SYSTEM flag (default off) keeps mid-conversation role: "system" turns in place on plain anthropic-compatible-* relays instead of hoisting them into the top-level system, but only when the client itself negotiated the mid-conversation-system-2026-04-07 beta (forwarded as-is, never invented). Every hoisted turn (environment updates, <total_tokens> budget notices) grew system on each request and forced the provider to re-write the whole cached history (≈20–30% cache hit on long Claude Code sessions). With the flag on and the beta present only the leading run before the first real turn is hoisted and the global system prompt never lands on a conversation turn. Flag off, beta absent, native claude and the CC bridge keep the previous behavior (#16236) — thanks @MikeTuev
  • combo: a request-scoped 401/403, a 401/403 that names the model as unsupported, and the synthetic stream-early-EOF 502 no longer mark the whole connection exhausted, so the remaining targets that share it stay eligible in the same request (#14313, thanks @asdlombardi4-lang). Opt-in: COMBO_EMPTY_POOL_NO_ELIGIBLE_TARGETS_ENABLED makes an auto combo whose candidate pool is already empty answer 503 no_eligible_targets instead of the legacy 400 tool-calling capability_mismatch (off by default). (#16256)
  • fix(copilot-m365): preserve trusted current and regional ChatHub hosts when importing HAR credentials, share parsing between validation and execution, and reject malformed or unsafe imported URLs. (#16327) (#16755)
  • Accept the exact Microsoft 365 Copilot ChatHub host substrate.svc.cloud.microsoft when importing or pasting credentials, preserving the host through validation and execution while rejecting untrusted hosts and malformed URLs (#16327). (#17137)
  • fix(providers): OpenRouter video models are callable again. POST /api/v1/videos answers 202 with a job id, and the synchronous OpenAI video handler reported that as success with an empty data array, so a combo stopped instead of trying its next target; expired jobs were polled until a misleading 504. openrouter now uses a submit→poll preset, the auth-gated unsigned_urls result is downloaded with the provider credential and re-served from GET /api/v1/media/{id}, and that URL is built from the configured public base URL, a trusted proxy's forwarded host, or the Host the caller dialed (never an untrusted X-Forwarded-Host). Served media follows the files API ownership rule (the generating key, a manage-scope key or the dashboard session) and expires on the call-log retention window (#16372) — thanks @LuisAlejandro
  • Honor the configured thinking-summary display for native Claude OAuth requests while preserving explicit display settings and disabled thinking modes. (#16407) (#16772)
  • Discover Muse Code models from the authenticated Meta API and expose synchronized models through the CLI catalog with the shared access filters. (#16774)
  • Load ChatGPT Web browser dependencies only when a browser operation starts, and resolve installed browser-runtime packs explicitly so absent optional packs do not break metadata and storage helpers. (#16771)
  • Preserve Windows CLI .cmd/.bat shell selection in Linux-built bundles without routing direct executables through the shell (#16547). (#16768)
  • fix(opencode): add the OpenCode v2 CLI client fingerprint (v2 User-Agent/x-opencode-client, v2 session headers without x-opencode-request, the v2 CLI read/shell placeholder schemas and no tool_choice: "none" on free-tier requests) behind the opt-in OPENCODE_V2_CLIENT_FINGERPRINT_ENABLED flag, default off (#16571) (#16998) — thanks @Junior-HJ
  • fix(sse): opt-in opencode-zen egress-IP lockout (OPENCODE_ZEN_EGRESS_IP_LOCK_ENABLED) and opt-in full-text error-rule bodies for the opencode family (OPENCODE_FULL_TEXT_ERROR_RULES_ENABLED) and cloudflare-ai (CLOUDFLARE_AI_FULL_TEXT_ERROR_RULES_ENABLED), all off by default so the release behavior is unchanged (#16797) — thanks @maxmad64bis
  • fix(opencode): opt-in OPENCODE_PUBLISHED_CATALOG_ROUTING_ENABLED (default off) lets each key-based OpenCode Zen/Go model without an explicit registry endpoint take the request surface the published models.opencode.ai catalog declares (Responses for @ai-sdk/openai, Messages for @ai-sdk/anthropic); off keeps the registry heuristics and never fetches the catalog (#16798) — thanks @maxmad64bis
  • fix(models): Keep the unified model catalog cached when OpenRouter capabilities are unchanged, preserve dashboard retry metadata, and retry transient catalog failures once in model pickers without losing saved custom values. Real capability changes still invalidate the cache; pending reads are cancelled when the picker unmounts or changes source. (#16845)
  • feat(providers): Complete the capability-aware model catalog loop: classify Gemini transcription/TTS/Lyria models, retain declared/inferred/defaulted endpoint provenance through import and Auto-Sync, expose editable endpoint badges with endpoint-only reset across dashboard model lists, preserve operator choices and discovered metadata, reject trusted wrong-endpoint requests locally before provider execution, and avoid advertising known-unrunnable specialty models. Surface provider adapter availability in model tests and row badges. Existing unsupported Gemini AI Studio transcription still requires the separate provider adapter tracked in #16887 (refs #16929) (#16956) — thanks @fabioKoruja00
  • fix(workbuddy): Adapt WorkBuddy AI reverse-proxy headers and tool/turn compliance to prevent gateway errors (10085, 11101, 11128, 11133, 11153). Enforces mandatory WorkBuddy/5.5.6 client identity, stable turn/conversation tracking header family (X-Conversation-*), first-turn system prompt injection, Gemini tool-result turn alignment, and deep tool schema normalization for DeepSeek/Kimi/GLM models. Adds alias wbai. (#16978) (#17046) — thanks @Junior-HJ
  • fix(gamification): advanceStreak read the stored streak with an await before writing it back, so two overlapping first-of-day calls both saw yesterday's date, and both reported extended: true (the caller pays the streak bonus on that flag); the read and the write now have no await between them (#17009) — thanks @pacocartones
  • fix(api): POST /api/plugins and PATCH /api/relay/tokens/[id] called request.json() with no guard, so a body that is not JSON made the handler reject instead of answering; both now answer 400 with Invalid JSON body (#17010) — thanks @pacocartones
  • fix(sse): a browser context is no longer served without its assigned proxy when the proxy lookup fails — the request fails visibly instead of egressing from the real IP (#17011) — thanks @maxmad64bis
  • fix(gamification): the stored streak counters were returned as found, so a negative or fractional currentStreak or longestStreak in the key-value row was passed on to the level route, the streak badge check and the MCP gamification tool; they are now read as finite, non-negative whole day counts (#17012) — thanks @pacocartones
  • fix(gamification): the early-adopter badge read the first-activity timestamp (SQLite UTC text without a zone marker) as server-local time and added the window with setDate, so on a server whose time zone is behind UTC the window ended hours late and a user 30 days and 1 hour in could still unlock it; the timestamp is now read as UTC and the window added in UTC (#17013) — thanks @pacocartones
  • fix(sse): a Microsoft 365 Copilot turn whose WebSocket frames never terminate now aborts at the same 1 MiB pending-buffer cap as the Claude Web reader instead of growing memory without bound (#17014) — thanks @maxmad64bis
  • fix(gamification): POST /api/gamification/invite accepted a fractional maxUses such as 1.5 (201); the schema now requires a whole number, so a fractional value answers 400 (#17016) — thanks @pacocartones
  • fix(gamification): the transfer history is ordered by created_at DESC, which has a one-second resolution, so transfers recorded in the same second had no defined order; ties are now broken by id DESC, newest first (#17019) — thanks @pacocartones
  • fix(gamification): validateScoreChange only compared recentXp + amount with the per-window limit, so a NaN amount passed (every comparison with NaN is false) and a negative amount offset the window total; both are now refused with Invalid XP amount. The XP amounts come from the XP_REWARDS table today, so this guards the function itself (#17020) — thanks @pacocartones
  • cli: omniroute doctor, status and config no longer print Node 24's DEP0190 warning on Windows: the CLI tool detector now probes .cmd/.bat shims through cmd.exe with an argument array instead of shell: true plus arguments (#17021) — thanks @pacocartones
  • fix(chat): a null entry in tools (or in input/messages for the session fingerprint) made generateSessionId and the OpenAI→Claude tool conversion throw a TypeError instead of ignoring the entry. messages: [null] was already answered with a 400 (#13755); the same entries are now skipped in the session fingerprint and in the tool conversion (#17023) — thanks @pacocartones
  • fix(models): stop treating Kilo variable-price auto routing as unconditionally free; verify ten existing Kilo and Gemini free-tier evidence records (#17028). (#17138)
  • fix(models): require per-model source evidence for new recurring-uncapped free catalog entries, with a shrinking legacy audit list (#17028). (#17124)
  • fix(providers): Preserve SiliconFlow live pricing and free/paid flags through model discovery and retain explicit paid verdicts in the discovery cache, so free-only synchronization rejects paid models even with optimistic free hints while retaining explicitly free entries. (#17028) (#17121)
  • fix(providers): Stop billing-risk misclassification of ten paid SiliconFlow flagship models as unlimited free. Only trust live free evidence; explicit paid flags and positive pricing veto free-only imports and badges (#17028). (#17031) — thanks @fabioKoruja00
  • fix(oauth): Kimi Coding retains its refresh token after a rejected OAuth refresh so the bounded health-check retries can recover instead of becoming stuck in no_refresh_token; Codex/OpenAI token-rotation handling is unchanged (#17042) — thanks @fabioKoruja00
  • fix(sse): opt-in PROVIDER_SAFETY_STOPS_TERMINAL flag (default off): preserve explicit provider safety-stop codes through HTTP, WebSocket and completed-stream diagnostics; stop combo retries and cancel admitted siblings without treating a refusal as successful fallback. Legacy stops retain a local classification without inventing an upstream detector. The check:any-budget:t11 checker no longer counts property calls such as AbortSignal.any as any types (#17049) — thanks @aldoeliacim
  • fix(zcode): the opt-in image-runtime helper (scripts/ops/zcode-image-runtime.mjs) gains a second hash-guarded step, numeric-provider-code-v1: the ZCode finish-record detector now treats a code as a provider business error only when it is numeric, so GLM tool-call text that lands in that field no longer kills a valid long agentic turn with 502 Provider returned a business error. Explicit failure markers still fail as before, and the receipt records both steps (#17062) — thanks @megamen32
  • fix(vision): When Claude Code's Read tool embeds an image inside tool_result.content, a confirmed text-only target no longer receives raw image bytes if captioning fails or exceeds the configured image cap. The bridge inserts an explicit unavailable-text marker, keeps tool-result context intact and still preserves raw images when a target's vision capability is unknown or native. (#17069) (#17087) — thanks @fabioKoruja00
  • fix(oauth): the dashboard OIDC login now sends a PKCE challenge and a nonce, and the callback refuses an id_token that does not echo the nonce (#17072) — thanks @maxmad64bis
  • fix(dashboard): Model test no longer sends image-generation custom models (e.g. Agnes) as chat completions; it honours the Image Generation API format and reports the probe as skipped (#17085) — thanks @quantum-shubham
  • fix(i18n): Preserve source-owned <think> code examples during documentation translation while continuing to reject added reasoning and malformed or missing examples (#17088).
  • fix(quota): A locked database (SQLITE_BUSY) during the quota snapshot cleanup no longer crashes the server; the background tick retries briefly, logs and skips, and the next tick tries again (#17101) — thanks @tiagovilasboas
  • perf(api-manager): load the per-key request, cost and last-used columns from a new GET /api/usage/analytics?view=apiKeys view that runs one grouped query, instead of the full all-time analytics report plus 1,000 call logs. On instances with a few hundred thousand usage_history rows the full report held the event loop for 10+ seconds on every API Manager visit, stalling all proxied streams; last-used times now also cover keys outside the newest 1,000 call logs. (#17056) — thanks @fouadSalkini
  • fix(api): new opt-in CATALOG_MIRROR_REFINEMENTS_ENABLED flag (default off — /v1/models is unchanged until an operator turns it on) stops advertising mirror ids that cannot be used: claude/ discovery aliases also mirror reasoning-effort variants already in the catalog (e.g. claude/codex/gpt-6-sol-high; listing of those variants still follows OMNIROUTE_DISABLE_THINKING_LEVEL_VARIANTS), skip built-in auto/* ids the request path cannot materialize and never duplicate an alias already listed; functional-gateway mirrors are no longer emitted for combos, for OmniRoute-synthesized ids (claude/…, no-think/…), for a model whose owner is connected under an alias prefix (e.g. cx/… with a codex connection), for the gateway's own models, through a gateway disabled in blockedProviders, or for models outside the gateway's authoritative live catalog (matched on the full id the gateway receives). Built-in auto/* materialization now goes through one pure resolver with unchanged request-path behavior (#16255) — thanks @aldoeliacim.
  • Forward adaptive thinking for Kiro Claude Opus 5.5 and preserve its native max effort through Responses conversion without changing other providers' effort normalization. (#16098) — thanks @Kizuno18
  • fix(typecheck): remove duplicate KeyHealth export (direct commit 550f64211d)
  • fix(onboarding): make Auto Combo ToS inclusion explicit (direct commit a75077fa7c)
  • fix(arena): preserve sync freshness across restarts (direct commit dadd89f851)
  • fix(models): ignore inactive seats when computing combo catalog limits (direct commit fc7f47f9a8)
  • fix(luna): disable reasoning for Chat Completions tools (direct commit b91a10c677)
  • fix(opencode-go): align verified model context limits (direct commit 38afc5bead)
  • fix(cli): repair authenticated key policy commands (direct commit 5348c217e4)
  • fix(mitm): resolve the antigravity bridge cert like the AgentBridge server (direct commit 8189d4d2b9)
  • fix(github): strip temperature for Copilot GPT-5.4+ and GPT-6 models (direct commit 51d40dae86)
  • fix(grok-cli): keep hosted tools and the forced function when capping at 200 tools (direct commit 9a35f7a8ea)
  • fix(translator): keep Gemini literal unions typed and bound $ref expansion (direct commit 65bd855254)
  • fix(proxy): give relay proxy tests 30s to survive free-plan cold starts (direct commit 842f478a2f)
  • fix(codebuddy): normalize empty finish_reason / function_call shells in CodeBuddy SSE (direct commit dead01403d)
  • fix(kiro): preserve upstream output-token limit in live model discovery (direct commit 325547b8a9)
  • fix(logging): retain upstream error bodies in call logs (direct commit dc62f1c747)
  • fix(docs): restore public catalog and canonical guide links (direct commit f1ab2657e6)
  • fix(streaming): drain upstream after terminal client close (direct commit e50b4848aa)
  • fix(providers): clarify static-only model discovery warning (direct commit 2c9c8e45eb)
  • fix(cli): avoid volatile env warnings for generated defaults (direct commit 685a3025b5)
  • fix(free-tiers): remove unsupported Bytez and Nous free claims (direct commit 6aec22c1ca)
  • fix(docs): validate aggregate provider counts across published surfaces (direct commit 3e704ce8d6)
  • fix(skills): synchronize cli-keys key-policy flags (direct commit 38fdfa28ae)
  • fix(docs): widen the version-claim gate and correct AGENTS.md's coverage floor (#15966) — thanks @jonlwheat2-gif
  • fix(resilience): limit Anthropic missing-model lockouts (#16406)
  • fix(sse): recover rejected empty streams on eligible sibling accounts (#16438) — thanks @fouadSalkini
  • fix(quality): reject incomplete complexity outcomes on main (#16536)
  • fix(ci): deliver ONNX install skip through explicit env (#16782)
  • fix(docs): validate aggregate provider counts on main (#16809)
  • fix(dashboard): only switch timeline to pan on an actual drag (#17024) — thanks @for-the-zero
  • fix(compression): fold role:"developer" into system for the engines (#17029) — thanks @DarknessKiller
  • fix(model-test): probe speech models through audio endpoint (#17030) — thanks @edosulai
  • fix(autoCombo): hold browser-backed and local-CLI transports out of the keyless pool (#17036) — thanks @opensource-elearning
  • fix(opencode): rotate Zen free DeepSeek id to v4.1 (#17044) — thanks @coqfu
  • fix(providers): keep uncloseai's context window across /v1/models rotations (#17067) — thanks @opensource-elearning
  • fix(release): reconcile-changelog reads fragments from disk and stops counting closing issues as coverage (#17068)
  • fix(release): drain v3.8.52 base-reds — #16956 fixtures, single customModels read in /api/models, stryker tap.testFiles (#17090)
  • fix(quality): drop two eslint suppressions that no longer match (#17126)
  • fix(dashboard): give nested sidebar items and subtitles AA contrast on the sidebar background (#17154)

📝 Maintenance

  • chore(quality): the open-sse → src boundary ratchet is base-relative on PR events (measure-boundary.mjs --ratchet --base-ref <sha>, wired in quality.yml fast-gates and the ci.yml lint job): the ceiling becomes max(baseline ceiling, edges on the PR base), so edges other merges added after the seed no longer red innocent PRs, while edges a PR adds still block. Push and workflow_dispatch keep the absolute ceiling; an unresolvable base SHA falls back to it with a warning. (#16968)
  • chore(quality): re-seed the open-sse → src boundary ceiling at the release tip (771 → 774): two merges landed between the seed measurement and #16874's merge, so the new ratchet was red on the base from its first run. The ceiling only goes down from here. (#16904)
  • ci(quality): the open-sse → src boundary is now a ratchet. scripts/quality/measure-boundary.mjs only measured the engine → app import edges (the v4 remediation list), and the count drifted 745 → 768 between 2026-10-10 and 2026-10-16 unnoticed. New --ratchet fails when edges.openSseToSrc exceeds quality-baseline.json → metrics.boundaryOpenSseToSrc (seeded at the measured 771, direction: down, dedicatedGate: true, no velocity-phase headroom) and --update lowers the ceiling but refuses to raise it. npm run check:boundary:ratchet runs in the ci.yml lint job next to check:cycles:ratchet and in the quality.yml fast-gates ratchet loop; the Markdown/JSON report is unchanged. (#16874)
  • refactor(sse): chatCore slice 2 — open-sse/handlers/chatCore.ts drops from 3,948 to 3,286 lines by moving request-entry guards (managed-lease fence, custom system prompt, plugin gate, device tracking), the Claude passthrough / Claude Code bridge translation branches, the translation-failure handler, post-translation finalize and OmniGlyph compression, the Quota Share pre-hook, the final output budget, the stream controller wiring and compression helpers into 20 leaves under open-sse/handlers/chatCore/, with no behaviour change; the file-size cap is lowered by hand to 3945 (design after #15709 by @HouMinXi). (#16901) — thanks @HouMinXi
  • chore(dev): two read-only PR babysitting helpers replace the shell pipelines rebuilt by hand during every campaign — scripts/dev/ci-discriminate.sh <PR> [--ref-pr N | --base-red issue] pulls each failing job log, reduces it to signature lines and sorts them into own / inherited / unknown (exit 1 only on own reds, pending checks called out), and scripts/dev/audit-worktree.sh <path> [--base ref] reports dirty files, commits and diff since the merge-base, protected-surface hits, removed assert/test lines, eslint-suppression counts that went up, new test-masking allowlist keys, leftover tsconfig.*tmp*.json and whether the PR head on GitHub matches the local HEAD (#16903)
  • chore(dev): two developer scripts replace shell recipes that every 2026-10 campaign session re-typed by hand. scripts/dev/typecheck-touched.sh [--base <ref>] [file ...] typechecks only the touched .ts/.tsx files through a temporary tsconfig (extends ./tsconfig.json, types: ["node"] so narrow includes stop inventing TS2591 Cannot find name 'node:test'), runs tsc under a 6 GB heap instead of the project-wide run that OOMs the shared devbox, keeps the full log under _artifacts/typecheck-touched/, always deletes the temp tsconfig, and exits 1 for errors in touched files, 2 for errors only in imported files or a tsc crash. scripts/dev/new-worktree.sh now serialises concurrent invocations on <git-common-dir>/omniroute-new-worktree.lock (flock, mkdir fallback, 600 s wait with a visible message) instead of racing on .git/config (exit 255), and rolls back the branch it created when git worktree add fails before the worktree exists, so a retry needs no manual git branch -D. (#16905)
  • chore(quality): config/quality/file-size-baseline.json loses the 69 file-path keys a 2026-09-03 merge resolution (#12626) left at its top level, outside frozen. check-file-size.mjs only ever read cap/testCap/frozen/testFrozen, so those keys did nothing except show a second, different ceiling for files like combo.ts and chatCore.ts; the removal is 69 deleted lines with no other byte changed and the gate output is identical before and after. A new shape test (tests/unit/build/file-size-baseline-shape.test.ts) keeps the top level to those four keys plus _ notes. The architecture docs, REPOSITORY_MAP.md, OPEN_SSE_ARCHITECTURE.md and CONTRIBUTING.md now give the measured 150 executor modules instead of the stale 67/89/107/108/148, which clears the soft check:docs-counts drift, and no longer claim every module in open-sse/executors/ is a BaseExecutor subclass (the per-provider helper modules are not). (#16876)
  • ci: Lock released branch and OpenSSF Scorecard now run at most once per ref (concurrency with cancel-in-progress), so the ~10 pushes/hour on the active release branch no longer queue 20 hosted runs each against the account's concurrent-job limit. (#16967)
  • release(lts-cut): the 3.9.0 cut dry-run/rehearsal gains a blocking source-full-history precondition — a shallow clone cannot push stable/v3 (the receiving side rejects the pack at the first grafted parent), so the script now refuses to run the cut from one and says how to unshallow. (#16875)
  • refactor(mcp): createMcpServer now registers tools instead of also implementing them (#15159 M-01). The audit found open-sse/mcp-server/server.ts mixing a 784-line registration function with twelve handler bodies, and prescribed "file = register + wrap. handle* → tools/canonical/*.ts". Two of M-01's three parts had already landed with M-06 (#15468), which moved the internal hop into internalFetch.ts and closed graph cycles 193/194; this completes the third. The twelve handlers — handleGetHealth, handleListCombos, handleGetComboMetrics, handleSwitchCombo, handleCreateCombo, handleCheckQuota, handleRouteRequest, handleCostReport, handleListModelsCatalog, handleWebSearch, handleXSearch, handleWebFetch — move verbatim into tools/opsTools.ts, tools/inferenceTools.ts and tools/webTools.ts, grouped by the surface each talks to rather than one file per handler, matching how advancedTools.ts already holds sixteen. server.ts drops 1538 → 1068 lines and reads as registration. The shared coercion helpers moved to a new leaf, coercions.ts: withScopeEnforcement wraps every registration and also calls toRecord, so leaving the helpers inline would have forced tools/* to import server.ts — reintroducing exactly the edge M-06 removed to close those cycles. check-cycles holds at 13, unchanged. No behavioural change: 493/493 MCP vitest tests and 25/25 focused MCP unit tests pass, and 11 of the 12 handler bodies are byte-identical to the originals after comment and whitespace normalisation — the twelfth differs only by the rename below. The local toNumber is now toFiniteNumber because no-restricted-syntax (#7879) bars that name in favour of src/shared/utils/numeric.ts, and the old server.ts carried a suppression for it; renaming removes the suppression instead of relocating it into a fresh file. Its semantics are unchanged and it is still not the canonical helper, which also parses numeric strings — swapping that in is a real behaviour change and needs its own PR with a test. (#15695) — thanks @jonlwheat2-gif
  • ci(llm-security): the nightly promptfoo/garak eval servers now boot on a fresh DATA_DIR under $RUNNER_TEMP with auto-backups off, so the self-hosted runner's stale ~/.omniroute database no longer trips the mass-migration guard and kills the security verdict (dead since 2026-08-27). (#16847)
  • ci: the nightly LLM-security guard eval runs its pinned npx promptfoo from runner.temp, so the repository's own (currently unimportable, #17017) promptfoo install can no longer shadow the isolated npx tree. (#17034)
  • ci: the nightly LLM-security guard eval runs a network-isolated npx --yes [email protected] (same version as the devDependency pin) instead of the lockfile binary, which cannot import on the release branch because the lock nests js-yaml 4.3.2 under @apidevtools/json-schema-ref-parser (#17017). (#17018)
  • ci: the nightly LLM-security guard eval now runs the lockfile's promptfoo (pinned exactly at 0.123.0) instead of promptfoo@latest; releases ≥ 0.123.1 throw result?.error?.includes is not a function on the guard's 4xx JSON reply, which turned every blocked request into an eval error. (#17005)
  • ci(resilience): the nightly k6-soak job now runs on the self-hosted omni-light pool (the hosted 7 GB runner died in the backend-only build 8 nights out of 8), boots its server on a fresh DATA_DIR under $RUNNER_TEMP on its own port 21776 with auto-backups off, and fails when the server never comes up instead of soaking nothing (#16853). (#16873)
  • test(auth): pin the invariant that makes the 8 provider-auth routes safe (#15159). The agy-auth / claude-auth / codex-auth import, apply-local and export routes forward a typed error straight to the client — if (error instanceof AgyAuthFileError) return NextResponse.json({ error: error.message, code: error.code }, …) — while their generic branch already routes through sanitizeErrorMessage. That is safe only because every message those classes are built with is a static app-authored literal: all 34 throw sites across src/lib/oauth/utils/{agy,claude,codex}Auth*.ts were verified to interpolate nothing, so no upstream body, token or filesystem path can reach the client through those routes. The new test asserts that, naming the offending file and line the moment someone adds a throw site like `refresh failed: ${upstream.body}` — which would otherwise silently start leaking through eight routes that read as safe on inspection, past a check:error-helper that trusts the whole file because it imports the sanitizer. The right fix for such a site is at the throw site, not eight rounds of route sanitization. (#15394) — thanks @jonlwheat2-gif
  • chore(skills): regenerate the omni-providers and omni-inference endpoint reference files with the official agent-skills generator so check:agent-skills-sync is green again on release/v3.8.52 (drift left behind by the call-site spawn gate wording of #16366 and the management-only auth notes). (#16890)
  • quality: the two circuit-breaker tests added by #16594 (circuit-breaker-degraded-decay, connection-test-closes-degraded-breaker) are now listed in stryker.conf.json tap.testFiles, so check:mutation-test-coverage --strict passes again and their mutant kills count. (#17052)
  • test(auth): allowlist the assert reduction from deleting the dead isProviderFailureCode export (#15159 B-04). check:test-masking flagged tests/unit/account-fallback-service.test.ts for a net drop of 11 asserts (333 → 322) — the anti-masking gate cannot tell a retired-surface removal from weakening-to-go-green, so the reduction is recorded with its evidence in config/quality/test-masking-allowlist.json. isProviderFailureCode and PROVIDER_FAILURE_ERROR_CODES had zero importers on release/v3.8.52 (the only hits were the definition at open-sse/services/accountFallback.ts:1349 and the stale pointer comment corrected by B-05), and the removed case asserted isProviderFailureCode(429) === true — the exact opposite of the shipped breaker policy, which scopes 429 to connection cooldown / model lockout and reserves whole-provider trips for 408/500/502/503/504. The deleted test therefore pinned wrong behaviour on dead code. The live surface is now covered more strictly by tests/unit/breaker-status-set-single-source-15159.test.ts, which asserts the export is gone, that no production module imports it, and that both the combo copy and the live PROVIDER_BREAKER_FAILURE_STATUSES deep-equal [408, 500, 502, 503, 504] with 429 excluded. (#15381) — thanks @jonlwheat2-gif
  • tests(translator): Isolate the provider-scoped thinking-output regression from changing provider catalogs, while separately checking the real OpenCode Go output cap and operator-override precedence (#10139, #15306). (#16807)
  • Align auto-combo inspector and skipped-variant test fixtures with the default ToS eligibility policy, preserving strict-cost, account-lockout, and warning assertions. Related to #15180. (#16249)
  • docs(routing): pin the Auto-Combo scoring diagram weights to the code with a parity test that compares each diagram weight against the scoring defaults (#15225) — thanks @maxmad64bis
  • test(db): five tests/unit/db files no longer fail at import when the better-sqlite3 native addon cannot load — the migration 185/186/191 fixtures run on node:sqlite through the SqliteAdapter path, and the health-check and omp suites, which need the native driver, skip with a documented reason (#15252) — thanks @yourspraveen
  • test(autoCombo): two AutoCombo Vitest tests no longer time out under parallel load — autoCombo.test.ts and tieredRotation.test.ts warm the task-fitness DB in a scoped beforeAll, so the lazy first-call init no longer counts against a pure-computation test's 5 s budget (#15253) — thanks @yourspraveen
  • test(api): the chat completions keepalive correlation test now guards the route wiring for real — it reads the wrapper's buffer under the caller's X-Correlation-Id instead of re-recording the bytes itself (red without correlationId: reqId, green with it), and drops the 5 s real sleep (#15254) — thanks @yourspraveen
  • docs(compression): the Compression Guide names cavemanOutputMode.enabled as the switch for the legacy caveman output mode, states where its back-compat injection differs from the legacy one (marker line, ja/zh spacing, Hungarian fallback), lists the boundaries clauses as written, and documents what the Auto-Clarity Bypass toggle does and that it lives on the Caveman page (#15297) (#15596) — thanks @woodsonl
  • docs: QUALITY_GATES.md describes the check:file-size PR mode as the whole-tree, base-relative check it runs, and its gate-table row states the source and test-file caps (#15345) — thanks @woodsonl
  • docs(contributing): align the Co-Authored-By checklist line with Hard Rule #16 so reused human work keeps its standard credit (#15389) — thanks @maxmad64bis
  • ci: scheduled workflows now run only in the upstream repository, so forks with Actions enabled skip the inherited crons (#15398) — thanks @maxmad64bis
  • fix(test): stream-readiness-fallback no longer fails the whole file on asynchronous activity after a test ends — two cases cancelled the served body instead of draining it, so the feeding pipeline settled post-test and tripped an unhandledRejection (#15428). — thanks @BenjaminAronsson
  • perf(db): speed up the provider health matrix with a covering call-log index (#15435) — thanks @maxmad64bis
  • perf(i18n): syncing UI keys across locales no longer rebuilds each target node per key — missing keys merge in linear time with identical output (#15466) — thanks @maxmad64bis
  • docs(compression): the Compression Guide's Caveman Output Mode, Output Styles, Tool Result Compression and language-selection sections now match the sources that implement them, and a verification sweep corrected the Ultra, Aggressive, Cache-Aware, Progressive Aging, Stacked Pipeline and Combo Overrides sections (independent ultra engine, truncation not summarization, provider-triggered cache downgrade, shipped aging thresholds, stacked math, string compressionOverride, 55 RTK filters) (#15596) — thanks @woodsonl
  • refactor(db): the engines-map derivation no longer calls the unreachable aggressiveEnabled helper, which always returned false because the aggressive config normalizer never emits an enabled key; the derived default is now pinned by a unit test (aggressive off unless defaultMode is "aggressive") (#15601) — thanks @woodsonl
  • test(compression): the Auto-Clarity chat-pipeline tests from #14551 move to tests/unit/compression and cover the legacy caveman switch, the Output Styles panel, and compression-combo output mode with the toggle off, on, left unset, and a benign turn; the compression guide (and its 66 locale mirrors) now say the toggle gates the bypass (#15610) — thanks @woodsonl
  • docs(ops): document the native systemd watchdog in the monitoring guide. A unit written by hand (Type=simple) never detects a frozen event loop; the new section gives the four unit lines, why WatchdogSec must be 120 or more, the foreground-only rule, how to check WatchdogUSec and WatchdogTimestamp, and how to turn the pings off (#15748) — thanks @maxmad64bis
  • fix(skills): generate a concrete archive filename for the current sync pull [outPath] download contract and keep the example free of unsupported source/merge/replace/dry-run flags (#15945). Regression coverage compares both generated and published examples with the real CLI registry. Thanks @BenjaminAronsson for the generator-drift diagnosis and original repair.
  • maintenance(skills): synchronize the generated provider and inference references with the current OpenAPI descriptions. Environment-independent generation remains a separate fix in #15948. (#15945) — thanks @BenjaminAronsson
  • fix(i18n): translate 63 untranslated and half-English Spanish UI strings (#15989) — thanks @pacocartones
  • fix(i18n): translate 33 untranslated Brazilian Portuguese UI strings (#15991) — thanks @pacocartones
  • fix(i18n): translate 42 untranslated European Portuguese UI strings (#15992) — thanks @pacocartones
  • fix(i18n): translate 32 untranslated French UI strings for Codex, providers, proxy subscription and tunnels (#15993) — thanks @pacocartones
  • fix(i18n): translate 27 untranslated French UI strings for combos, quota windows, CLI guides and menus (#15994) — thanks @pacocartones
  • fix(i18n): translate 31 untranslated Italian UI strings for Codex, providers, proxy subscription and tunnels (#15995) — thanks @pacocartones
  • fix(i18n): translate 28 untranslated Italian UI strings for combos, quota windows, CLI guides and menus (#15996) — thanks @pacocartones
  • fix(i18n): translate 27 untranslated German UI strings for Codex, providers, proxy subscription and tunnels (#15997) — thanks @pacocartones
  • fix(i18n): translate 29 untranslated German UI strings for combos, quota windows, CLI guides and menus (#15999) — thanks @pacocartones
  • docs(api): document the per-key permission fields accepted by PATCH /api/keys/{id} (#16001) — thanks @pacocartones
  • test(plugin-v2): the plugin keeps the four catalog filter defaults when relaying parsed options, and the guide names * as the full-catalog escape hatch (#16163) — thanks @maxmad64bis
  • ci(build): Use the measured Webpack + 10 GB V8 heap path in the manually dispatched hosted artifact workflow, matching the memory-safe release CI configuration instead of retrying the Turbopack OOM path (refs #16169) (#16938) — thanks @fabioKoruja00
  • test(providers): Keep the Gemini Web validation fixture aligned with the signed-in session-token requirement while retaining signed-out rejection coverage (#16170).
  • test(combos): Give positive default Auto Combo route fixtures an eligible official API account while preserving family fixtures, empty-pool checks and shipped ToS protections (#16173).
  • test(providers): align the Gemini Web validation success fixture with the authenticated-page contract and cover rejection of a signed-out HTTP 200 page through the public validator, without changing production authentication checks (#16174).
  • test(mcp): Follow quota validation through the extracted operational handler and verify its scope-protected registration instead of requiring the endpoint literal in the registration module (#16175).
  • Exercise quota consumers through the registered MCP RPC tools and exported A2A skill, including HTTP failure propagation and a disconnected-handler mutation control, without changing production handlers or weakening structural wiring checks. (#16203)
  • perf(sse): large-context requests serialize the outbound body once (the fingerprint path previously paid a discarded full stringify), compute the final input-token breakdown once per body snapshot, and skip compression-stat clones on image-free bodies; final-guard estimation measured 33-43% faster on 0.5-3.2 MiB fixtures with byte-identical outputs (#16277) — thanks @woodsonl
  • perf(sse): context estimation measures structured bodies in a single pass without building the media-substituted placeholder tree, and target-driven image pruning applies exact per-image deltas instead of re-estimating the full history before every removal; pruning measured 3.7-4.4x faster on 24-100 image histories with estimates byte-identical (own __proto__ keys now match JSON.stringify instead of being undercounted) (#16278) — thanks @woodsonl
  • perf(db): speed up call-log artifact lookups with a partial path index (#16350) — thanks @maxmad64bis
  • refactor(shared): process-wide state shared through one helper instead of ten hand-written copies (#16382) — thanks @maxmad64bis
  • docs(routing): document that the Auto-Combo quality tracker is in-memory and restarts cold with a neutral 0.5 until first observations (#16408) — thanks @maxmad64bis
  • refactor(security): remove unused caller-controlled disable-list resolution from mandatory prompt-injection middleware; regression tests ensure body, metadata and both disable headers cannot bypass block mode (refs #16413) (#16934) — thanks @fabioKoruja00
  • test(guardrails): characterize the guardrail pipeline before the v4 module extraction — tests/unit/guardrail-pipeline-characterization.test.ts pins the five disabledGuardrails channels (API key, body, metadata, x-omniroute-disabled-guardrails, x-disabled-guardrails) including their order/dedupe and the either/or header precedence, fail-open on a throwing pre/post-call guardrail, per-API-key semantic-cache signatures and the vision-bridge describe self-loop headers; tests/unit/guardrail-pipeline-wiring-characterization.test.ts pins with source assertions which modules call runPreCallHooks/runPostCallHooks, which routes import the chat handler (image/embeddings/audio do not), that post-call runs only on the non-streaming path, and the memory → tools → post-call → semantic-cache order in nonStreamingResponse.ts (#16415) — thanks @yourspraveen
  • chore(keys): remove the unused open-sse/services/keyGroupAuth.ts. It had no importers, and it duplicated the key-group model check that isModelAllowedForKey already enforces through checkKeyModelAccess (src/lib/db/apiKeyGroups.ts). (#16418) — thanks @yourspraveen
  • docs: align the CONTRIBUTING Node.js floor with engines.node (#16534) — thanks @trabgarmarjc-maker.
  • docs(resilience): clarify rateLimitProtection: false semantics under the rate-limit auto-enable safety net, and identify POST /api/rate-limits as the supported write path (#16535) (#16683) — thanks @pedrohx26 / @PrinceThummar011
  • docs(guardrails): GUARDRAILS.md now documents where the guardrail hooks actually run (pre-call on chat-family routes, post-call on non-streaming responses only, streaming PII via the SSE transform), what is not covered (media/embeddings/data routes, streamed responses), that both disable headers are merged since #16412, prompt-injection block mode as a mandatory boundary that disables cannot bypass (the middleware's body-only disable list is ignored by evaluatePromptInjection()) (#16413), and what sees a response before post-call (#16564) — thanks @yourspraveen
  • fix(build): Windows local builds no longer die at next build with Failed to load native binding under Next 16.3.8 — the SWC binding cache is pinned outside the temp-home sandbox (#16605) — thanks @Theadd
  • perf(logs): Isolate each Node log-export destination run in a short-lived OS child process so payload hydration, JSON work and destination I/O do not block the server event loop; preserve acknowledged cursors across crashes, retain environment-proxy/TLS routing and enforce per-run timeout and child heap limits with actionable OOM errors. Keep Bun's compatibility exporter in-process with a one-time warning. (#16663) — thanks @xiaoyaner0201
  • fix(types): Restore the core TypeScript check for OpenAI-to-Claude translators and DSML match offsets without changing protocol output (#16825) — thanks @jktan0504.
  • skills: regenerate 22 drifted skills/** mirrors from docs/openapi.yaml so check:agent-skills-sync passes again (base-red #15306). (#16496)
  • docs(agents): document Bun's postinstall-placeholder failure as distinct from bun: not found — when lifecycle scripts are skipped, node_modules/bun/bin/bun.exe is still the ~450-byte bash placeholder the package ships (the real binary is already present under node_modules/@oven/bun-*), and on Windows it surfaces as the misleading This version of … is not compatible with the version of Windows you're running, which reads like an OS/architecture fault. Fix: cd node_modules/bun && node install.js (#15522) — thanks @jonlwheat2-gif
  • chore(quality): raise every file-size ceiling by 20% so contributor fixes a few lines over a zero-slack freeze can land during the v3.8.52 drain. Coverage, complexity, and eslint budgets are unchanged. (#16831)
  • fix(release): scripts/release/merge-train.sh now runs the blocking check:cycles:ratchet instead of the advisory check:cycles, fails fast with an actionable message when the install is broken (missing tsc, stray node_modules/node_modules, un-run bun postinstall), and classifies the i18n-drift and agent-skills gate output lines so an inherited base-red is no longer UNCLASSIFIABLE (#15689)
  • Version the batch-merge workflow alongside the shared quality contracts and detect missing or drifted installations. Combined focused validation remains diagnostic; each actual landing follows current-candidate admission and contributor-credit rules without administrative or blanket rebaseline exceptions. (#16854)
  • Include the referenced protected-instruction and contributor-credit rules in the drift contract. Keep per-PR instruction approval and contributor ownership explicit, without automatic closure or direct protected-branch fallback exceptions. (#16854)
  • Reject interrupted, invalid, or incomplete compiler results in the dashboard, API, and open-sse typecheck gates before comparing or updating frozen baselines; invoke the candidate's installed TypeScript directly. (#16849)
  • Preserve the Windows no-shim execution contract, including checkout paths containing spaces and shell metacharacters, with real gate subprocess regression tests. (#16849)
  • docs(readme): list the auto/subscription and auto/thrifty subscription-first auto combos (#11146) in the README's zero-config auto table, linking the subscription ladder guide (#15467) — thanks @yourspraveen
  • docs(release): the README footer and llm.txt (and their 66 locale mirrors) now carry the v3.8.52 version the cycle-open bump set in package.json, clearing the strict docs-sync base-red on release/v3.8.52 (#15100) (#15185)
  • chore(release): ledger the v3.8.51 -> v3.8.52 CHANGELOG sync-back (direct commit 0b62441dbc)
  • docs(changelog): carry the final [3.8.51] section into release/v3.8.52 (direct commit d9fcf09932)
  • test(providers): include Copilot's auto model in the pinned GitHub lineup (#16743)
  • chore(ci): rebaseline file-size ceilings from the owner merge batch (#16777)
  • Enforce API-key model policy after combo and connection resolution (#13052) — thanks @Sheshiyer
  • Feat/session sync extension (#13127) — thanks @avichal15
  • docs(quality): synchronize skills with exact-candidate verdicts (#14404)
  • test: use peer-aware batch server and await process shutdown (#14424)
  • ci: make quarantined test debt bounded and executable (#14426)
  • ci(docker): fail fast when the omni-build runner is offline (#14701) — thanks @gonisulaimann
  • refactor(sse): split handleChatCore into four response-path leaves (#14725) — thanks @HouMinXi
  • open-sse: per-key 429 cooldown for api-key rotation (#14573) (#14884) — thanks @HouMinXi
  • docs(i18n): refresh 19 drifted docs across 66 locales and fix the db module count (#15123)
  • docs(i18n): retranslate 11 mirrors whose code fences had drifted from the source (#15194)
  • docs(guides): add common provider combinations section (#13735) (#15195) — thanks @alltomatos
  • test(plugins): isolate plugin dir per test process (#15296) — thanks @woodsonl
  • docs: correct the db module count to 137 and re-translate stale CLAUDE.md/GEMINI.md mirrors (#15329)
  • chore(quality): reconcile the jonlwheat2 audit wave — freeze providers.ts, register spawn-gate test in stryker (#15330)
  • docs(i18n): refresh 6 drifted docs and re-translate stale mirrors found by a full audit (#15335)
  • test(api): lock the filtered models response against a stale content-length (#15369) — thanks @HouMinXi
  • test(executors): lock the thinking budget against raising a caller max_tokens (#15418) — thanks @HouMinXi
  • search: make the search deadlines configurable (#15495) — thanks @HouMinXi
  • docs: clarify npm 11 allow-scripts for global installs (#15541) — thanks @lorenzozanee
  • build: stamp the service worker after the public directory is recopied (#15553) — thanks @HouMinXi
  • docs: clarify npm allow-scripts for global installs (#15570) — thanks @Siva-2517
  • test(compression): make output-style checks able to fail (#15614) — thanks @woodsonl
  • chore(skills): regenerate omni-providers endpoints reference from openapi.yaml (#15652)
  • docs(i18n): refresh QUALITY_GATES and CONTRIBUTING mirrors across 66 locales (#15666)
  • chore(quality): freeze gateways.ts growth from the freetheai domain move (#15679)
  • chore(quality): freeze the wave-2 file-size growth (#15680)
  • chore(i18n): record the wave-3 doc hashes (#15684)
  • chore(quality): record wave 5 file-size and i18n hashes (#15692)
  • chore(ci): clear two inherited lint violations (#15693) — thanks @ekinnee
  • chore(quality): record wave 6 file-size growth and i18n hashes (#15694)
  • chore(quality): record fixes-board file-size growth (#15696)
  • chore(quality): record wave8 file-size and i18n drift (#15697)
  • test(authz): cover isLocalOnlyPath path-normalization variants (#15712) — thanks @Sahas2711
  • chore(quality): record Onda A file-size and i18n drift (#15728)
  • docs: correct Auto-Sync interval (6h), RATE_LIMIT_AUTO_ENABLE default and Copilot CLI version (#15736)
  • chore(quality): record Onda B file-size drift (#15743)
  • chore(quality): raise the gateways.ts file-size ceiling to 1574 (#15747)
  • docs(i18n): refresh hidden-stale mirrors (ENVIRONMENT, ERROR_SANITIZATION, OPENCODE-V2-PLUGIN, COMPRESSION_GUIDE, FEATURE_FLAGS) (#15749)
  • docs: sync release counts for v3.8.52 (#15770)
  • test(ui): unmount FlowCanvas roots before jsdom teardown (#15779)
  • test(resilience): pin the #15690/#15698 provider-breaker fixes landed by #15594 (#15818)
  • docs(i18n): refresh hidden-stale and drifted mirrors, round 2 (#15840)
  • test(release): align health-autopilot breaker target to canonical provider id (#15594) (#15910)
  • test(release): classify the #14269 SYNTX image-edit credential site in the hard-lease inventory (#15937)
  • usage: show Anthropic API-key rate limits (#15943) — thanks @HouMinXi
  • ci(mergify): batch size 5 and checks timeout at 2×p95 for the release queue (rail G11) (#16066)
  • chore(quality): ratchet file-size baseline down after the combo/chatCore decompositions (rail G2) (#16070)
  • build(docker): let the build point apt, npm and node-gyp at a mirror (#16071) — thanks @HouMinXi
  • chore(release): dormant LTS release infra — labels, PR templates, RELEASE_STRATEGY (rail 3.8.54) (#16073)
  • ci: expose change domains from the PR classifier (one-CI-policy step 1, rail 3.8.52) (#16078)
  • test(providers): assert LMArena auth payload reaches TLS transport (#16080) — thanks @lorenzozanee
  • test(auto): seed eligible accounts in chaos combo fixture (#16090)
  • test(notion): cover sanitizeNotionAssistantText (#16119) — thanks @pacocartones
  • test(providers): cover pickCanonicalAlibabaFreeTierConnection (#16123) — thanks @pacocartones
  • test(db): cover describeReclaim (#16124) — thanks @pacocartones
  • docs(api): clarify raw versus cutoff-adjusted usage percentages (#15530) (#16155) — thanks @PrinceThummar011
  • ci(release): dormant LTS workflows, nightly version/dist-tag resolver and release channel API (rail 3.8.54) (#16172)
  • test(characterization): fix the public surface of memory, guardrails, skills, a2a and oauth before the v4 extraction (rail 3.8.55) (#16179)
  • test(characterization): fix the public surface of cloudAgent, evals, webhooks and gamification before the v4 extraction (rail 3.8.55) (#16192)
  • test(rail): boundary coupling measurement script and multi-format provider-journey contract (rail 3.8.55) (#16205)
  • refactor(combo): route strategy-specific branching through a typed strategy registry (rail 3.8.52 R0.4a) (#16206)
  • ci(release): preview-artifact build-once workflow, LTS cut dry-run script and GO/NO-GO checklist (rail 3.8.58–.59) (#16210)
  • docs(roadmap): mark the 3.8.50→.54 rail as shipped and refresh the README growth table (2026-10-10) (#16214)
  • test(quality): complementary gate and TIA cases for path-literal edges (#16068) (#16265)
  • refactor(zed-hosted): use nullish coalescing for cleaner is_error defaults (#16308) (#16314) — thanks @innonazarene
  • test(electron): cover playwright-core resolution after packing (#16433) — thanks @lorenzozanee
  • docs(env): document four CLI environment variables (#16441) — thanks @pacocartones
  • docs(i18n): fix the Bosnian AGENTS.md mirror, keep literal code spans verbatim, sync base drift (#16503)
  • test(security): compare the parsed host instead of a URL substring (CodeQL #1136) (#16616)
  • test(security): escape backslash too when building the IDE-version regex (CodeQL #1144) (#16832)
  • ci: make the .113 omni-build slot opt-in (USE_VPS_BUILD_RUNNER) (#16833)
  • refactor(security): route token fingerprints through one opaqueTokenFingerprint() helper (#16838)
  • refactor(security): compare parsed hostnames through one urlHost helper instead of substring checks (#16839)
  • docs(roadmap): refresh rail status for 2026-10-16 (#16851)
  • docs(security): point the fingerprint helper at the real managementPassword module path (#16855)
  • ci: build with webpack in the nightlies and in the Build job (Turbopack OOMs the .113) (#16870)
  • docs(i18n): reconcile release guide translations (#16924)
  • ci: Turbopack back on the self-hosted build slot (64 GB); webpack stays elsewhere (#16941)
  • test: align release regressions with current catalog and UI contracts (#16944)
  • ci(nightly-schemathesis): give the webpack backend-only build time to finish (#16953)
  • test(golden): add the daoxe and nan providers to the translate-path golden (#16962)
  • docs(guardrails): describe the injection middleware after #16934 removed its dead disable call (#17048)
  • deps: 17 Dependabot bumps — bump dompurify from 3.4.14 to 3.4.16 (#15302); bump @grpc/grpc-js from 1.14.4 to 1.14.5 (#15304); chore(deps): bump trunk-io/analytics-uploader from 2.1.3 to 2.1.5 (#15370); chore(deps): bump github/codeql-action/init from 4.38.1 to 4.38.2 (#15371); chore(deps): bump github/codeql-action from 4.38.1 to 4.38.2 (#15372); chore(deps): bump github/codeql-action/analyze from 4.38.1 to 4.38.2 (#15373); bump electron from 44.4.3 to 44.5.1 in /electron (#15577); chore(deps-dev): bump http-cache-semantics (#15648); bump http-cache-semantics (#15649); bump the development group across 1 directory with 21 updates (#15732); build(deps): bump trunk-io/analytics-uploader from 2.1.5 to 2.1.6 (#16097); bump oven/bun from 1.4.2-slim to 1.4.3-slim (#16560); bump electron from 44.5.1 to 44.7.0 in /electron (#16561); chore(deps): bump github/codeql-action from 4.38.2 to 4.38.3 (#16882); chore(deps): bump github/codeql-action/init from 4.38.2 to 4.38.3 (#16883); chore(deps): bump github/codeql-action/analyze from 4.38.2 to 4.38.3 (#16884); chore(deps): bump ossf/scorecard-action from 2.4.4 to 2.5.0 (#16885)
  • fix(docs): sync v3.8.52 version in readme and llm mirrors (#15113) — thanks @doramirdor
  • ci: the nightly a11y axe job boots its prebuilt dashboard on a hosted runner — on the light self-hosted pool the 1.2 GB build artefact took 1,416 s to download and the 30-minute timeout cancelled the job before the axe suite ran; hosted runners pull the same class of artefact in seconds, and the build itself stays on the build pool. (#17122)
  • ci: the nightly a11y job no longer builds the dashboard on the light runner pool — a new a11y-build job builds it once on the build pool and hands the artefact over; the a11y job boots it on a fresh DATA_DIR and pipes the server output into the log (#16853). (#17078)
  • test(logs): call-log readers in ten more suites wait on one shared CALL_LOG_SETTLE_TIMEOUT_MS bound instead of 1.5–10 s budgets the artifact worker's cold start outran on a loaded machine. (#17099)
  • test(logs): the upstream-error-call-log-16558 reader waits for the call-log save itself (waitForCallLogSaves()) instead of a fixed ~4 s poll, which the artifact worker's cold start outran on a loaded machine. (#17091)
  • ci: the two hosted webpack heap budgets are now recorded with their evidence — build.yml 10240 MB (validated 2026-10-18) and docker-publish 12288 MB (three green publishes) — and a test keeps every Docker heap override at or above the native floor. (#17089)
  • docs(i18n): the 66 locale mirrors of AUTHZ_GUIDE.md and REMOTE-MODE.md carry the current open-mode credential rules (the five changed sections retranslated), which also turns remote-token-privilege-locales green. (#17098)
  • i18n: the translation pipeline treats inline code as a literal — the prompt now says so explicitly, and a chunk whose output loses any of the source's inline code spans (a translated error message, MemoryHigh=4G split in two, a dropped backtick that re-pairs the rest of the line) is rejected and retried instead of written; in the 2026-10-18 resync every pass had translated such literals in 7–12 locales. (#17128)
  • docs: SECURITY.md now states the v3.9 LTS support window: stable/v3 receives security fixes and critical regressions for 12 months after the 3.9.0 cut, or until 90 days after v4.0.0 becomes latest, whichever is later (T-GAP-3). (#17086)
  • ci: pull-request and merge-group Build jobs run on the hosted pool with webpack instead of queueing for the single self-hosted Turbopack slot, where 18 Builds waited up to 18 h and the release line's own push Build got the slot once in 40 jobs; push and dispatch Builds keep the Turbopack validation of the bundler npm-publish ships. The runner box document records the restored 56 GB build-slot budget and how to clear a build whose job GitHub already gave up on. (#17100)
  • ci: the scheduled full Release-Green sweep runs its heavy lanes one after another instead of in one parallel wave: unit as ci.yml's own 8 shards (test:unit:ci:shard, as many at once as the slot has cores for), then integration, vitest and the package build. The 2026-10-18 and 10-19 sweeps killed unit, integration and the build at their ceilings inside that wave, and on the 8-core, 20 GB omni-release slot the webpack build pushed the cgroup past memory.high until the app's own pressure guard answered 503 inside unit tests. (#17156)
  • ci: the .113 runner box is reorganised by role — omniroute-113-5 keeps only omni-build (Turbopack builds), Release-Green moves to its own omni-release slot on webpack and no longer cancels a started run on every push, the light pool keeps the nightlies and lint-guard, and every runner has a memory and CPU budget (docs/ops/RUNNER_BOX.md). The VPS deploy job gets a 60-minute timeout. (#17082)
  • test(video): the OpenRouter off-origin test compares parsed hosts with isUrlHost() instead of a URL-prefix match (CodeQL #1159, #1160). (#17093)
  • Correct the guardrail coverage guide to reflect that prompt-injection middleware no longer receives caller-controlled disable lists; mandatory block-mode behavior is unchanged. Refs #16413. (#17111)
  • test(providers): align Gemini validation fixture with signed-in contract (direct commit 83cbf3b96c)
  • test(providers): include Copilot's auto model in the pinned GitHub lineup (direct commit 3d6be13368)
  • chore(ci): rebaseline file-size ceilings from the owner merge batch (direct commit 9dae75f291)
  • docs: slim AGENTS.md under the 40k context limit (62.3k -> 39.4k) (#12743) — thanks @wahidsadik371-coder
  • ci: make the .113 omni-build slot opt-in on main (USE_VPS_BUILD_RUNNER) (#16835)
  • docs(i18n): refresh approved documentation mirrors for v3.8.52 (#16980)
  • fix(deps): sync the js-yaml parser lockfile entry (#17017) (#17022) — thanks @fabioKoruja00
  • docs(release): v3.8.52 credit audit — CHANGELOG reconciliation and README refresh (#17066)
  • ci: stop starting the full CI and Quality matrices on every push to release/** (hosted queue diet) (#17092)
  • docs(i18n): resync nine drifted sources across all 66 locales (#17117)
  • docs(ci): keep the Build routing comment true after the main-only push trigger (#17127)
  • test(nightly): wait for the fake server before reading its environment (#17130)
  • docs(i18n): re-translate README, CONTRIBUTING and 11 more drifted sources across 66 locales (wave 1018) (#17142)
  • docs(i18n): restore the issue-reference literals in the Uzbek QUALITY_GATES mirror (#17148)
  • ci(release-green): sweep main once a day, keep three daily sweeps for the release branch (#17160)

🙌 Contributors

Thanks to everyone whose work landed in v3.8.52:

Contributor PRs / Issues
@0xH3li0 #16281
@123abab123789 #16468, #16479, #16491
@47vigen #16284
@66666-design #16290
@70nyIT #15193
@AbdullahPesteli #16686
@Abhijeet1520 #13016
@adivekar-utexas #15565
@afonsoft #15397, #16109
@aldoeliacim #15259, #15841, #16255, #17049, #17096, #17097
@AlexanderMakarov #16640
@alexdrgpy06 #14719
@aliepratama #16153
@alltomatos #15195, #15197, #15578, #15631, #16038, #16744
@alvinveroy #13003, #13005, #15640
@alysson-souza #16886
@amirho1 #16729
@amirsalaar #16717, #16738, #16909
@AmirSys-dev #16289
@amsjavan #13788
@AnandkumarMall #16537
@andikadevs #16530
@AndreAlmeidaDC #16654, #16955, #16959
@AndresDiagoM #16546, #16648
@Android-Login-Analysis #16297
@andyvandaric #16713
@anhtahaylove #14988
@anndev-69 #16892
@ape1121 #16034, #16106
@arashashrafii #14103, #15448
@ArchdukeViel #16296
@arcslash #16543
@Ardem2025 #15513
@aref-alapour #14803, #14821, #14835
@artickc #14131, #14269, #16559
@asdlombardi4-lang #14313
@askidmobile #16524, #16841
@asuan-dev #16735, #16950
@atescivitci-cmd #12917, #15868, #15918
@avichal15 #13127
@Azzt17 #16510
@backryun #12608, #12759
@bayhnf #16537
@bbmy85552 #15179
@BenjaminAronsson #15428, #15945, #15948
@bilalsafdardev #14889
@Bl0ck154 #16623
@branben #15498
@brashdaniel-droid #16322
@brimob-sowax #16521, #16675
@c4lyp5o #15333
@caniko #15920
@castocolina #16985, #16986, #16987, #16988
@Ch3fUlrich #14901, #14992
@chenli971014-sketch #16283
@chirag127 #12564
@christianjgilman #16601
@christianmahardhika #15559
@Chumr26 #16954, #16965
@ClaudioASantana #13387
@claw-io #15492
@claytontavaresdan #16539
@cmbergamo #16610
@CodeSavvy77 #15886, #16060
@coozgan #16537, #16894
@copystring #17000
@coqfu #17044
@cryptiklemur #13868, #15102, #15103, #16866
@D4rk4 #15533, #15750, #15787, #16006, #16017, #16260
@dajiaohuang #15708
@DangTAnh #14199
@daniadasaya #16642
@DarknessKiller #16999, #17029
@dasepmoch #16871
@DavyMassoneto #16447, #16448
@Deepleakk #16644
@defyma #16703
@developerjillur #16522
@doramirdor #13051, #15113
@dpozimski #15093
@dqphong0302 #16541
@dungartoriaaa #16287, #16511
@DuqueSouza #16357
@ebouther #15755
@edosulai #15359, #15566, #17030
@ekinnee #15668, #15693
@elidrissi-amine #16879
@elielsousa-pathbit #16621
@ElioBonfim #15615, #15935
@ElissonNadson #16337
@elvis460 #16840
@engomarzaied-prog #15425
@erickkkyt #14875
@erikyao0125 #16733
@ersaasyhar #16618
@espokaos-ops #13345
@exodus2512 #15669
@f-liva #16293
@fabioKoruja00 #16819, #16856, #16859, #16860, #16863, #16868, #16888, #16934, #16937, #16938, #16947, #16951, #16952, #16956, #17001, #17022, #17031, #17042, #17087
@fabioluissilva #15178
@farshadasadpour #16085
@fasilu #16529
@feci #13337, #15624, #16586
@FerryAr #16512
@fidelix #15299, #15310
@FitBBC #13191
@fomathpro-arch #15711
@for-the-zero #15255, #15508, #17024
@fouadSalkini #14799, #14880, #14885, #14887, #15628, #15630, #15637, #15700, #15990, #16438, #17056
@foxy1402 #16403
@gabrielalberton #14907
@ggdayup #15778
@ggiak #14546
@GGsXPosiedon #16234
@GimingRao #16908
@gonisulaimann #14701
@grapeslush #15450
@gsmarcil #16478, #16818
@guanbear #12893
@hamedniroomand #16628
@hartmark #15757, #15758, #15761, #15881
@haumanto #16542
@HDBR #15262, #15532, #15629, #15941
@hongkongkiwi #15904, #15932
@HouMinXi #13603, #14449, #14546, #14548, #14725, #14884, #15035, #15049, #15090, #15099, #15158, #15280, #15355, #15356, #15361, #15363, #15364, #15365, #15368, #15369, #15406, #15407, #15408, #15415, #15416, #15418, #15430, #15461, #15472, #15494, #15495, #15505, #15510, #15516, #15535, #15553, #15624, #15642, #15643, #15674, #15701, #15710, #15830, #15943, #15944, #16018, #16028, #16033, #16054, #16071, #16423, #16673, #16901
@huohua-dev #16708
@hussmanne #15237
@huynle-dev #17040
@iamzidar #13311
@IEatCodeDaily #16675
@igorgomes3 #16285
@imnotscare #16113
@initguru #15192
@innonazarene #16314
@insoln #15557, #15597, #15598, #15608, #15685, #15772, #15784, #15786, #15891, #15893, #15895, #15897, #15917, #16007, #16011, #16014, #16015, #16037, #16041
@jackjinke #12343
@jackyfan01 #16966
@jacobsparts #12486, #16076, #16309, #16549, #16784
@james-martinez #12690
@jeanbispo #12534
@JeanRGW #15670
@jefripunza #16852
@JessenReinhart #16917
@jiweiyeah #15131
@JJAbrams-eng #13330
@jktan0504 #16248, #16825
@jonlwheat2-gif #12740, #15215, #15217, #15221, #15222, #15226, #15227, #15233, #15242, #15244, #15281, #15282, #15283, #15290, #15291, #15376, #15379, #15381, #15382, #15384, #15388, #15394, #15468, #15520, #15522, #15695, #15766, #15767, #15768, #15771, #15773, #15949, #15950, #15951, #15952, #15964, #15965, #15966, #16056
@Jordannst #16493, #16625
@JunDev76 #16682
@Junior-HJ #15572, #15575, #15602, #15791, #15792, #15797, #15799, #16118, #16396, #16427, #16627, #16632, #16922, #16998, #17046
@juniorgaudencio00-code #16704
@Juriseagle #15561
@JxnLexn #13554, #15132, #15133, #15135, #15138, #15481, #15486, #15487, #16299, #16404, #16474
@kabeza #16281
@kalfian #16910
@kalpakprod #15357
@kang-heewon #15219
@khanghack222 #16622
@KiaroSama #15574, #16420, #16424, #16599
@Kizuno18 #15318, #15319, #16098, #16099, #16100, #16101
@KooshaPari #4165
@krzemienski #16523
@KrzysiekSko #16434
@kveita #15858
@kvyb #16711
@lacijowo #15703
@Laksopan23 #15156
@LazyGatto #15308, #15752, #15759, #16274, #16752
@lepkov #15082
@likongkong666 #16646
@lorenzozanee #15476, #15478, #15479, #15480, #15499, #15541, #15542, #15543, #15576, #15739, #16080, #16105, #16213, #16426, #16433, #16601
@Louayzouaoui1 #16204, #16216, #16425
@louisphamdev #16294, #16295
@lsilveirab39 #16902
@LuisAlejandro #15975, #15976, #15977, #16372, #16378
@m00nie #16725
@M3stryX #16250
@mad-gooze #17004
@MakdumIbrohim #16531
@mannnrachman #16949
@matheusgcdj #14910
@maxmad64bis #14554, #14805, #15003, #15140, #15224, #15225, #15236, #15239, #15240, #15248, #15314, #15316, #15353, #15354, #15362, #15374, #15375, #15378, #15383, #15386, #15389, #15390, #15392, #15395, #15396, #15398, #15399, #15402, #15432, #15433, #15434, #15435, #15436, #15443, #15449, #15451, #15455, #15462, #15463, #15466, #15469, #15474, #15475, #15482, #15483, #15484, #15560, #15562, #15564, #15737, #15738, #15741, #15748, #15885, #15888, #15908, #15978, #15980, #15981, #16051, #16067, #16082, #16102, #16104, #16162, #16163, #16164, #16166, #16167, #16193, #16194, #16212, #16228, #16267, #16268, #16270, #16271, #16272, #16273, #16275, #16276, #16291, #16292, #16300, #16301, #16302, #16303, #16304, #16305, #16331, #16333, #16334, #16335, #16336, #16339, #16340, #16344, #16350, #16353, #16354, #16361, #16362, #16363, #16367, #16371, #16373, #16374, #16376, #16377, #16379, #16382, #16386, #16389, #16390, #16391, #16408, #16419, #16504, #16759, #16760, #16769, #16773, #16775, #16779, #16785, #16786, #16793, #16797, #16798, #16925, #16927, #16932, #17011, #17014, #17072
@mdc2122 #15024
@mdigitalbh81 #15277
@megamen32 #16926, #17062
@miguelenes #15278
@MikeTuev #16236
@minhtuancn #16282
@minthanthtoo #16877
@mohammadraufzahed #15431, #15446
@morpheus9393 #13891
@mrnim94 #16700
@MRQC-st #16598
@MumuTW #15127, #15128
@mwpn #16490
@n1klaus #16685
@Neuron-Mr-White #15187, #16567, #16568
@nhan14t1 #16963
@nitsuah #16280, #16867
@Notaloop763 #13700
@ntdat812 #16519, #16520, #16521, #16523, #16862
@ntdatt812 #16510, #16521, #16539, #16540, #16891, #16893
@octo-patch #16508
@octrow #16387
@omar-nahhas #16483, #16484
@omnitrixAG #16498
@onexer88-blip #16077
@oooohhoo #16974
@opensource-elearning #16323, #16324, #16341, #16342, #16346, #16355, #16430, #16436, #17036, #17067
@Owaid0723666 #14297
@pacocartones #15946, #15989, #15991, #15992, #15993, #15994, #15995, #15996, #15997, #15999, #16001, #16005, #16008, #16009, #16043, #16047, #16117, #16119, #16123, #16124, #16168, #16441, #16453, #16456, #16501, #16504, #16507, #16517, #16544, #16556, #16606, #16607, #16609, #16614, #16615, #16620, #16630, #16633, #16636, #16639, #16641, #16647, #16649, #16650, #16651, #16656, #16657, #16658, #16659, #16661, #16668, #16671, #16730, #16991, #16992, #17009, #17010, #17012, #17013, #17016, #17019, #17020, #17021, #17023
@patrykkopycinski #15647, #15724, #15725
@pedrohx26 #16683
@PIKACHUIM #16911
@PixmaNts #14546, #15427
@piyush97 #15485
@Poid-ZA #15676
@pompousman #16613
@potatosips #15419
@Primexz #16629, #16679
@PrinceThummar011 #16155, #16683, #16764
@qirh #15573
@QuangBlue #14973, #15117, #15160, #15168, #15186, #15261, #15331, #15332, #15412, #15413, #15423, #15702
@quantum-shubham #17085
@rafi-zadanly #16961
@raflyazf #16626, #16710, #16719, #16906, #16915, #16920, #16946
@raheemuddin786 #11837
@RaviTharuma #15745
@Reinazhard #16713
@rifuki #16710
@royhanmh #16916
@rqzbeh #12967
@rusak47 #16513
@ryanngit #16288, #16469, #16475, #16480, #16484, #16497, #16728, #16742
@sadSanta-07 #16262, #16269
@Sahas2711 #15712
@sandypratama2030-gif #16921
@sasazemzulin058-debug #16532
@scottchristian #16705
@seanford #14286, #15549
@semihisikman #16628, #16699, #16704, #16713, #16715, #16720, #16907
@SemonCat #16495
@seven7763 #16889
@shannonlowder #15358
@Sheshiyer #13052
@ShiDai567 #16638
@shipsfromrio #14952, #15023, #15154, #15155, #15164, #15169, #16108, #16146, #16151, #16525, #16550, #16566, #16580, #16581, #16584, #16588, #16589, #16592, #16594
@Shurelia #16617
@sirwhy #16494
@Siva-2517 #15570
@SkiiOS #16665
@skygunner #15054, #15231, #15457, #15458, #15459, #16091, #16316, #16359, #16360, #16375, #16388, #16395, #16504, #16528, #16582
@SoloDrex52 #15742
@Spoon94 #16865, #16914
@sprintberlin #14301
@StevanusPangau #16684
@suhanrain #16821
@tasarren #16912
@TechNickAI #16397
@Theadd #15340, #16605, #17079
@TheCrabe #16481, #16499
@TheDemonTuan #13739
@themirzaalibaig #16858, #16895
@TheViniAlmeida #15594
@thienpvt #16541
@thotypous #16086, #16087, #16088
@tiagolabsit #16462, #16463, #16464, #16465, #16466, #16467, #16514
@tiagovilasboas #15800, #15925, #17101
@tikhomirov #15900
@timootten #15754, #15756
@tommyyzhao #16960
@TPOHH #13224
@trabgarmarjc-maker #16534
@trespassmk #15699, #16144, #17002
@turbolego #16223, #16227, #16810
@Tushar49 #16021
@unifically-dev #14182
@velizarofff #16826
@vianhanif #16511, #16539, #16645, #16945
@VihaanR #15301
@W25X80 #16181
@wahidsadik371-coder #12743
@wakqasahmed #15905
@waleedkkk #15878
@wallerjenifernk530-hash #16722
@webzidev #14988
@wedreamer #15548
@WellArtDev #16295
@woodsonl #15137, #15206, #15209, #15296, #15298, #15345, #15346, #15571, #15583, #15587, #15589, #15590, #15591, #15592, #15593, #15596, #15599, #15600, #15601, #15603, #15604, #15605, #15607, #15609, #15610, #15612, #15613, #15614, #15617, #15618, #15682, #15801, #15857, #16277, #16278, #16366, #16392, #16445
@wray-lee #12914, #16351
@wusaki0723 #16674
@WYSTAF #16698, #16702, #16722, #16913
@x-adam #16732
@xiaoyaner0201 #14603, #14619, #14640, #14976, #15084, #15086, #15087, #15092, #15094, #15171, #15189, #16473, #16505, #16574, #16585, #16591, #16655, #16663, #16693, #16829
@Xxx91n #11902
@Yashxp1 #15114
@Yi-111-a #15328
@YoannDev90 #16919
@yourspraveen #15252, #15253, #15254, #15452, #15453, #15465, #15467, #16399, #16401, #16410, #16415, #16418, #16564, #16565, #16931
@yugui923 #12869, #15213
@ZaimMarzuki #13019
@zasasamar2129 #16637
@zebrajaeger #15337
@zeeshanhaque21 #14212
@zeydsama #16222
@zitus91 #15663
@diegosouzapw maintainer

[3.8.51] — 2026-09-29

Living section — reconciled 2026-09-29 from all cycle commits (091589089c → 4aed4b4a08, 1,972 non-merge commits). Bullets carry the merged PR and its author; direct pushes are listed with their commit hash. Regenerated at each /generate-release phase.

📊 Release by the numbers

👥 People who contributed 325
📝 Commits in the cycle 1,972
🔀 Pull requests referenced 1,942
📋 Changelog entries 2,022
🙌 Contributors credited in entries 322
🤖 Automated dependency commits 33

Entries by type

Type Count
🐛 Fixes 1443
✨ Features 228
🧹 Chore 135
🧪 Tests 90
📚 Docs 72
⚙️ CI 11
♻️ Refactor 11
⚡ Performance 10
🏗️ Build 8
🔒 Security 7
📦 Dependencies 5
⏪ Reverts 2

🏆 Top 25 contributors this cycle

By commits in 091589089c..4aed4b4a08, author identities consolidated via .mailmap and the merged PR's GitHub login. Bots excluded.

# Contributor Commits
🥇 diegosouzapw 589
🥈 Dizzle (@maxmad64bis) 175
🥉 Bob.Hou (@HouMinXi) 167
4 Paco Cartones (@pacocartones) 93
5 Koosha Paridehpour (@KooshaPari) 50
6 Ravi Tharuma (@RaviTharuma) 43
7 Markus Hartung (@hartmark) 33
8 Nguyen Thanh Dat (@ntdatt812) 31
9 QuangBlue 31
10 Webman (@jonlwheat2-gif) 23
11 Goni Sulaiman (@gonisulaimann) 22
12 lorenzozane (@lorenzozanee) 22
13 anhtahaylove 19
14 backryun 18
15 Fouad Salkini (@fouadSalkini) 16
16 Abhishek Sharma (@abhisheksharma2411) 15
17 Nguyen Thanh Dat (@datrixlab) 15
18 Patryk Kopyciński (@patrykkopycinski) 15
19 Rafa Martins (@rafacpti23) 15
20 Caio Figueiroa (@shipsfromrio) 15
21 initguru 12
22 Jan Leon (@JxnLexn) 12
23 Syed Raheemuddin (@raheemuddin786) 12
24 Xmon Dai (@xiechimon) 12
25 Nguyễn Viết Tuấn (@TheDemonTuan) 11

✨ New Features

  • feat(audio): proxy native ElevenLabs voices, text-to-speech, and speech-to-text HTTP routes through stored OmniRoute credentials, preserving query strings, multipart uploads, binary responses, and upstream errors (#10556). (#11434) — thanks @hartmark
  • Added Google AI Studio Gemini batch text-to-speech support through POST /v1/audio/speech. (#11434) — thanks @hartmark
  • Run synchronous RTK and Caveman request compression in a bounded worker-thread pool, keeping large /v1/responses compression heaps outside the HTTP isolate while preserving strict fail-open behavior and per-engine telemetry. (#11434) — thanks @hartmark
  • feat(routing): subscription-first auto groupings — auto/subscription routes only through plan-included connections with a documented hard-stop overage and fails closed on exhaustion, while auto/thrifty orders the pool subscription → keyless → free → cheap → premium and steps up one rung at a time as each is exhausted. Billing class comes from a curated per-connection catalog (uncurated is treated as metered, never plan-included), both reuse STRICT_ZERO_COST's per-connection verification, and a quota reading whose resetAt has passed is now refreshed regardless of TTL so routing returns to plan capacity as soon as it resets (#11146) — thanks @yourspraveen
  • feat(providers): publish a management-authenticated versioned web-session credential contract from OmniRoute's canonical browser credential metadata (#11340) — thanks @Zartharas
  • feat(video bridge): harden the optional drill-down cache substrate with exact-path broker policy, canonical principal/session/media isolation, independent retained-byte quotas, cancellation-safe commits, rejection of excess or non-canonical Base64 padding and non-JPEG/truncated media, warning-sensitive full JPEG canonicalization that strips trailing polyglot bytes, server-derived dimensions, and auditable derivation metadata; production tenant binding and multi-resolution selection remain follow-up work (#11369) (#11434) — thanks @hartmark
  • feat(search): Add Xquik X search with typed results, credential validation, REST routing, and MCP selection (#11370) — thanks @kriptoburak
  • feat(video): add an opt-in focused analysis mode that safely uses a normalized, 500-code-point latest-user hint for task-aware frame captions while preserving full-mode prompts, temporal-window isolation, and cache identity without storing raw task text (#11383). (#11434) — thanks @hartmark
  • feat(dashboard): surface durable exclusive managed leases in the existing Sessions view, keeping leased clients visible across idle gaps while marking connections with in-flight work as active (#11389) — thanks @KaspaPulse
  • build(bun): allow Turbopack bundler flag on Bun 1.4+ with configurable Webpack fallback (#11471) — thanks @TheDemonTuan
  • feat(api): add an opt-in modelVisibilityAllowlist/modelVisibilityDenylist settings pair to curate exactly which models /v1/models advertises, mirrored into every auto/* combo candidate pool so a denied model cannot be routed to via combo selection either (#11481) (#11997)
  • feat(rankings): the Free Provider Rankings page shows what each provider actually served over the last 24 h. It ranked by ELO alone, which left a provider that answers every call with an error in first place; the usage data was already served by the API but never requested. A provider with too small a sample shows a dash, not a number (#11546) (#11553) — thanks @maxmad64bis
  • feat(guardrails): enforce a bounded, deterministic contract for Video Bridge transcripts — 256 cues, 4096 input code units and 4 KiB UTF-8 per cue, 64 KiB total text, malformed-Unicode rejection, focus-window scoping, cross-source reconciliation with contributing-source metadata, and a structural provenance trust boundary so caller JSON can never self-assert embedded/audio-bridge provenance (#11652) (#12009)
  • feat(video): orchestrate optional Video Bridge audio extraction and Audio Bridge STT behind a dual opt-in (operator setting AND per-request signal) — a new loopback-only broker mode=audio operation shares the frame path's exact process queue, deadline, AbortSignal, and byte budgets to extract a bounded mono 16 kHz PCM WAV from the same already-downloaded video, then reuses the existing Audio Bridge transcription boundary; provider segment timing is preserved when available and marked coarse otherwise, and every failure degrades to a visual-only-safe partial instead of throwing (#11654). (#12012)
  • Add a tenant-bound Video Bridge drill-down lifecycle on top of the existing secure cache substrate: opaque hashed handles (never raw session/video identifiers), preview/standard/detail multiresolution variants resampled on read, response pagination capped at 8 frames and 32 MiB, and a new authenticated /api/v1/video-bridge/drilldown consumer route that stays disabled for remote access by default and denies cross-key access with the same response as a nonexistent handle (no existence oracle). (#12006)
  • test(video): Add the Video Bridge FU-07/FU-09 promotion-evidence harness (#11656) — a frozen Zod manifest schema covering the 8 required scenario kinds (static scenes, rapid cuts, late facts, fades, blur, small text, close events, visual prompt injection) with a minimum of 3 repetitions per case, deterministic declarative fixture recipes (videoBridgePromotionFixtures.ts), a pure medians/p95 metrics aggregator, a pure FU-07/FU-09 promotion-verdict evaluator applying the ticket's exact thresholds (missing token usage always holds), a digest-only persistence layer that never retains raw media or raw model responses, and a versioned per-model promotion allowlist shipped empty with every model defaulting to hold. The FU-07/FU-09 promotion verdicts themselves remain HOLD — they require a real evidence run against real models on VPS 192.168.0.15. (#12008)
  • feat(video bridge): "embedded" transcript provenance can now be legitimately earned instead of merely asserted — a bounded, allowlisted (mov_text/subrip/webvtt) subtitle probe runs through the loopback-only Video Bridge broker (at most 2 streams, 10s subdeadline bounded by the request deadline, 256 KiB output, 4096-code-unit lines), normalized through a bounded, ReDoS-safe WebVTT parser and Zod-validated end to end. The adapter always resolves to an explicit success/absent/transient_failure outcome — a subtitle failure never breaks the visual description path, and only a fingerprint-verified broker response (never a caller-declared label) can produce embedded cues (#11659). (#12011)
  • Default new Antigravity-family connections (agy CLI imports and Antigravity OAuth connects) to model auto-sync, so live model discovery lands in the synced catalog and /v1/models picks up freshly released upstream models (e.g. Gemini 3.7 Flash tiers) without code changes. Existing connections keep their current setting; the per-connection dashboard toggle remains the opt-out. (#11685 — thanks @MumuTW) (#11887)
  • feat(ui): enable React Compiler (reactCompiler: true + babel-plugin-react-compiler) for automatic memoization at build time (#11783) — thanks @jonlwheat2-gif
  • feat(providers): add SeekAi (seekai.cc) as an OpenAI-compatible New-API gateway — catalog id seekai (alias ska), https://seekai.cc/v1, live /v1/models via passthroughModels, aggregator-list membership so New-API balance detection can opt in. No referral/aff codes. (#11786) (#12557) — thanks @HouMinXi
  • feat(zai): add GLM-5.3-Flash Coding Plan support (1M context, 128K output, vision, low|high|max reasoning) and route zai GLM-5.3-family API-key traffic through the OpenAI-compatible Coding Plan endpoint with native thinking defaults (#11801) — thanks @Neuron-Mr-White
  • feat(combo): choose how combo models are ordered — manual, provider, score, or name — via a sort control in the dashboard builder, persisted in config.modelSort and re-applied on load and after add (#11812) — thanks @maxmad64bis
  • feat(free): custom models can be marked free-tier via customModels[].isFree; isFreeModel() is the first door and hidePaidModels respects it even for providers outside the free budget (#11843) — thanks @maxmad64bis
  • feat(nodejs): add 5dive as a configure target — omniroute configure 5dive / omniroute setup-5dive write a 5dive auth profile that points an agent fleet's claude seats at OmniRoute, with the root-only write, the loopback-vs-https endpoint rule and the per-seat model pin handled explicitly (#11852) — thanks @5dive-bot
  • feat(sse): treat max as a first-class reasoning-effort tier and clamp per model family (GLM 5.1+/DeepSeek V4+/Kimi K3+ keep native max; o1/MiniMax/Grok/Muse Spark clamp to their upstream ceiling) (#11875) — thanks @Chewji9875
  • feat(providers): the provider plugin manifest now advertises a usage-fetch capability for the 40 providers that have a wired usage/quota fetcher, so external dashboards can read it from GET /api/v1/provider-plugin-manifest instead of parsing open-sse/services/usage.ts after every release. Discovery only — no new fetcher, no quota change, and the Dashboard quota widget stays gated by USAGE_SUPPORTED_PROVIDERS. USAGE_FETCHER_PROVIDERS moved to a zero-dependency leaf (open-sse/services/usage/fetcherProviders.ts) and is re-exported from services/usage.ts, keeping the manifest module a light leaf instead of pulling the ~490-module usage dispatcher into the manifest route. (#11903) — thanks @maxmad64bis / @pacocartones
  • feat(plugins): OMNIROUTE_PLUGINS_DIR sets the directory the runtime plugin scanner reads — and the root the plugin manager installs into — overriding the HOME-derived default, so a Docker/K8s deployment can point straight at its bind-mounted plugin tree instead of moving HOME just to relocate the scan path. An image that exports no home no longer scans /tmp/.omniroute/plugins in silence: the resolved directory is logged once at startup as scanner.dir_resolved, naming the input that won. Unset, behaviour is unchanged. Distinct from the CLI-only OMNIROUTE_PLUGIN_PATH, which finds omniroute-cmd-* command packages and never reached this scanner (#11906) — thanks @amaleta / @pacocartones
  • feat(leases): add an explicit owner-authenticated status action that returns only the active lease's privacy-safe configured connection and provider labels, with generation fencing and no credential or internal-id disclosure (#11910) — thanks @KaspaPulse
  • feat(providers): import-from-file modal shows per-row API errors and ships a downloadable CSV template (#12071) (#12504) — thanks @HouMinXi
  • feat(providers): dashboard search matches connection name and baseUrl so imported OpenAI-compat nodes surface on the provider card (#12108) (#12495) — thanks @HouMinXi
  • feat(routing): add a score Auto router strategy that selects the highest configured weighted score and reuses explorationRate. (#12155) — thanks @jacobsparts
  • Add a runtime feature flag to disable universal context handoffs globally without changing the default behavior. (#12167) — thanks @jacobsparts
  • feat(providers): the provider plugin manifest now also advertises a usage-supported capability for the 46 providers whose usage API is accepted by the server and Dashboard routes, so integrators can distinguish "the server will serve quota for this provider" from "a fetcher is wired" without reading TypeScript. Discovery only — no fetcher or quota change. usage-fetch resolves on id or alias (the usage dispatcher accepts both); usage-supported resolves on id alone, matching the runtime guard USAGE_SUPPORTED_PROVIDERS.includes(providerId). USAGE_SUPPORTED_PROVIDERS moved to a zero-dependency leaf (open-sse/services/usage/supportedProviders.ts) and is re-exported from providers.ts, mirroring the fetcherProviders leaf from #11903 and keeping the manifest a light module. (#12214) — thanks @maxmad64bis
  • feat(rankings): order Free Provider Rankings by what each provider actually served — GET /api/free-provider-rankings?sortBy=reliability and a "Most reliable first" toggle on the page. Providers with too few calls to state a success rate keep their score order below the measured ones; the default order is unchanged (#12218). — thanks @maxmad64bis
  • perf(sse): defer cloneLogPayload() in the structured SSE collector until after the maxEvents/maxBytes cap check, eliminating ~9,800 wasted structuredClone calls per streaming response (65–71% faster push()). Reducer snapshot isolation restored for OpenAI and Responses summaries (#12241) — thanks @PauloHSOliveira (#12243)
  • feat(usage): Devin CLI agentic quota (Codeium seat-management GetUserStatus) and OpenRouter key limits plus account credits now surface in Provider Limits (#12256 — thanks @Neuron-Mr-White)
  • feat(settings): persist headroomUrl through Settings so status/start use the operator URL instead of only HEADROOM_URL (#12306) (#12487) — thanks @HouMinXi
  • feat(auto-combo): Auto-Combo scoring can now weigh how often a provider/model has actually succeeded. The engine already carried that number on every candidate — 24 hours of usage history behind a ten-sample floor, real-time metrics otherwise — and the scoring function never read it, while the weight table described stability as if it did. reliability (1 - failureRate, with the same field precedence and the same rate-bounding the speed ranking already uses, so a corrupt reading means "nothing observed" rather than "fails every call") is now a declared factor shipping at weight 0, so routing is unchanged until an operator gives it one, and the stability description now matches what that factor computes (#12317) — thanks @maxmad64bis
  • feat(routing): With freeAccessPolicy: "strict", the read-only candidate listing (GET /v1/auto-combo/{channel}/candidates) no longer hides the candidates the zero-cost guard excludes — the same read-only transparency the resilience filter already honours (#9133). Each candidate now carries freeAccessExclusion saying why it would be kept out, and it tells an exhausted allowance apart from a quota reading that never arrived or went stale, which used to look identical from the outside. Routing is unchanged: the listing reports, it never enforces. The separate excludeTosAvoid guard still drops its candidates without a reason; that gap is now documented rather than closed (#12319) — thanks @maxmad64bis
  • feat(radar): The Radar catalog table now shows two facts it was already receiving from the feed and dropping on the floor: the per-model rate limits (requests and tokens, per minute and per day) in a new column, and a badge when a provider's terms state it may train on the prompts you send. A limit of zero renders as zero rather than "rate-only" — for a ceiling those are opposite facts — and a model with no training statement gets no badge, because an absent statement is not a guarantee (#12320) — thanks @maxmad64bis
  • feat(radar): explain Community, single-use, contributor, supporter, recovery, abuse, offers, and privacy rules before either Radar activation action, and remove the superseded fixed-PR grant promise from every UI locale (#12342)
  • feat(gamification): the dashboard Profile page now shows the real daily streak — /api/gamification/level returns streak: { current, longest } (per key with apiKeyId, operator-wide maximum otherwise) and the streak card reads it instead of a hard-coded 0 (#2403) (#12377) — thanks @pacocartones
  • feat(gamification): the dashboard leaderboard now shows each API key's display name under the Name column instead of a truncated key id; GET /api/gamification/leaderboard attaches name per entry (name only — no key material), while the shared ranking helper and the federation leaderboard stay id-only — thanks @pacocartones (#12385)
  • feat(gamification): enforce the documented 1000 XP/min per-API-key anti-cheat rate limit on the XP award path; over-limit awards are logged and skipped instead of persisted, and the sliding window now matches the timestamp format stored in xp_audit_log (#2403) (#12390) — thanks @pacocartones
  • feat(admin): localize the gamification anomalies page — the loading state, the Status column and the Suspicious badge now come from the common catalog (new common.suspicious key propagated to every locale) — add it to the Gamification sidebar group as gamification-admin (/dashboard/gamification/admin), and expose the loading and empty states as polite role="status" live regions (#12401 — thanks @pacocartones)
  • feat(providers): skip GitHub combo members missing from the live synced catalog, and drop Copilot models that are policy-disabled or hidden from the model picker (#12473) — thanks @RaviTharuma
  • feat(i18n): locale-expansion tooling — npm run i18n:add-locale adds a language to config, dashboard, docs mirrors, CLI, README/indexes and site in one command; browser, CLI and cookie detection resolve uk, fil/tl, zh-Hant (and the retired legacy in → id) via config aliases; config/i18n.json now ships in the npm package so the published CLI can read it; new real-translation ratio gate (npm run i18n:check-ratio, advisory) with a per-locale ratchet baseline; run-translation --adopt restores .i18n-state.json; sync-language-bars generates 🌐 bars from config; validate_translation.py loads its allowlist again. Retires the duplicate in locale (Indonesian mislabelled as Hindi) — 42 honest locales; saved NEXT_LOCALE=in / OMNIROUTE_LANG=in keep working. (#12496)
  • feat(api): Emit gateway-measured tokens_per_second (TTFT excluded) on streaming usage and X-OmniRoute-Tokens-Per-Second when first-token latency is known (#12616) (#12631) — thanks @RaviTharuma
  • feat(opencode-plugin): map gateway cost/usage/tok/s onto OpenCode inference payloads — the official plugin copies X-OmniRoute-Response-Cost, token counts, X-OmniRoute-Tokens-Per-Second / usage.tokens_per_second, TTFT, and the winning X-OmniRoute-Model onto the JSON/SSE body OpenCode already consumes. Missing tok/s is left unset (never tokens / latency). (#12636) — thanks @RaviTharuma
  • feat(combo): add combo strategy quota-weighted: skip exhausted accounts, then weighted-draw by leftover / in-flight load; existing conversations stay pinned until the account is empty (#12789) — thanks @HouMinXi
  • feat(models): Account live listings become the chat catalog source for Claude, Codex, Copilot, and AGY; public metadata only fills prices on IDs those accounts already list. (#12866) — thanks @HouMinXi
  • fix(models): Union agy and antigravity live catalogs so agy/gemini-3.8-flash-high is not rejected after the prefix folds to antigravity. (#12866) — thanks @HouMinXi
  • feat(opencode): Add @omniroute/opencode-plugin-v2, an opencode v2 plugin that gives a v2 install a route to an OmniRoute gateway: it publishes models, combos and auto-combos into the host catalog, refreshes them lazily behind a 300s TTL, and keeps serving the last known catalog from an on-disk snapshot when the gateway is unreachable. The v1 plugin (@omniroute/opencode-plugin) is untouched and keeps working as before. The plugin adapts to the host instead of assuming one shape: it reads the shape the host seeds into the catalog draft and publishes the api block, the pre-api top-level fields, or both. The gateway key comes from the host credential store when one is connected, so no secret needs to sit in a config file; a source the gateway refuses is named in a warning instead of silently degrading the catalog to raw model ids; and tool schemas bound for a Gemini model are stripped of the JSON-Schema keywords Gemini rejects, so tool calling survives that route (#12870) — thanks @maxmad64bis
  • feat(services): show sanitized CLIProxyAPI account health from its authenticated management API without exposing credentials, file paths, or raw account metadata (#6342) (#11434) — thanks @hartmark
  • feat(search): Add AnySearch free web search + URL extract (webFetch) with typed results, credential validation, REST routing, and MCP selection - fallback-only (#11690) — thanks @Xxx91n
  • feat(dashboard): display clamped [0, 100]% cached input token ratio in request logs table (#11970) (#11970) — thanks @b3nw
  • feat(catalog): add OMNIROUTE_DISABLE_THINKING_LEVEL_VARIANTS feature flag to optionally filter out thinking level variants from model catalog (#11971) (#11971) — thanks @b3nw
  • feat(dashboard): continuously export call logs to external analytics stores. A pluggable destination registry ships the full Logs-tab record set on an hourly JobRegistry cron, with a persisted per-destination cursor, batched inserts, a config UI rendered from each destination's own field descriptors, and a REST layer (/api/log-export/*) for CRUD, a connection test, and an on-demand run. A destination can opt into includeBodies to also ship the request and response payloads shown in the Logs detail pane, including the client and provider views of each call; this is off by default, and payloads inherit the dashboard's PII sanitisation, secret redaction and noLog handling. Google BigQuery is the first destination, using a service-account key stored encrypted at rest and streaming inserts keyed by call-log id, into a table that is day-partitioned on timestamp and clustered on api_key_name, provider, model and status. (#11945) — thanks @dpozimski
  • feat(providers): add MaxAI as a signed, OpenAI-compatible provider serving its 13 paid chat models (GPT-5.6 / Luna / Thinking, Claude 5 Sonnet, Claude Haiku 4.5, Gemini 3.1 Pro / Flash-Lite, Grok 4.1-fast / 4.5, DeepSeek V3.2 / R1, Llama 3.3 70B) through OmniRoute's /v1 endpoint, with per-request HMAC-SHA1→SM3→AES request signing, live model + context-window discovery from /models/get_config, and prompted tool-calling translated to OpenAI tool_calls (#11461) — thanks @arminanton
  • feat(providers): MaxAI vision input — image_url content parts are forwarded inline in message_content to the 6 vision-capable models (GPT-5.6 / Luna / Thinking, Claude Haiku 4.5, Gemini 3.1 Pro / Flash-Lite) (#11461) — thanks @arminanton
  • feat(providers): MaxAI image generation — 6 image models (gpt-image-1, dall-e-3, flux-1-schnell/dev/pro, sd3-medium) exposed through POST /v1/images/generations (#11461) — thanks @arminanton
  • feat(providers): MaxAI document RAG — inline base64 file/document attachments are uploaded to MaxAI (content-addressed doc_id) and attached to the chat via doc_list (#11461) — thanks @arminanton
  • feat(providers): browserless MaxAI onboarding — email device-pair login (/api/providers/[id]/login) and signed access-token refresh, so a connection can be created and kept fresh without a real browser or Google OAuth (#11461) — thanks @arminanton
  • feat(providers): per-provider TLS impersonation profile (MaxAI presents a Windows Firefox-150 client fingerprint) so its bot-sensitive endpoints accept OmniRoute traffic (#11461) — thanks @arminanton
  • feat(providers): add Nimble as a web-search and web-fetch provider (nimble-search) — /v1/search routes to Nimble's search API at lite depth (locale, freshness and include/exclude domain filters mapped onto the shared request contract), and /v1/web/fetch routes to Nimble Extract, which covers all four fetch formats (markdown, html, links, screenshot) from a single call. One API key serves both surfaces. (#11620) — thanks @wildcard
  • The npm publish is automatic again, through npm Trusted Publishing (OIDC): the hosted stage-npm job publishes with a short-lived credential minted from GitHub's id-token — no NPM_TOKEN, no 2FA prompt, provenance attached. publish_mode=staged (owner approves with 2FA) and direct (token) remain available on workflow_dispatch. (#11931)
  • feat(providers): Add Opper as an API-key gateway provider — EU-hosted AI gateway with 700+ models from 30+ providers behind one OpenAI-compatible API and one key (OPPER_API_KEY); model ids use provider/model format (e.g. anthropic/claude-sonnet-4-6, openai/gpt-5); live model catalog at https://api.opper.ai/v3/compat/models; entry mirrors requesty (same shape, passthroughModels: true, no static seed) (#11629)
  • feat(dashboard): the /dashboard/orchestration snapshot hook now subscribes to the agents WebSocket channel (agent.task.updated) instead of requests as its refetch trigger, and relaxes its background poll from 5s to 30s while that WS connection is up — falling back to the tighter 5s cadence, reprogrammed live on any connect/disconnect transition, whenever the socket is down. (#12409)
  • feat(dashboard): Orchestration canvas quick wins — search box plus state/source/provider filter chips with a one-click clear, and per-source collapse/expand, all reflected in the URL so a filtered/collapsed view is shareable and survives a refresh; the detail drawer gained a "copy trace JSON" action and hardened error/empty-state and accessibility handling; the Agents-tab edges now animate traveling particles along active (running) connections; and the canvas node/edge status colors moved off fixed hex values onto theme-aware --orch-status-* CSS custom properties, so they adapt correctly to light/dark mode. (#12393)
  • chore(dashboard): Orchestration UI hardening pass and the missing component/model test coverage it called for — OrchestratorNode/ActivityNode/OverflowNode rendering, the ?node=/overflow-click page routing, the Agents-tab orchestrator-click no-op and showCompleted toggle, and the overview kanban's done-column sort order (#12270, #12271). (#12393)
  • New /dashboard/orchestration page: live unified view of everything running — Cloud Agent, A2A and Conductor as a real-time graph (Agents tab), the combo cascade (Routing tab, reusing the Combo Live Studio) and a state kanban (Overview tab), with a detail drawer (trace, cost, approve/cancel). Read-only over existing APIs — no new backend. Canvas concept credit: PR #11815 design (#12261)
  • feat(dashboard): new "History" tab on /dashboard/orchestration — an Airflow-style grid of finished runs over a 24h/7d/30d preset window, one row per (source, identity), clicking a cell opens the existing detail drawer. It is backed by real persistence: A2A task lifecycle transitions are now written to the a2a_tasks table (purged after 30 days, configurable via OMNIROUTE_A2A_HISTORY_RETENTION_DAYS) and served by the new GET /api/a2a/tasks/history listing endpoint, with the task-detail route falling back to persisted history once a run leaves the in-memory snapshot. Conductor runs stay remote and are not persisted locally — the tab says so instead of silently omitting them. (#12479)
  • feat(dashboard): the orchestration detail drawer gained a "Repeat" action for Cloud Agent, A2A and Conductor tasks — a two-click confirm (click once to arm, click again within the confirm window to fire) re-submits the original prompt/input as a new run. The button is disabled with an explanatory tooltip whenever the original input can't be recovered from the loaded task detail (e.g. it never carried a prompt, or the detail failed to load). Two limitations of the A2A variant, by design: it targets the /a2a JSON-RPC endpoint, which authenticates with an API key only (REQUIRE_API_KEY=true or a configured OMNIROUTE_API_KEY makes a dashboard-session repeat answer HTTP 400 — surfaced verbatim in the drawer's error line, never as a success), and the message/send call is SYNCHRONOUS: the POST blocks for the whole skill run, so the success confirmation only appears once the repeated task finishes. A dashboard-authenticated A2A creation path is deliberately left to a follow-up — widening the endpoint's auth posture is an operator decision, not a side effect of this feature. (#12508)
  • feat(a2a): A2A task execution now records which memories were consulted for the task's last user message as metadata.memoryHits (id/key/type/content-snippet) plus a memory_hits history event, purely for observability — the retrieved memory is never injected into a skill's prompt or behavior. Gated by the OMNIROUTE_A2A_MEMORY_HITS kill-switch (default enabled; set to 0 to skip the recall lookup entirely). The drawer's new "Memory used" section lists these hits for a2a tasks and is omitted whenever there are none. Known limitation: recall only resolves under the keyless posture — a keyed caller's task owner is a SHA-256 prefix of the API key, while memory rows are keyed by the database api-key id, and no hash→id lookup exists today, so the hit list stays empty for keyed callers. The recorded hits are also kept out of the task's own input (and therefore out of the persisted input and of the "Repeat" request body), so repeating a task never re-sends the previous run's memory snippets. (#12508)
  • feat(providers): add a Perplexity Agent API provider (perplexity-agent / pplx-agent) for Perplexity /v1/responses, including the documented Anthropic, OpenAI, Google, xAI, DeepSeek, Z.AI, Moonshot/Kimi, NVIDIA, and Perplexity model IDs plus Anthropic-model max_output_tokens compatibility. (#12103) — thanks @AIB1TAL0S
  • feat(providers): add RPD (Requests Per Day) limit to provider rate limit overrides across UI, schemas, DB, and i18n (#12147) (#12147) — thanks @b3nw
  • feat(providers): add UC Direct (uncensored.com Developer API), the metered OpenAI-compatible surface. A standard OpenAI-compatible passthrough (default executor) for uncensored.com's official REST API at https://api.uncensored.com/api/v1: X-api-key auth (never-expiring uai_sk_live_ key), POST /chat/completions with streaming SSE and native tool-calling, and the full live metered catalog (82 models across 15 providers, discovered from the public GET /v1/models). Registered as provider uc-direct (alias ucd). Complements the un-metered uc persona provider — same models, metered credits and a plain API key instead of a subscription session. (#11513) — thanks @arminanton
  • feat(providers): add UC (uncensored.com), the un-metered subscription "persona" chat as an OpenAI-compatible provider. A WebSocket web-app port: a durable Clerk credential mints a short-lived session token per connect (browserless — no API key), driving UC's persona socket. Ships the browserless email-code login (request → verify → harvest), the 19 verified persona models (Claude Opus, Gemini, Grok, GLM, Kimi, DeepSeek, MiniMax, incl. the uncensored variants), prompted <tool> tool-calling with a per-model code-style dialect + auto-cure retry for guardrailed models, live <think>/reasoning split, streaming + non-streaming OpenAI responses, and full quota/auth error surfacing (paywall / message-limit / rate-limit → 429, invalid session → 401 re-login). Full multimodal parity via the persona blob-upload layer: vision (image input, 15 vision-capable models), document RAG (PDF/doc upload, server-side extraction), image generation (22 models), video generation (14 models, async signed-url → poll), and TTS (streaming MP3). Registered as provider uc (alias ucn). The metered OpenAI-compatible Developer API is a separate uc-direct provider. (#11513) — thanks @arminanton
  • feat(logs): add copy-all button to request detail modal (#11083) — thanks @AndrianBalanescu
  • feat(check): make serviceKinds required and add the reverse-walk provider consistency gate (#11392) — thanks @Tushar49
  • feat(providers): only advertise verified-working connections by default (#11448) — thanks @alltomatos
  • feat(resilience): add hierarchical concurrency admission (#11493)
  • feat(providers): derive + surface expiry for JWT-bearing web cookies (#11497) (#11505) — thanks @oyi77
  • feat(tunnels): support named cloudflared tunnels via CLOUDFLARED_CONFIG (#11561) — thanks @ggiak
  • feat(compression): make proactive context-compression threshold a live setting (#11564) — thanks @ragnar-claude
  • feat(resilience): add provider peak-hour protection (#11622) — thanks @Neuron-Mr-White
  • feat(providers): add 1min.ai provider (#11631) — thanks @zero-executioner
  • feat(dashboard): collapsible JSON tree viewer for request/response payloads (#11703) — thanks @hartmark
  • feat(routing): add LiquidAI LFM2.5-2.6B free tier via OpenRouter (#11752) — thanks @brick30llc-ctrl
  • feat(pricing): add GLM-5.3-Flash pricing, model specs, and catalog registration (#11830) — thanks @AndrianBalanescu
  • feat(routing): add relayMode for schema-locked context handoffs (#11839) — thanks @raheemuddin786
  • feat(config): add support for runtime system prompt configuration and hot-reloading (#11841) — thanks @raheemuddin786
  • feat(routing): expand connection-aware quota prefilter across combo strategies (#11682) (#11850) — thanks @HouMinXi
  • feat(models): reactive discovery sync after Antigravity model-not-found 404 (#11888) — thanks @MumuTW
  • feat(models): auto-aliases and token limits from synced Antigravity discovery (#11889) — thanks @MumuTW
  • feat(settings): raise sticky round-robin limit caps to 1000 (#12015) — thanks @adivekar-utexas
  • feat(resilience): operator-configurable global credential health check interval (#12043) — thanks @adivekar-utexas
  • feat(quota): use official OpenCode Go usage API (#12124) — thanks @ddarkr
  • feat(resilience): default the credential health check sweep to 60 minutes (#12138) — thanks @adivekar-utexas
  • feat(quality): new-code mode for the complexity and dead-code ratchets (Clean as You Code) (#12142)
  • feat(dashboard): orchestration canvas — unified model + snapshot hook (part 1/2) (#12156)
  • feat(usage): add Kilo Code balance and Kilo Pass quotas (#12178) — thanks @killer30001000
  • feat(providers): refresh vendored ChatGPT Web connector to v4.0.7 (#12181) — thanks @backryun
  • feat(providers): manual "Clear cooldown" action in the cooling panel (#12224) — thanks @patrykkopycinski
  • feat(providers): restore ChatGPT Web via clean-room browser transport (#12239) — thanks @backryun
  • feat(sse): wire the PROVIDER_PROFILES window gate into the global provider cooldown (#12247)
  • feat(providers): modernize CLOVA Studio chat and embeddings (#12277) — thanks @backryun
  • feat(browser): adopt Obscura as primary headless browser engine with Chromium fallback (#12286) — thanks @opensource-elearning
  • feat(sse): full language parity for output styles (es/de/fr/it/ru/zh + autoDetect) (#12289)
  • feat(quality): complete test:scoped — --full map rebuild, stdin selection, CI loader parity (#8084 D1) (#12353)
  • feat(providers): refresh Fable, Cursor, and Devin catalogs (#12367) — thanks @backryun
  • feat(video): redact transcript text from logs and durable memory (#12150 P1) (#12427)
  • feat(dashboard): parent-link, genuine-continuation badge, and modal perf fixes (#12448) — thanks @hartmark
  • feat(dashboard): adaptive context-budget dial on compression panel (#12488) — thanks @HouMinXi
  • feat(providers): refresh Z.ai Web models and browser transport (#12524) — thanks @backryun
  • feat(video): redact raw client-snapshot transcript fields in the detailed log (#12150 P2) (#12528)
  • feat(providers): refresh NVIDIA hosted models (#12538) — thanks @backryun
  • feat(quota): Moonshot Open Platform balance and TPD lock for custom nodes (#12590) — thanks @HouMinXi
  • feat(video): redact transcript in the in-memory pending-request snapshot (#12430 item 6) (#12596)
  • feat(dashboard): link a log entry's Conversation Context to its owning conversation (#12646) — thanks @hartmark
  • feat(catalog): eligibility-gated free-tier bucket (#12669)
  • feat(db): fail-closed previous_response_id continuation for redacted video turns (#12150 P2b) (#12707)
  • feat(gamification): pay the documented streak_bonus and badge_unlock XP rewards. XP_REWARDS listed both rewards but the award pipeline never paid them: the private reward table in events.ts omitted them, updateStreak() did not report when a streak extended, and badge unlocks carried no XP. Every request that extends a daily streak now pays streak_bonus × streak length once per UTC day (guarded by a same-day xp_audit_log check), and every badge unlocked through the pipeline pays badge_unlock once per badge (guarded by the user_badges primary key; unlockBadge() now reports whether it inserted). Bonus XP flows through the same addXp + level sync + global/weekly/monthly leaderboard path as action XP, so level-ups and rankings include it. The Radar supporter recognition unlock stays XP-free. (#12522 — thanks @pacocartones)
  • feat(providers): advertise a free-tier capability in the provider plugin manifest for every provider with documented free models, so sidecars and dashboards can filter free-capable providers without reading the quota catalog (#12786) — thanks @maxmad64bis
  • feat(providers): Added EURouter as an OpenAI-compatible API-key gateway (https://api.eurouter.ai/v1), with live model discovery via passthroughModels. Its copy states that models are served by third-party upstreams listed per model, so an EU-based router is not read as EU data residency for inference. (#13025) — thanks @ntdatt812
  • feat(providers): Added GreenPT as an OpenAI-compatible API-key provider (https://api.greenpt.ai/v1), with live model discovery via passthroughModels. No free-inference badge: the published docs describe a free API subscription billed per token, not a free tier. (#13024) — thanks @ntdatt812
  • feat(i18n): 9 new locales — Greek (el), Croatian (hr), Serbian (sr), Lithuanian (lt), Estonian (et), Latvian (lv), Slovenian (sl), Maltese (mt), Irish (ga) — across the dashboard, docs mirrors, CLI, README and the site; OmniRoute now ships every official EU language (51 locales). (#13044)
  • feat(i18n): 8 new locales — Kannada (kn), Malayalam (ml), Odia (or), Punjabi (pa), Nepali (ne), Sinhala (si), Burmese (my), Khmer (km) — across the dashboard, docs mirrors, CLI, README and the site (59 locales). The translator now restores the ICU literal escape around angle placeholders and splits oversized docs sections before translating. (#0000) (#13660)
  • feat(providers): list Agnes 3.0 Flash as the current free chat model, drop retired 1.5 Flash, add Image 2.0/2.5 Flash plus Video 2.5/2.5 Flash, and discover the live /v1/models catalog (including the CN host api.agnes-ai.cn). agnes-1.5-flash now forwards to agnes-3.0-flash. Video 2.5 polls GET /v1/videos/{id} (not the V2.0 /agnesapi contract). Live /v1/models (2026-09-09) no longer serves 1.5; the wiki marks it deprecated. 3.0 Flash is 512K context / 65,536 max output, same window as 2.5. CN-region keys use the existing per-connection base-URL field, default stays apihub.agnes-ai.com. (#13120) — thanks @HouMinXi
  • feat(sse): Codex and OpenAI catalogs list GPT-6 Astra with effort aliases (-low through -ultra); Codex CLI identity pins @openai/[email protected] in lockstep with the image (#13026) — thanks @HouMinXi
  • feat(dashboard): the orchestration History tab gained a "Compare runs" mode — toggling it turns each grid cell into a 2-item selection queue (a 3rd click drops the oldest pick), and picking two cells opens a side-by-side comparison panel instead of the usual detail drawer. The panel fetches both runs' detail the same way the drawer does (falling back to persisted history once a run leaves the live TTL window) and shows, per side: identity/source/state, start time, a signed right - left delta for duration/cost/event count, the event timeline aligned by index, and any memory hits. One side's fetch failing never blocks the other, and a delta is only ever computed when both sides have a finite value — otherwise it renders "—", never NaN. Comparing two runs from different sources or skills still works; a banner marks the deltas as informational rather than hiding them, since the two runs aren't a strict apples-to-apples pair. (#12677)
  • feat(dashboard): badge a conversation that never reached a clean stop (#12717) — thanks @hartmark
  • feat(release): reconcile-changelog tool + version-anchored fragment aggregation (#12987)
  • feat(mcp): carry context_length through MCP list_models_catalog projection (#12776) (#13406) — thanks @KooshaPari
  • feat(cli): make startup readiness budget configurable (#13369) (#13433) — thanks @KooshaPari
  • feat(docs): every Markdown page under docs/ is now mirrored in all 65 dashboard locales, not only the 22-page core set — 152 sources × 65 locales = 9,880 mirrors (6,208 new), with the 🌐 language bar of every mirror rewritten for the full locale list. The docs drift gate (npm run i18n:check, blocking in CI) derives its scope from the tree, so it now guards all 152 pages. Found and fixed by the run in scripts/i18n/run-translation.mjs: a markdown table or tight bullet list with no blank line inside it (PROVIDER_REFERENCE.md's 244-row table, FREE_TIERS.md's 71-item list) was sent as one 16–40 KB request that outlived the backend socket for verbose scripts (Greek, Amharic); oversized runs of table rows or list items are now cut at item boundaries and rejoined without a blank line, so no chunk exceeds 6 KB across the docs tree. 48 older mirrors whose tables had lost rows were retranslated with the fixed chunker. (#14106)
  • feat(usage): openai-compatible-* connections can now report billing/quota in Provider Limits. The connection declares its own quota endpoint, auth mode and a dot-path mapping onto UsageQuota in providerSpecificData.quotaEndpoint, so no upstream-specific code is needed per service — a mapping that resolves nothing reports no quota rather than an exhausted-looking 0/0 (#13616) (#13673) — thanks @abhisheksharma2411
  • feat(sse): track LLM Gateway DevPass quota — the llmgateway provider now reads its monthly plan-credit and weekly premium-model allowance from GET /v1/key and surfaces both windows in Dashboard › Limits and quota-aware preflight (#12462). — thanks @PixmaNts
  • feat(providers): register gemini-3.8-flash (#12638) — Gemini 3.8 Flash (DeepMind 2026-09-02) with tool calling and vision support (#12663) — thanks @toor11
  • feat(proxylogs): proxy log columns and detail pane now show the registry proxy name instead of a bare host:port when several registry entries share the same gateway (#12814) — thanks @tiangao88
  • compression: add Hungarian Caveman language pack with Hungarian-specific rules, language detection, localized output instructions, and language-pack tests. (#12825 - thanks @botii16)
  • feat(sse): parse/scrub DSML tool-call markers embedded in reasoning and recognize adaptive thinking on the response side — dsmlToolCalls.ts module + translator/stream/handler wiring (#12905) — thanks @initguru
  • feat(codex): safely discover compatible models by classifying upstream models before activation to keep hidden, unsupported, retired, or newer-client models out of the active catalog, exposing candidate diagnostics while persisting only active models, adding GPT-6 Astra fallback definitions, and bumping the tested Codex CLI version to 0.153.4 (#12933) — thanks @TheDemonTuan
  • feat(api): POST /api/keys accepts expiresAt (ISO datetime, nullable) with the same semantics as the key-update path, so automation can create an expiring key in one operation instead of create-then-update. Omitted/null preserves the current non-expiring behavior; enforcement reuses the existing expiry policy (#12952) — thanks @caniko
  • feat(build): add build:fast and start:fast to bypass standalone tracing (#13021) — thanks @tuandinh0801
  • feat(sse): OMNIROUTE_DISABLE_CONVERSATION_TRACKING=1 turns off conversation-history collection for operators who do not use the dashboard's conversation view. resolveConversationId() returns an untracked result before it reads SQLite or parses message history, and the switch also covers client-supplied session IDs. Routing-session handling is unchanged, tracking stays on by default, and existing records are not deleted. One reporting install held 5.97 million turn records at about 4.26 GB (#13150) — thanks @cryptiklemur
  • feat(providers): Add auto/kimi, auto/qwen, auto/deepseek, auto/gpt, and the auto/claude-haiku fast variant to the built-in routing catalog, including bare k3 models on Kimi coding and web backends (issue #13214). (#13709) — thanks @keii-2596
  • feat(usage): Claude OAuth usage now shows the separate weekly Fable limit next to the shared five-hour and weekly meters. Anthropic reports it as a weekly_scoped entry in limits[], which OmniRoute ignored, so the pool was invisible. The provider-limits cache keeps modelQuotas and restores it on stale-data fallback. The Fable meter is display-only and does not affect routing, account selection, or cooldowns (#13266) — thanks @cryptiklemur
  • feat(playground): copy an individual Compare column's response. Each column in the Compare tab now has a copy button beside the remove button, reusing the existing useCopyToClipboard hook to copy that column's response text and show a checkmark while disabled on an empty response. (The independent-scrolling half of this PR was already fixed separately in #13532.) (#13317 — thanks @ventulus95)
  • feat(models): add Gemini 3.8 Flash tiers to Antigravity and AGY catalogs (#13318) — thanks @tuandinh0801, with credit to #12499 (@Abhishekchhetri020)
  • feat(reasoning): adaptive reasoning effort (auto) — the gateway resolves the thinking budget per user turn from deterministic request-shape signals (stateless per-turn pin) instead of forwarding a literal auto, applied at the gateway pre-translation for any harness (Claude Code, Cursor, Codex, opencode, Hermes) whose request dispatches to an OpenAI Chat-Completions-shaped upstream (targetFormat === FORMATS.OPENAI — reasoning_effort is an OpenAI-shaped field, so a Claude- or Gemini-targeted request is unaffected). Opt in via X-OmniRoute-Effort: auto or a model's defaultReasoningEffort: "auto" (now a valid ModelSpec value); any explicit client reasoning field always wins (#13448) — thanks @patrykkopycinski
  • feat(dashboard): Add a dedicated, full-width API-key routing editor with explicit model/combo choices, searchable selectors and protection for unsaved rule drafts. (#13555) — thanks @JxnLexn
  • feat(proxies): proxy pools and opencode's per-account rotation stop re-serving a proxy that just failed (refused TCP probe, or a 429 through it) for a period that doubles on each repeat up to a cap, without writing any proxy status; with every candidate set aside the choice is unchanged. Opt-in via the PROXY_SKIP_RECENTLY_FAILED feature flag (default off: selection unchanged) (#13578) — thanks @maxmad64bis
  • feat(proxy-logs): proxy log rows keep the HTTP status the provider actually returned (upstream_status, null when no response arrived), so a throttled egress IP (429), a refused one (403) and a provider outage (500) are no longer the same "error" line, and a 429 generated locally is no longer mistaken for one from the provider (#13580) — thanks @maxmad64bis
  • feat(proxies): the proxy pool editor shows, for the last 24 h, how many distinct egress IPs actually served the pool's members, how many connections went through them and the most seen behind one IP, read from the proxy log through a separate route so it can never break the pool screen; opt-in with the PROXY_POOL_EGRESS_OBSERVATION feature flag (default off) (#13581) — thanks @maxmad64bis
  • feat(sse): learn hard request caps stated in 429 bodies and pace under them. Providers such as TokenRouter reject bursts with prose like Maximum 5 requests within 1 minutes and no rate-limit headers, so the limiter never learned the ceiling and kept racing into it; every 429 also tore the limiter down and rebuilt it with no pacing. updateFromResponseBody now parses that phrasing (and N requests per minute, N requests per M seconds, N RPM) into a per-window cap, applies it to the limiter as an empty reservoir that refills N every window with calls spread window / N apart, and records it in learnedRateLimits. A learned cap is reapplied whenever the limiter is rebuilt after a 429 and when limits are restored at startup, unless the connection has an explicit RPM override. Fixes #13594. (#13895) — thanks @costajohnt
  • feat(proxies): a proxy pool stops re-serving a member the provider just refused through it and tries another member instead, reusing the existing skip cooldown; a later success through the member clears it. Opt-in with the PROXY_SKIP_RECENTLY_FAILED feature flag (default off: pool selection unchanged) (#13602) — thanks @maxmad64bis
  • feat(flags): add DB_HEALTHCHECK_STARTUP_DEFERRED_ENABLED (default off) — opt-in deferral of the startup DB health/integrity check past process boot via setImmediate; off keeps the pre-#13717 behavior of blocking getDbInstance() until the check has already run (#13717). — thanks @HouMinXi
  • feat(i18n): 7 new locales — Hausa (ha), Yoruba (yo), Igbo (ig), Amharic (am), Uzbek (uz), Georgian (ka), Armenian (hy) — across the dashboard, docs mirrors, CLI, README and the site (66 locales, the full planned expansion from 43). (#13727)
  • feat(api): POST /v1/rerank (and the memory engine's loopback rerank step) can route to OpenAI-compatible provider nodes on a LAN/Tailscale host — not only loopback — behind the new RERANK_REMOTE_PROVIDER_NODES feature flag (default off), subject to the provider outbound URL policy; the loopback host check is consolidated into @/shared/network/loopbackNodeHost shared by rerank, audio, and the local health checker (#13732) (#13733) — thanks @seanford
  • feat(security): OmniRoute now warns at boot when the server that answers /v1 inference is bound to a non-loopback interface while REQUIRE_API_KEY is disabled. The guard added in #12568 covered the API bridge (API_HOST, default loopback) and the live dashboard WebSocket, but not the Next server that actually serves /v1/chat/completions and /v1/responses — which binds HOST || 0.0.0.0, every interface by default. That matters because GET /v1/models follows the dashboard login posture (requireAuthForModels) while inference follows REQUIRE_API_KEY, so an instance with an admin password and REQUIRE_API_KEY=false answers 401 to the probe an operator naturally runs while inference stays open to anyone who can reach the port. The bound host is resolved from OMNIROUTE_BOUND_HOST (published by scripts/dev/run-next.mjs) then Next's own HOSTNAME (the Docker path); HOST is deliberately excluded because the standalone server ignores it and a warning naming the wrong interface is worse than none. New docs/security/INFERENCE_AUTH_POSTURE.md documents the split, how to actually probe inference, and the #2257 caveat that an invalid bearer degrades to anonymous. (#13820) — thanks @abhisheksharma2411
  • Expose combo wall-clock timeout (comboTimeoutMs) next to Target timeout in the combo editor and Combo defaults. Empty keeps the 10-minute hang-stop; a positive value replaces it. (#13857) — thanks @HouMinXi
  • feat(i18n): retranslate-site rewrites the site catalogs' verbatim-English leaves (2,059 across 63 catalogs; mean English residue 10.3 % → 6.3 %, the rest being brand names kept on purpose). (#13886)
  • feat(compression): Lite tool-result truncation length is configurable (lite.maxToolLength, env OMNIROUTE_LITE_MAX_TOOL_LENGTH). Default stays 2000. An out-of-range step cap no longer hides a valid global cap; a toggle-only settings write keeps a stored cap; maxToolLength: null clears it. Dashboard copy no longer hard-codes 2,000 characters. (#13915 — refs #13178) — thanks @HouMinXi
  • feat(proxy): support multiple local core endpoints, one per line (#13923 — thanks @maxmad64bis)
  • feat(providers): Added Agnes AI (China) as agnes-cn pointed at https://api.agnes-ai.cn/v1. Keys issued for apihub.agnes-ai.com stay on the existing agnes card. Live /v1/models on that host lists agnes-3.0-flash (same id as intl); the CN seed matches 2.0/2.5/3.0 and not retired 1.5. (#13399) — thanks @HouMinXi
  • feat(api): add per-key allowAutoCombos (default true) to gate the built-in auto/* combos, which previously bypassed a key's allowedCombos/allowedModels/blockedModels restrictions entirely — a restricted key could still reach any model through auto/best-fast. Also adds a per-key catalogScope (all/combos/models) to control what /v1/models advertises, and the dashboard gained an Auto Combos toggle and a catalog scope selector in the API key permissions UI. (#13670) — thanks @fouadSalkini
  • feat(sse): Claude OAuth connections can opt in (per account, Edit connection → Claude section) to Claude Code's lower-priority lane and once-a-week session-limit reset. After the first 5-hour usage-wall 429 carrying anthropic-ratelimit-unified-slow-offer: treatment, OmniRoute retries the same account with anthropic-usage-limit: slow and keeps sending it until the window resets — the account keeps serving past the limit instead of being cooled down (slot_busy/529 wait the server's slow-retry-after, bounded by slow-max-wait). With auto-reset on, the wall first tries POST /api/organizations/{org}/reset_rate_limits (juniper_tide) and retries at full speed when the server grants it. Both default off; nothing is sent before the limit is hit. (#13074) — thanks @davidebaraldo
  • feat(providers): update Fish Audio for S2.1 Pro Free, validated advanced TTS controls, and provider-scoped persistent voice-clone management. (#13090) — thanks @Bl0ck154
  • feat(usage): redeem GLM Coding Plan Reset Cards (glm / glm-cn / glmt / zai) from the Provider Limits UI — clear an exhausted 5-hour or weekly coding-plan window before it rolls over, via the new /api/usage/glm-reset-card route (GET lists, POST redeems). List and redeem requests egress through the connection's proxy and honor exclusive-lease isolation; z.ai's requestId is reused for retries of an ambiguous (transport-failed) redemption so a lost response cannot double-consume a card (in-memory, best-effort — restart the server and a fresh key is required). Responses are validated fail-closed (HTTP 200 alone is never treated as success), unavailable/expired cards are filtered and the list is sorted by earliest expiry, and the post-redemption quota refresh is best-effort: a refresh failure still reports the successful reset. (#12754) — thanks @insoln
  • feat(routing): self-hosted unified OpenAI-compatible entry (/v1/chat/completions). When OMNIROUTE_SELF_HOSTED_PROVIDERS (inline YAML) or OMNIROUTE_SELF_HOSTED_PROVIDERS_FILE is set, the existing /v1/chat/completions route diverts through the self-hosted provider adapters (open-sse/services/providerAdapters.ts) — OpenAI / Anthropic / local-compatible — instead of the cloud pipeline. Provider is auto-routed via the x-omniroute-provider header, a provider/model (or provider::model) model prefix, or the first configured provider; upstream credentials stay runtime-only and are stripped from echoed responses. Optional OMNIROUTE_SELF_HOSTED_API_KEY guards the entry with Authorization: Bearer (reserved for the D5 quota-key system); unset = open loopback/trusted-network route. Upstream failures return the standard OpenAI error shape (including a normalized 502 for unreachable providers). One OpenAI SDK snippet can now traverse multiple self-hosted providers without changing the client. (#RIC-738 / RIC-697 D4) (#13611) — thanks @luyuehm
  • feat(routing): deterministic routing strategies for the self-hosted entry (strategy: block, M2/RIC-740). The unified /v1/chat/completions entry (RIC-738) now accepts a declarative strategy: block — inline in the providers YAML or via OMNIROUTE_SELF_HOSTED_STRATEGY / OMNIROUTE_SELF_HOSTED_STRATEGY_FILE — expressing five explainable, non-predictive routing policies: blacklist / whitelist (hard filters), cooldown circuit breaker (consecutiveFailures + cooldownMs), cost-priority (cheapest costPer1MInput first), latency-aware (fastest recent average first), and an explicit fallbackChain order. 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 x-omniroute-route-decision — the one-line "why this model / why not that one" audit trail (D3). 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; malformed strategy config returns 500 rather than silently becoming a no-op. (#RIC-740 / RIC-697 D3) (#13611) — thanks @luyuehm
  • feat(dashboard): add sidebar pinned items shortcut section with individual item pin toggle and localStorage persistence (#12891) — thanks @ZaimMarzuki
  • feat(providers): share the existing api.x.ai/v1/models discovery config with xai-oauth so SuperGrok OAuth connections pick up new Grok ids without a registry seed edit. (#13518) — thanks @HouMinXi
  • feat(core): improve observability for dual-auth fallback execution (#11828) — thanks @raheemuddin786
  • feat(chat-admission): expose admission tunables via dashboard settings (#12038) — thanks @oyi77
  • feat(dashboard): show exact token counts on hover in usage analytics cards and tables (#12553) — thanks @ZaimMarzuki
  • feat(providers): add xKiro (#12648)
  • feat(reasoning): opt-in min output budget floor for thinking models (#12742) — thanks @patrykkopycinski
  • feat(sse): retry transient 5xx backend errors with jitter (global-fallback call site) (#12695) (#13143) — thanks @KooshaPari
  • feat(services): add open-wa as a 6th embedded service (#13222) — thanks @birdleandro-bit
  • feat(providers): update Openference free models and add Deyin to compatible agents (#13378) — thanks @AnhLead
  • feat(docker): add self-host compose + 5-minute deploy doc (RIC-739) (#13639) — thanks @luyuehm
  • feat(antigravity): expose physical send telemetry (#13659) — thanks @domenicomassafra
  • feat(i18n): blocking key-completeness gate — every locale carries every en.json key (#13827)
  • feat(i18n): new-key gate rejects MISSING markers; skills translate new keys in parallel (#13996)
  • feat(mitm): dynamically inject configured models into Antigravity model catalog (#14006) — thanks @steve25060
  • feat(cache): configurable dual-layer semantic caching with Redis/In-Memory vector stores (re-land of #12630) (#14159) — thanks @BillyOutlast
  • feat(providers): add Lyceum (lyceum.technology) as an OpenAI-compatible, pay-per-use provider — chat, embeddings, and live /models discovery through https://api.lyceum.technology/openai/v1, plus a credit-balance quota fetcher (GET /api/v2/external/billing/credits) surfaced in Dashboard › Limits and quota-aware preflight (#12470). (#12474) — thanks @PixmaNts
  • Add a bounded, request-correlated Auto candidate evaluation trace so pre-dispatch exclusions can be diagnosed without changing routing behavior or recording request content. (#14444) — thanks @keeltrace
  • feat(dashboard): combo editor shows the effective round-robin sticky limit and its source, a connection-aware expansion inherit/on/off toggle, and a prompt-cache locality hint on pinned RR/weighted steps; AUTO-COMBO docs cover the sticky batch and the single dynamic-account recommendation. Save keeps stickyRoundRobinLimit and connectionAwareExpansion when the current strategy has no editor for them, matching sanitizeComboRuntimeConfig (#13370) (#14014) — thanks @HouMinXi
  • feat(dashboard): request timeline plots MCP tool invocations from mcp_tool_audit next to LLM calls, with a per-API-key filter covering both row kinds. GET /api/mcp/audit (and /stats) is exempt from the local-only MCP prefix so a tunnel-served dashboard can poll it; SSE/stream stay loopback-only. Audit rows stamp the caller's API-key id from the HTTP principal, not OMNIROUTE_API_KEY_ID. (#13941 - refs #13898) — thanks @HouMinXi
  • feat(api-manager): collapsible and searchable connection groups with provider select-all in permissions modal (#13952) — thanks @fouadSalkini
  • feat(plugins): the editor plugin registers the host sdk telemetry hook behind an opt-in flag that defaults to off, marking matching inference calls without touching anything else. (#14235) — thanks @maxmad64bis
  • feat(kiro): expose Kiro's provider-native Opus 5 Max effort tier — <base>-max in the Claude effort catalog, max in the Kiro effort values, and the adaptive-thinking envelope for claude-opus-5 (#14284) — original change by tarciorick, thanks @bufftop25
  • feat(sse): forward the auto mode classifier beta (dangerous-tool-use-2026-09-03) to Anthropic-format upstreams, and let anthropic-compatible-* providers forward client-negotiated betas at all, so Claude Code sessions behind the gateway stay eligible for server-side auto mode classification (#14312) — thanks @dpozimski
  • feat(oauth): Muse Code (Meta) device-authorization login with CLIProxyAPI-parity credential lifecycle — RFC 8628 against auth.meta.com, mint the subscription inference key at /muse-code/key, persist the durable dca: token, remint on 401 / DCA-only records, honor error.resets_at quota cooldowns, and import CLIProxyAPI type: meta auth files. Dual-auth keeps pasted META_API_KEY on the same card. (#14329) — thanks @RaviTharuma
  • feat(routing): add opt-in planner/executor orchestration for two-model pipeline combos, with protocol-aware tool-loop detection, dynamic final ownership, native parallel tool calls, and bounded sequential rounds (#14346). (#14616) — thanks @potatosips
  • feat(proxies): show each pool member's last seen egress IP next to the pool totals (#14364) — thanks @maxmad64bis
  • test(cli): add isolated real-binary smoke contracts with local sentinel authentication, response receipts, and bounded process-group cleanup. (#14382)
  • test(cli): allow explicit inheritance of the harness's freshly sanitized environment so metadata controls reach real clients without exposing caller credentials or loader hooks. (#14382)
  • feat(cli): Add remote/context-aware manual model CRUD with metadata validation, dry-run, protected override removal and persistence readback; prefer the public model catalog and keep JSON/JSONL complete. Manual models remain inference-unverified. (#14392)
  • feat(sse): track Tavily API credit usage — tavily-search and tavily providers now read plan and account credit usage from GET https://api.tavily.com/usage, surfacing monthly credits in Dashboard › Limits and avoiding exhausted connections during search rotation (#14442) — thanks @HouMinXi
  • feat(providers): add Cloudflare Workers AI image generation (FLUX.1 Schnell, free tier) — reuses the existing cloudflare-ai chat provider's Account ID + API Token connection with no new credential type (#14443). — thanks @brick30llc-ctrl
  • feat(sse): Context7 rate-limit quota and usage in Provider Limits — reads the ratelimit-* headers of a keyed GET /api/v1/search probe (60 s cache, fail-open) so combo/round-robin rotation can skip exhausted context7 keys (#14454) — thanks @HouMinXi
  • feat(cli-tools): assign models to the review and Mixture-of-Agents roles from the dashboard (#14455) — the Hermes agent already routes work to a dedicated review model and to the MoA reference/aggregator pair, but the config generator had no slots for them, so picking a different model meant hand-editing a file the dashboard would overwrite on the next save. The roles are now first-class and translated across all locales. — thanks @aldoeliacim
  • feat(dashboard): quota pools can now be scheduled. Each pool takes a list of time windows (days of the week + hours, in an IANA zone of its own) saying when OmniRoute may use its accounts: an allow window turns the pool into a whitelist, a block window is a blackout that always wins over an overlapping allow. Inside an open window two independent limits apply, either of which may be left off — keep free, one floor per metered quota window (5h, weekly, …, or any for whichever is most consumed), measured on the whole-account saturation signal so a human burning the same credentials pushes OmniRoute out first; and OmniRoute budget, a requests/tokens/usd allowance measured on OmniRoute's OWN recorded consumption, so somebody else's traffic on the account neither eats nor inflates it. Edited from Costs → Quota Share on each pool card, served by GET/PUT /api/quota/pools/{id}/schedules. (#14466) — thanks @davidebaraldo
  • feat(cli): add a copy-only dashboard companion with local/remote-context runbooks, explicit server-host status and credential-free command previews. (#14494)
  • fix(compression): keep lossy compression off the default path, following HouMinXi's plan (dedup and whitespace stay on; summaries, relevance filters, and style rewrites are per-request) and KooshaPari's #8251 plan (combo 400s are a decision table, so model-scoped wrappers advance). Opt-in header: x-omniroute-compression: allow-lossy, engine:<id>, or a named combo (#14529) — thanks @HouMinXi @KooshaPari — thanks @RaviTharuma
  • fix(resilience): follow HouMinXi's per-connection circuit breaker, so one account's failure no longer opens the breaker for every account on that provider. Proxy failures stay provider-wide. Model-scoped 400 patterns now live in one module, which is KooshaPari's #8251 plan, shared by combo dispatch and account fallback (#14530) — thanks @HouMinXi @KooshaPari — thanks @RaviTharuma
  • feat(providers): add expiry-first, a per-provider account fallback strategy that spends the quota closest to being lost. It ranks each account by how much it must burn per hour to avoid wasting its leftover at the next reset (usable / hoursUntilNearestReset), so a full account whose window closes soon outranks an equally full one that holds for days, while a nearly empty account never wins on its near reset alone. fill-first drains the top-priority account and lets the rest roll over unspent; on a four-account Codex pool that left an 88% account untouched through a reset. Distinct from reset-aware, whose resetUrgency * (1 - remaining) term is a recovery signal and saturates to zero outside the nominal window length. Session stickiness and prompt-cache affinity are untouched; accounts within expiryFirstTieBandPercent rotate least-recently-used. (#14533) — thanks @feci
  • perf(logging): stop walking whole request bodies several times per request when PII sanitization is off: feature flag overrides are cached instead of one SQLite SELECT per string leaf, sanitizePayloadPII returns the payload by reference when the flag is off, pending-request previews are pruned to the preview shape before they are protected (strings are still protected whole and cut last), and the PII masker guardrail no longer JSON-clones every body. On 2 MB agentic bodies at 20 concurrent streams this removed ~12,800 flag lookups per request and cut TTFT p50 from 2.4 s to 1.7 s (#14535) — thanks @dpozimski
  • feat(opencode-plugin): features.diskCacheMaxAgeMs is an opt-in bound for the disk-cache fallback. Unset or 0 stays unbounded. Past a positive bound the snapshot is still served and the fallback log escalates from warn to error (#14540) — thanks @RaviTharuma
  • feat(providers): correlate CLIProxyAPI X-CPA-TRACE-ID auth_index with usage history and the sanitized account-health label (#14544) — thanks @RaviTharuma
  • feat(opencode): bound the Responses headers wait with an opt-in rotation budget (#14558) — thanks @maxmad64bis
  • feat(sse): tell stalled content streams apart from severed transport connections in stored stream error codes (#14582) — thanks @maxmad64bis
  • feat(sse): Codex catalog lists GPT-6 Sol (efforts -low through -ultra) and GPT-6 Luna (-low through -max) with Codex pricing; explicit max effort is no longer clamped to xhigh, and catalog effort tiers for every GPT-6 Codex model follow the live levels (#14677) — thanks @QuangBlue
  • feat(dashboard): Add a unified, searchable model catalog across all providers (#14684) — thanks @Dattuog
  • feat(dashboard): pool egress observation gains a per-exit failure breakdown by error family (#14749) — thanks @maxmad64bis
  • feat(proxy): a local-core entry can steer its selector to the next member when the existing set-aside sets it aside, without restarting the core (#14752) — thanks @maxmad64bis
  • feat(admin): pool members show set-aside state (reason, expected end, repeat count) and preference order (#14754) — thanks @maxmad64bis
  • feat(webhooks): emit proxy.set_aside when a pool member is temporarily set aside, plus proxy.pool.exhausted declarations for wildcard subscribers (#14755) — thanks @maxmad64bis
  • New PROXY_WEBHOOK_REBOUND_MS env (default 5 min, 1s–1h) caps repeat alerts per member so a burst of refusals sends one event (#14755) — thanks @maxmad64bis
  • feat(dashboard): proxy log journal shows received code, exit address, correlation and served account (#14777) — thanks @maxmad64bis
  • feat(calls): record rendered-content presence and usage provenance per request in the call log (#14795) — thanks @maxmad64bis
  • feat(proxy-health): the proxy health cell shows the last sweep verdict with cause and age (#14807) — thanks @maxmad64bis
  • feat(call-logs): show per-request recovery actions in the request log (#14810) — thanks @maxmad64bis
  • feat(proxy-logs): record per-attempt upstream headers duration on proxy log rows, plus opt-in first-chunk duration (PROXY_LOG_FIRST_CHUNK_TIMING), and show them in the proxy log detail (#14892) — thanks @maxmad64bis
  • feat(proxyHealth): a proxy that relays but whose target refuses the egress address now keeps a bounded per-proxy refusal streak (count, first and last seen, last cause and status) across sweeps and restarts, shown in the proxy health cell and on both health routes; a sweep that sees the proxy healthy ends the streak, and entries of deleted proxies are pruned on the next sweep. Observability only — selection is unchanged. (#14938) — thanks @maxmad64bis
  • feat(providers): request Claude's model list with limit=1000, enable xAI API-key model passthrough, default XAI_OAUTH_LIVE_MODEL_DISCOVERY on, and persist autoSync: true on new Grok Build OAuth connections so newly published chat ids can appear without a seed bump. (#14237) — thanks @HouMinXi
  • feat(i18n): New locale — Bosnian (bs) — across the dashboard, docs mirrors, CLI, README and the site; (#14187) — thanks @jasminsehic
  • feat(dashboard): synced/imported models can now carry a manual context-window override. Rows imported through a provider's /models import had no edit affordance at all, so the override #4125 added for custom models was unreachable for them — even though PUT /api/provider-models has always accepted contextWindowOverride for a model with no customModels row, and the runtime already prefers the override. The read path was the other half of the gap: GET /api/provider-models attached overrides only to custom-model rows, so a value stored for a synced model could never be read back. The GET now returns the provider's overrides directly, and the passthrough row renders the same 🪟 badge and inline editor, with #4125 semantics unchanged — a positive whole number sets the override, blank clears it and the model falls back to the discovered value. This is the correction path for third-party OpenAI-compatible aggregators whose discovery metadata understates context length (#14337). (#14356) — thanks @abhisheksharma2411
  • feat(db): per-request cost ledger + per-key tpm/rpm/monthly quota (RIC-741) (#13610) — thanks @luyuehm
  • feat(cli-tools): add Oh My Pi (omp) CLI detection and configuration support (#14178) — thanks @navanshjagetiya7-eng
  • feat(api): per-agent AgentBridge restore-default endpoint (re-land of #13846) (#14730)

🐛 Bug Fixes

  • security(streaming): sanitize generic mid-stream error messages before emitting OpenAI, Responses, or Claude SSE failure frames and before diagnostic logging, while preserving raw failures for internal classification and keeping client disconnects out of provider failure state. (#12457)
  • fix(claude): blockExtraUsage: false no longer lets 5h quota preflight skip the account; extra usage is billed after the session bar is gone, so the request must reach Anthropic (#12803) — thanks @HouMinXi
  • Sanitize HuggingChat conversation-creation and message-send transport failures before they reach client error bodies or provider logs. (#12467)
  • fix(dashboard): Keep local and theme-aware provider SVG icons at a definite layout size so Chromium does not collapse them to 0×0 after the v3.8.50 image-rendering change (#12054) — thanks @ponkcore
  • fix(security): Sanitize provider and runtime failures before public API, SSE and MCP responses and before persistent request, proxy and usage logs, preventing credentials, stack traces and host filesystem paths from crossing those boundaries while preserving stable error codes and useful diagnostics. (#12506)
  • fix(cli): omniroute tunnel create no longer crashes with Cannot read properties of undefined (reading optsWithGlobals) — removed the duplicate positional argument that caused Commander.js to misalign the action callback parameters (#12295) (#12368) — thanks @gonisulaimann
  • Stop advertising Gemini Live-only models as supported audio endpoints until OmniRoute proxies the bidirectional Live protocol. (#11710) — thanks @RaviTharuma
  • fix(kie): reroute flux/kontext off the KIE Market createTask flow — it is catalogued with isMarket: true but has no Market catalog page, so KIE rejected it with "model name not supported"; it now hits the dedicated POST /api/v1/flux/kontext/generate / GET /api/v1/flux/kontext/record-info endpoints instead (#11296). (#11985) — thanks @hartmark
  • fix(kie): correct 12 more KIE Market catalog ids that were sent to createTask unchanged but diverge from KIE's documented upstream model values — GPT Image 2 T2I/I2I (drops the gpt/ prefix), GPT Image 1.5 T2I/I2I (gpt-image/ namespace), Seedream 5.0 Lite T2I/I2I (drops the .0), all 4 Flux 2 variants (flux-2/ namespace, generic variant renamed flex), and Wan 2.7 Image / Image Pro (dash instead of dot) — each verified individually against the literal example request published on docs.kie.ai. #11326's "everything else already matches" claim was wrong a second time (#11296); z-image/4.0-*/z-image/4.5-* and flux/kontext remain open, documented as unresolved in KIE_MARKET_UPSTREAM_MODEL_IDS's comment pending further verification. (#11667)
  • fix(dashboard): useApiKeySave.handleSaveApiKey no longer forces a full upstream /models catalog sync on every non-curated provider connection save — callers can now pass skipModelSync: true to opt out, so a workflow that only wants to add one manual model no longer floods the provider's available-models list with hundreds/thousands of synced entries. The flag is a client-side intent signal only and is stripped before the connection payload is POSTed to /api/providers; default behavior (full sync on save) is unchanged when the flag is omitted (#11324) (#11664)
  • Provider connections: keep tokenExpiresAt when a connection is created. The create-path allowlist omitted it, so every insert stored NULL and the dashboard token badge could read a fresh connection as expired until its first background refresh (#11368). — thanks @ntdatt812
  • fix(combos): the combo builder's precision-select, global-model-search, and manual-entry flows now serialize a model step's model string using the provider's already-computed routing-alias prefix (e.g. oc/) instead of rebuilding it from the raw canonical providerId, fixing the no-auth "OpenCode Free" provider (opencode) being routed to the unrelated paid "OpenCode Zen" provider (opencode-zen) because opencode doubles as a manual routing-prefix override (#11433). (#11670)
  • fix(radar): the catalog feed cache now keeps generatedAt, the date the feed's data was built, next to fetchedAt, the date this install downloaded it (#11435). The feed schema requires that date and the sync path validates it, but the cache dropped it — so a feed fetched minutes ago and one carrying weeks-old figures looked identical to everything downstream, including the dashboard's "Last fetched" line. getRadarCatalog().meta and GET /api/radar/status now report both dates, the latter as its own field rather than folded into version — and omitted entirely for the offers and intel caches, which keep no build date, where a null would read as "unknown" rather than "never stored". The dashboard still shows only the fetch time; surfacing the build date there needs a new translated label and is left to a follow-up. Rows cached before migration 163 read back as null: unknown stays unknown instead of borrowing the fetch time. The referrals cache has persisted the same date since migration 142. — thanks @maxmad64bis
  • fix(cli): postinstall no longer fills JWT_SECRET and API_KEY_SECRET in the installed package's .env (#11436). .env.example ships both blank on purpose: the server restores them from its durable store, or generates and persists them there, in ensureSecrets(). Pre-filling them defeated that — the file lives inside the package directory, so npm i -g replaced it and postinstall wrote different values, while ensureSecrets() (which only acts on an empty variable) never got to restore the real ones. Both secrets rotated silently on every update, invalidating dashboard sessions and API-key CRCs. STORAGE_ENCRYPTION_KEY left the same list for the same reason in #1622; its comment pointed at a function that no longer exists and now names the real provisioning path. — thanks @maxmad64bis
  • fix(cli): the CLI now says when a loaded .env lives inside the installed package directory (#11437). It already announces every env file it reads, without distinguishing the ones that survive an update from the one that does not: npm i -g replaces the package directory wholesale, so values set there are gone at the next update, silently. The warning names the durable path to move them to, and fires only when that file actually supplied a value — a file entirely shadowed by a durable one supplied nothing. A development checkout stays silent: there the same path is stable and documented in SETUP_GUIDE.md. — thanks @maxmad64bis
  • fix(free-models): the shared isFreeModel() predicate no longer reports catalog entries marked freeType: "discontinued" as free, so hidePaidModels can't route to Pollinations' seven premium models that now require a paid key (#11441) — thanks @maxmad64bis
  • fix(ui): let AnySearch use the normal provider-icon fallback when LobeHub has no matching icon (#11449)
  • Fixed the v3.8.50 Costs and Analytics dashboards so flat-rate Claude Code usage can be shown as an explicitly requested token-price estimate without changing default billed-cost semantics. (#11460) — thanks @xiaoyaner0201
  • Fixed archived usage retention so each request is priced individually instead of pricing a day's summed tokens once, which understated archived cost whenever a day mixed cache-heavy and ordinary requests. (#11460) — thanks @xiaoyaner0201
  • Fixed the Costs dashboard so it discloses when displayed figures include flat-rate token-price estimates instead of labelling them as billed spend, using the flag the analytics API already returns; the month-end projection and the CSV/JSON exports carry the same marker, and billed-cost mode is unchanged. (#11460) — thanks @xiaoyaner0201
  • fix(combo): attach the same combo-diagnostics trace (poolSize/attemptOrder/excluded/terminalReason, plus x-omniroute-combo-* headers) to the round-robin strategy's and the nested pipeline/fusion runtime-unit loop's "Maximum combo retry limit reached" 503 that the priority-strategy path already attaches for the identical terminal condition — previously those two paths returned a bare, contextless 503 (#11462). (#11665)
  • fix(bun): use native bun:sqlite in bootstrap-env and sync-env to avoid loading better-sqlite3 N-API addon during Bun startup (#11468) — thanks @TheDemonTuan
  • fix(docker-bun): make Dockerfile.bun install reliable with pre-install script helpers and native bun:sqlite smoke check (#11470) — thanks @TheDemonTuan
  • fix(docker-bun): remove every vendored better-sqlite3 native addon from the Bun runtime image so startup cannot abort before the bun:sqlite fallback (#11482) — thanks @TheDemonTuan
  • fix(autoCombo): effort/alias model variants (gpt-5.6-sol-xhigh, gpt-5.6, cursor's claude-4.6-opus-high) inherit their base model's task fitness instead of falling to the wildcard 0.5 (#11492) — thanks @MumuTW
  • fix(db): rate-limit repeated Arena ELO leaderboard fetch-failure warnings instead of logging one per sync attempt (#11500) (#11989) — thanks @hartmark
  • fix(db): dedupe the raw [Encryption] Decryption failed... log line emitted by the lazy-decrypt views (createLazyRowProxy/createLazyConnectionView), which power getProviderConnections() and were re-triggering that line on every CredentialHealth/model-sync cycle for the same corrupt or stale-key credential — a fresh Proxy over a fresh row on every cycle meant the per-proxy memoization never suppressed it, unlike the dedup decryptConnectionFields() already had since #9927. Now shares that dedupe tracking so the line logs at most once per credential (#11500). (#11678)
  • fix(intelligence): Arena ELO sync stores a synthesized base-model row for every effort/harness variant (claude-opus-5-max → claude-opus-5, gpt-5.6-sol-xhigh (codex-harness) → gpt-5.6-sol), so bare model requests reach the synced intelligence layer instead of the static fallback table; the hand-maintained MODEL_ALIAS_MAP — which collapsed generations (gpt-5.5 → gpt-5) — is gone (#11506) — thanks @MumuTW
  • fix(autoCombo,sse): catalog hygiene — the static task-fitness table no longer ranks retired models above live ones (dead rows dropped, patterns matched on segment boundaries), the 7 deprecation aliases that forwarded to retired or non-existent models now point at the vendors' published replacements, an alias is no longer rewritten when the serving provider offers the id as-is, and a new check:model-lifecycle gate diffs both tables against a vendor lifecycle snapshot (#11507) — thanks @MumuTW
  • fix(sse): stop dropping resolved thoughtSignature values on parallel (multi tool-call) turns sent to Gemini 3.x — the claude→gemini and openai→gemini translators previously kept the signature only on the first function call of a message, causing Gemini to reject subsequent calls in the same turn with HTTP 400 "Function call is missing a thought_signature"; each function call now keeps its own resolved signature (#11510). (#11669)
  • Treat Docker publishing from a non-default release branch as an expected skip instead of a failed workflow, while preserving the guard that prevents it from overwriting the next channel. (#11527) — thanks @pacocartones
  • fix(search): prefer credentialed providers over duckduckgo-free fallback (#11524) — the fallback-only loop ran before the credentialed-providers loop in executeWebSearch, making configured providers unreachable when duckduckgo-free was available. (#11565) — thanks @vsd2807
  • fix(sse): cap the upstream headers-wait phase for STREAMING requests to a client-realistic ceiling (110s, under Codex's own ~120s hard client-abort window) instead of the flat 10-minute FETCH_TIMEOUT_MS default — that default was 5x longer than the body-phase readiness watchdog's own adaptive bound, so a request whose upstream never returned any response at all (not even headers, e.g. a stalled NVIDIA target behind a tool-heavy Responses→Chat translation) kept the client connection alive on keepalives only, guaranteeing the client's own patience ran out first with an opaque 499 instead of OmniRoute detecting and failing the stall fast. Non-streaming requests are unaffected — they keep the existing flat default (open-sse/utils/fetchStartTimeoutPolicy.ts) (#11526) (#11668)
  • fix(ci): protocol-clients E2E harness boots the peer-stamped custom server (run-next.mjs dev) instead of the bare next CLI, so LOCAL_ONLY locality resolves from the real TCP peer stamp and /api/mcp/audit answers 200 (open bootstrap preserved via an OMNIROUTE_E2E_BOOTSTRAP_MODE=open credential clear in run-next.mjs, pinned to loopback HOST) instead of a deterministic 403 (#11535) (#11549) — thanks @jonlwheat2-gif
  • fix(docs): the Auto-Combo engine is described with the number of factors it actually declares. Four different counts were in circulation — 6, 9, 13 and 14 across AGENTS.md, four guides and six code comments — while DEFAULT_WEIGHTS declares 15. check:docs-counts now reads that number from the source, so adding a factor without updating the prose fails CI (#11545) (#11553) — thanks @maxmad64bis
  • fix(quota-share): in-flight leases are per request, so an aborted request's slot ages out instead of being kept alive by later traffic on the same connection (#11547) — thanks @abhisheksharma2411
  • The free-tier summary route now serves the refreshed Radar catalog when the feed is active — the same numbers the dashboard shows — and states which catalog answered plus its real build date, instead of always reporting release-frozen figures with a stale curation date. The supporter-key live feed stays reserved to authenticated callers of the instance. (#11550) — thanks @maxmad64bis
  • sse: fix LiveWS/embed-WS servers crashing at startup under the Node/tsx runtime — liveServer.ts, embedWsProxy.ts and apiBridgeServer.ts imported @/shared/utils/httpClientAbortGuard without the .mjs extension, so the client-abort crash guard added by #11556 was unreachable and every dependent test failed with ERR_MODULE_NOT_FOUND (#11556). (#11671) — thanks @jonlwheat2-gif / @yxyxy
  • fix(chat): shed any of the matched provider-node's routing identifiers (public prefix or internal id) before model lookup, so composites like <connId>/<connId>/<model> no longer reach upstream verbatim (#11557) — thanks @maxmad64bis
  • fix(cache): a semantic-cache entry now expires on its TTL instead of surviving until the next UTC midnight, and dbEntries stops counting expired rows (#11573) — thanks @ntdatt812
  • fix(api): /v1/models schedules its stale-while-revalidate rebuild through Next's after() again, so a stale catalog reaches the client before the rebuild blocks the event loop (#11574) — thanks @ntdatt812
  • fix(check): the known-symbols executor conformance gate awaits the now-async getExecutor(), so it stops reporting all 142 registered aliases as dead symbols and can detect a lazy import that fails (#11582) — thanks @ntdatt812
  • fix(pack): register bin/cli/utils/volatileEnvPath.mjs in PACK_ARTIFACT_REQUIRED_PATHS, so the prune can no longer silently drop a module the published CLI imports on every boot (#11588) — thanks @ntdatt812
  • i18n(vi): translate the five reliability-column strings on the free-provider rankings page, which rendered as __MISSING__: markers in the Vietnamese UI (#11589) — thanks @ntdatt812
  • resilience: restore the expired-connection retry-budget probe in the token-health sweep — the !isGitHubAccessTokenOnlyConnection carve-out reintroduced by #11608 contradicted the boundary pinned by #11592, so a GitHub connection parked at expired with retry budget remaining was never probed and could never self-heal (#11592). (#11672) — thanks @jonlwheat2-gif / @ntdatt812
  • fix(dashboard): Provider Stats now displays combo success rates on their native 0–100 percentage scale instead of multiplying them twice (#11595) — thanks @pacocartones
  • fix(providers): route DuckDuckGo Web Chat through the verified duck.ai endpoint, refresh its free model catalog, and avoid forwarding unsolved anti-abuse challenges (#11597) — thanks @solstxce
  • fix(ui): Console log Refresh and Copy controls now expose localized accessible names, keep copy actions visible on keyboard focus, and announce copy completion safely (#11599) — thanks @pacocartones
  • Dropped the stale-.eslintcache restore-keys fallback from both "Restore ESLint file cache" steps in ci.yml, so the blocking Lint job can no longer be served per-file verdicts computed under a different lint config, suppressions file or lockfile. quality.yml had already dropped it in #11963; ci.yml — the workflow that actually gates PRs — had not (#11600). (#11996)
  • fix(ci): pass --pass-on-unpruned-suppressions in run-eslint-json.mjs so the CI Lint job no longer fails when a suppression is merely orphaned by a genuine fix, mirroring the identical fix already in validate-release-green.mjs (#11600) (#11983) — thanks @hartmark
  • fix(handoff): enforce provider allowlist for universal handoff (#11602) — universal handoff now skips summarization when the selected summary model's provider is not included in the configured provider allowlist. (#11619) — thanks @vsd2807
  • fix(dashboard): Show a stable error state when Search Analytics returns an HTTP error or malformed data (#11603) — thanks @pacocartones
  • test(gamification): pin the aggregate profile level to the XP-derived semantics of #11604 — getAggregateXp() now derives currentLevel from the summed XP (calculateLevel(sum)), not MAX(stored current_level), and the #3484 fixture levels are aligned with the XP curve (#11604). (#11676) — thanks @jonlwheat2-gif / @pacocartones
  • fix(dashboard): Keep the Profile level and progress aligned with aggregate XP, including bounded handling for invalid totals (#11604) — thanks @pacocartones
  • fix(dashboard): Prevent locked hidden badges from revealing their icon or opening private badge details before they are earned (#11605) — thanks @pacocartones
  • fix(dashboard): Restore keyboard focus after shared modals close and cancel delayed autofocus during cleanup (#11607) — thanks @pacocartones
  • fix(dashboard): Enable Enter and Space activation for clickable data-table rows without hijacking nested controls (#11610) — thanks @pacocartones
  • fix(autoCombo,sse): vendor-retired catalog ids are dropped from the auto-combo candidate pool and no longer win on leftover arena_elo / user_override rows; getModelLifecycleDecision consults model-lifecycle.json (prefix-stripped) so aggregator traffic is not untracked for ids the snapshot already knows (#11625) (#11626) — thanks @MumuTW
  • test(chatcore): move Codex/Claude combo fixtures off models the lifecycle guard now rejects — gpt-5.1-codex/gpt-5-codex are vendor-retired (snapshot, #11626) and claude-3-5-sonnet-20241022 is shut down, so native-passthrough and combo-fallback tests switched to gpt-5.6-sol and claude-sonnet-4.6 (#11626). (#11675) — thanks @jonlwheat2-gif / @MumuTW
  • fix(opencode-plugin): stop a free-tier budget that rounds up across a magnitude from rendering as 1000K/1000M in the model picker — fmtTokens chose its unit from the raw token count and then rounded with toFixed(1), so 999,950–999,999 printed as 1000K rather than 1M and just under a billion printed as 1000M rather than 1B (#11684) — thanks @abhisheksharma2411
  • Cap prose-derived Antigravity quota resets at the configured model cooldown maximum while preserving authoritative Retry-After headers and Google RetryInfo hints. (#11695) Thanks @pacocartones. (#11823)
  • Compare _rebaseline_* notes structurally in the ratchet-bank verifier so an object-valued note no longer reports itself as rewritten on an unmodified tree — the nightly bank-ratchet-shrinks lane had been aborting on every run since 2026-08-11, leaving 41,672 lines of earned file-size shrinks unbanked (#11702). (#11892) — thanks @MumuTW
  • fix(build): prepublish.ts bundles the ChatGPT Web (Codex) MCP bridge through runBuildTool() instead of spawning npx.cmd raw, fixing the build crash on Node ≥ 20/Windows where .cmd shims cannot be spawned without a shell (EINVAL) (#11704) (#11706) — thanks @jonlwheat2-gif
  • fix(codex): keep parallel_tool_calls:false on the translated Codex Responses Lite path (#11707) (#11984) — thanks @hartmark
  • fix(tests): the new Gemini Web Images retirement test (#11708) called the now-async getExecutor() synchronously, throwing TypeError: getExecutor(...).getProvider is not a function instead of asserting Gemini Web chat stays available. Awaited to match the R0.3 executor-registry refactor (#11220). (#11868)
  • fix(search): Allow REST and MCP web-fetch requests to select the advertised anysearch-search provider explicitly (#11716) — thanks @pacocartones
  • fix(providers): Keep overnight peak-hour protection active after midnight according to the window's configured start day (#11718) — thanks @pacocartones
  • fix(providers): correct all five g4f.space gateway descriptions to document proof-of-work credits or member API keys, variable limits, and the remote prompt boundary instead of promising free keyless access at five requests per minute (#11738).
  • fix(db): model_capabilities is created by a migration instead of lazily on the first models.dev sync, so a clean install and an upgraded install converge on the same schema regardless of which features have run (#11845)
  • fix(ci): check:install-upgrade now fails on an npm install truncated by ENOSPC (npm reports it as a warning and still exits 0), authenticates its health probe so the version assertion works against the hardened health payload, frees the clean-install tree before the upgrade phase, and no longer reports a schema divergence computed from a boot that never served (#11845)
  • fix(routing): harden session-affinity key derivation against oversized explicit IDs and empty structural-payload collisions, while recognizing Gemini and common text request shapes (#11746) — thanks @MumuTW
  • fix(packages/browser-pool): regenerate package-lock.json so the packages/browser-pool workspace's locked playwright/@types/node (and transitives) match its package.json specs, fixing cache-only/offline installs (npm ci --offline, Nix buildNpmPackage) that previously failed with ENOTCACHED (#11747) — thanks @benjaminkitt (#11784) — thanks @geekyNads
  • fix(assets): Stop distributing 79 provider and CLI logos without verified source/license provenance, keep every provider/tool available through the internal generic icon, and prevent those IDs from falling through to an external icon CDN. (#11750)
  • fix(providers): retire the common chatgpt-web / cgpt-web provider because its pre-key and proof-of-work provenance could not be cleared (#11754) — explicit chat and image requests, slash-prefixed models, persisted aliases, proxy overrides, and combo targets now fail closed with a sanitized HTTP 410 / PROVIDER_RETIRED before upstream dispatch. Migration 163 tombstones matching connections and invalidates their active leases without deleting history or widening API-key allowlists. The independent chatgpt-web-codex / cgpt-codex provider and its MIT-noticed vendored implementation remain available.
  • fix(usage): quota and usage refresh no longer 409 when an exclusive lease reserves the connection (#11758) — thanks @TheDemonTuan
  • Keep the embedding registry's vector width and embedding type on models when a synced model exists for the same id, so /v1/models no longer reports registry-described embedding models widthless or untyped (#11761) — thanks @rahilmavani
  • Correct google/gemini-embedding-001 on the OpenRouter route to 3072 dimensions, the width it returns when dimensions is not sent (#11761) — thanks @rahilmavani
  • fix(vision-bridge): Restrict automatic vision routing and fallbacks to models available in authoritative live provider catalogs, including aliases and effort variants (#11767) — thanks @pacocartones (#11813)
  • fix(providers): reclassify Cerebras as a one-time $5 signup credit (payment method required, 30-day validity), not a recurring no-card 1M tokens/day trial (#11773) (#12591) — thanks @HouMinXi
  • fix(cli): stop prepublish from re-rebuilding the already-built ESM-only opencode-plugin dist (#11787) (#11990) — thanks @hartmark
  • fix(db): Fresh installs no longer log a non-fatal no such table: compression_run_telemetry warning when retention cleanup runs before the lazily-created telemetry table exists (#11802) — thanks @RaviTharuma
  • fix(sse): set X-OmniRoute-Selected-Connection-Id on successful combo dispatches so downstream consumers stop falling back to an empty connection id (#11810) (#11986) — thanks @hartmark
  • fix(services): embedded CLIProxyAPI lifecycle checks now use public /healthz, while model discovery uses the configured dedicated data-plane API key instead of the management password (#11811) — thanks @RaviTharuma
  • Fix persisted model steps returning an error in the playground route simulator, and warn when structural steps cannot be simulated. (#11862) — thanks @pacocartones
  • fix(providers): Antigravity's dynamic mitmAlias table no longer routes gemini-3.7-flash-{high,medium,low} to a literal tier-suffixed upstream id just because one connected account's own discovery listed it directly — those display ids always resolve through the safe gemini-3.7-flash-tiered static alias, so one account's Google-provisioned access no longer 404s every sibling account of the provider (#11824, #11651) (#11988) — thanks @hartmark
  • fix(dashboard): Expose Profile loading, errors, page structure, and XP progress to assistive technologies (#11838) — thanks @pacocartones
  • fix(socks): Agent.connectTimeout now bounds both SOCKS handshake and TLS connect, and family === null no longer falls back to fetch-socks (#11842) — thanks @maxmad64bis
  • perf(compression): OOM mitigations for large payload hashing, memoization, and token estimation (#11844 — thanks @AndrianBalanescu)
  • Excluded Next.js Node File Trace manifests (*.nft.json) from the published npm tarball. They are build-time metadata and are never read while serving, but had grown to 668.7 MB — 61% of the package — which pushed the upload past the registry limit and made npm publish fail with 413 Payload Too Large. (#11864)
  • fix(cli): keep the global OmniRoute server --base-url separate from provider-node --endpoint values in add, update, and validate commands (#11860) — thanks @pacocartones
  • fix(config): Nous Research's Hermes-4-405B model now displays as "Hermes 4 405B (Nous Research)" in both the provider registry and the free-model catalog, instead of the mislabelled "Hermes 4 7B" (#11861) — thanks @Karan825 (#11993)
  • fix(provider/nous): inject required user= tag into Nous Research inference requests to resolve upstream 400 "missing tags" error (#11861) — thanks @Karan825 (#12044)
  • fix(build): npm run build now fails in one second with a named package and a copy-pasteable fix when npm silently drops an externalised optional native dependency, instead of dying four minutes in with Module not found: Can't resolve 'better-sqlite3' (#11863) — thanks @ujjawalkaushik1110
  • fix(install): postinstall no longer throws ReferenceError: isAndroid is not defined — failing the whole npm install — when the better-sqlite3 rebuild fallback times out; the manual-fix guidance is reachable again (#11863) — thanks @ujjawalkaushik1110
  • Fixed the Alibaba free-tier allowlist test that went red on its own once the shipped catalog's validUntil (2026-08-27) passed, leaving every PR and main with a failing Unit Tests (1/8). The test now builds its own packs with dates it controls, and covers the expired-pack fallback that production has actually been serving. (#11867)
  • Split the npm registry upload into its own GitHub-hosted job. npm refuses --provenance from a self-hosted runner (422 ... Only "github-hosted" runners are supported), which blocked the v3.8.50 publish; the heavy verification cannot move to a hosted runner, so it now hands the proven tarball over instead. (#11868)
  • fix(ui): the shared DataTable loading state no longer reads its decorative ⏳ glyph out to assistive technology, and now carries the same role="status" / aria-live="polite" / aria-busy="true" semantics as PageLoading (#11869) — thanks @pacocartones
  • fix(api): Bound the timeout accepted by POST /api/providers/[id]/login to the same 15s–600s window the Adobe Firefly and Conol browser-login paths already enforce, so a request can no longer pin a headful browser and the single active-login slot with an arbitrarily large poll budget. (#11871) — thanks @pacocartones
  • fix(api): Stop the tunnel and MITM routes from returning raw child-process error text, which disclosed host paths, binary install locations and Tailscale tskey-* credentials in a body some of these routes serve to non-loopback callers, and return a real 400 instead of a framework 500 when a tunnel request body fails validation. (#11872) — thanks @pacocartones
  • fix(open-sse): stream TTFT/ITL are sampled from a monotonic clock (performance.now) instead of Date.now, so an NTP correction or wall-clock jump can no longer inflate a genuine TTFT or produce a negative one that the ttft >= 0 guard silently drops — protecting the router's OTel/EWMA/usage_history/speed-ranking signals (#11873) — thanks @pacocartones
  • fix(providers): xAI reasoning_effort: "xhigh" now reaches grok-4.6+ instead of being silently clamped to "high" (#11879) — thanks @NoxzRCW
  • fix(dashboard): the providers page no longer crashes into the error boundary when a provider id collides with an Object.prototype member (constructor, __proto__); icon lookups are own-property guarded (#11880) — thanks @NoxzRCW
  • fix(skills): injected skill tools declared in shorthand ({"content": "string"}) now forward valid JSON Schema, unblocking providers that validate tool schemas strictly such as Zhipu GLM on the Console Go tier (#11881) — thanks @NoxzRCW
  • fix(api): POST /api/playground/simulate-route now surfaces combo-ref and provider-wildcard persisted combo steps with a specific warning (naming the referenced combo, or the unresolved provider/model wildcard) instead of folding them into a generic "unsupported step" count; a provider-wildcard step is also now included as an unresolved target so the operator can see it is in the route (ported from #11882 — thanks @NoxzRCW). (#11930)
  • fix(translator): the streaming OpenAI→Claude translator keeps upstream usage, including prompt-cache tokens, when it arrives on a trailing choices: [] chunk (Fireworks and any upstream using stream_options.include_usage) (#11883) — thanks @NoxzRCW
  • fix(cli): omniroute update --apply now tells you explicitly whether a running server was detected and, if so, that you must run omniroute restart to apply the update — it never restarted anything and previously implied the update was already live once files were installed. The dashboard's npm-mode Update flow (/api/system/version) now tries OmniRoute's own PID-file-managed supervisor before falling back to pm2, and reports an honest "restart required" step instead of a silent pm2-only "skipped" that read like a completed update. The server-side latest-version lookup backing the dashboard's update banner also gained --prefer-online, closing the same stale-npm-cache class already fixed in the CLI's own copy for #4376 (#11885). (#12005)
  • fix(resilience): clear persisted LKGP pins when a target suffers connection/provider exhaustion or is skipped before dispatch due to cooldown/exhaustion/unavailability, preventing subsequent requests from repeatedly prioritizing known-dead providers (#11911). (#12013) — thanks @HouMinXi
  • fix(sse): the OpenAI→Claude stream translator now defers the terminal message_delta/message_stop emission until the real usage block has arrived (or a genuine end-of-stream flush forces it) instead of emitting it immediately on finish_reason — previously, when the trailing usage-only chunk ({"choices":[],"usage":{...}}) arrived after the finish_reason chunk (the normal order for Fireworks/vLLM/Together and other stream_options.include_usage upstreams), the client-visible message_delta still carried stale/zero usage even though state.usage was internally corrected too late to matter (ported from #11915 — thanks @HouMinXi). (#11933)
  • fix(sse): stripStore() now forces store=false for stateless OpenAI-compatible Responses-API targets (unless the connection has explicitly opted in via providerSpecificData.openaiStoreEnabled), instead of only handling the openai/agentrouter cases — a client-supplied store value previously passed through untouched to backends that don't actually persist responses server-side (#11916 — thanks @HouMinXi).
  • fix(models): custom provider-node models (synced, custom, and alias-backed) now appear under their configured prefix in the unified catalog when the operator's model-id prefix mode is canonical, instead of being dropped whenever alias-inclusion was otherwise disabled (#11918 — thanks @HouMinXi).
  • fix(models): the unified model catalog now suppresses stale static registry models (including effort-tier variants) for any provider whose active connection has an authoritative live synced catalog, not only providers already using exclusive-synced-listing — a connection with providerUsesAuthoritativeLiveCatalog previously kept serving both the live-synced models and the stale static rows side by side (#11919 — thanks @HouMinXi).
  • fix(ui): ProviderIcon's three remaining unguarded lookups (PROVIDER_ICON_ALIASES, LOCAL_SVG_ALIASES, THEMED_SVGS) now use Object.hasOwn() own-property checks — a provider id such as constructor or __proto__ previously resolved through the prototype chain instead of falling through to the unknown-provider CDN fallback (getLobeProviderIcon() itself was already guarded by #11880); ProviderPageHeader's color field is also now optional, matching the rest of the component's defensive typing (ported from #11920 — thanks @HouMinXi). (#11935) — thanks @NoxzRCW
  • fix(ollama): preserve multi-byte UTF-8 content split across stream chunks in the Ollama NDJSON transform, which previously corrupted CJK/emoji into U+FFFD (#11921) — thanks @pacocartones
  • fix(providers): OrcaRouter chat requests now target /v1/chat/completions instead of the bare /v1 API root, fixing the upstream 404 Invalid URL (POST /v1) (#11923). — thanks @echel0nn
  • fix(plugins): deliver the onStreamComplete event to disk-installed plugins. The event shipped in v3.8.50 (#9669) was emitted internally but had no plugin-facing wiring, so no plugin could ever subscribe: the manifest schema silently dropped hooks.onStreamComplete, and the loader/manager only knew the seven legacy hooks. onStreamComplete is now a declarable manifest hook, wired through the loader and registered by the manager like the other hooks, and its payload carries a requestId so consumers can correlate the stream-completion event with the originating request (#11934) — thanks @amaleta / @pacocartones
  • fix(db): the Qdrant embedding-model dropdown now lists local/self-hosted providers (Ollama, LM Studio, vLLM, etc.) — an active connection is treated as "configured" when the provider allows an optional API key, not only when it has a real key or OAuth, so a running local embedding provider is no longer hidden from the picker (#11949) (#11995)
  • fix(providers): Vertex AI Anthropic partner-model discovery now calls the Model Garden v1beta1 publisher list (/v1beta1/publishers/anthropic/models, global) and parses its publisherModels envelope, so Claude models auto-synced from Vertex populate the active live catalog and route at request time instead of returning Model '<id>' is not available in the active live catalog (#11991) — thanks @fabioluissilva (#11998)
  • fix(cli): support --base-url alongside --endpoint in omniroute nodes add, update, and validate subcommands to prevent global --base-url shadowing issues (#11999). (#12033) — thanks @HouMinXi
  • fix(providers): cloudflare-ai no longer refuses image content parts for every Workers AI model (#12002) — the plain-string content requirement behind #2539 is carried by the model schema, not by the /ai/v1/chat/completions endpoint (measured: an all-text part array returns 200 on @cf/mistralai/mistral-small-3.1-24b-instruct, @cf/meta/llama-4-scout-17b-16e-instruct and @cf/meta/llama-3.3-70b-instruct-fp8-fast, and 400 on the text-only @cf/qwen/qwen2.5-coder-32b-instruct). transformRequest() flattened every array and threw on the first non-text part (#6390), so image input was refused for vision-capable Cloudflare models that accept it. All-text arrays are still flattened — the one shape every model accepts — while an array carrying a non-text part is passed through untouched, so the attachment is still never silently dropped. Regression guards: tests/unit/cloudflare-ai-image-parts-6390.test.ts. — thanks @davidlinfr
  • fix(resilience): decouple the limiter-managed execution backstop from the queue-wait budget — new requestQueue.executionMaxWaitMs (env RATE_LIMIT_EXECUTION_MAX_WAIT_MS, default 600000 = 10 min) now feeds Bottleneck's post-dispatch expiration, while requestQueue.maxWaitMs keeps its documented queue-wait semantics. Previously the queue-wait budget doubled as the execution expiration, so legitimate long-running calls on non-incremental gateways (whole generation buffered before the first upstream byte, e.g. Console Go / Command Code tiers serving GLM models) were killed mid-flight at the queue budget with a false 504 RATE_LIMIT_EXECUTION_TIMEOUT — the local limiter undercut the provider-aware upstream fetch-start timeouts. The surfaced 504 message now names requestQueue.executionMaxWaitMs; the error keeps the #4165 guarantees (disclaims an upstream timeout, preserves the Bottleneck error as cause, branded code + trusted provenance, classified request-scoped so combo falls back). A real queue-wait bound (the Promise.race around limiter.schedule() sketched in #9533) remains future work. (#12025) (#12027) — thanks @alvinveroy
  • Call logs: keep the error field when an artifact exceeds the storage cap, instead of replacing it with the omission marker. The error is the only field that says why a request failed and is typically ~90 bytes next to the multi-hundred-KB bodies that trip the cap, so dropping it left a size-limited row undiagnosable — a provider outage, a local timeout and an upstream 400 all rendered identically. It is now preserved at every fallback stage, truncated to 4KB if it is itself large (#12026). (#12095) — thanks @ntdat812
  • fix(diagnostics): preserve the error field (truncated to 4KB with a [truncated: …] suffix) in every call-log artifact size-limit fallback stage. Previously the minimal fallback replaced the error with [omitted: call log artifact size limit exceeded], so an oversized artifact row showed nothing about WHY the request failed — e.g. 91 of 847 opencode-go 504 rows on one production instance were undiagnosable from the dashboard. Oversized request/response bodies are still omitted exactly as before; the error cap is independent of the payload sizes that tripped the fallback. (#12026) (#12027) — thanks @alvinveroy
  • fix(sse): OpenAI Responses clients that declare the native web_search tool now receive a spec-shaped web_search_call output item with action.sources alongside the preserved function-call round-trip, so search results executed through OmniRoute's own search backend are consumable by standard Responses clients (Codex, pi-web-access, …). (#12031) — thanks @watchingdogs
  • fix(cli): use in-thread alias resolver hooks on modern runtimes to avoid deprecation noise and improve Node.js forward compatibility (#12073). (#12083) — thanks @xiaoyaner0201
  • fix(combo): an operator-set Agent Features → Context length on a combo is now honored at request time. The value was persisted and advertised through /v1/models, but resolveComboContextLimit() never consulted it — so a multi-target combo whose members carry no per-model window fell through to the provider's generic defaultContextLength (openrouter 128000, command-code 200000) and rejected large requests with Input exceeds context window … limit 128000 despite the combo being explicitly sized much larger. An identical single-target combo worked, because it collapses to its concrete target before the guard runs. Invalid values (0/negative/NaN/Infinity) are ignored, so the existing target → combo-min → fallback order is unchanged. (#12090) — thanks @adivekar-utexas
  • fix(sse): passthrough streams now estimate usage on finish when upstream closes without usage even with stream_options.include_usage — avoids 0 tokens / 0% for providers that stay silent (and correctly handles trailing empty-choices usage) (#12151) — thanks @maxmad64bis
  • fix(combos): clearing an agent feature in the combos editor now persists — unchecking context cache protection, or emptying the system message or tool filter, sends an explicit null instead of dropping the field from the PUT body, which the update merge read as "leave unchanged" (#12177) — thanks @foreveryh / @ntdatt812
  • fix(memory): self-hosted embedding endpoints now vectorize — the vector width is measured from the first embedding that comes back instead of being read from a registry that cannot describe them, so vec_memories is created and memories stop piling up unvectorized behind a green health check (#12180) — thanks @kanade-hoshino / @ntdatt812
  • fix(translator): the leading system message now reaches Responses-API upstreams when its content is a content-part array — it was read as typeof content === "string" ? content : "", so a prompt-caching client (Anthropic cache_control, the shape LiteLLM and the Anthropic SDK emit) had its entire system prompt replaced by an empty instructions. The request was still accepted with a normal prompt_tokens count, so the model answered with no instructions and nothing in the response said they were missing. Mid-conversation system turns already handled the array shape (#7056); only the first one did not (#12206). Regression guard: tests/unit/translator-openai-responses-system-content-parts.test.ts. (#12207) — thanks @Gorillaz322
  • fix(free-tier): /api/free-tier/summary no longer computes its totals from a Radar feed built before the catalog the running release ships. When the cached feed is older — or carries no build date at all — the route answers from the shipped catalog, resolved through the operator's local model state so disabled and tombstoned models stay out of the numbers (#12215). — thanks @maxmad64bis
  • fix(oauth): Keep a Claude personal workspace and a Team organization as separate connections — they share the same email and accountUUID, so the email-only OAuth dedup let the second login overwrite the first account's tokens; organizationUUID now disambiguates them, the way workspaceId does for Codex (#12222) — thanks @davidebaraldo
  • fix(resilience): a 402 on a single paid model of a passthrough/gateway provider (e.g. kilo-gateway, ollama-cloud) no longer terminalizes the whole connection with a never-auto-recovered credits_exhausted status — only the paid model is locked out, so free models on the same key keep serving. 402 variant of #3027. Single-credential providers are unaffected — a 402 there is still treated as the key being genuinely out of credit (#5239 / #10616) (#12242) — thanks @brick30llc-ctrl (#12266) — thanks @geekyNads
  • fix(sse): trust finish_reason: "length"/"max_tokens" over the reasoning-consumed-token ratio in response quality validation, so a reasoning model truncated below the old 90% threshold correctly fails and retries instead of returning empty content as a silent "success" (#12262) — thanks @brick30llc-ctrl
  • fix(settings): PUT /api/settings/cache-config now persists alwaysPreserveClientCache to the flat general settings the runtime cache-control policy actually reads; previously the value landed in the databaseSettings "cache" section and was silently ignored, so the endpoint had no effect on cache_control passthrough (#12304) — thanks @davidebaraldo
  • fix(grok-cli): treat omitted SuperGrokPro creditUsagePercent as 0% used so Provider Limits still renders a weekly bar (proto3 zero-elision) (#12312) — thanks @HouMinXi
  • fix(combo): Expose the two Auto-Combo scoring factors nobody could set — the combo validation schema and the dashboard weight sliders both declared 13 of the scorer's 15 factors, so connectionDensity (spreads load across a provider's connections) and quality were dropped on save and offered nowhere. The sliders also shipped their own default table that differed from the engine's on every non-zero factor and summed to 1.05, so the percentages shown next to them added up to 105%. Both lists now match DEFAULT_WEIGHTS, and a test keeps them there. Note that a combo whose stored weights omitted the two keys was effectively running with them at zero and the other thirteen renormalized upward; it now runs with the engine's intended distribution, so its routing does shift (#12314) — thanks @maxmad64bis
  • fix(docs): The free-tier reference no longer says its numbers come "confidence tagged per row" — no catalog entry carries a confidence tag and the API serves none, so every figure on that page is an estimate of the same, unstated quality. The page now states what an entry does vouch for: an independently documented hard stop (set by hand with the source in a comment, never defaulted to true) and a prompt-training disclosure, both with live counts the check:docs-counts gate keeps honest (#12318) — thanks @maxmad64bis
  • fix(usage): adobe-firefly and firefly have had a working usage fetcher since Adobe Firefly landed, but neither was ever added to the registration list, so the provider-plugin manifest, genericQuotaFetcher and the free-access quota cache all reported them as having no usage support — while USAGE_SUPPORTED_PROVIDERS said the opposite. Both are now declared, which also means their credit balance is fetched like any other declared provider's: registerGenericQuotaFetchers now registers a generic quota fetcher for them, and resolveFreeAccessState no longer returns early. A test holds the registration list to the dispatcher's switch in both directions, which is what the module's own docstring already asked for in prose (#12321) — thanks @maxmad64bis
  • fix(quota): drop the generic quota cache (agy / Antigravity / Claude OAuth) on an upstream 429 so reset-aware scoring does not keep a 60s stale snapshot, and force-refresh the next usage fetch so inner provider caches cannot recache the same window (#12325) — thanks @HouMinXi
  • fix(combos): deleting a combo now clears its persisted LKGP pins instead of leaving unreachable key_value rows behind (#12326) (#12425) — thanks @KrzysiekSko
  • fix(sse): Keep ZWNJ (U+200C) and ZWJ (U+200D) in assistant text, reasoning and tool-call arguments — Persian/Kurdish half-space (ارائه‌دهنده), Arabic/Indic shaping and emoji sequences no longer lose them; the response de-obfuscation now removes joiners only between ASCII word characters, where the request side inserts them (#12186) — thanks @rezjalibd (#12359) — thanks @pacocartones
  • fix(resilience): count resolved upstream 5xx results against the provider circuit breaker on the chat path — CircuitBreaker.execute() no longer reads a resolved { success: false, status: 5xx } as a success that cancels the call-site failure, so a provider answering 503s now trips its breaker instead of staying CLOSED at failureCount: 1; single-model and combo dispatches are each accounted exactly once (#12254) (#12360) — thanks @pacocartones
  • fix(providers): resolve the Codex quota auto-ping model from the live provider catalog and lifecycle registry instead of the retired gpt-5.1-codex-mini, and pause the ping with one actionable warning when no selectable Codex model exists rather than retrying a shut-down id every cooldown window (#11905) (#12361) — thanks @pacocartones
  • fix(api): keep the {created, data} wrapper on combo-routed /v1/images/generations responses and default Codex image results to b64_json on both /v1/images/generations and /v1/images/edits so Codex CLI's built-in image_gen can decode them (#12268) (#12362) — thanks @pacocartones
  • fix(sse): Name the shadowed custom provider node when a built-in provider id/alias (e.g. openference → of) reserves the prefix of an existing OpenAI/Anthropic-compatible node, so the runtime No active credentials for provider: <built-in> error explains that the prefix routed to the built-in and never reached the node's healthy connections, instead of contradicting the dashboard (#11943) — thanks @morpheus9393 (#12365) — thanks @pacocartones
  • fix(i18n): wrap ccOnboardingKeyPlaceholder in ICU single quotes across all 43 locales so angle brackets render literally instead of being parsed as rich-text tags, which crashed the Claude Code onboarding block with INVALID_MESSAGE: INVALID_TAG (#12302) (#12369) — thanks @gonisulaimann
  • fix(models): publish effort_tiers on Kimi K3's synced base-model entries (kmca/k3, kmca/k3-256k) so catalog-only clients (OpenCode, plain SDK pickers) can see and select the reasoning tiers (low/high/max) the synced metadata already carried — the isSkippedEffortProvider gate no longer suppresses tier visibility on those base entries, while synthetic <id>-<tier> variant generation stays prevented and Codex/GLM base models remain excluded unchanged (#12299) (#12371) — thanks @gonisulaimann
  • fix(guardrails): keep auto/auto/* virtual combos exempt from the Vision Bridge fixedModel credential guard so a combo target is passed through instead of silently falling back to global auto-selection (#12237) (#12373) — thanks @pacocartones
  • fix(combo): capability-filter exhaustion caused by max_tokens above every target's known output limit now reports that reason (requested max_tokens vs the pool's highest known ceiling) instead of the unrelated "supports structured output" message (#12229) — thanks @DW-MediaLab (#12374) — thanks @pacocartones
  • fix(auth): the least-used account strategy now prefers accounts without backoff before falling back to oldest lastUsedAt, the same tie-break round-robin already applies, so a failover no longer lands on a just-rate-limited account for a single request (#12279) — thanks @tenshiak (#12375) — thanks @pacocartones
  • fix(docker): the chatgpt-web-codex-browser image now finds the Chrome binary under chrome-linux64/ (Chrome for Testing layout in playwright:v1.62.0-noble) as well as the legacy chrome-linux/, so the container no longer crash-loops with exec: --headless=new: not found (#12024) (#12376) — thanks @pacocartones
  • fix(providers): declare groq/compound and allam-2-7b as non-reasoning models in the curated Groq registry so reasoning_effort / output_config.effort / thinking from Claude Code are stripped instead of forwarded, which Groq rejected with HTTP 400 (#12134) (#12379) — thanks @pacocartones
  • fix(executors): OpencodeExecutor no longer forces a direct connection when the connection has a proxy assigned in Proxy Management but no per-account proxies: the single-account fast path used to wrap the upstream dispatch in the direct-egress sentinel, discarding the ambient proxy context the chat handler had pinned from proxy_assignments, so API-key opencode/opencode-go connections egressed from the host IP (and hit geoblocks) despite the assignment. The direct pin is now applied only when no ambient proxy context exists (#11894 — thanks @hizzt) (#12380) — thanks @pacocartones
  • fix(api): GET /v1/models with MODELS_CATALOG_PREFIX_MODE=canonical (or ?prefix=canonical) now lists providers whose registry alias is undefined or equal to their own id (Antigravity, Antigravity CLI and other self-aliased built-ins) — their single provider/model id was dropped by the alias/canonical duplicate guard in the static, synced, custom and alias-backed catalog loops (#12058) — thanks @cheynetom (#12381) — thanks @pacocartones
  • fix(translator): Drop replayed thinking blocks that carry no signature (the shape produced from cross-provider reasoning_content) instead of stamping the default Claude signature on them, which Anthropic rejected with 400 Invalid signature in thinking block on the next turn served by an Anthropic rung (#12105) — thanks @atescivitci-cmd (#12386) — thanks @pacocartones
  • fix(cli): Resolve Bun's --preload polyfill path against the package root instead of dist/, so omniroute installed with bun install -g no longer crashes at startup with error: preload not found …/dist/open-sse/utils/setupPolyfill.ts (#11980) — thanks @joglomedia (#12387) — thanks @pacocartones
  • fix(providers): gemini-business now publishes its model catalog — /v1/models and /v1/providers/gemini-business/models list the 12 enterprise Gemini ids the executor understands instead of returning an empty list (#12107) (#12389) — thanks @pacocartones
  • fix(db): install busy_timeout before the SQLite connection's first statement so a process opening the database while another one closes its WAL connection waits out the transient EXCLUSIVE lock instead of dying with database is locked, and recognise the drivers' real BUSY/PROTOCOL/IOERR errors as transient in the corruption probe so the same lock no longer renames the database away as corrupt; deflakes cross-process contenders never both acquire the same connection (#12394 — thanks @pacocartones)
  • fix(chat-admission): derive the chat_admission_busy 503 Retry-After from observed heavyweight-lease occupancy — the larger of the exhausted OMNIROUTE_CHAT_ADMISSION_QUEUE_MS window and the time since capacity last turned over, capped at 60 s — instead of a fixed 1 s (structural) / 2 s (byte-stage) hint that invited Codex/agent fan-out clients to re-send ~1 MiB /v1/responses bodies every second into a gate held for the whole SSE lifetime; an idle gate keeps the historical floors (#12135) (#12395 — thanks @pacocartones)
  • fix(api-manager): the API key permissions modal no longer silently drops allowedCombos entries its Combo picker cannot render — routing-rule names such as rt-*, which the backend already honours — when "All" is clicked and the key is switched back to "Restrict"; those entries now survive the toggle, are listed read-only under the combo list so the count and the list agree, and are saved back verbatim instead of persisting [] (deny-all) (#12397 — thanks @pacocartones)
  • fix(catalog): write the NUL separator of the catalog connection memo key, the provider serviceKind memo key, the Video Bridge promotion group key and a JSON-exactness test fixture as the \u0000 escape instead of a raw byte — same runtime value, but the raw byte made git, GitHub and ripgrep treat those files as binary (hidden PR diffs, silently skipped searches); a guard test now keeps raw NUL bytes out of src/, open-sse/ and tests/ (#12403 — thanks @pacocartones)
  • fix(providers): add CLAUDE_CODE_CLIENT_VERSION and GITHUB_COPILOT_CLI_VERSION env overrides so Anthropic/Copilot client-version gates can be unblocked without a rebuild (#12417) (#12632) — thanks @HouMinXi
  • fix(db): back-fill last_ping_at and last_pinged_reset_key on provider_connections during schema reconciliation so divergent lineages that skipped 123_quota_auto_ping still accept quota auto-ping writes (#12470 — thanks @KooshaPari)
  • fix(ci): document MIT exceptions for @eloqnt/{config,format-json,format-po} (next-intl transitive; locked tarballs omit license) and keep the A2A lifecycle vitest off the real SQLite persistence seam (#12581) (#12595)
  • OpenCode plugin /v1/models catalog fetch now waits 30s by default and attaches HTTP statusCode on 401/5xx so host fallback plugins can hop instead of seeing an untyped AbortError/UnknownError. (#12607) — thanks @RaviTharuma
  • CI: the OpenAPI security-tier gate now mirrors isAlwaysProtectedPath() in full — it also reads ALWAYS_PROTECTED_API_PATTERNS, so the pattern-gated credential routes (/api/providers/{id}/{claude,codex}-auth/{export,apply-local}, GHSA-5926-2w35-7h4q) no longer report as unannotated. (#12605)
  • fix(api): GET /v1/models no longer waits forever on a hung coalesced catalog rebuild; cold-path waits are bounded (CATALOG_BUILD_TIMEOUT_MS, default 8s) and a last-good 200 is served when the rebuild times out (#12627). (#12628) — thanks @RaviTharuma
  • fix(cursor): a non-array tool_calls on an assistant message no longer crashes the cursor request translator with a TypeError; both loops now require an array (#12689) (#12691) — thanks @soroush5
  • fix(chat): Chat Completions no longer return empty content after a server-owned memory or skills tool; the first provider send and account/model recovery now share one pipeline so a follow-up round-trip can fill the reply (#12696) — thanks @HouMinXi (#12867)
  • fix(dashboard): batch-deleting provider keys no longer toasts failure after a successful delete when the confirm button's click event is forwarded as onAfter (#12711) — thanks @HouMinXi
  • fix(glm): drop the extra 16th argument to createSSETransformStreamWithLogger that TypeScript rejected (TS2554) and that never reached the TransformStream (#12711) (#12770) — thanks @HouMinXi
  • fix(dashboard): pass nodeMap into Runtime QuotaGroup so a quota monitor in error/exhausted/alerting no longer throws ReferenceError: nodeMap is not defined. (#12868) — thanks @HouMinXi
  • Restricted API keys whose allowedModels lists a combo name no longer skip every combo member at pre-dispatch (ALL_TARGETS_SKIPPED). Inner-target filtering still applies when the allow-list is a provider prefix or disableNonPublicModels is on (#12899). — thanks @HouMinXi
  • fix(combos): Effort-suffixed combo members inherit the base model's model_context_overrides row so priority order is not inverted on large requests (#12926) — thanks @HouMinXi
  • fix(catalog): Picker-added customModels enter the dispatch-time live catalog so combo members and bare inference no longer 400 (#12934) — thanks @HouMinXi
  • fix(docker): re-derive the Docker build's worker-pool memory budget from the MEASURED ~4.5 GB per-process RSS (the issue owner's own VPS dmesg OOM-killer reproduction) instead of the stale 2560 MB/worker inference #11419 shipped, and lower OMNIROUTE_BUILD_WORKERS 3 → 2 so 1 parent + 1 page-data worker (2 processes × 4.5 GB = 9 GB) fits the 12.288 GB (75%) budget on a 16 GB GitHub Actions runner — the previous default (1 parent + 2 workers = 13.5 GB) still overcommitted the runner and kept "Publish to Docker Hub" failing with cannot allocate memory after #11419 merged (#7518). (#11663)
  • fix(electron): the packaged Windows desktop build now passes the #7592 cold-restart smoke end to end. Five defects were found and fixed on the way: (1) optional-pack staging failed on any machine whose tar is GNU tar — it reads the drive letter in an absolute -f C:\... archive path as a remote rsh host (Cannot connect to C:), so staging now passes a bare filename with cwd at the tarball directory and surfaces tar stderr on failure; (2) the Electron files allowlist was missing lib/loginHeaderCapture.js, so the packaged main process crashed at startup with Cannot find module './lib/loginHeaderCapture'; (3) electron-builder ≥26 injects an !**/node_modules/** ignore into every extraResources pattern list that no positive filter can override, which silently dropped the staged runtime node_modules (including the better-sqlite3 N-API prebuild) from resources/app — a fresh v3.8.50 desktop build would have shipped with no native SQLite driver and reproduced the #7592 sql.js fallback on every machine; an afterPack hook now restores the staged node_modules after packing; (4) the packaged-app smoke harness redirected USERPROFILE into its temp DATA_DIR, but Electron resolves the Roaming profile from %USERPROFILE%\AppData\Roaming\<name> (USERPROFILE takes precedence over the APPDATA env var) and the path service throws instead of creating it, so requestSingleInstanceLock() returned false and the app exited(0) silently before app.whenReady() — the harness now pre-creates the derived tree, and ensureSmokeEnvDirs is exported and unit-tested; (5) the #7592 driver assertion parsed a [DB] Driver: ... line that the server's primary DB path never printed (only the unused openDatabaseAsync() did) — getDbInstance() now emits the same line on open so the guard can actually assert the native driver. Also: the smoke env-allowlist unit test hardcoded POSIX paths and could only pass on Linux/macOS; expectations are now host-agnostic, with new regression tests for the USERPROFILE-derived tree and tarPack under absolute Windows-style paths (#11443) — thanks @jonlwheat2-gif
  • fix(sse): stop the auto-combo candidates inspector from silently dropping model-locked/cooled-down rows (#9133) (#11994) — thanks @hartmark
  • fix(providers): Perplexity Web no longer turns upstream stream failures into successful assistant text; pre-content failures remain eligible for fallback, partial output ends with a structured sanitized error, and failed sessions are not persisted (#12465)
  • fix(video-bridge): preserve chronological video and transcript fusion, source-frame timestamps, transcript provenance, full contact-sheet intervals, and pre-abort callback isolation (#11681)
  • Z.ai Web: HTTP 200 streams carrying an upstream error now terminate with a structured failure instead of assistant text plus a normal stop, preserving partial output while allowing pre-content combo fallback. (#12454)
  • Absorb Error [AbortError]: request_signal_aborted and DOMException AbortError shapes in the process-level client-abort crash guard so routine client disconnects no longer kill the server (exit code 7). (#12165) — thanks @alvinveroy
  • fix(sse): Treat Adapta Web type:error SSE events as sanitized non-stream failures instead of empty HTTP 200 completions. (#12459)
  • fix(autoCombo): layer 3 (models_dev_tier) no longer scores retired models as premium, merge cross-provider capability rows deterministically, and inherit tier scores through variant ids (#11508). getModelsDevTierFitness now vetoes ids present in config/quality/model-lifecycle.json with status retired before any other signal; loadModelCapabilities() aggregates per model_id (any-non-null-true for booleans, max for limit_context) instead of last-write-wins over SQLite's undefined row order; a capabilities miss retries through resolveScoresAs()'s base id and reports models_dev_tier:inherited, matching layers 1–2 (#11492). Adds getModelsDevTierFitnessWithSource() and an invalidateCapabilitiesCache() test/ops hook. (#11598) — thanks @Meet6338-X
  • fix(sse): accept Claude content:[] + stop_reason: max_tokens/tool_use in detectMalformedNonStream (match isEmptyContentResponse) so Claude Code /model probes with max_tokens: 1 no longer become a false 502 (#11568) — thanks @Seramicx
  • fix(security): sanitize request.failed diagnostics before publishing them to live dashboard listeners and replay history, while keeping status, model, provider, latency, and internal call-log diagnostics intact. (#12469)
  • fix(pwa): do not intercept dashboard navigations so Chrome can fall back from a stale HTTP/3 Alt-Svc advertisement (UDP :20128 is unpublished; F5 on a long-lived tab hung until a new tab opened a fresh TCP connection). (#12767) — thanks @HouMinXi
  • fix(executors): handle DuckDuckGo ERR_BN_LIMIT (418) without retrying — when the upstream returns 418 ERR_BN_LIMIT (rate-limit/ban), the executor now returns the error immediately instead of burning another VQD acquisition that would only count against the IP limit. The retry logic for 418 ERR_CHALLENGE (unsolved challenge) remains unchanged. (#11598) (#12110) — thanks @oyi77 / @Meet6338-X
  • Fixed DuckDuckGo streaming responses losing JSON lines and UTF-8 characters split across network chunks. (#11528) — thanks @pacocartones
  • Harden generated Cloudflare, Vercel, and Deno relays by stripping hop-by-hop, framing, and proxy authentication headers before forwarding requests upstream. (#11533) — thanks @pacocartones
  • feat(grok-cli): Provider Limits shows grok-cli banked reset credits from GetRemainingResets (including a real zero; a failed RPC omits the row) and the existing View credits button now calls ConsumerUiSvc/RedeemReset for grok-cli. Live tokens use nested fields 10/20/30 (id + Timestamp), not compact 1/2/3. (#12805) — thanks @HouMinXi
  • fix(memory): Embedding Model Quick select, Embedding Source remote dropdown, and Rerank selector now list every configured provider with embedding/rerank support instead of only chat-catalog text matches plus OpenRouter live discovery; a generic OpenAI-compatible /embeddings + Cohere-compatible /rerank runtime fallback resolves any configured chat provider's embedding/rerank endpoint, so unlisted providers no longer fail with "Unknown embedding provider"; both memory selectors gained a free-text model override (#11390) — thanks @rqzbeh
  • Harden SQLite upgrades around the historical migration-074 version collision: missing discovery and inspector tables are replayed atomically, pre-existing databases (including setup-created skeletons) receive reusable content-addressed safety snapshots, and Node test/eval probes without DATA_DIR are isolated from the operator database. (#12435)
  • fix(dashboard): Moonshot/Kimi Open Platform voucher and cash leftover percentages follow the bucket balance, so an empty wallet no longer paints those rows as 100% while Available is 0% (#12733) — thanks @HouMinXi
  • fix(combo): return non-retryable HTTP 400 when all candidates for a pinned native Codex turn are unavailable due to model-scoped lockout, terminating the turn cleanly while preserving turn continuity and enabling standard Combo routing on subsequent turns (#12240) — thanks @mdigitalbh81
  • Fixed auto-router strategies preserving the selected connection through ranking and dispatch when multiple connections share the same provider and model. (#11530) — thanks @pacocartones
  • Preserve Chat Completions JSON-object response formats when translating requests to the Responses API. (#11529) — thanks @pacocartones
  • Fix idempotency fingerprints so Responses API requests include input and other semantic generation fields using stable serialization, while excluding credentials and request metadata. (#11532) — thanks @pacocartones
  • fix(api): Generated API CLI commands now enforce required OpenAPI request bodies; Combo test commands forward the required comboName body, while API keys created by older writers after migration 149 preserve legacy allow-all Combo access without widening explicit empty allowlists — thanks @marcelokarval (#12070)
  • fix(grok-web): treat upstream streaming failures as failures instead of successful assistant text: error-only streams now fail readiness with HTTP 502, while failures after legitimate content preserve that partial output and terminate through the sanitized stream failure path without a normal stop completion. (#12458)
  • HuggingChat now turns HTTP 200 JSONL generation failures into a sanitized 502 before content, or a fixed public stream failure after partial output, so fallback and request persistence no longer record a false successful stop. (#12456)
  • fix(providers): keep 1min.ai HTTP 200 stream errors out of assistant content, preserve partial output, and expose sanitized terminal errors so pre-content failures can fall back. (#12466)
  • fix(sessions): preserve legacy Sessions data when the additive exclusive-session projection is unavailable, returning an empty projection and warning only once per contiguous outage instead of failing the endpoint; the Sessions badge intentionally reflects the merged legacy and exclusive row count introduced by #11389 (#11469) — thanks @KaspaPulse
  • Render Nimble Search with the generic provider icon and serve Opper's proven logo locally. (#12415)
  • fix(quality): config/quality/provider-assets-provenance.jsonl was stale against the physical public/providers/ tree (still listing 225 assets and 5 duplicate-content groups after several provider-retirement PRs removed 85 icon files), which would have made check:provider-asset-provenance fail on the next full run. Reconciled the manifest to the real 142-file snapshot (142 asset records, 1 duplicate-content group, 2 newly-registered files marked unresolved pending provenance research), repinned auditedCommit, and updated the regression test's hardcoded counts to match. (#11876)
  • fix(ui): wrap quota card footer action buttons to prevent clipping Refresh button on narrow cards — when extra action buttons such as View credits are present on Codex cards, allow the footer row to flex-wrap with shrink-0 buttons so Refresh now remains visible and clickable across all responsive card widths (#11464) — thanks @adevwithpurpose
  • Keep Antigravity Gemini usable when the same connection's Claude weekly quota is empty; generic quota cache stays per-connection for every other provider. (#12637) — thanks @HouMinXi
  • Fixed Tailscale tunnel setup to reject ports outside the valid 1-65535 range and safely install Windows MSI packages from paths containing apostrophes. (#11534) — thanks @pacocartones
  • Classify Cloudflared and Tailscale process-management routes as local-only and spawn-capable while preserving remote access to read-only tunnel status endpoints. (#11531) — thanks @pacocartones
  • Electron release workflow: the publish-npm job now grants actions: read to the reusable npm-publish.yml it calls (its publish job requests it), which is what made GitHub refuse the whole v3.8.50 run at startup and ship the release with zero desktop assets; a workflow_dispatch now builds the requested tag instead of the dispatching branch and can skip the npm leg (publish_npm=false) when only re-attaching assets (#11974)
  • fix(dashboard): The Combos page usage guide now reads its dismissal through useSyncExternalStore instead of correcting SSR state inside an effect, removing an extra commit of the page tree on every load (and the react-hooks/set-state-in-effect error it raised). (#12671)
  • Electron release: electron/package-lock.json regained the optional electron-builder-squirrel-windows subtree (13 entries) that npm ci had been refusing as out of sync, electron-release.yml gained a build_ref dispatch input and stops regenerating release notes on a re-attach dispatch, and the npm publish workflow attaches the SBOM to the GitHub Release on dispatch publishes too — so the v3.8.51 tag ships every desktop asset and the SBOM like v3.8.49 did (#12022)
  • Drained the reds every PR against release/v3.8.51 was born with: documented the three service-worker build-id variables, registered the six retirement/tunnel tests with the mutation gate, approved eslint-plugin-react-hooks in the dependency allowlist, added the six combo.sort.* strings to vi and pt-BR, pointed the ChatGPT Web doc at the real migration-168 test, worded the g4f hint around the member key, and realigned four tests to the retired-provider catalog and the legacy-schema fixtures the retirement migrations touch. (#11940)
  • fix(providers): Claude, Grok, LMArena, Notion, and Perplexity web-cookie transports now use pooled wreq-js 3.2 instead of the native sidecar, with all nine supported bindings pinned and audited, and the applicable platform binding plus native-license evidence included in each release artifact (#12429, supersedes #11753).
  • fix(providers): Zed Hosted streaming failures now trigger fallback before content and end partial streams with a sanitized structured error instead of fake assistant text and a normal-success stop. (#12455)
  • fix(test): capture hook result after render (direct commit b85d0abb0b)
  • fix(sync): repair the two regressions the v3.8.50 sync-back left on release/v3.8.51 (direct commit 5ade9e0851)
  • revert(models): drop the memo reset hook ported from #10055 (direct commit 282de841ae)
  • fix(typecheck): guard the optional connectionId before the reactive model sync (direct commit f2825bad75)
  • fix(quota-share): release the winner's reserved in-flight slot (#11371) (#11408) — thanks @oyi77
  • fix(providers): stop silently dropping reasoning effort on opencode families (#10788) (#11409) — thanks @oyi77
  • fix(combo): enable genuine quota-aware routing for generic providers (antigravity, claude, etc.) (#11411) — thanks @HouMinXi
  • fix(oauth): allow expired token retry before deactivation and activate expired retry loop (#P0) (#11414) — thanks @HouMinXi
  • perf(executors): lazy-load the executor registry — defer class imports + construction to first use (#11220) (#11421) — thanks @oyi77
  • fix: make release/v3.8.50 CI gates green (run 32786966560 root causes) (#11450) — thanks @jonlwheat2-gif
  • fix(sse): exact-domain cookie match and origin-equality URL assertions (#11452) — thanks @hartmark
  • fix(providers): send Responses API payload during OpenAI-compatible validation (#11454) — thanks @tuandinh0801
  • fix(command-code): fallback to /alpha/generate for Go plan without Provider API access (#11455) — thanks @Chewji9875
  • fix(responses-continuation): fail closed on a log-truncated stored input/output array (#11473) — thanks @hartmark
  • fix(cli): write jcode settings as [providers.omniroute] in config.toml (#11484) — thanks @hongnoul
  • fix(logging): raise chat-log array truncation cap to 1000 and unify duplicate implementations (#11499) — thanks @hartmark
  • fix: restore live websocket URL helpers (#11509) — thanks @PauloFH
  • perf(core): eliminate async waterfalls across home, cache, and db operations (#11396) (#11512) — thanks @HouMinXi
  • fix(kiro): add runtime.us-east-1.kiro.dev as first-attempt endpoint (#11517) — thanks @rifqiawl
  • fix(providers): remove duplicate 128-tool truncation in OpencodeExecutor (#11444) (#11518) — thanks @HouMinXi
  • fix(providers): poll onboardUser LRO before classifying BYOP (#11519) — thanks @rifqiawl
  • fix(translator): preserve exact tool name casing for Claude tool_use payloads (#11487) (#11520) — thanks @HouMinXi
  • fix(kimi-web): align health probe with executor domain (#11521) — thanks @pacocartones
  • fix(cli): preserve existing admin password during setup (#11522) — thanks @pacocartones
  • fix(mitm): keep Windows DNS elevation off the POSIX sudo path (#11430) (#11542) — thanks @official-burak
  • fix(sse): scale chat admission by ingest byte budget instead of a fixed request count (#11548) — thanks @TheDemonTuan
  • fix: accept Gemini Web cookie exports (#11554) — thanks @rafacpti23
  • fix(i18n): complete Brazilian Portuguese translations (#11558) — thanks @rafacpti23
  • fix(sse): explicit types for openai-responses pureHelpers — clears last failing core typecheck gate (#11567) — thanks @jonlwheat2-gif
  • fix(opencode): hydrate Proxy Pool references for no-auth connections (#11584) — thanks @AStupidBear
  • fix(ci): resolve inherited release/v3.8.51 base-red CI failures (#11608) — thanks @jonlwheat2-gif
  • fix(db): fallback load STORAGE_ENCRYPTION_KEY from env files (#11614) — thanks @benzntech
  • fix(providers): mark Brave Search as no longer having a free tier (#11621) — thanks @turbolego
  • fix: pass tool result content through as string in OpenAI/Claude -> Gemini translation (#11624) — thanks @Hsia97
  • fix(sse): scope credential health to exact key (#11628) — thanks @KaspaPulse
  • fix(install): guard prepare script against missing husky (#11634) — thanks @Prajeeth-12
  • fix(security): redact NVIDIA_API_KEY at the log sink in the ad-hoc diag script (#11639)
  • fix(types): clear the two type regressions red on release/v3.8.51 (#11640) — thanks @jonlwheat2-gif
  • fix(build): resolve workspace: protocol dependencies to concrete semver versions during prepublish (#11540) (#11641) — thanks @HouMinXi
  • fix(search): properly resolve configured search connection in /v1/responses pipeline (#11524) (#11642) — thanks @HouMinXi
  • fix(authz): trust cf-connecting-ip for via-proxy detection when peer is Cloudflare edge (#11514) (#11643) — thanks @HouMinXi
  • fix(security): compare the OIDC state cookie in constant time (#11646)
  • fix(security): namespace the dedup hash by the calling API key (#11649)
  • fix(models): gate the Codex-native catalog loop on prefix mode (#11632) (#11662) — thanks @xiaoyaner0201
  • fix(dashboard): project Codex routing quota in account pools (#11666) — thanks @KaspaPulse
  • fix(ddgw): normaliza system no upstream da release v3.8.51 (#11692) — thanks @RhianB14
  • fix(memory): honor category filter in GET /api/memory (#11699) — thanks @Sabeekhann
  • fix(ci): detect stale workspace lockfile entries (#11714) — thanks @pacocartones
  • fix(ci): stamp dist/BUILD_SHA before the npm publish provenance gate (#11721)
  • fix(docker): decouple the best-effort Bun image from the release manifest (#11724)
  • perf(compression): avoid cloning memo misses twice (#11727) — thanks @RaviTharuma
  • fix(docker): let the bun digest artifact be absent, not fatal (#11740)
  • fix(shared): block cloud-metadata hosts under default remote-image guard (#11755) — thanks @santosraju99-hub
  • fix(api): preserve caller-provided X-Correlation-Id on chat completions (#11760) — thanks @aniruddhaadak80
  • fix(resilience): lock permanently retired models instead of short backoff (Gemini ban prevention) (#11762) — thanks @turbolego
  • fix: mark Vercel AI Gateway as passthroughModels (#11771) — thanks @Rahulsharma0810 / @Rahulsharma0810
  • fix(resilience): stop hammering permanently-moved endpoints and billing-suspended accounts (#11774) — thanks @turbolego
  • fix(lease): remove global static reservation and gate routing on live active lease occupancy (#11775) — thanks @TheDemonTuan
  • fix(ci): size the install-upgrade gate to a measured run, and log the pack cost (#11776)
  • fix(pwa): stop serving the stale shell after deploys (#11779) — thanks @HouMinXi
  • fix(combo): lock GitHub models rejected as "not supported" for future requests (#11781) — thanks @turbolego
  • fix(cli): probe both IPv4 and IPv6 loopback for server readiness (#11766) (#11794) — thanks @vermasomesh835
  • fix(sse): preserve CLIProxyAPI wrappers per connection (#11797) — thanks @RaviTharuma
  • fix(providers): honor model auto-fetch opt-in on create (#11798) — thanks @RaviTharuma
  • fix(dashboard): make model auto-fetch order independent (#11805) — thanks @RaviTharuma
  • fix(docker): make GHA cache exports best effort (#11806) — thanks @RaviTharuma
  • fix(kiro): do not permanently ban on 'User is not authorized to make this call' (#11809) — thanks @Deftera186
  • fix(sse): keep cache-write tokens in OpenAI-shaped usage (#11814) — thanks @TheDemonTuan
  • fix(docs): resolve relative markdown and wiki links across Fumadocs and GitHub wiki (#11834) — thanks @AndrianBalanescu
  • fix(chat): preserve unstripped model string for passthrough provider routing (#11840) — thanks @raheemuddin786
  • fix(mcp): avoid async initialization inside sync esm wrappers (#11569) (#11849) — thanks @HouMinXi
  • fix(sse): preserve coherent markdown token boundaries across stream translation chunks (#11606) (#11851) — thanks @HouMinXi
  • fix(ci): run the install-upgrade gate on disk, not on the /tmp tmpfs (#11855)
  • fix(skills): expand shorthand property types in injected tool schemas (#11857) — thanks @alvinveroy
  • fix(ci): publish npm from a hosted runner so provenance is accepted (#11877)
  • fix(lint): pin react hooks plugin to 7.0.1 (#11886)
  • fix(ci): surface and clear release-green ESLint failure (#11890) — thanks @MumuTW
  • fix(guardrails): prevent duplicate prompt-injection-guard log output (#11936) — thanks @patrykkopycinski
  • fix(auth): downgrade expected transient states from warn to debug (#11937) — thanks @patrykkopycinski
  • fix(codex): fail over image generation for imported free plans (#11948) — thanks @Bl0ck154
  • fix(codex): normalize non-stream responses (#11951) — thanks @Bl0ck154
  • fix(images): forward Antigravity image size (#11952) — thanks @Bl0ck154
  • fix(db): invalidate connection cache after upsert (#11953) — thanks @Bl0ck154
  • fix(codex): restore imported account state (#11954) — thanks @Bl0ck154
  • fix(sse): give extended-thinking targets the reasoning readiness budget (#11959) — thanks @abhisheksharma2411
  • fix(dashboard): use opaque background and readable text color on cost chart tooltips (#11960) — thanks @ZaimMarzuki
  • fix(antigravity): send complete loadCodeAssist metadata (ideType/platform/pluginType as numeric enums) (#11969) — thanks @rifqiawl
  • fix(core): resolve DB init race condition and reasoning translation (#12003) — thanks @raheemuddin786
  • fix(admission): exclude reclaimable page cache from the cgroup pressure ratio (#12017) — thanks @HouMinXi
  • fix(build): prune non-production trees in NFT trace excludes and tsconfig (#12028) — thanks @Chewji9875
  • fix(migrations): add renamed migration compatibility for 056/073/077/101 (#12036) — thanks @oyi77
  • fix(sse): default crash-guard logger to console.warn, not console (#12042) — thanks @HouMinXi
  • fix(catalog): derive vision/modalities for built-in auto combos from effective target pool (#12046) — thanks @Prajeeth-12
  • fix(release): the packaged-app smoke verifies the database opened, not a driver line the primary path never prints (twin of #12032) (#12047)
  • fix(barrel): migrate src/app/ off the @/lib/localDb barrel import (#11795 Phase 2) (#12051) — thanks @jonlwheat2-gif
  • fix(barrel): migrate src/lib/ off the localDb barrel to direct db imports (#59) (#12052) — thanks @jonlwheat2-gif
  • fix(barrel): migrate open-sse, src/shared, src/sse, src/models, src/domain off the @/lib/localDb barrel import (#11795 Phase 4) (#12053) — thanks @jonlwheat2-gif
  • fix(barrel): delete the @/lib/localDb barrel — every consumer migrated (#11795 Phase 5) (#12055) — thanks @jonlwheat2-gif
  • fix(radar): restore D12 public boundary (#12057)
  • fix(dev): bound webpack and Tailwind scans (#12075) — thanks @backryun
  • fix(dev): isolate root layout settings reads (#12076) — thanks @backryun
  • fix(dev): reduce instrumentation executor fan-out (phase 3) (#12078) — thanks @backryun
  • fix(dev): make logging resources HMR-singleton (#12079) — thanks @backryun
  • fix(sse): keep unavailable forced connections scoped (#12080) — thanks @keeltrace
  • fix(dev): isolate batch dispatch from instrumentation (#12081) — thanks @backryun
  • fix(ci): reconcile release test contract drift (#12082) — thanks @backryun
  • fix(ci): accept CVE-2025-68121 in the prebuilt tls-client .so, auto-close base-red issues, guard Scorecard on the default branch (#12085)
  • fix(db): drop three consumer-less 1proxy exports — dead-code base-red on release/v3.8.51 after the barrel deletion (#12055) (#12087)
  • fix(api): clear the six API-route TypeScript regressions the new gate landed red on (#12094)
  • fix(release): never let the tag-push Create Release append auto notes to the curated body (#12096)
  • fix(sse): honor CLIProxyAPI environment API key (#12099) — thanks @RaviTharuma
  • fix(oauth): bind Google refresh to the client that issued the token (#12106) — thanks @HouMinXi
  • fix(dashboard): make RequestLoggerDetail loadable outside Next — CSS via globals.css + CJS/ESM interop (#11703 base-reds) (#12114)
  • fix(leases): project status lease row to lease columns so joined connection PII never escapes (#12115) — thanks @geek007git
  • fix(plugins): do not kill the plugin process when a fire-and-forget hook times out (#12116) — thanks @geek007git
  • fix(guardrails): restore injection-guard logging on middleware-only routes (#12117) — thanks @geek007git
  • fix(plugins): refresh stored manifest from disk on activate so new hook fields reach existing installs (#12120) — thanks @geek007git
  • fix(codex): preserve existing provider state when bulk-import upserts a matching connection (#12122) — thanks @geek007git
  • fix(ci): clear the two base-reds the 2026-08-30 merge batch left on release/v3.8.51 (round 3) (#12123)
  • fix(sse): strip type:'custom' from Claude tools on agentrouter dispatch (#12126) — thanks @wahidsadik371-coder
  • fix(sse): stop advertising video providers the dispatcher cannot run (#12131) — thanks @quiterunner-commits
  • fix(search): treat HTTP 432 and plan limit errors as transient cooldown (#12139) — thanks @HouMinXi
  • fix(sse): refuse an AI Horde queue that cannot fit the request budget (#12143) — thanks @quiterunner-commits
  • fix(ci): clear the base-reds the afternoon merge batch left on release/v3.8.51 (round 5: provider count 352, TS2554/TS2677) (#12144)
  • fix(translator): strip neutral tool_choice when tools absent in Responses-to-Chat (#12141) (#12166) — thanks @HouMinXi
  • fix(guardrails): pass providerId to getResolvedModelCapabilities in checkComboVision (#12112) (#12169) — thanks @HouMinXi
  • fix(translator): strip plaintext reasoning content for opaque responses backends (#12128) (#12171) — thanks @HouMinXi
  • perf(stream): compile hot-path regexes once, bound token caches, fix quadratic buffering (#12179) — thanks @opensource-elearning
  • fix: resolve compression worker file using runtime anchors instead of… (#12183) — thanks @NightStalker-87
  • fix: prevent Claude Code session kills via liveness-aware readiness + auto model echo (#12189) — thanks @opensource-elearning
  • fix(translator): preserve falsy primitive values in Gemini and Antigravity function response results (#12191) — thanks @Siva010
  • fix(oauth): mark empty Antigravity projectId as degraded and clear stale errors (#11284) (#12205) — thanks @HouMinXi
  • fix(providers): gate the Codex auto-ping usage read on the shared quota throttle (#12209) — thanks @abhisheksharma2411
  • fix(combos): prioritize SQLite row id over inner JSON id and notify delete errors (#12213) — thanks @HouMinXi
  • fix(providers): mark featherless.ai as no longer having a free tier (hasFree: false) (#12216) — thanks @turbolego
  • fix(deepseek-web): stop Turbopack dev panic in the PoW worker path resolver (#12221) — thanks @hartmark
  • fix(chat-admission): clarify local 503 source (#12223) — thanks @keeltrace
  • fix(dev): silence webpack runtime module warnings (#12228) — thanks @backryun
  • fix(db): use module.require for CommonJS runtime driver loading (#12230) — thanks @raheemuddin786
  • fix(memory): honest probe-driven FTS5 keyword status + memory_id rowid sync (#12231) — thanks @raheemuddin786
  • fix(system): propagate abort signal to stream reader in HTTP version checks (#12232) — thanks @raheemuddin786
  • fix(resilience): isolate local host execution errors from provider circuit breakers (#12233) — thanks @raheemuddin786
  • fix(sse): sort injected tools deterministically for prompt caching (#12234) — thanks @raheemuddin786
  • fix(combo): always clear the loop-safety timer, not just on the happy path (#11804) (#12245)
  • fix(quality): base-red pair — stryker tap registration + turn-pin suites aligned to the window gate (#12255) — thanks @backryun
  • fix(dev): qualify Turbopack runtime boundaries (#12258) — thanks @backryun
  • fix(combo): fall back to full pool when collapsed sole survivor is context-too-small (#12278) — thanks @opensource-elearning
  • fix(combo): bound the pre-dispatch unavailable skip so a stale label cannot dark a pool (#12168) (#12285)
  • fix(security): strip Qwen/Alibaba console-session cookies from provider API responses (#12287)
  • fix(usage): console-aware Token Plan guidance + subscription hint on bailian 401 (#12288)
  • fix(resourcePressure): log numeric detail on every rejection, recover faster (#12293) — thanks @hartmark
  • fix(release): drain the 2026-09-01 base-red window — passthrough usage regression + radarPage i18n keys (#12327)
  • fix(ci): openapi-security-tiers checker must honor routeGuard patterns + imported prefixes (#12350) — thanks @ggiak
  • perf(ci): cache node_modules in the npm-ci-retry composite (#8084 D3) (#12408)
  • revert(ui): point CTAs back at their real destinations (#12410)
  • fix(release): drain the 2026-09-02 base-red — rerank-providers import + api-typecheck baseline ratchet (#12414)
  • fix(types): make system prompt injection noImplicitAny-safe (#12416)
  • fix(quality): validate typecheck baseline schema (#12419)
  • fix(memory): resolve rerank provider node cache import (#12421)
  • fix(ci): repair release-root regressions — pack dedup, web-session syntax, uc-image ids (#12423) — thanks @backryun
  • fix(providers): separate MaxAI and UC credential contracts (#12431)
  • fix(cli): preserve Claude settings on config set (#12432) — thanks @KrzysiekSko
  • fix(providers): repair the maxai credential block truncated by merge auto-resolve (#12433)
  • fix(quality): record the 2026-09-02 merged growth in the file-size baseline (#12434)
  • fix(uc): route UC error strings through sanitizeErrorMessage; allowlist the retired codex id (#12437)
  • fix(adapta): redact streamed upstream errors (#12438)
  • fix(sse): remap non-contiguous upstream tool_calls index to a gap-free output_index (#12445) — thanks @hartmark
  • fix(quality): detect duplicate tool_calls entries in one response (#12446) — thanks @hartmark
  • fix(conversations): resolve turn content OmniRoute never sends back to the client (#12447) — thanks @hartmark
  • fix(combo): do not treat credits-exhausted 401 as auth skip (#12449) — thanks @RaviTharuma
  • fix(auth): do not park healthy quota accounts as expired (#12452) — thanks @RaviTharuma
  • fix(responses-continuation): fail closed on a collector-truncated, empty output array (#12460) — thanks @hartmark
  • fix(logging): raise the SSE payload collector's default cap (#12461) — thanks @hartmark
  • fix(resilience): surface Responses failed.error.message in 502s (#12472) — thanks @RaviTharuma
  • fix(security): unbiased maxai X-Random nonce + stricter URL/regex assertions (#12502)
  • fix(video): re-anchor transcript log-redaction so PII/credential maskers can't reopen the leak (#12150 P1 follow-up) (#12503)
  • fix(monitoring): serve cached credentialHealth off the request path (#12533) — thanks @RaviTharuma
  • fix(quota): keep Antigravity Gemini usable when Claude weekly is empty (#12566) — thanks @HouMinXi
  • fix(authz): hard-gate every credential export and CLI-config write (GHSA-5926-2w35-7h4q) (#12600)
  • fix(docs): restore the Next build — REMOVED_PROVIDERS.md had no frontmatter (base-red #12581) (#12610)
  • fix(quality): re-point the zcodeProtocol public-creds allowlist to line 313 (#12615)
  • fix(security): close 3 advisories — search baseUrl exfil, sk- in the error sanitizer, bifrost relay header leak (#12620)
  • fix(providers): sanitize boolean required and nested bare maps for Gemini (#12269) (#12624) — thanks @HouMinXi
  • fix(resilience): keep Overloaded STREAM_EARLY_EOF off the provider breaker (#12626) — thanks @HouMinXi
  • fix(catalog): re-audit free-tier quotas against official pages (#12649)
  • fix(dashboard): keep a request's pending-tracking id stable across combo target retries (#12650) — thanks @hartmark
  • fix(tests): retire dead model ids from the chat-pipeline integration suite (base-red #12581) (#12670)
  • fix: resolve SqliteError no such table compression_run_telemetry during cleanup (#12682) — thanks @groovecityJO
  • fix(combo): treat a pin-only step as implicit connection allowlist (#12697) — thanks @HouMinXi
  • fix(ci): ship .npmrc in published package so legacy-peer-deps applies to consumers (#11544) (#12699) — thanks @KooshaPari
  • fix(docker): pin 4 CLI tools to exact versions (#12576) (#12703) — thanks @KooshaPari
  • fix(docker): document and harden cli profile trust boundary (#12570) (#12706) — thanks @KooshaPari
  • fix(sse): redact video transcript in pre-guardrail rejected-request logs (#12150 P2 item 7) (#12710)
  • fix(i18n): quote placeholder in OMNIROUTE_AUTO_SYNC_CLAUDE_PROFILES description (#12505) (#12769) — thanks @KooshaPari
  • fix(api/pricing): surface validation error message as string, not raw object (#12494) (#12771) — thanks @KooshaPari
  • fix(sse): re-enable prompt compression for native Codex passthrough (#12834) — thanks @thomasmaerz
  • fix(dashboard): model health tests for a provider node set to the Responses API now call /v1/responses with a Responses-shaped body instead of /v1/chat/completions — those models were reported as Provider returned HTTP 200 but no text content even though the same model answered normally through /v1/responses (#13070) (#13087) — thanks @ntdatt812
  • fix(routing): stop a round-robin combo's "opencode" targets from collapsing onto the opencode-zen connection (#11912) (#13283)
  • fix(routing): stop the reactive-compaction debug log from lying when compression is globally disabled (#11977) (#13276)
  • fix(electron): relativize standalone-bundle symlink targets so Stage 8 manifest verification stops failing on Windows (#11979) (#13251)
  • fix(dashboard): surface a visible error when Compression Studio's combined preview run fails (#12061) (#13288)
  • fix(dashboard): make the compression "Effective pipeline" preview honor the active profile and warn when the master switch is off (#12063) (#13287)
  • fix(providers): scope TinyCMS Web signer's DOM shims to each call instead of leaking them for the process lifetime, and surface a clean HTTP status on a non-JSON interception-toggles error (#12072) (#13254)
  • fix(guardrails): stop Vision Bridge from re-selecting a model locked after a 404 (#12111) (#13259)
  • fix(sse): require Responses-shaped body before native OpenAI-compatible passthrough (#12129) (#13278)
  • fix(providers): minimax-m3 now collapses manual thinking.type:"enabled" to adaptive, preventing upstream 400 (2013) (#12132) (#13249)
  • fix(db): scope model visibility overrides by modality so hiding a Chat model no longer hides an identically-ID'd Image/Embeddings/etc. model (#12172) (#13258)
  • fix(db): scope local-provider apiKey dedup to matching base URL so LM Studio/Ollama-style connections support multiple accounts (#12173) (#13250)
  • fix(providers): refresh Trae's stale Referer/Origin and forward user timezone so imported connections stop failing with 401 (#12190) (#13255)
  • fix(providers): route opencode-go/gpt-5.6-luna to /responses instead of /chat/completions (#12196) (#13275)
  • fix(dashboard): allow deleting the last extra-upstream-header row even when invalid (#12251) (#13252)
  • fix(i18n): translate pre-existing __MISSING__: keys for combo.sort, requestLogger.detail expand/collapse, common.profile, and settings.resilienceCredentialHealth* across 39 locales (#12272) (#13034) — thanks @HouMinXi
  • fix(cli): run the Node.js runtime compatibility guard before the heavy tsx/esm + Commander import chain so an unsupported runtime gets a clear message instead of a raw Invalid regular expression flags crash (#12296) (#13253)
  • fix(dashboard): refresh the providers list after deleting a compatible provider node (#12298) (#13256)
  • fix(usage): fail closed on API-key budget enforcement when a provider's auto routing alias has no pricing row, instead of silently counting it as $0 (#12341) (#13257)
  • fix(providers): include the submitted Agnes video model when polling by video_id (#12356) — thanks @pacocartones
  • fix(routing): custom OpenAI-compatible nodes now honor the saved Chat/Responses API type after edits instead of letting the node's original ID prefix override the live connection setting (#11884). (#12358) — thanks @pacocartones
  • fix(sse): surface an error instead of a silent empty 200 when a Claude stream closes with zero bytes (#12398) (#13285)
  • fix(oauth): warn before the dead localhost:8080 redirect in antigravity/gemini oauth start (#12413) (#13265)
  • fix(providers): stop devin-cli spawn error from double-closing the SSE controller (#12517) (#13261)
  • fix(rerank): clamp Voyage top_k to the documents actually sent after empty-string filtering, and honor return_documents: false in the NVIDIA response adapter (#12523 — thanks @pacocartones)
  • fix(audio): /v1/audio/translations now resolves combo names the way /v1/audio/transcriptions already does, so a combo that GET /v1/models advertises is fanned out to its targets instead of being rejected with 400 Invalid translation model: <combo>. Use format: provider/model; literal provider/model ids and unknown bare names behave as before (#12536 — thanks @pacocartones)
  • fix(gemini): strip the JSON-Schema-2020-12 prefixItems keyword from Gemini tool schemas at every nesting level, so Claude Code tool definitions no longer fail with 400 Unknown name "prefixItems" on Gemini models (#12540 — thanks @pacocartones)
  • fix(api-manager): Expose an accessible loading status while API keys are fetched instead of an empty accessibility tree (#12541 — thanks @pacocartones)
  • fix(video): Clamp estimateJpegFrameBytes at zero for padding-only payloads and build the three encode-side frame data URIs from JPEG_FRAME_DATA_URI_PREFIX instead of a repeated literal (#12543 — thanks @pacocartones)
  • fix(devin): accept Windows DEVIN_AGENTIC_HOME sandbox paths (C:\...\.sandbox\...) in the isolated-home check so the Devin Claude Bridge no longer fails closed on Windows (#12405) (#12545 — thanks @pacocartones)
  • fix(db): Add conversation_turn_nodes and orphaned agentic_conversations to the auto-cleanup cycle under the existing retention.callLogs window, so identity nodes whose call-log content has already been purged no longer accumulate without bound in storage.sqlite (#12548 — thanks @pacocartones)
  • fix(i18n): Wrap the ~/.claude/profiles/<name>/settings.json placeholder in ICU single quotes in the OMNIROUTE_AUTO_SYNC_CLAUDE_PROFILES feature-flag description across all 42 locales and the TypeScript default, so next-intl no longer fails with INVALID_MESSAGE: UNCLOSED_TAG and the Feature Flags card shows the description instead of the raw key (#12549 — thanks @pacocartones)
  • fix(orchestration): updateCloudAgentTask now publishes the task's real status on agent.task.updated when an update only touches result, activities or error, instead of the fabricated "updated" state, and stays silent when no row matched the id (#12550 — thanks @pacocartones)
  • fix(i18n): the home "Recent Requests" panel and the Provider Topology legend are now translated instead of rendering English copies on non-English dashboards; the legend reads its own home.topologyLegend* labels with consistent casing rather than borrowing the memory-settings "Recent" and analytics "Error" strings (#12551 — thanks @pacocartones).
  • docs(reference): bring the FEATURE_FLAGS.md catalog back to 1:1 with featureFlagDefinitions.ts — 20 missing flags added, the two *_BLOCK_THRESHOLD env-only knobs moved out of the flag tables, category/total counts and the Live WS port corrected, guarded by a static test (#12552 — thanks @pacocartones)
  • fix(i18n): backfill missing usage.kiloPass* strings in 39 locales and restore featureFlagOmnirouteDisableThinkingLevelVariantsDescription in pt.json (#12561) (#13027) — thanks @HouMinXi
  • fix(docker): default docker-compose app ports (dashboard/API/live-WS) to loopback instead of 0.0.0.0, closing the anonymous /v1 LAN/WAN exposure gap left open by REQUIRE_API_KEY=false (#12568) (#13240)
  • fix(api): close DNS-rebinding SSRF gap in webhook outbound-URL guard (#12569) (#13243)
  • fix(docker): require a per-session token on the VNC browser CDP bridge and isolate it on a dedicated Docker network (#12571) (#13241)
  • fix(open-sse): write Adobe Firefly session tokens and cookie jars with 0700/0600 permissions instead of the process umask (#12572) (#13242)
  • fix(api): remove hardcoded wildcard CORS in openai-to-gemini-sse.ts so the centralized fail-closed CORS gate is the sole source of Access-Control-Allow-Origin (#12573) (#13239)
  • fix(api): enforce API key policy (budget/rate-limit/schedule/endpoint scoping) on the ElevenLabs speech-to-text, text-to-speech and voices proxy routes (#12574) (#13244)
  • fix(sse): cap HuggingChat NDJSON body size and bound the read loop with the fetch timeout so a stalled or hostile upstream cannot buffer unbounded memory (#12577) (#13282)
  • fix(docker): scope the cliproxyapi/qdrant/bifrost sidecars to loopback by default and forward CLIPROXYAPI_MANAGEMENT_KEY into the cliproxyapi container so its management API is not left both unauthenticated and LAN-published (#12578) (#13240)
  • fix(api): create DB export temp paths with fs.mkdtempSync instead of predictable timestamps (#12579) (#13237)
  • Cline 401 bodies that say "re-authenticate your Cline account" classify as a refreshable OAuth token, not a terminal expired key. The cooling panel no longer labels every cooldown as a 429; it shows the recorded last error instead. (#12594) (#13060) — thanks @HouMinXi
  • fix(catalog): degrade unknown combo targets instead of dropping LCD modalities, and persist OpenRouter architecture.input_modalities into the capability snapshot (#12613) (#13042) — thanks @HouMinXi
  • fix(providers): send x-api-key instead of Authorization: Bearer for OpenCode Zen's /v1/responses endpoint (Muse Spark Contributor models), fixing a 401 on OmniRoute's auth header (#12633) (#13247)
  • fix(sse): surface the actionable "Auggie CLI not found" message when the shell reports a missing auggie binary via exit code instead of a spawn error (#12645) (#13263)
  • fix(config): Persist deletions of built-in background-degradation entries — when a stored settings record exists its degradationMap is now authoritative instead of being merged under the defaults, so an entry the user removed in the dashboard no longer reappears on the next apply or restart (#12424) (#12647) — thanks @pacocartones
  • fix(gamification): action-count badge milestones (First Token, Token Consumer, Token Machine, Token Whale, and the token-sharing tier) are now backed by a durable xp_action_counts counter incremented in addXp(), instead of a live COUNT(*) over xp_audit_log. The audit log is pruned by retention.xpAuditLog (default 30 days), so on a default install those "lifetime" milestones were really "actions in the last 30 days" and unlocked badges could stop unlocking once old rows aged out. getActionCount() and checkActionCountBadges() now read the same durable source, and a migration backfills existing totals from the surviving audit rows (#12546) (#12651) — thanks @pacocartones
  • fix(images): /v1/images/edits now iterates a combo's targets the same way /v1/images/generations does (#9239) instead of flattening a bare combo to its first target. A combo whose first target is not edit-capable — or lacks credentials — now falls through to a later edit-capable target rather than hard-erroring, and missing credentials are skipped (not a hard 401) to match the generations path. The per-target skip/terminal classification is extracted into a shared runImageComboTargets loop, so generations behavior is unchanged (#12547). (#12653) — thanks @pacocartones
  • fix(sse): add first-byte watchdog to the TLS-fingerprint transport so a stalled wreq body falls back instead of hanging for minutes (#12656) (#13272)
  • fix(sse): exempt tiny-budget reasoning probes from combo quality failure and surface persisted-cooldown skips in ALL_TARGETS_SKIPPED diagnostics (#12659) (#13269)
  • fix(models): declare the real ~1M contextLength for OpenCode Zen's Muse Spark 1.2 models instead of falling back to the 200000 provider default (#12681) (#13247)
  • fix(oauth): align codebuddy-cn OAuth User-Agent with the chat/usage CLI version to avoid WAF false positives (#12702) (#13264)
  • fix(dashboard): surface an authentication-required banner instead of silently blanking database settings for a guest session (#12709) (#13270)
  • fix(resilience): stop unbounded queue that hangs 6min until Aborted — gate, provider slot, and Bottleneck queue now share a per-connection maxWaitMs budget; fail-closed on exhaust (503); execution backstop executionMaxWaitMs overridable per connection with upstream clamp (#12715) — thanks @maxmad64bis (with thanks to @Tushar49 for surfacing the slow-provider need in #12635)
  • fix(auto-combo): every mode pack now carries quality and reliability weights (quality 0.02, 0.03 in quality-first; reliability 0.03, 0.04 in reliability-first), so selecting a pack no longer silences either signal; DEFAULT_WEIGHTS is unchanged (#12731) — thanks @maxmad64bis
  • fix(sse): the provider execution pipeline keeps the upstream error code/type again — both were lost when the non-streaming leg was extracted, so a config-class 422 (Antigravity missing project) degraded into a generic account cooldown instead of reaching the gate that recognises it (#12732) (#13349)
  • fix(db): a legacy call_logs table no longer aborts startup — the provider-stats index is created after the column healing runs, not before it, so an install predating request_type opens instead of failing with no such column (#12732) (#13349)
  • fix(catalog): a malformed row in the operator-writable custom-models blob no longer kills every auto/* pool (#12732) (#13349)
  • fix(tests): realign the guards that the same day's merges left asserting the old behavior — the provider-count total after GreenPT and EURouter, the GLM stream buffer slot that #12925 turned into a declared parameter, and the injection scan bound that #13104 reshaped from a head-only window into head-plus-tail (#12732) (#13349)
  • fix(cache): fold tool_choice/tools/response_format into the semantic cache signature so a cached tool_calls response can no longer be replayed for a request whose tool policy forbids it (#12734) (#13267)
  • fix(models): stop treating provider-supplied isFree:true/:free/0/0 flags as trusted unless the provider has a documented free tier — only the shipped free-tier catalog decides otherwise; fetched isFree:true trusted only for a free-tier provider, custom isFree:true via a trusted path (#12744) — thanks @maxmad64bis
  • fix(memory): authenticate the internal /v1/rerank loopback call so memory reranking no longer silently degrades to unranked order when REQUIRE_API_KEY=true (#12745) (#13268)
  • fix(sse): parse Ollama Cloud's current usage markup ($X of $Y used aria-label, nested width style) (#12749) (#13260)
  • fix(cli): setup-opencode no longer sends an active context's management token to /v1/models when --api-key/OMNIROUTE_API_KEY is supplied — an explicit flag or the env var now always outranks the context's token, and the flag itself is no longer swallowed by the parent program's global --api-key option (#12783) (#13246)
  • fix(sse): register Arcee AI in the executor provider registry so requests reach api.arcee.ai instead of silently falling back to OpenAI (#12784) (#13277)
  • fix(quota): share one upstream quota read when identical saturation checks arrive at the same time, so traffic bursts don't multiply provider API calls (#12787) — thanks @maxmad64bis
  • fix(db): ignore expired or invalid rate-limit cooldown writes so a stale timestamp can't lock a connection that should be usable — clearing still works as before (#12788) — thanks @maxmad64bis
  • fix(providers): single source for provider order with xao ranking alongside xai-oauth (#12790) — thanks @maxmad64bis
  • fix(routing): off-table models check the free-model catalog before inheriting premium prices, latency bootstraps from the observed pool median, async tiers read live database pricing with a 90-day freshness gate, and the tier cache invalidates on every pricing write (#12792) — thanks @maxmad64bis
  • fix(auto-combo): snapshot scoring uses the observed per-connection account tier and per-model quality instead of neutral constants, and circuit-open providers rank lower at build time (#12794) — thanks @maxmad64bis
  • fix(pool): empty auto/* pools now say why they are empty, and the models listing uses the same paid check as routing (#12795) — thanks @maxmad64bis
  • fix(routing): recognize CLIProxyAPI's 'unknown provider for model' 400 as fallback-worthy (#12800) (#13284)
  • fix(sse): translate-mode streams now emit the estimated token counts as a trailing usage-only chunk before [DONE] when the upstream stays silent, so metered chat clients see totals instead of nothing (#12828) — thanks @maxmad64bis
  • fix(db): provider stats stay truthful on empty and legacy databases: fallback counts default to 0 instead of null, latency averages read null (not 0) when no durations were recorded, failures logged before the error-type column existed group under pre_migration instead of unclassified, and per-provider queries use two new composite indexes (#12832) — thanks @maxmad64bis
  • fix(nvidia): fail open when a synced model catalog goes stale instead of gating forever (#12849) (#13248)
  • fix(db): WAL maintenance lives in its own module: a TRUNCATE checkpoint that hits a busy database now warns with its streak and retries once via PASSIVE instead of logging success, and closing-time checkpoints no longer report success on builds without a database file, with the busy totals visible in the authenticated monitoring health payload (#12853) — thanks @maxmad64bis
  • fix(health): quota with no snapshots reads empty instead of a contradicting 0% (#12857) — thanks @maxmad64bis
  • fix(a2a): accept the dashboard's own session cookie on /a2a so "Run message/send" no longer fails with "Unauthorized: missing or invalid API key" (#12888) (#13271)
  • fix(logs): the Logs grid's in-memory filter pass no longer discards rows the SQL query already matched — selecting an API key from the dropdown (which sends the key's id) returns its calls again, the Combo tab shows every combo instead of only those whose name contains a "1", and the model filter and search cover the same columns as the query (#12896) — fixes #12873 — thanks @ntdatt812
  • fix(a2a): /api/a2a/status now builds the agent card from the request that asked for it, so a gateway reached at a non-localhost host no longer advertises http://localhost:20128 as its A2A URL (#12918). — thanks @ntdatt812
  • fix(compression): progressive aging now appends its [COMPRESSED:aging:…] annotation after a turn's tool_result blocks instead of in front of them, so Anthropic no longer rejects aged conversations with "tool_use ids were found without tool_result blocks immediately after" (#12920). — thanks @ntdatt812
  • fix(bedrock): model import now resolves context limits for every vendor prefix instead of only anthropic.*, so global.openai.gpt-5.6-* no longer imports with a null inputTokenLimit and gets rejected pre-flight at the 200k default (#12921). — thanks @ntdatt812
  • fix(stream): the 64 KB stream buffer GLM asks for is honoured instead of dropped, and the type error it caused no longer fails the API Route Typecheck gate on every open PR (#12925) — thanks @ntdatt812
  • fix(guardrails): mask PII inside a tool_result's nested content array, which the masker walked past while redacting its sibling block (#12930) — thanks @ntdatt812
  • fix(dashboard): the Radar catalog table no longer leaves absent data unexplained — empty limits, unknown context windows, and unreported capabilities each explain themselves on hover, and a new check keeps it that way (#12937) — thanks @maxmad64bis
  • fix(sse): geo-blocked opencode requests rotate to the next account proxy instead of failing, so one refused egress no longer aborts the whole chain (#12941) — thanks @maxmad64bis
  • fix(codex): dashboard "clear cooldown" and the CAS recovery path now drop nested codexScopeRateLimitedUntil maps in the same write that nulls rate_limited_until, and fresh quota snapshots with headroom lift fallback-sourced scope cooldowns parked by quota preflight (#12817, #12860, #12951) The dashboard reset-credit button (consumeCodexResetCredit) now exercises that same snapshot path after a successful redeem, so a filled quota bar is enough to unpark a leftover fallback Codex child without another manual clear. — thanks @HouMinXi
  • fix(db): resolve sql-wasm.wasm across global npm install and hoisted layouts, ensuring OmniRoute can boot cleanly on Node 24 when native better-sqlite3 is uncompiled. (#13035) — thanks @HouMinXi
  • fix(api): DELETE /v1/batches/delete-completed now sweeps only the calling API key's own completed batches (batches with no owner stay out of a key-scoped sweep on purpose), with an explicit instance-wide mode reserved for authenticated dashboard sessions, a 401 for a presented key that is unknown, revoked, deactivated, banned or expired (never falling through to the session branch), audit logging of every sweep, a sanitized 500 on failure, an owner-scoped file half (a key-scoped sweep never nulls a file another tenant owns) and an atomic sweep — chunked in 200-batch transactions in instance mode — so a mid-way error never leaves a batch pointing at a nulled file (GHSA-wvxc-jp3v-5mg5) (#12969)
  • fix(combo): quota-weighted routing stops drawing on an out-of-credit connection — a 402 now invalidates the stored quota snapshot instead of leaving its stale remaining percentage in place, and a snapshot older than 10 minutes no longer counts as confident headroom for the primary pool (#12972) (#13006) — thanks @HouMinXi
  • fix(vision): Custom Models with "Vision capable" checked no longer have image requests swapped to glm/glm-4.6v when the client sends the advertised alias (vllm/path/...) or the bare path-shaped id — Vision Bridge now matches the stored override for all three id forms (#12758) (#12974) — thanks @HouMinXi
  • fix(sse): transient opencode upstream failures rotate to the next account proxy instead of failing, so one flapping egress no longer aborts the whole chain (#12975) — thanks @maxmad64bis
  • fix(azure): Deployments from GPT-6 onward now send max_completion_tokens instead of max_tokens, which Azure rejects with HTTP 400. The rule matched a literal gpt-5, so each new generation arrived broken; it now matches the generation range, while gpt-35-turbo still keeps max_tokens. (#13007) — thanks @ntdatt812
  • fix(db): add an opt-in self-restart circuit for sustained critical memory pressure, gate post-cleanup VACUUM behind a minimum freed-rows threshold, and checkpoint the SQLite WAL every 5 minutes with a size guard that escalates to TRUNCATE, so a growing WAL can no longer stall the event loop into a full outage. (#13011) — thanks @HouMinXi
  • fix(auth): an explicit connection pin may probe a quota-disabled row once and re-enable it on success (#12874) (#13017) — thanks @HouMinXi
  • fix(api): thread X-OmniRoute-Fallback-Attempts through combo chat completions so streaming and non-streaming responses report how many prior legs were attempted (#13038) — thanks @HouMinXi
  • Restore provider failure classification and credential refresh on non-streaming requests: classify non-2xx failures to lock models on per-model quota exhaustion, update connection rate limits from headers and body, and pass credential refresh handlers to pipeline execution so 401 tokens can be refreshed and retried (#13043). (#13069) — thanks @HouMinXi
  • fix(responses): wrap forced-non-streaming web_search fallback JSON as Responses SSE so Codex still sees response.completed (#13050) — thanks @HouMinXi
  • fix(antigravity): Preserve upstream thought-token usage in normalized completion and reasoning token counts (#13055) — thanks @pacocartones
  • Fix a concurrency-slot leak in the DeepSeek PoW solver: a worker that failed to spawn (for example a missing worker script) never released its slot, so MAX_CONCURRENT_WORKERS failures disabled the solver until restart. (#13097) — thanks @anhtahaylove
  • fix(acp): bound the ACP session output buffers — stdoutBuffer and stderrBuffer now cap at 1 MiB keeping the most recent output behind a visible [...output truncated...] marker, and stderrBuffer is reset per prompt instead of accumulating for the lifetime of the session. (#13100) — thanks @anhtahaylove
  • fix(acp): release the stdout/exit listeners and the idle timer that a sendPrompt timeout used to leave attached to the acpManager singleton, and drop sessions that exited on their own from the session map instead of keeping them forever. (#13096) — thanks @anhtahaylove
  • fix(security): the prompt-injection and PII scanners now read the text a tool_result block carries on content (string or nested block list), in messages and in system blocks, so tool output is judged by the same rules as user text (#13101) — thanks @ntdatt812
  • fix(gamification): close the badge notification SSE stream when the request signal is already aborted before the stream starts — a client that disconnects while the route is still awaiting auth used to leave both the 2s unlock poll and the 15s heartbeat running for the lifetime of the process. (#13106) — thanks @anhtahaylove
  • fix(security): the prompt-injection scan now spends its 16 KB budget on both ends of the request instead of the first 16 KB only, so system, instructions, query, documents and the newest turns are no longer hidden behind one long message (#13104) — thanks @ntdatt812
  • fix(dashboard): expose the Volcano Ark console cookie on quota scraping and unwrap connect-error objects so the dashboard shows the upstream message (#13107) — thanks @HouMinXi
  • fix(db): release the beforeExit/SIGINT/SIGTERM handlers when a node:sqlite adapter closes, so a closed adapter and its database handle are no longer pinned to process for the lifetime of the run — the same treatment #7494 gave the sql.js adapter. (#13109) — thanks @anhtahaylove
  • fix(translator): contentSchema and unevaluatedItems are now treated as subschema positions by the tool-schema sanitizer, so a truncation placeholder in either is replaced with a permissive schema instead of being forwarded as a string (#13110) — thanks @ntdatt812
  • fix(cli-helper): clear the createLogStream timeout on the abort path — stop() aborts the in-flight fetch and returned through the signal.aborted branch, which skipped clearTimeout and left an armed timer per stopped stream. The stream reader is now also cancelled when the read loop exits early. (#13114) — thanks @anhtahaylove
  • fix(docs): drop TLS-impersonation, MITM-decrypt, supply-chain attestation, and XOR-mask writeups from the public /docs catalog and Docker image. Files stay in git for engineers; operators who need them open the repo, not the website. (#13136) — thanks @HouMinXi
  • fix(combo): parse numeric-epoch rate_limited_until in the combo cooldown read path (#13141) — thanks @maxmad64bis
  • fix(opencode): both OpenCode plugins now reject a gateway address typed without http:// at configuration time, instead of publishing every model with an api url no client can call, and the v2 plugin no longer publishes a model card whose api url is blank or relative (#13142) — thanks @maxmad64bis
  • fix(providers): lock opencode model on upstream 400 model-unavailable (#13146) — thanks @maxmad64bis
  • fix(logging): keep the provider exchange rather than the raw client bodies when a call log exceeds its size budget, and show that recovered payload in the request-detail panel instead of replacing it with the stored response body (#13147) — thanks @maxmad64bis
  • fix(traffic-inspector): the WebSocket route no longer leaks a traffic-buffer subscriber and a 30s ping timer when the client socket is already closed at handler time — listeners are attached before any resource is acquired, a destroyed socket bails out early, and the ping interval stops on a dead socket where write() never throws (#13155) — thanks @anhtahaylove
  • fix(telegram): bound the per-user API key cache in the Telegram chat proxy so a burst of distinct chat ids can no longer grow the process heap without limit (#13165) (#13166) — thanks @anhtahaylove
  • fix(stream): cancel the upstream response body when the JSON-to-SSE sniff unwinds on a body timeout, so a stalled upstream no longer pins the connection (#13169) (#13171) — thanks @anhtahaylove
  • fix(telegram): authenticate webhook deliveries with Telegram's secret_token so an unauthenticated caller can no longer mint API keys or spend upstream quota (#13172) (#13175) — thanks @anhtahaylove
  • fix(models): give discoverable Gemini 3.8 Flash ids their own 65536 output spec so Antigravity no longer clamps them to 16384 (#13195) — thanks @HouMinXi
  • fix(dashboard): leftover catalog-removed provider rows (gemini-cli) can be listed and purged from the providers page (#13067) (#13197) — thanks @HouMinXi
  • fix(auth): a dashboard session now requires the authenticated: true claim that login, OIDC and the session refresh already emit — a JWT merely signed with JWT_SECRET (for example the Cursor CLI passthrough token, which any API-key holder can obtain) no longer verifies as the auth_token cookie on any route, the WebSocket handshake or the live server; existing sessions keep working (#13298) (#13375)
  • fix(api): DELETE /v1/batches/delete-completed now applies the caller's API-key policy (endpoint allowlist, schedule, usage cap, rate limit) like every other /v1 route, commits the key-scoped sweep per 200-batch chunk instead of holding one write lock for the whole sweep, refuses to loop without progress, and rejects a scope that names both a key and allTenants (#13684)
  • Restore the API-route typecheck gate: the non-streaming leg result lost its discriminated-union narrowing after the server-owned tool loop reassignment, producing 13 new TS2339 diagnostics in chatCore.ts. (#12963)
  • Restore four non-streaming behaviours the server-owned tool loop refactor dropped: client aborts map to 499 with the fixed Request aborted message, an aborted request no longer logs a synthetic clientResponse, Claude prompt-cache telemetry is recorded again, and a body that cannot be canonicalized no longer throws when the tool loop is off. (#12990)
  • fix(skills): The CLI registry parser now reads positionals declared with .addArgument(), not only those written inline in .command(). tunnel create [type] was being published as tunnel create, so the agent-skills sync gate reported drift on every branch and regenerating would have deleted the argument. (#13009) — thanks @ntdatt812
  • fix(combos): dashboard combo test uses a short prompt, serial probes, and a 60s timeout so reasoning models and rate-limited free pools do not fail the health check (#13001) — thanks @HouMinXi
  • fix(compression): terminate idle worker threads on eviction so long-running instances stop leaking OS threads and MessagePorts (#13091) — thanks @anhtahaylove
  • Redact Google API keys of any length in error bodies: the pattern required exactly 39 characters, so shorter or longer AIza… credentials passed through unredacted. (#12964)
  • fix(grok-cli): a 402 "Grok Build usage balance exhausted" parks that Grok login as out of credit (Grok Build CLI, grok.com cookie, and xAI OAuth share the weekly pool). Combo routing then tries the next login instead of locking the model for every account in the pool (#13061) — thanks @HouMinXi
  • fix(compression): spawn the LLMLingua worker with a file URL object so compression actually runs instead of silently failing open on Node (#13093) — thanks @anhtahaylove
  • fix(test): run the local test and test:unit scripts at concurrency 4 so a full-suite run no longer exhausts the machine's commit charge and kills unrelated processes (#13187) — thanks @anhtahaylove
  • fix(monitoring): GET /api/monitoring/health credentialHealth now includes a bounded failedConnections list (connectionId, status, sanitized lastError) when the probe-cache gauge failed>0, plus source: probe-cache and a cheap staleDbNonOkCount for sticky SQLite test_status on active rows. Documents that the live gauge is not provider_connections.test_status. (#12876) — thanks @RaviTharuma
  • fix(resilience): an openai-compatible multi-upstream gateway no longer marks the whole connection credits_exhausted when a single upstream (or the /models / representative-model health probe) returns 402. Quota failures stay model-scoped; true connection-wide auth failures are unchanged. Single-credential 402 key disable (#5239) is preserved. (#12875) — thanks @RaviTharuma
  • fix(dashboard): nine fixes on the Orchestration Canvas, all diagnosed in the Phase 2 reviews. A source that fails now keeps the timestamp of its FIRST failure instead of being re-stamped every poll — the stale line said "since the last poll" no matter how long the source had been down, and the churn also defeated the snapshot's stable identity (it serializes the source list), so the canvas re-rendered on every tick while anything was broken. A source that HAD data and only then started failing is flagged too: previously only a source with no node at all got the warning, so a source that went down mid-session kept a healthy-looking node forever. The rest are pointwise: clicking a filter chip cancels the pending search debounce (left armed, it fired ~300ms later and silently reverted the chip); the search input carries an accessible name; ?state=running, failed parses like the unpadded form instead of dropping the padded value; the CSV toggle helper is defined once in model/urlParams.ts rather than twice; the agents tab tells "nothing running" apart from "the filter matched nothing", offering a clear-filters button instead of setup links that would be wrong advice there; edges stop emitting SMIL particles above 40 simultaneously active edges, keeping the colored stroke; and the drawer's error banner clears when a retried action succeeds. (#12983)
  • fix(i18n): the nine locales added with the EU-language batch (Greek, Estonian, Irish, Croatian, Lithuanian, Latvian, Maltese, Slovenian, Serbian) were missing the eleven Orchestration Canvas keys that Phase 3 introduced, so the compare-runs panel and the "no runs match these filters" empty state fell back to English in those languages (deepMergeFallback substitutes English for an absent key, so nothing rendered blank — it rendered untranslated). The coverage gate does not catch this: it enforces an 80% floor per locale, and eleven missing keys out of ~13,000 leaves coverage at 99.9%. Translated for real in each language, calibrated against the wording each file already uses for "run", "filter" and "skill". (#13294)
  • fix(dashboard): five follow-ups from the Phase 2 Orchestration Canvas review. The History drawer now shows the memory section for runs that already left the live TTL window: the persisted memory_hits event is parsed into metadata.memoryHits with the same defensive validation the drawer applies, and — because that event is observability rather than a state transition — it no longer leaks into the timeline, where it had been inheriting the task's state and rendering as a duplicate transition. The A2A memory recall runs against its own 1.5s deadline instead of inheriting the memory backend's 30s one; overshooting degrades exactly like any other recall failure (no hits, task proceeds), and the timer is cleared on both paths. Repeating a Conductor run carries its requirements.cli/requirements.model forward, so the new run is pinned to the same runner profile and model rather than drifting to whatever the fleet picks. A successful repeat from the Agents tab now focuses the run it created, instead of leaving the operator on the finished one. And the auth test for POST /api/conductor/tasks moved into the shared ROUTES array rather than restating the pattern. (#12988)
  • fix(plugins): stop leaking an exit listener per plugin hook timeout, which triggered MaxListenersExceededWarning on plugins that ignore SIGTERM (#13092) — thanks @anhtahaylove
  • fix(validation): Provider node edits no longer fail with a generic "Invalid request" when the optional daily-quota reset fields are left blank. The dashboard sends dailyQuotaResetTimezone and dailyQuotaResetHour as null, and only the hour accepted it. (#13066) (#13083) — thanks @ntdatt812
  • fix(combo): quota-aware expansion drops banned, inactive, missing, and wrong-provider connections before quota fetch or model dispatch; pins and allowlists stay selectors, not a bypass. Antigravity automatic exhaustion now requires a reported zero remaining, so a positive balance below 1% stays eligible. (#13006) — thanks @HouMinXi
  • Fix an intermittent failure in the quota-weighted routing test suite: scores are a function of Date.now(), so peers with identical quota scored microseconds apart never tied and swapped order. (#12946)
  • fix(volcengine): resolve quota window mapping gaps, safe multi-connection console binding, and response cookie sanitization (#12950) — thanks @HouMinXi
  • fix(test): resolve the WebDAV handler path with fileURLToPath so the suite's 37 WebDAV tests run on Windows instead of failing with a doubled C:\C:\ drive prefix (#13196) — thanks @anhtahaylove
  • fix(ui): fall back from invalid AnySearch icon alias (direct commit 9587e07b69)
  • fix(sse): stop rebuilding a truncated summary from the collector's cap-dropped event array (#12718) — thanks @hartmark
  • fix(conversations): show a pending spinner for unresolved tool nodes instead of "(empty)" (#12727) — thanks @hartmark
  • fix(streaming): fail fast when an upstream stream produces only lifecycle/heartbeat events, never real content (#12741) — thanks @hartmark
  • fix(responses-continuation): bridge the write-in-flight window with an in-memory pending store (#12854) — thanks @hartmark
  • fix(translator): drop or map agent_message on Chat Completions fallback (#12880) — thanks @RaviTharuma
  • fix(api): page and stream-complete GET /v1/models for large catalogs (#12882) — thanks @RaviTharuma
  • fix(quota): align AUTH window labels with usage API and clarify 503 (#12884) — thanks @RaviTharuma
  • fix(ci): clear two base-reds on release/v3.8.51 — mutation-coverage gate + image-only-model guard (#12945)
  • fix(api): scope batch bulk-delete to the calling API key (#13211)
  • fix(tests): align model-sync log assertion to shipped path redaction (base-red) (#13227)
  • fix(test): make the ToS heading guard actually require the parentheses (#13228)
  • fix(quality): register 4 drifted covering tests in stryker tap.testFiles (#13229)
  • fix(logging): mask *-api-key headers in request-log pipeline (#13273) (#13401) — thanks @KooshaPari
  • fix(tests): correct stale Jina catalog prefix in custom-models test (#13403) — thanks @KooshaPari
  • fix(backend): prune old backups after health-check-repair snapshots (#13308) (#13404) — thanks @KooshaPari
  • fix(backend): guard memory_fts_au trigger with WHEN clause to stop FTS bloat (#13326) (#13405) — thanks @KooshaPari
  • fix(free-tiers): remove obsolete Together $25 signup credit (#12526) (#13407) — thanks @KooshaPari
  • fix: skip non-chat generation models in Test all models (#13409) — thanks @KooshaPari
  • fix(providers): de-duplicate agy and antigravity model catalogs (#13410) — thanks @KooshaPari
  • fix(dashboard): send provider in body for mini-playground web search (#13411) — thanks @KooshaPari
  • fix(resilience): add CJK quota-exhaustion patterns to 429 classifier (#13412) — thanks @KooshaPari
  • fix(backend): eval runner: failed calls no longer score as passed + dotAll regex (#13413) — thanks @KooshaPari
  • fix(backend): eval runner opts out of output-style and memory injection (#13414) — thanks @KooshaPari
  • fix: chain serial quarantine step in npm test script (fixes #13286) (#13418) — thanks @KooshaPari
  • fix(antigravity): derive stable session ID from credentials for prompt-cache reuse (#13305) (#13424) — thanks @KooshaPari
  • fix(memory): merge memory into top-level system field instead of unshifting at messages[0] (#13425) (#13427) — thanks @KooshaPari
  • fix(compression): preserve instruction blocks from lossy engines (#13523) — thanks @KooshaPari
  • fix(compression): preserve polarity words and newlines in ultra heuristic (#13524) — thanks @KooshaPari
  • fix(compression): run user preservePatterns before built-in patterns (#13457) (#13525) — thanks @KooshaPari
  • fix(reasoning): expand deepseek-v4 regex to match v4.1+ and add command-code provider (#13430) (#13528) — thanks @KooshaPari
  • fix(providers): add alwaysReasons metadata for reka-flash-3 (#13531) — thanks @KooshaPari
  • fix(playground): allow Compare columns to scroll independently (#13532) — thanks @KooshaPari
  • fix(security): empty whitelist allows through to prevent admin lockout (#13534) — thanks @KooshaPari
  • fix(tests): add DISABLE_IOREG_STRATEGY env var for macOS test compatibility (fixes #13467) (#13539) — thanks @KooshaPari
  • fix(zed-hosted): map developer role and Gemini enums for Zed proxy (#13543) — thanks @KooshaPari
  • fix(analytics): normalize provider aliases to canonical id in usage_history (#13545) — thanks @KooshaPari
  • fix(combo): use traceId for call log id to prevent UNIQUE constraint failure (#13546) — thanks @KooshaPari
  • fix(auth): check model lockout before returning synthetic noauth connection (#13547) — thanks @KooshaPari
  • fix(guardrails): skip credential redaction for base64 image data URLs (#13550) — thanks @KooshaPari
  • fix(resilience): increase requestQueue.maxWaitMs default from 15s to 30s (#13553) — thanks @KooshaPari
  • fix(docs): document OMNIROUTE_READY_TIMEOUT_MS and allowlist the test-only DISABLE_IOREG_STRATEGY (#13692)
  • Electron release workflow: the publish-npm job now grants actions: read to the reusable npm-publish.yml it calls (its publish job requests it), which is what made GitHub refuse the whole v3.8.50 run at startup and ship the release with zero desktop assets; a workflow_dispatch now builds the requested tag instead of the dispatching branch and can skip the npm leg (publish_npm=false) when only re-attaching assets (#11974)
  • fix(api): /v1/files and /v1/batches now enforce one ownership rule everywhere — a dashboard session is the instance operator, an API key acts on its own records only, and a record with no owner is denied to every non-session caller. Previously a file or batch whose api_key_id was null (a dashboard-session or anonymous upload, or a batch artifact inheriting one) could be read, downloaded, deleted, cancelled or used as a batch input by any other key or by an unauthenticated caller (GHSA-2jm2-mpx8-6523), and GET /v1/files / GET /v1/batches returned every tenant's records to an anonymous or invalid-bearer caller under the default REQUIRE_API_KEY=false (GHSA-m3hp-hq9g-fpmv) — both lists now fail closed with a 401, and only a dashboard session without a key reads the whole instance. The same shared rule lets the dashboard cancel any batch, not just unowned ones. Behaviour change: the anonymous upload → batch → download flow no longer works without an API key, since a null owner cannot be attributed. Subsumes #13683 (#13749) — thanks @hartmark
  • fix(i18n): every dashboard catalog other than pt-BR (64 locales) went through the same quality review pt-BR received in #13885 — each leaf changed by the 2026-09 retranslation was checked against its English source by the translation backend and rewritten where the meaning, placeholders, register or product terminology were off: 73,586 corrections net (75,263 applied, 1,677 that had turned a real translation into the plain English term reverted so the real-translation ratio gate stays where it was). scripts/i18n/review-locale.mjs now survives an upstream hiccup (per-batch retries with backoff, skipped batches listed in _artifacts/i18n-review/<code>.skipped.json), checkpoints the catalog every 25 batches instead of writing only at the end, and writes leaves whose own key contains a dot (compliance.eventTypes["apiKey.ban"]) instead of crashing. (#14078)
  • fix(auth): closed the JWT_SECRET bootstrap chain (GHSA-7pq4-8pvv-rx7r). The fresh-install bootstrap gate in isAuthRequired() now decides "loopback" from the trusted peer — the token-stamped real TCP peer the custom server writes, the pipeline's own locality verdict, or a real socket — and never from the client-controlled Host / nextUrl.hostname whenever a stamping server is in front (every supported runtime), so Host: localhost from a remote address no longer opens the window; the anonymous first-password write (POST /api/settings/require-login) is under the same loopback constraint instead of being open to every network peer, and managementPolicy hands its peerContext verdict down explicitly. /api/settings/obsidian (incl. /webdav, which mints reusable WebDAV Basic credentials for a caller-chosen root served before Next.js) joined ALWAYS_PROTECTED_API_PATHS, and enableObsidianVaultSync() refuses a vault that is, sits inside, or contains the data directory (realpath-resolved), so the WebDAV file service can no longer be pointed at server.env / storage.sqlite (#13791)
  • Combo routing: a context-cache-pinned model that returns 401 now falls through to the normal combo fallback loop instead of terminating the request, allowing other eligible connections or providers to serve it. (#12818) — thanks @keeltrace
  • fix(sse): the Antigravity account picked for a request can now be reserved for that request's streaming lifecycle, so a concurrent retry or the credential handoff cannot re-pick an account already committed to an in-flight stream; a fully leased pool answers with a structured 503 antigravity_pool_busy carrying a bounded Retry-After. Opt-in behind the new ANTIGRAVITY_ACCOUNT_LEASE_ENABLED flag (default off) (#10011) (#13929) — thanks @Ardem2025
  • fix(docker): bump the Bun image to 1.4.0, enable Turbopack on Bun, and port the node image's build memory guards so the -bun container builds fit the 16 GB GitHub runner instead of dying with cannot allocate memory (#11719). Both images now default OMNIROUTE_BUILD_WORKERS to 2 (1 page-data worker) against the measured ~4.5 GB per-process RSS budget (#7518/#11663). (#13140) — thanks @ozeas
  • security(runtime): fail closed on hostile thrown values and keep upstream text out of public error surfaces — the chat pipeline now reads rejection metadata through a safe accessor, sanitizes the message before it reaches call logs and console, and projects the failure-usage code onto the bounded public vocabulary; Perplexity's non-streaming quota/upstream error body sanitizes the upstream message and projects the provider-supplied error code; Arena (lmarena) maps every public failure onto a fixed vocabulary instead of echoing the upstream error; Notion's TLS transport failure sanitizes the transport error before it reaches the response body (#11742).
  • fix(resilience): only clear the combo-level LKGP pin when it names the target that actually failed, so an unrelated target skip under auto/round-robin no longer discards a valid pin for a healthy provider (#12235) — thanks @abhisheksharma2411
  • fix(sse): 429 bodies phrased as N API calls / month (Cohere trial keys) now classify as quota_exhausted instead of a short transient rate_limit, so a spent monthly allowance is no longer retried every few seconds for the rest of the billing cycle (#12252) — thanks @brick30llc-ctrl
  • fix(cache): fold response_format/Responses-API text.format into the semantic cache signature so a temp=0 request can no longer be served a stored response body with a different output schema (#12307) (#12309) — thanks @amirrezakm
  • fix(gemini): preserve response-schema nullability across union flattening so a model with nothing to say returns a valid null instead of the string "null" or a fabricated value (#12308) (#12310) — thanks @amirrezakm
  • fix(combo): a priority combo whose steps are different models on one Claude OAuth connection now falls through to the next step — a model-specific 404 or 5xx is scoped to the model instead of retiring the whole account, while a 429 stays account-wide (#12334) (#12340) — thanks @Kizuno18
  • fix(api): restore the name field on non-streaming /v1/responses function_call output items — a plain (non-namespace) tool call's identity restore was blindly applying the _toolNameMap alias-table fallback as a {namespace, name} object, silently blanking name to undefined (dropped entirely by JSON.stringify) and leaving Codex unable to dispatch the call, so it re-narrated its intent in a loop instead (#12370) (#13824)
  • fix(memory): extracted facts and oversized extraction input are now truncated at a word or sentence boundary instead of at a hard character offset. sanitizeMatch() (500-char fact cap) and capExtractionText() (64KB extraction-input cap) previously sliced at the exact limit, which could cut a fact mid-word or mid-clause; both now back the cut index off within an 80-char lookback window, preferring sentence-ending punctuation (. ! ?), then a plain word boundary, and only falling back to the original hard cut when neither is found — the same pattern already used for compressToolResults (#8169) — thanks @kareem-jalal / @LeMonBLOCK (#12383)
  • fix(chatCore): stop executeWithUpstreamStartTimeout leaking its abortPromise listener onto the long-lived client/stream signal, and stop mergeAbortSignals leaking per-attempt abort listeners, so a later hedge cancellation or client disconnect cannot reject an orphaned promise and take the process down (Error [AbortError]: hedge-cancelled). The crash guard also absorbs combo abort reasons (hedge-cancelled, combo-per-model-timeout) and raw string disconnect reasons as a last-resort net (#12406 — thanks @Beexly)
  • fix(db): give conversation_turn_nodes its own independent retention knob (retention.conversationTurnNodes, default 30 days — matching callLogs so upgrading changes nothing until an operator overrides it) instead of sharing callLogs, and sweep orphaned agentic_conversations after the nodes expire (#12453). (#13344) — thanks @HouMinXi
  • fix(usage): Render OpenRouter PAYG account credits as a metered quota when no per-key spending limit is set (#12468) — thanks @killer30001000
  • fix(cli): omniroute serve no longer reports "Server did not respond within 60s" for a server that is actually up: the readiness probe's per-attempt timeout now escalates (2s, 4s, 8s, 15s, clamped to the time left in the budget) instead of aborting every attempt at a fixed 2s, so a health route that needs more than 2s for its first response is observed rather than repeatedly torn down. The timeout diagnostic now also states whether the port was accepting connections. (#12484) — thanks @dmlanday
  • fix(cli): omniroute serve now checks whether the port is already owned before spawning anything, and reports the conflict with the owning PID plus the two ways out (omniroute stop, or --port). Previously it handed the conflict to the child process, which died with EADDRINUSE and was retried twice on the supervisor's restart budget, printing three identical raw Node stack traces without ever saying that another instance held the port. Because that happened after the pid files were written, the doomed second instance also de-registered the healthy running one, leaving supervisor/.pid pointing at the dead starter and server/.pid deleted. (#12485) — thanks @dmlanday
  • fix(sse): Codex WebSocket transport (including the app-server) no longer fails to load in the Next.js standalone Docker runtime — the wreq-js loader now resolves its module name dynamically instead of a literal Turbopack could rewrite to an unreachable build-time symlink (#12491) (#13756) — thanks @marshalfevzi
  • fix(devin): treat Devin CLI model ids as literal — never strip or synthesize effort suffixes (#12492 — thanks @Neuron-Mr-White)
  • fix(command-code): floor a tiny caller-set max_tokens (e.g. 64) to MUSE_SPARK_MIN_OUTPUT_TOKENS = 512 for muse-spark ids, detected through the prefix-aware MUSE_SPARK_PATTERN so provider-prefixed forms (meta/muse-spark-1.2-contributor, cmd/meta/muse-…) are covered in both buildOpenAiBody (the /provider/v1/chat/completions path from #12130) and buildCommandCodeCliBody (the /alpha/generate fallback) — the hidden server-side reasoning phase can no longer consume the whole output budget and answer HTTP 200 with null content (out=64, reasoning=61), mirroring the #11214 mitigation already shipped for opencode-go; a caller that sent no budget is left without one and budgets at or above the floor pass through untouched (#12497) — thanks @Stazyu
  • Fix keys regenerate/keys reveal in the CLI to fall back to the dashboard /api/keys route when an ID from keys list does not exist in the registered-keys store, closing an ID-namespace drift between the two API key families. (#12520) — thanks @Gaulnews
  • fix(cli): Windows dashboard no longer reports Claude Code as settings_found_binary_unresolved when npm-global detection fails inside Electron. A failed npm config get prefix is no longer cached as permanent "" (which deleted every npm-derived candidate for the process lifetime), Windows lookup PATH is enriched with npm-prefix / %APPDATA%\npm / nvm / %ProgramFiles%\nodejs, and stock Node MSI .cmd shims under Program Files remain an explicit safety net. Separate from the #7831 .ps1 / known-path fix for #7774. (#12563) (#12565) — thanks @drmikecrypto
  • fix(sse): strip temperature/top_p on native Codex Responses passthrough so combo codex-review traffic no longer 400s with Unsupported parameter: temperature (#12585) — thanks @fouadSalkini
  • fix(pricing): saving model pricing from the dashboard no longer fails with a 400 / [object Object] — sync-written pricing fields round-trip through PATCH and validation errors surface actionable details (#12629) — thanks @wofiporia
  • fix(chat): requests with null/non-object entries in messages[] are now rejected with a clear 400 instead of crashing translators with an HTTP 500 (#12643) (#13755) — thanks @soroush5
  • fix(sse): Claude-native context handoffs now land in Anthropic's top-level system parameter instead of a leading role: "system" message, and the final Claude executor dispatch hoists any remaining leading prompt system/developer messages and relocates directive-only output_config envelopes away from messages[0], preventing the messages.0: use the top-level 'system' parameter HTTP 400 on model switches (#12668). — thanks @insoln
  • Honor a model's declared reasoning_efforts vocabulary in the reasoning-routing rule gate: a model-scoped or connection-scoped rule forcing max/ultra is now treated as supported when the model's resolved capabilities list that tier (operator overrides apply to models without a static registry declaration), instead of being rejected by the hardcoded gpt-5.6-* regex. Custom OpenAI-compatible providers whose models accept max natively (for example Merge Gateway zai/glm-5.3-flash, which accepts low|high|max) can now use forced-max rules without the request failing with Reasoning effort 'max' is not supported by the configured target. (#12686) — thanks @woodsonl
  • fix(open-sse): reasoning_details[].text is now promoted to reasoning_content even when reasoning is also present, so OpenRouter thinking models (GLM-5.3-Flash, DeepSeek-V4-Flash, Kimi K3) no longer lose their thinking traces in clients that only read reasoning_content (#12688 — thanks @thomasmaerz)
  • fix(providers): xAI requests no longer silently drop an assistant tool call sent in the legacy OpenAI function_call shape (instead of tool_calls[]) — the call is now translated into the xAI request the same way modern tool calls are (#12692) (#13753) — thanks @soroush5
  • fix(providers): xAI responses no longer report total_tokens/totalTokenCount as 0 when upstream usage uses the legacy prompt_tokens/completion_tokens names instead of input_tokens/output_tokens (#12700) (#13753) — thanks @soroush5
  • fix(dashboard): the Modal provider connection form now shows a Base URL field (placeholder https://<workspace>--<app>.modal.run/v1), so bring-your-own-deploy Modal connections can be validated and saved instead of failing outright — the server-side validator already required providerSpecificData.baseUrl (#12704) (#12736) — thanks @gonisulaimann
  • fix(cursor): Kimi-k3 / kimi-k3-high on the Cursor provider sometimes emit tool calls by imitating the executor's own history narration ("Assistant called tool … with arguments: …") instead of using structured tool calls, so clients received raw narration text plus native generation delimiters with finish_reason: "stop" — and the leaked turn compounded on every subsequent request via history re-send; the cursor executor now detects this shape and reassembles it into a structured tool_calls entry in both streaming and non-streaming finalization paths, gated on "no structured tool calls yet" so healthy turns are untouched (#12723) — thanks @patrykkopycinski
  • fix(sse): route the dario and 9router request bodies through the internal-marker strip before they are serialized upstream — both executors override transformRequest() without calling the base implementation, so the internal context-relay / universal-handoff markers (_omnirouteSkipContextRelay, _omnirouteInternalRequest, _omnirouteSkipUniversalHandoff) reached strict OpenAI-compatible gateways and got the call rejected with HTTP 400 "Unsupported parameter(s)" (#12729) (#12735) — thanks @gonisulaimann
  • fix(providers): OpenAI-compatible model discovery now parses per-vendor-route effort_values (nested under vendors.<vendor>.capabilities.reasoning in /v1/models), intersected across vendor routes so a synced level is always honored on every route the model can land on; re-syncing a connection whose catalog declares this shape no longer silently resets the synced supportedThinkingEfforts/defaultThinkingEffort data (#12730) — thanks @woodsonl
  • fix(models): a cold GET /v1/models on a large deployment no longer blocks the event loop for about a second at a time or overruns the 8s cold-build bound: since #12046 the built-in auto/* combos resolved catalog metadata for every target of every combo without memoizing or yielding, and they all draw on the same candidate pool, so 720 synced models took the build from ~4s to ~18s. Each distinct target is now resolved once per build, with a yield between misses (#12732) (#13678)
  • fix(combo): a quota-share combo whose steps carry no weight now rotates across its targets again instead of sending every request to the first one — the resolver turns an unset weight into 0 and #10881 made 0 mean "disabled", so an all-unweighted combo had no quanta and fell back to definition order; an explicit 0 still disables a target next to weighted siblings (#12732) (#13678)
  • fix(codex): the Codex WebSocket transport now emits a terminal response.failed (code upstream_websocket_closed) when the upstream socket closes before a terminal response event, instead of ending the client stream as if it had completed normally — preventing silent output truncation and allowing fallback/retry to trigger (#12737). — thanks @insoln
  • fix(models): preserve free-model metadata (isFree) discovered live from a provider through synced-model normalization, so free models no longer lose that flag before reaching the UI/consumers (#12763) — thanks @keeltrace
  • /v1/models combos whose merged capabilities.vision is true now also advertise input_modalities: ["text","image"] / output_modalities: ["text"] (synced modality intersections keep precedence), so models.dev-shaped clients no longer see a vision combo as text-only. (#12799 — thanks @aref-alapour)
  • fix(db): the background cleanup scheduler no longer runs a blocking full VACUUM after pruning rows (it froze every route, /healthz included, for minutes on large databases — 30 s after every start and every 6 h); freed pages are now reclaimed with paced PRAGMA incremental_vacuum batches plus a WAL checkpoint, and on auto_vacuum=NONE a full VACUUM is deferred to the Storage page's scheduled window via vacuumScheduler.requestFullVacuum() (#12821) (#12830) — thanks @insoln
  • fix(sse): stop over-escaped tabs from gpt-5.6-luna-xhigh corrupting Codex tool-call arguments — \\t is now collapsed back to a real tab instead of a literal \t text (#12841) — thanks @rafacpti23
  • fix(resilience): a recoverable direct-fetch response-start timeout (DIRECT_RESPONSE_START_TIMEOUT) could, in a narrow timer/promise-settlement race, escape as an unhandledRejection → uncaughtException and kill the server process — even though proxyFetch already retries this exact condition on a fresh socket. Guarded the timer callback so it can no longer fire against an already-settled attempt, and extended the process-level crash guard (already used by the WS/API-bridge servers) to recognize and swallow this code if it ever escapes anyway. Also installs that same guard in the production server entrypoint (dist/server-ws.mjs), which never had it even though the dev server already did (#12861) (#13636) — thanks @insoln / @HouMinXi
  • fix(translator): Gemini to Claude usage no longer double-counts the cached prompt prefix — input_tokens now excludes cache_read_input_tokens, matching the Anthropic Messages semantics (#12863) — thanks @ThiagoMafra-Integrare
  • fix(sse): An Anthropic OAuth 403 "Request not allowed" no longer bans the Claude connection on the first response — it is a per-request refusal on an otherwise healthy token, so it is now classified as the non-terminal request_rejected type, the connection is excluded for a growing cooldown (5 min, then 15 min) and only three consecutive refusals with no success in between escalate to banned; previously a single such response flipped the only Claude connection to banned and every later request was short-circuited with "All 1 connection(s) banned by upstream" until an operator reconnected (#12859, #12864 — thanks @insoln)
  • fix(cache): never write a truncated completion (finish_reason: "length"/max_tokens) into the semantic cache — a partial answer cached under a temperature:0 signature was served to every later identical request, permanently returning a mid-sentence reply that no retry cleared (#12885) (#12885) — thanks @patrykkopycinski
  • fix(providers): vLLM connections now advertise the real context window: model discovery reads max_model_len instead of falling back to the 128K default (#12897, closes #12858) — thanks @ntdat812
  • fix(combos): A combo's visibility can be changed through the API again: updateComboSchema accepts isHidden, so a visibility-only update is no longer rejected as empty and a mixed update no longer drops it (#12898, closes #12836) — thanks @ntdat812
  • fix(guardrails): Vision Bridge now extracts and replaces base64 images nested inside a tool_result.content array (the shape Claude Code uses), not just top-level content parts — previously these requests silently reached a vision-incapable provider and returned a 400. Also resolves a provider prefix that has no known alias (e.g. a custom OpenAI-compatible connection's model prefix) to the node id its credentials are actually stored under, instead of discarding the operator's fixed model and falling back to a no-auth candidate that fails (#12903) — thanks @initguru
  • fix(sse): inject the operator's global system prompt once, after request translation, for every target shape (Claude, Gemini, OpenAI Responses, OpenAI/Codex messages) instead of before translation — the pre-translation injection could be lost, repositioned, or duplicated 2-3× depending on the target format, and never reached the Responses API path at all. The new injectSystemPromptPostTranslation() is idempotent per request via a non-enumerable marker (#12904) — thanks @initguru
  • fix(sse): strip echoed system/directive preamble on /v1/messages responses and preserve large analysis/summary blocks in systemPreambleStripper to stop autocompact empty-response (#12905) — thanks @initguru
  • fix(sse): thread the client's thinking intent into the non-streaming translation path so the same request answered with stream:false no longer leaks a thinking block that stream:true withholds (#12905) — thanks @initguru
  • fix(sse): gate thinking block emission on requestedThinking (streaming + non-stream) and flush reasoning-only responses as text to stop reasoning leak, autocompact loops, and 502 (#12905) — thanks @initguru
  • fix(sse): make the system-preamble stripper opt-in (OMNIROUTE_STRIP_SYSTEM_PREAMBLE=1) and flush both preamble strippers at stream end, so English-prose heuristics no longer delete a legitimate section of every openai→claude reply and an unterminated echo block no longer reaches the client as an empty message (#12905) — thanks @initguru
  • fix(sse): make the direct response-start timeout reasoning-aware — detect reasoning_effort high/max in the body and raise the ceiling to 180s to stop 504 on high-effort TTFB (#12906) — thanks @initguru
  • fix(sse): retry 0-byte empty_response 502 like STREAM_EARLY_EOF to stop autocompact 502 — RETRYABLE_STREAM_EMPTY_CODES + shouldRetryStreamEarlyEof wiring (#12906) — thanks @initguru
  • fix(compat): preserve GPT and Claude Code tool-call history when translating OpenAI Responses API requests to Chat Completions — role:"tool" items and role-based assistant tool_calls are no longer dropped, fixing incomplete multi-turn history (#12909) — thanks @initguru
  • fix(usage): finalize semantic cache hits by exact request id — checkSemanticCache now uses finalizePendingScope(pendingScope, ...) instead of an ambiguous (model, provider, connectionId) tuple, fixing wrong-request finalization when connectionId is null or multiple requests are in flight on the same connection (#12910) — thanks @initguru
  • fix(sse): the Codex Responses WebSocket bridge now fails fast to another eligible account instead of queueing behind a saturated one, releasing its per-account lease exactly once when the session ends (#12911 — thanks @initguru). This also fixes accountSemaphore's maxQueueSize: 0 handling, which previously behaved as an unbounded queue instead of failing over immediately — benefiting every caller that configures queueDepth: 0 (for example combo routing), not just the Codex WS bridge.
  • fix(chatcore): block a client's own duplicate retry (same idempotency key) from opening a second upstream turn while the first is still in flight, returning 409 turn_in_progress instead of wasting quota on a redundant execution (#12912) — thanks @initguru
  • fix(sse): bound streams that keep sending raw upstream bytes forever without ever emitting a terminal event — a new STREAM_ACTIVE_TIMEOUT_MS watchdog (default 1260000ms/21min — the largest per-model timeoutMs in the registry plus a one-minute margin, 0 disables) tracks the stream's total lifetime independently of the existing byte-stall watchdog, so a continuously-active non-terminal stream can no longer occupy a connection indefinitely (#12913) — thanks @initguru
  • fix(providers): correct Magnific API key validation, which reported every valid key as invalid due to a GET probe against a POST-only endpoint (#12927) (#13754) — thanks @hubo1989
  • fix(routing): Auggie now fails over to the next combo model instead of returning the quota-exhausted CLI warning as a successful reply (#12949) (#13751) — thanks @honeypot55
  • fix(combo): A weighted combo whose every target was excluded before dispatch by a resilience timer (model lockout, open circuit breaker, provider cooldown) now answers 503 all_targets_cooling_down with Retry-After set to the earliest exclusion to lapse, the excluded targets and reasons in diagnostics.excluded, a wait recovery hint, and a [COMBO] warning naming the reasons; previously the pool was dropped silently and the host answered 404 "Combo has no executable targets" (recovery hint "switch combo / reconnect the missing providers") for a pool that was configured, connected and merely cooling down — which clients such as Claude Code render as "this model may not exist". A pool with nothing to run keeps its 404 (#12954, #12956 — thanks @insoln)
  • fix(resilience): A 5xx model-lockout failure — a transport error (terminated, EHOSTUNREACH, connect timeout), an upstream server error, or OmniRoute's own synthesized 502 from quality validation — now locks only the exact provider/connection/model tuple instead of the quota family; previously one empty response on a single gpt-5.6-* model removed every gpt-5* model of the codex connection from routing for 2–30 min (escalating) while its quota was untouched. Quota statuses (429/403/402) keep the family scope; success-decay and the Model Cooldowns card now handle exact-scope locks too (#12955, #12957 — thanks @insoln)
  • fix(providers): GitLab Duo Retest and chat requests now fall back to the public Code Suggestions endpoint for ANY direct_access 403 (not only the "direct connections are disabled" tenant-config message), and surface the real upstream error body instead of a generic "Access denied" when both endpoints reject the token (#12958) (#13758) — thanks @Rahulsharma0810
  • fix(sse): stop misclassifying a truncated Anthropic-compatible max_tokens probe response (content:[{type:"text",text:""}]) as an empty upstream response (#12968) (#13771) — thanks @pranay-gpt
  • fix(images): image-combo legs now fall back when an upstream provider returns HTTP 2xx with an empty or malformed image payload. fetchImageEndpoint previously normalized any successful HTTP response to success: true (data.data || []), so executeImageCombo stopped on the first leg and handed the client an image-less 200. The OpenAI-compatible normalization now requires at least one usable item (non-empty b64_json or url) in data[]; an empty/malformed 2xx becomes a retryable 502 with a sanitized error, so priority image combos advance to the next leg. Valid payloads and direct image-model requests are unchanged. #12982 — thanks @tiangao88
  • fix(claude): forward client-negotiated thinking-binding-controls-2026-08-01 and thinking-display-updates-2026-08-18 betas so Fable 5.1 thinking.block_binding / thinking.display requests are no longer rejected upstream with Extra inputs are not permitted (#12989) — thanks @fidelix
  • fix(cli): skip the POSIX CLI path lookup during Windows autostart setup, preventing a bogus path error before successful enablement (#12993) — thanks @zachary-frederich
  • fix(i18n): stop the "Saving..." hang on /dashboard/combos — BuilderIntelligentStep.tsx's exploration-rate hint called t("explorationRateHint") with no ICU values even though the key requires {percent}, and next-intl's default error handling throws FORMATTING_ERROR for that call, unmounting the whole builder step and looking like a silent save hang. Fixed across all three affected call sites (BuilderIntelligentStep.tsx, AgentBridgeMaintenanceCard.tsx, RawJsonPanel.tsx), not just the one that was reported (#12995). — thanks @hartmark
  • fix(api): restore the MCP namespace field on streamed and non-streamed Responses tool calls in follow-up turns of a session that don't re-declare their type:"namespace" tools (#12996) (#13769) — thanks @rolemiaster
  • fix(db): add an opt-in automatic sweep for terminal batch checkpoints and expired file content, behind the new BATCH_AND_FILE_AUTO_CLEANUP_ENABLED feature flag (default off) — batch_item_checkpoints had grown to 182K rows / 5.25 GB with no batch ever explicitly deleted by an operator: the manual delete-completed route existed but nothing called it automatically, and it never covered failed/cancelled/expired batches either. Extracts a shared deleteBatchesMatching() (age-gated, every terminal status, keeping deleteCompletedBatches()'s exact existing contract) and adds pruneExpiredFiles() for uploaded file content past its own expires_at (1,874 rows / 5.19 GB observed live, most long past expiry). With the flag off (the default), every existing install keeps this data exactly as before; an operator must opt in via the dashboard or BATCH_AND_FILE_AUTO_CLEANUP_ENABLED=true before the sweep deletes anything (#12999). — thanks @hartmark
  • fix(providers): stop @omniroute/opencode-plugin combo context limits from downgrading to the raw Math.min(member) lower bound after a restart — the static catalog now honors the server-computed computed_context_length (mirroring the dynamic hook), and a background refresh with a degraded /api/combos response backfills the field from the last-known-good disk snapshot instead of overwriting it (#13000) (#13759) — thanks @morpheus9393
  • fix(streaming): allow a per-provider override of the fetch-start (headers-wait) timeout cap so providers that buffer the full generation before the first byte (e.g. command-code, opencode-go) are not cut off at the global 110s cap; the same two entries also gain a reasoning-safe requestDefaults.maxTokens of 16384 so thinking models such as z-ai/glm-5.3-flash are not cut off mid-reasoning (#13002) — thanks @alvinveroy
  • fix(resilience): An apikey-category 429 whose body explicitly says a long-window quota was exhausted no longer skips the quota cache — shouldPreserveQuotaSignals() (open-sse/services/quotaResetParsing.ts) gained an errorText parameter in the #6638 fix, but only one of its two call sites was updated: checkFallbackError() passes the upstream body while shouldMarkAccountExhaustedFrom429() (open-sse/services/accountFallback.ts) still called it with the provider alone. With errorText undefined the helper's Boolean(errorText) && looksLikeQuotaExhausted(errorText) branch can never be true, so for every apikey-category provider without per-model quotas the connection was never marked quota-exhausted. errorText is now threaded through the helper and passed at the src/sse/handlers/chat.ts call site. Plain rate limits (Rate limit exceeded, retry in 20s, Too Many Requests) still fall through to the short generic cooldown. Regression guard: tests/unit/quota-signal-errortext-threading.test.ts. (#13008) — thanks @Rick7C2
  • fix(cli): omniroute mcp restart no longer 404s — the missing POST /api/mcp/restart route now exists — and new omniroute mcp enable/mcp disable [--transport] subcommands give the CLI a way to turn the MCP server on without the dashboard (#13012) (#13770) — thanks @ricardusx
  • fix(mitm): add catch-all (*) model mapping fallback for Agent Bridge (#13013) — thanks @tuandinh0801
  • fix(providers): Antigravity connection Retest probes Cloud Code envelope (#13015) — thanks @tuandinh0801
  • fix(dev): allow Ctrl+C to promptly kill dev server by closing active connections (#13020) — thanks @tuandinh0801
  • fix(skills): repair nested malformed skill-tool schemas (bare property maps, boolean required: true) for OpenAI-compatible providers, not just the schema root (#13022) (#13772) — thanks @ftevxk
  • fix(sse): reasoning replay now works for Chat Completions and Anthropic Messages clients on Responses-API reasoning targets such as opencode-go/deepseek-v4-flash: plain (non-tool-call) assistant turns are captured against the same normalized transcript the read side digests (the Responses body carries input, not messages, so the write side digested only the assistant message instead of the full transcript and every replay missed), and the replay pass runs on the OpenAI pivot for every source format, so Anthropic Messages clients are replayed too. Fixes the intermittent 400 The reasoning_text in the thinking mode must be passed back to the API from Console Go for clients that drop reasoning_content (#13031) — thanks @jmche
  • fix(friendliai): FriendliAI's free-tier credit-exhaustion 403 ({"detail":"You've exhausted all your credits..."}) is now classified as QUOTA_EXHAUSTED instead of AUTH_ERROR, so omniroute treats it as depleted credits rather than a credential problem (#13040) — thanks @turbolego
  • fix(oauth): kimi-coding/github device-flow pollToken no longer rejects with TypeError: Body is unusable when the token endpoint returns a non-JSON error page (CDN/anti-bot/proxy interstitial) — the body is now read once and parsed, preserving the graceful invalid_response fallback instead of a generic 500 (#13046 — thanks @ysntony)
  • fix(providers): lazily load chatgpt-web-codex admin helpers in PUT /api/providers/[id], mirroring #12355's exact pattern for the sibling POST route — a source-inspection test pins the import-graph invariant, since Turbopack's client-bundle boundary can't be exercised from the test harness directly (#13071). — thanks @hartmark
  • fix(providers): honor an operator-set endpoint override (PUT /api/provider-models) for a local model whose own /v1/models response carries no capability data of its own (llama.cpp included) — the override previously only took effect when a matching customModels entry already existed, so declaring a brand-new local model as embeddings-capable silently did nothing. updateCustomModel() gains an opt-in createIfMissing mode; every other caller's existing contract is unchanged. Also accepts a collapsed single-slash id for path-based local models (#13078). — thanks @hartmark
  • fix(quota): keep Kiro active while any _freetrial pool has quota (#13088) (#13324) — thanks @giauphan
  • fix(routing): round-robin combos now show up in Combo Studio's Live dashboard — they were completing successfully but never publishing the attempt/success/failure events the dashboard listens for (#13089) (#13776) — thanks @adityadwi21
  • fix(sse): stop rejecting a Responses API tool_choice.type: "custom" (e.g. Codex CLI forcing functions__exec) with a 400 unsupported_feature error (#13122) (#13775) — thanks @phamtienduceng-eng
  • fix(translator): recognize tool_choice.type: "custom" in Responses→Chat translation and propagate custom tool names (including namespace-flattened ones) across both the streaming and non-streaming provider legs, so non-streaming Responses clients get custom_tool_call/raw input instead of function_call/JSON arguments (#13128) — thanks @ducphamtien-fonos
  • fix(providers): remove the chipotle/pepper provider — its upstream (amelia.chipotle.com) now 404s on every route and is fully decommissioned (#13131, #4037) (#13913) — thanks @Falco20100
  • fix(db): authenticated GET /api/db/health polls no longer run a SQLite quick_check. The health dashboard polls every 15 seconds, and that scan ran synchronously on the request-serving event loop, blocking it for the length of the scan. Reference and state checks still run, and explicit repair requests keep integrity checks unless OMNIROUTE_SKIP_DB_HEALTHCHECK=1 is set (#13149) — thanks @cryptiklemur
  • fix(sse): fail over once to a sibling connection on stream early EOF (the original STREAM_EARLY_EOF 502 is kept when no sibling can serve the request), gated behind STREAM_EARLY_EOF_SIBLING_FAILOVER_ENABLED (default off) (#13153) — thanks @maxmad64bis
  • resilience: a Cloudflare managed challenge (cf-mitigated: challenge / challenge HTML on 403) is classified as a fingerprint rejection and retried on another account/transport instead of banning the connection (#13161 (#14090) — thanks @anhtran-ai)
  • fix(cli): OpenCode config generator preserves catalog display names — custom names win, then display_name/native name (with the owned_by/ prefix stripped once), then a readable label for auto/* ids, instead of always showing the raw model id (#13168) — thanks @domenicomassafra
  • fix(sse): Fable 5 and 5.1 keep their prompt-cache prefix across mid-conversation system messages. OmniRoute treated only Opus as capable, so every Fable system turn was hoisted into the top-level system prompt and moved the cached prefix, which reported system_changed from turn 2 on. Fable now has its own mid-conversation-system capability path, and context-1m stays limited to Opus so Fable is never sent an unrelated beta header (#13173) — thanks @cryptiklemur
  • fix(combo): auto-resume a pinned native Codex turn on a healthy sibling connection or model when the pinned provider becomes unavailable for a model-scoped reason (quota, model lockout) instead of failing the turn outright — provider-wide circuit-breaker/cooldown state, pending tool calls, opaque continuation state, and partial streams still block resume, and at most one auto-resume happens per logical turn (#13180) (#14162) — thanks @mdigitalbh81
  • fix(evals): an eval case whose model call errored is no longer scored as passed — a case that never reached a model has no measured behaviour to grade (#13201) — thanks @aaustinhuang / @dajiaohuang
  • fix(evals): the eval runner now sends x-omniroute-compression: off and x-omniroute-no-memory: true on every case, so a graded case measures the model instead of the operator's injected output style, retrieved memory and memory_* tools (#13139, #13206) — thanks @aaustinhuang / @dajiaohuang
  • fix(combos): stop dropping live keys and persisting dead ones (#13217) — thanks @maxmad64bis
  • fix(db): the WAL checkpoint busy counter reported by /api/monitoring/health now survives restarts — busy checkpoints are counted in memory and persisted from the next clean maintenance tick or at shutdown, never with a write while the database is contended (#13218) — thanks @maxmad64bis
  • fix(vertex): preserve Claude prompt-cache breakpoints for Vertex and Vertex Partner, use the documented five-minute ephemeral TTL by default, and forward cache usage metadata through streaming responses (#13220) — fixes #13219 — thanks @SIGTERM-015
  • fix(mcp): load the audit better-sqlite3 driver via the shared runtimeRequire() helper instead of createRequire(import.meta.url), which broke when the Next.js standalone build emits the module as a CommonJS chunk (#13223) — thanks @chatchawan-simplewish
  • fix(sse): classify a missing Playwright Chromium install on the Z.ai web transport as an actionable 503 host/config cooldown instead of a generic 502 that trips the provider circuit breaker (#13232) (#13777) — thanks @oleksandr1811
  • fix(combos): testing a combo aborts in-flight probes when the client disconnects instead of probing on after the dashboard navigates away (#13279) — thanks @maxmad64bis
  • fix(quota): in-process routing and quota caches (quality tracker, saturation and rate-limit header caches, quota-fetcher cache, learned rate limits, account buckets) are now size-bounded through one shared boundedMap — caps sit far above normal deployments, evictions are logged once per minute per cache instead of per entry, and state whose loss would change routing (live saturated quota buckets, evaluator quality scores) is never evicted (#13280) — thanks @maxmad64bis
  • fix(call-logs): call-log error types are now a versioned vocabulary (ERROR_TYPE_CONTRACT v1) with explicit unknown instead of ambiguous null, and free-text history reads back as unclassified (#13281) — thanks @maxmad64bis
  • Fixed tests leaving temp DATA_DIR folders behind on Windows by closing the SQLite handle before removing the directory (#13290). (#13292) — thanks @anhtahaylove
  • fix(providers): honor the selected Alibaba workspace and region endpoints for custom embedding and qwen3-rerank requests (#13293) — thanks @xiaoyaner0201
  • fix(backend): error messages are no longer truncated after a path — redactErrorPaths treated any slash-bearing span as an unequivocal filesystem path and swallowed the rest of the line, so the image-model 400 lost the Use POST /v1/images/generations instead. hint it exists to give, and a redacted diagnostic lost its with api_key='[REDACTED]' tail. Only a Windows path, file URI or known POSIX root with no determinable end swallows the line now (#13144) (#13295) — thanks @abhisheksharma2411 / @ggiak
  • fix(providers): Fetch Qwen and Alibaba Token Plan model catalogs through their authenticated console gateways, with public-only URL validation and local-catalog fallback when discovery is unavailable. (#13299) — thanks @JxnLexn
  • fix(db): defer process.exit(0) on graceful shutdown by one macrotask, avoiding a Windows-only libuv abort when the sql.js fallback driver has a statement in flight (#13306) (#13778) — thanks @anhtahaylove
  • fix(cli): omniroute serve now surfaces a fatal [STARTUP] Fatal: ... boot diagnostic (e.g. a DB driver init failure) to the console immediately, even without --log, instead of only when the process later crashes or restarts (#13314) (#13779) — thanks @Orion1943
  • fix(sse): A transient error on a round-robin combo target no longer resets its concurrency limit to 3, so a target capped at 1 is not sent three queued requests at once when its cooldown ends (#13320) — thanks @datrixlab
  • fix(resilience): Rate-limit reset headers in RFC 3339 form (Anthropic) and with fractional seconds (2m59.56s) are parsed correctly, instead of an Anthropic reset seconds away throttling the connection for about 34 minutes (#13321) — thanks @datrixlab
  • fix(cli): contexts export --no-secrets now leaves the access tokens and API keys out, and chat --no-history and serve --no-recovery take effect; all three flags were accepted and ignored (#13322) — thanks @datrixlab
  • fix(providers): Switching "Allow Private Provider URLs" off in the dashboard now takes effect when OMNIROUTE_ALLOW_PRIVATE_PROVIDER_URLS=true is set in the environment (#13323) — thanks @datrixlab
  • fix(cli): omniroute restart now comes back on the port set by PORT (shell or <DATA_DIR>/.env) instead of always 20128, like serve and dashboard (#13327) — thanks @datrixlab
  • fix(redis): Warmup circuit-breaker keys now honor REDIS_KEY_PREFIX like every other OmniRoute Redis key, instead of always using omniroute:warmup:cb: (#13328) — thanks @datrixlab
  • fix(api): Setting DISABLE_SQLITE_AUTO_BACKUP=true no longer makes the API-key rate limiter and auth cache skip Redis, which let every replica enforce the full per-key limit on its own (#13329) — thanks @datrixlab
  • fix(translator): tool_choice: "none" is now sent to Claude-format providers as { type: "none" } (and back to OpenAI as "none") instead of auto, so the model can no longer call tools the client switched off (#13333) — thanks @datrixlab
  • fix(translator): Gemini tool results sent without an id (the usual case, since OmniRoute's own Gemini responses never emit one) now reach the model instead of being replaced by an empty result, on the Gemini, Antigravity and /v1beta request paths (#13334) — thanks @datrixlab
  • fix(translator): An image returned inside a Claude tool_result (Read on a PNG, an MCP screenshot) is now sent to Gemini as an image part instead of base64 text (#13335) — thanks @datrixlab
  • fix(combo): stop retrying a malformed-request-shape error across the entire fallback chain — a 400/422 that fails one target for a kind: "model" reason (exact status + error message match) will fail identically on every other target too, so a bad payload previously burned the full MAX_GLOBAL_ATTEMPTS budget instead of failing fast. Observed live: 41-44 identical combo decisions over 13+ minutes for a single request. Trips only on 3 consecutive identical kind: "model" failures, leaving transient and provider-side errors untouched (#13338). — thanks @hartmark
  • fix(db): getDbInstance() now closes the probe and primary SQLite connections on every failed initialization path, not just the happy path, fixing a handle leak that caused EPERM on Windows teardown. (#13303) (#13342) — thanks @voidstackloop
  • fix(dispatch): strip every _omniroute* internal marker at the shared pre-serialization chokepoint (cliFingerprints.ts) instead of a hand-maintained per-key allowlist, so internal routing/handoff markers (e.g. _omnirouteSkipContextRelay, _omnirouteResponsesStore) can no longer leak into serialized upstream request bodies and draw 400 Extra inputs are not permitted from strict Anthropic-compatible gateways (#12729, fixed in #13355) — thanks @patrykkopycinski
  • fix(providers): stop zed-hosted claude-haiku-4-5 extended-thinking requests from inflating max_tokens past the model's real 64000 output cap (#13364) (#13780) — thanks @ThiagoMafra-Integrare
  • fix(providers): gemini-web no longer drops the system instruction on single-turn requests or the tool contract when a client system message is present, and switches to an atomic composer insert so embedded newlines can't submit the message early (#13380) (#13784) — thanks @formilw
  • fix(providers): gemini-web now attempts to select and verify the requested Gemini UI mode (and Extended Thinking) before answering, and fails closed with a clear 400 instead of silently running the account default under a mismatched model label (#13381) (#13919) — thanks @formilw
  • fix(db): stop routine connection-backoff auto-recovery from busting the entire /v1/models response cache, which was causing intermittent 75-120s/502 responses on deployments routing many providers (#13389) (#13783) — thanks @RaviTharuma
  • fix(mitm): bound per-request SSE transcript retention to 1 MiB and stop the upstream read when the downstream disconnects — handler-side collected strings grew without bound before the inspector clamp, and abandoned streams kept the reader alive for the full upstream lifetime (#13395) (#13702) — thanks @oyi77
  • fix(opencode-plugin): The stale disk-cache fallback warning now reports the snapshot's age (using stale disk cache (N models, age 168h)), matching the existing warm-startup log. Previously a week-old catalog was indistinguishable from a five-minute-old one, so silent model drift went unnoticed. (#13426) — thanks @RaviTharuma
  • fix(api): GET /api/logs/export now streams rows from the database itself (a cursor for proxy-logs, a LIMIT-bounded generator for call-logs/request-logs) instead of buffering every matching row in memory before serializing, fixing a V8 heap OOM on large tables (#13428, #13123) — thanks @KooshaPari. Breaking: the limit query param now defaults to 10,000 rows (max 50,000) — exports that previously returned every matching row are silently truncated (with "capped":true,"totalAvailable":<n> in the response) unless the caller passes a larger explicit limit.
  • fix(compression): stop lite compression from dropping a role:"tool" message when it is byte-identical to the previous message, which orphaned a tool_call_id and triggered upstream 400 errors on parallel tool calls (#13429) (#13787) — thanks @tolgaaksoy
  • fix(sse): frame post-keepalive /v1/responses stream errors with a top-level type field so Responses clients (Codex) surface the real upstream error instead of reporting "stream disconnected before completion" (#13431) (#13785) — thanks @andrea-kingautomation
  • fix(db): reconcile auto_vacuum drift between the configured INCREMENTAL mode and the live SQLite pragma — detected at startup and reconciled out-of-request by the vacuum scheduler, which now also runs a bounded PRAGMA incremental_vacuum reclaim instead of an unconditional full VACUUM once INCREMENTAL is actually in effect (#13432) (#13786) — thanks @tolgaaksoy
  • fix(models): Reconcile provider dashboards with confirmed authoritative live catalogs, excluding retired built-in/imported rows while preserving manual custom models and partial-catalog fallbacks. (#13434) — thanks @JxnLexn
  • fix(build): the production build no longer breaks when a client component reaches a server-only module, and the client-bundle guard now discovers server-only modules instead of matching a fixed list (#13436) — thanks @maxmad64bis
  • fix(models): return a retryable 503 with Retry-After instead of a 500 when the first catalog build outlasts its time bound (#13438) — thanks @maxmad64bis
  • fix(combo): new opt-in flag PROTECTED_PRIORITY_INFRA_502_ENABLED (default off): when a priority target marked fallback-only-on-quota-exhaustion stops the combo because its provider circuit breaker is open or a predictive latency check rejected it — causes that are provably not quota — the response is 502 instead of a quota-looking 503; lockout, cooldown, unavailable, exhaustion, credential-gate and concurrency-cap stops keep 503 (#13439) — thanks @maxmad64bis
  • fix(resilience): non-TPD daily-quota cooldowns honor the provider node's configured daily-reset clock (timezone + hour) instead of server midnight, on single-model and combo (priority and round-robin) paths; timezone edits apply without a restart (#13440) — thanks @maxmad64bis
  • fix(call-logs): the call-log write point validates error_type against the versioned vocabulary with a Zod schema and stores unknown for any value outside it, so a classifier family that drifts from ERROR_TYPE_CONTRACT can never persist free text (#13441) — thanks @maxmad64bis
  • fix(oauth): token health check now parses a numeric epoch expires_at (number or string, seconds or milliseconds), so connections synced by external tools keep their expiry-driven refresh instead of being skipped forever — or refreshed on every sweep (#13444) — thanks @elielsousa-pathbit
  • fix(db): Arena ELO sync now fetches and validates the leaderboards before touching model_intelligence, and applies the upsert + prune of expired rows inside a single atomic transaction. Previously, an unavailable/rate-limited Arena API left the table pruned with nothing written back, and since the sync runs on every boot, repeated restarts against a rate-limited upstream permanently drained the table to zero (#13446) — thanks @CrashCartCapital
  • fix(analytics): the compression analytics writer now passes flatRateAsZero: true to calculateCost, matching /api/usage/analytics. Flat-rate subscription lanes (minimax, glm, kimi, bailian, xiaomi, web-cookie) no longer report a dollar "savings" figure that was never actually payable (#13446) — thanks @CrashCartCapital
  • fix(sse): stop an unhydrated openai-compatible-*/anthropic-compatible-* connection from silently routing chat requests (and its stored credential) to the real OpenAI/Anthropic API instead of the operator's configured provider-node endpoint (#13452) (#13798) — thanks @DenXio101
  • fix(sse): Pollinations and Perplexity-web requests now fail over instead of returning the provider's own "out of credits"/"account suspended" text as if it were a real answer — an HTTP 200 body whose short assistant message is dominated by a known credits-exhausted or account-deactivated phrase is now classified as a malformed response and triggers the existing combo/auto-fallback path (#13461) (#13910) — thanks @arjav1181
  • fix(resilience): background OAuth token refresh (proactive health-check sweep and the shared refresh helper behind refreshAccessToken/refreshClaudeOAuthToken/etc.) now fails closed like the interactive chat path when a connection's assigned proxy pool is entirely dead, instead of silently sending the refresh-token exchange out direct or via a stray HTTPS_PROXY (#13470) (#13793) — thanks @elielsousa-pathbit
  • fix(providers): Muse Spark 1.3 works on OpenCode Zen, OpenCode and OpenCode Go instead of failing with a 500, and gets its real 1M context window (#13471) — thanks @maxmad64bis (with thanks to @bacnh85, @shermzy and @atakhadiviom for #12675, #12973 and #13111)
  • fix(sse): forward Anthropic prompt-cache-creation tokens through the /v1/responses usage hop so cache-write counts stop logging as zero (#13472) (#13790) — thanks @fidelix
  • fix(opencode): opt-in OPENCODE_RESPONSES_STALL_ROTATION flag (default off): a streamed Responses reply that sends headers and then nothing is cut after RESPONSES_FIRST_BYTE_TIMEOUT_MS (default 15 s) instead of waiting for the stream readiness timeout — the account is cooled down and the request rotates to the next account once (proxied or proxy-less), a second stall fails fast; with the flag off nothing changes (#13484) — thanks @maxmad64bis
  • fix(sse): stop the streaming PII sanitizer from splicing OpenRouter metadata (provider, native_finish_reason, reasoning_details[].format) into the answer text buffer (#13488) (#13792) — thanks @Xore
  • fix(sse): opt-in OPENCODE_USER_BLOCKED_ROTATION flag (default off): an opencode 403 or 451 carrying a user_blocked refusal cools the refused account down and fails over to the next account at most once per request, cancelling the abandoned response body; with the flag off the refusal is returned unchanged (#13498) — thanks @maxmad64bis
  • fix(runtime): eliminate hardcoded 20128 port remnants and make loopback URLs dynamic (#13533) — thanks @ggdayup
  • fix(cli): redraw the CLI/Electron system tray icon with a dark outline and ship a native multi-res icon.ico so it is no longer a pure-white, nearly invisible glyph on the Windows light-theme taskbar and hidden-icons flyout (#13535) (#13797) — thanks @ProphetOfDoom-PoD
  • fix(cli): persist the supervisor's give-up crash record to <DATA_DIR>/server/crash.log (surfaced by omniroute doctor) instead of only printing it — the console output was discarded when --tray mode's detached worker exited, leaving no trace of why the gateway/tray disappeared (#13538) (#13908) — thanks @ProphetOfDoom-PoD
  • fix(api): /v1/audio/transcriptions, /v1/audio/translations and /v1/audio/speech requests now show up in Dashboard → Request Logs — the three routes never called the shared call-log pipeline, so every successful (and failed) transcription/translation/speech request was silently dropped from call_logs (#13544) (#13803) — thanks @delafu
  • fix(translator): Claude tool input_schema with a root-level anyOf / oneOf / allOf is flattened into a plain object schema instead of being forwarded verbatim. Anthropic refuses such a tool before inference (tools.N.custom.input_schema: input_schema does not support oneOf, allOf, or anyOf at the top level), so a single MCP/agent tool carrying one made every request fail with no combo failover possible (#13552) (#13561) — thanks @sprintberlin
  • fix(routing): Preserve forced reasoning effort across native requests, account defaults and combo fallbacks while keeping internal routing directives out of upstream payloads. (#13556) — thanks @JxnLexn
  • fix(providers): MiniMax-M3's inline <think>...</think> reasoning no longer leaks into message.content/delta.content on the minimax/minimax-cn routes — it is now stripped and surfaced as reasoning_content, in both streaming and non-streaming responses (#13558) (#13799) — thanks @pan17
  • fix(api): PATCH /api/settings now persists hideAutoCombos and hideNoThinkVariants instead of silently dropping them (#13562) (#13800) — thanks @texastoland
  • fix(api): resolve the codex-settings apiKey through the canonical key resolver instead of an inline 400 guard, so the dashboard Apply flow no longer fails with baseUrl, apiKey and model are required in cloud mode when no management key is selected (#13563) (#13566) — thanks @opensource-elearning
  • fix(sse): release the native Codex turn pin when the pinned model becomes model-scoped unusable, so a long-running Codex session falls back to the next healthy combo model instead of dying to a terminal 400 NATIVE_CODEX_PINNED_MODEL_UNAVAILABLE (#13564) (#13566) — thanks @opensource-elearning
  • fix(proxy): ordinary SOCKS5 data-plane requests no longer trigger the T14 speculative bare-TCP reachability probe — that probe opened and immediately closed a raw TCP connection, which a SOCKS5 listener (e.g. GOST) sees as an incomplete handshake and logs as unexpected EOF; HTTP/HTTPS fast-fail and the explicit directFallbackOnUnreachable control-plane probe are unchanged (#13571) — thanks @mdigitalbh81
  • fix(gemini): stop sending Gemini a 400 on combos ("function call turn comes immediately after a user turn") when conversation history opens on a functionCall turn — after mergeConsecutiveSameRoleContents, contents[] alternation is guaranteed for every index ≥ 1, but the one remaining violation was a history whose very first turn is a functionCall. Prepends a synthetic leading user turn in that case (#13573). — thanks @hartmark
  • fix(db): make the chat-path proxy resolver (resolveProxyForConnection) rotate a multi-member pool the same way the registry resolver already does — its per-connection cache was freezing on the first pool member forever instead of re-running the scope's round-robin/sticky/random strategy on each request, unless the connection needs a stable egress (opencode's egress-bucketed quota, grok-web's IP-pinned cf_clearance) (#13575) (#14044)
  • fix(proxies): a subscription refresh, a bulk re-import or an API update that omits the status no longer turns a disabled proxy back on, and a refresh no longer rewrites a manual proxy that shares a subscription node's address (#13577) — thanks @maxmad64bis
  • fix(cli): omniroute update now passes --legacy-peer-deps to npm install -g, suppressing the ERESOLVE / peer-dependency wall seen on fresh global installs; dry-run output reflects the same flag; troubleshooting guide documents the supported install form (#13579 — thanks @prabhu-omkar)
  • fix(api): a partial update no longer resets the fields the client did not send: renaming a disabled reasoning routing rule keeps it disabled with its priority, description and tags, renaming a playground preset keeps its params, and renaming or re-importing a proxy keeps its address family (#13582) — thanks @maxmad64bis
  • fix(providers): Antigravity error responses and logs now surface the real upstream message (e.g. Gemini field-path rejections) instead of the generic "Antigravity upstream error (400)" placeholder (#13591) (#13801) — thanks @afonsoft
  • fix(usage): the call-logs artifact worker's failure warning now includes the underlying error's message/code instead of the generic "detail omitted" — a crashed or non-zero-exit worker was previously undiagnosable in the logs (#13597) (#13802) — thanks @afonsoft
  • fix(providers): echo back reasoning_content on bai DeepSeek thinking-mode follow-up turns, fixing the upstream 400 "reasoning_content must be passed back" (#13599) (#13807) — thanks @afonsoft
  • fix(i18n): drop the second copy of featureFlagProxySkipRecentlyFailedDescription that the 2026-09-15 batch merges left in 59 dashboard catalogs (a scripted keep-both conflict resolution concatenated the key both PRs carried; JSON.parse silently kept the last copy) and the duplicated ERROR_TYPE_CONTRACT import in src/lib/db/callLogStats.ts (TS2300); adds tests/unit/i18n-catalogs-no-duplicate-keys.test.ts, a raw-text guard that fails on any key declared twice in one object of src/i18n/messages/*.json or bin/cli/locales/*.json (#13602, #13641) (#13816)
  • fix(proxy): proxy credentials holding a literal % (e.g. pa%ss) no longer break the proxy — HTTP(S) proxies now receive a correctly built Proxy-Authorization header instead of undici throwing URIError, SOCKS5 proxies get the raw credential, and the proxy registry, subscription import and legacy settings parsers keep the value instead of dropping the entry; correctly percent-encoded credentials decode exactly as before (#13605) — thanks @maxmad64bis
  • fix(connection-cooldown): skip connection cooldown for locally rejected token-budget 429s so a per-key limit never cools a healthy connection (#13606) — thanks @maxmad64bis
  • fix(opencode-plugin-v2): write the catalog snapshot to a temp file and rename it into place, ignore newer snapshot versions, and warn when a write is skipped or fails (#13607) — thanks @maxmad64bis
  • fix(proxy-health): a target-refused probe (401/403/429) can reset the consecutive-failure streak instead of staying neutral, behind the opt-in PROXY_HEALTH_BLOCKED_RESETS_STREAK feature flag (default off: refusals keep the #10654 neutral policy); a relayed 5xx stays inconclusive and a refusal never removes or disables a proxy (#13608) — thanks @maxmad64bis
  • fix(providers): opt-in MISTRAL_AMBIGUOUS_401_SOFT_LOCKOUT flag (default off): a bare Mistral 401 with no explicit auth signal (identical for a revoked key and an exhausted quota) cools the connection down instead of parking it as expired, at most 3 times per hour per connection before it parks, so a revoked key still converges; the ambiguity check is now one implementation shared by the connection test and the runtime (#13609) — thanks @maxmad64bis
  • fix(proxies): pool validation no longer rewrites proxies set to inactive or dead; only active and error statuses are updated (#13612) — thanks @maxmad64bis
  • fix(opencode): the v2 plugin reads the management token from OMNIROUTE_MANAGEMENT_API_KEY (plugin option wins) and warns once at startup when management calls fall back to the inference key (#13613) — thanks @maxmad64bis
  • fix(routing): a failed stale-pin (LKGP) clear on the combo fallback path now logs the combo and execution key while staying non-blocking, and a new opt-in npm run check:routing-error-guard script (not wired into CI) flags new swallowed catches and unanchored fire-and-forget async on routing paths, with frozen entries keyed by file and catch body instead of line numbers (#13614) — thanks @maxmad64bis
  • fix(opencode): opt-in OPENCODE_TRANSIENT_FAILOVER_BACKOFF flag (default off): once two consecutive opencode accounts fail with a transient upstream error, the rotation pauses before the next account (1.5 s doubling, capped at 6 s per pause and 10 s per request), releases the failed response body first and stops dispatching if the client disconnects during the pause; with the flag off failover stays immediate (#13615) — thanks @maxmad64bis
  • fix(devin): Fall back to the CLI probe (devin acp --agent-type summarizer) when the connection-test HTTP API rejects a CLI-format key, since routing authenticates against the local Devin CLI, not api.devin.ai (#13617) — thanks @patrykkopycinski
  • fix(sse): stream reasoning deltas from combo targets incrementally instead of buffering them into a single burst, and stop rejecting reasoning-only streams as an empty completion (#13620) (#13806) — thanks @NaNomicon
  • fix(models): keep OpenRouter's Batch-API-only :batch variants out of chat routing — ModelSync imported all 77 of them into the chat catalogue, where every request that landed on one was rejected with 404 This model is only available through the Batch API (#13622) — thanks @L4XB
  • fix(cursor): kv_after_text no longer settles away a trailing exec_mcp tool call in the same buffer, preventing Composer from dropping in-flight MCP tool invocations during KV checkpoint settling (#13627) — thanks @patrykkopycinski
  • fix(providers): restore grok-4.6/4.5 default reasoning effort so requests without an explicit effort keep reasoning enabled (#13628) — thanks @HouMinXi
  • fix(registry): declare supportedThinkingEfforts on claude-opus-5 and claude-fable-5 across the anthropic/claude/claude-web/ghe-copilot/github registries (#13628) — thanks @HouMinXi
  • fix(stream-recovery): opt-in STREAM_RECOVERY_TOOLCALL_ORDER_FIX (default off) makes mid-stream continuation tool-call safe — a cut stream is never resumed once a tool call was emitted, whether still in flight or already finished with finish_reason: "tool_calls" — and closes after one empty continuation instead of spending the whole budget (#13633) — thanks @maxmad64bis
  • fix(ci): clear the release/v3.8.51 base-reds on the PR fast path. The provider pipeline keeps the upstream error code and type again, so an Antigravity missing-project 422 stays fail-closed. The glued-prefix sk- credential pattern scans error text in linear time instead of quadratic. The provider detail page no longer bundles node:fs. /v1/models stops listing custom Jina models twice. The models/providers import cycle is gone, and opencode-plugin-v2 passes the pack policy. Stale guards, fixtures, locale keys and docs counts are aligned with their merged changes (#13635) — thanks @dpozimski
  • fix(tests): add dist/httpClientAbortGuard.mjs to the expected missing-paths list in tests/unit/pack-artifact-policy.test.ts — #13636 registered the file in PACK_ARTIFACT_REQUIRED_PATHS without updating the assertion, leaving the test red on the release tip for every PR that runs it (#12732) (#13872)
  • fix(compression): report compression-worker faults instead of silently sending the uncompressed body, and fall back to the in-process pipeline for fast faults (thread error, exit, engine throw); a dispatch timeout still degrades to uncompressed, but is now logged (#13637) — thanks @marcs7
  • fix(db): search stats and analytics no longer surface "ghost" rows — a NULL/- provider or a keyed search provider whose connection was deleted — while keyless providers (duckduckgo-free, searxng-search, anonymous context7) and credential-fallback providers (perplexity-search on a perplexity key) stay visible; the analytics totals apply the same filter, so total always matches the per-provider breakdown (#13641) — thanks @maxmad64bis
  • fix(sse): recognize reasoning_effort in the reactive 400 field-strip retry — strict OpenAI-compatible upstreams that reject the field are retried once without it instead of surfacing the 400 (#13642) — thanks @Moseyuh333
  • fix(dashboard): new opt-in flag FREE_BADGE_REQUIRES_PROVIDER_FREE_TIER (default off) makes the provider-page Free badge strict — it drops the display-name heuristic, non-boolean free fields and :free suffixes on registered providers without a documented free tier, while keeping catalogued free models, explicit free: true and :free on free-tier providers and compatible nodes; with the flag off the badges are unchanged (#13645) — thanks @maxmad64bis
  • fix(api): share the single SOCKS5 flag reader across the settings proxy routes so the dashboard and the dispatcher stay consistent (#13646) — thanks @maxmad64bis
  • fix(stream-recovery): log every mid-stream continuation outcome with its attempt N/MAX token — the stitched suffix, overlap rejection, terminal/empty continuation and tool-call refusals at debug, and a recovery that gives up (budget spent, or the continuation request returned no stream) at warn — without adding any warn line to a healthy or tool-call stream; the existing mid-stream continuation attempt N/MAX line is unchanged (#13650) — thanks @maxmad64bis
  • fix(sse): Kiro translator no longer re-prepends the full relocated tool-documentation block onto every subsequent turn of a multi-turn conversation; it now stays anchored to the turn that originally carried it. (#13652) (#13808) — thanks @KelvinKSPS
  • fix(sse): opt-in OPENCODE_RATE_LIMITED_429_EARLY_STOP flag (default off): an opencode 429 classified as a real rate limit (parseable Retry-After, or a body naming a rate/usage limit) stops the cross-account wave and returns that upstream 429 unchanged — body, Retry-After and quota headers intact, so the opencode quota error rules still apply; unclassified 429s keep rotating, and with the flag off every 429 rotates as before (#9611) (#13657) — thanks @maxmad64bis
  • fix(sse): a configured daily-quota reset hour that falls inside a daylight-saving gap (New York 02:00 on spring-forward, Havana/Santiago midnight) now resolves to the first wall-clock time that exists instead of landing an hour early, sometimes on the previous day (#13671) — thanks @maxmad64bis
  • fix(sse): new opt-in flag RETRY_AFTER_PROVENANCE_ENABLED (default off): aggregated 429/503 unavailable responses omit Retry-After when no concrete future retry time is known instead of sending a synthetic 1s, carry error.retry_after_provenance (signal | none), and combo drain paths read prose retry hints from JSON and plain-text upstream bodies; non-JSON upstream error pages no longer log at warn (#13672) — thanks @maxmad64bis
  • fix(docker): isolate the ChatGPT Web (Codex) CDP proxy sidecar onto its own Compose network, add an opt-in CDP_PROXY_TOKEN auth gate to cdp-proxy.mjs, and stop the VNC browser-login CDP bridge from starting when no token is configured (#13679) (#13811)
  • fix(security): the CLI/management bearer token is now derived from a random per-install salt persisted under DATA_DIR instead of the checked-in literal omniroute-cli-auth-v1 — since /etc/machine-id is commonly world-readable, any local user could previously derive the same token as every install that never set OMNIROUTE_CLI_SALT; the explicit env override still takes priority and rotation still works the same way (#13679) (#13909)
  • fix(auth): verifyCloudSignature() no longer accepts an unverifiable X-Cloud-Sig when OMNIROUTE_CLOUD_SYNC_SECRET is unset — a forged/garbage signature is rejected outright, and the new opt-in OMNIROUTE_CLOUD_SYNC_ENFORCE_SIGNATURE=true flag rejects unsigned Cloud-sync payloads too (default stays legacy pass-through for v3.8.x; the default flips in v3.9) (#13679) (#13804)
  • fix(security): default the published Docker image and fly.toml deployment to REQUIRE_API_KEY=true (npm/CLI local-dev default unchanged), and stop /api/free-tier/summary's wildcard CORS from leaking the operator's local token usage to unauthenticated callers (#13679) (#13911)
  • fix(security): removed the copy-pasteable placeholder JWT_SECRET/API_KEY_SECRET/INITIAL_PASSWORD values from the Podman Quadlet deploy manifest, and blocked remote dashboard logins with the well-known default INITIAL_PASSWORD=CHANGEME (#13679) (#13812)
  • fix(security): the internal self-loop admission-bypass bearer is now a random per-process secret instead of the checked-in literal "sk_omniroute" when no OMNIROUTE_API_KEY/ROUTER_API_KEY is configured (#13679) (#13813)
  • fix(db): DELETE /v1/batches/delete-completed now caps the work it does per request and reports hasMore so a caller can resume, and the sweep no longer deletes a file that another batch still references (#13680, #13681) (#13805)
  • fix(security): a restricted API key is now enforced on the alias spellings next.config.mjs rewrites onto /api/v1/… — a route handler sees the client's original URL, so POST /chat/completions, /responses, /responses/*, /models, /codex/* and the doubled /v1/v1/* prefix all skipped the endpoint-category lookup and let a key allowed only on search reach chat or any other endpoint (#13685) (#13741) — thanks @gonisulaimann
  • fix(usage): mark locally estimated token usage in the call log (_omniroute.usageEstimated on the logged response) so operators can tell estimated counts and costs from provider-reported ones — covers OmniRoute's own estimate for streams without upstream usage and web executors that report estimated: true; billing, API-key budgets, quota-share and client payloads are unchanged (#13686) — thanks @maxmad64bis
  • fix(openrouter): sync the :free 1000/day tier from /credits lifetime purchases instead of staying stuck at 50/day for $10+ accounts (#13689) — thanks @Notaloop763
  • fix(translator): prevent schema property name collisions (e.g. properties, required) in Gemini schema sanitizer (#13690) — thanks @zcrew0x
  • fix(providers): map thinking.type: "adaptive" to "enabled" for AgentRouter GLM models instead of forwarding it unhandled, fixing a 400 from AgentRouter's upstream GLM endpoint (#13696) (#14043)
  • fix(providers): the shared Responses-API input sanitizer now converts Codex's proprietary agent_message input items (used for multi-agent task/reply passing) into a plain message item before forwarding to any non-Codex-native Responses upstream. Previously such items reached third-party Responses endpoints untouched, and OpenCode Go Muse Spark 1.3 rejected the request with input[N] did not match any supported type (#13698). The real Codex/ChatGPT native passthrough path is unaffected and continues to receive agent_message items as-is. (#14041)
  • fix(sse): stop the direct (no-proxy) fresh-socket retry from reusing the pooled attempt's flat OMNIROUTE_DIRECT_HEADERS_TIMEOUT_MS response-start watchdog — the retry is a brand-new socket with no zombie to detect (#10214's rationale only applies to the pooled attempt), so when the caller already attached its own deadline signal (the resolved connection/model/provider/FETCH_TIMEOUT_MS cascade) the retry now defers to a generous, OMNIROUTE_DIRECT_RESPONSE_RETRY_TIMEOUT_MS-configurable backstop instead of an identical short flat window, fixing spurious 504s on healthy slow-TTFB reasoning models (#13703) (#14047)
  • fix(copilot): fall back to the copilot-chat identity once when a standard GitHub Copilot account rejects the CLI identity with a 403, without breaking Enterprise Copilot (#13705) — thanks @tuandinh0801
  • fix(dashboard): Saving or clearing a proxy on a provider page now refreshes the per-connection proxy badges immediately instead of leaving them stale until a manual reload (#13711) — thanks @xiaoyaner0201
  • fix(db): bound health scans and isolate native diagnostics so large quota histories no longer exhaust memory or block request handling; diagnostics now run in a cancellable child and are awaited at HTTP and MCP callers (#13717) — thanks @HouMinXi
  • fix(chat): preserve suffix-model reasoning effort across model attempts so a replacement model no longer inherits or drops the original suffix, and keep explicit reasoning choices in request dedup hashes (#13720) — thanks @HouMinXi
  • fix(api): POST /v1/rerank now actually works against native TEI / Infinity provider nodes: the /rerank fallback sends texts + return_text alongside documents, and bare-array or score-only upstream responses are normalized to the Cohere {results: [{index, relevance_score, document?}]} envelope (sorted, top_n-capped) so clients and the memory engine's rerank step see real scores (#13733) — thanks @seanford
  • fix(api): GET /v1/models types compatible-provider-node rows by the node's apiType when a discovered/added model carries no endpoint metadata — an embeddings node's models are type: "embedding" and a rerank node's models are type: "rerank" instead of surfacing as untyped chat models; a manual overlay's supportedEndpoints also re-types the merged row (#13734) (#13740) — thanks @seanford
  • fix(memory): the Embedding and Rerank selectors on Memory → Engine now list local provider nodes with the models they actually expose for that modality (synced + custom rows, typed the same way GET /v1/models types them) instead of by node apiType alone with an empty model list — a node typed embeddings that also serves a reranker now appears under Rerank with that model, and under Embeddings with its embedding model, so prefix/model no longer has to be typed by hand (#13740) — thanks @seanford
  • fix(usage): preserve nested prompt cache-read counters before cost calculation (#13760) — fixes #13746 — thanks @xiaoyaner0201
  • fix(db): Health-check-repair backup pruning now resolves maxFiles/retentionDays from the persisted Storage-page setting (not just env vars), matching manual/API/auto backups. (#13308) (#13773) — thanks @voidstackloop
  • fix(sse): CCR retrieve-tool detection now recognizes an MCP-gateway-namespaced tool name (e.g. Docker MCP Toolkit's mcp__docker__omniroute__omniroute_ccr_retrieve or a single mcp__<server>__omniroute_ccr_retrieve prefix) via a separator-bounded trailing-segment match instead of requiring an exact omniroute_ccr_retrieve string — previously an MCP-capable caller reachable only under a gateway-assigned name was treated as unable to retrieve, so the entire CCR compression engine was skipped for it (#13781). (#14028)
  • fix(i18n): retranslate the English strings that had been copied verbatim into the locale catalogs (Spanish alone carried 7,142) and turn the real-translation ratio gate into a blocking ratchet. (#13782)
  • fix(resilience): the chat admission gate's pressure check now actively re-samples instead of reading a passive cache, so the resource_pressure guard can observe recovery and stop shedding once real pressure clears, instead of requiring a full process restart (#13823) — thanks @pandudpn
  • fix(providers): the Zylo API key check now probes the authenticated chat route instead of the open catalog — Zylo serves GET /v1/models without authentication, so the account-setup dialog accepted any string as valid and the key was only rejected later, when a model test returned Zylo's own 401 "Key not found: zk-…" (#13828) (#13877)
  • fix(sse): when a gateway API key's allowed_connections / quota scope hides every connection of a provider, chat now answers 403 naming that scope instead of the generic No active credentials for provider: X — which was indistinguishable from "never configured" even though /test and /sync-models kept working on the same connection (#13832) (#13879)
  • fix(combos): Keep vision capability consistent for MiMo V2.5 and Step 3.7 Flash provider/free variants so /v1/combos no longer under-reports multimodal combos whose members are already advertised as vision-capable by /v1/models (#13847). (#13863) — thanks @smshagor-dev
  • fix(translator): Third-party tool names (e.g. GitHub Copilot's own web_fetch function tool) are no longer sent unprefixed to Claude-wire-format providers outside genuine first-party Anthropic traffic, fixing a rejected tool(s): web_fetch 400 for any gh/claude-* model (#13856) — thanks @dylanhaskins
  • fix(build): externalize @modelcontextprotocol/sdk in standalone server webpack config to prevent TDZ ReferenceError on MCP initialize (#13859) — thanks @HouMinXi
  • fix(combo): stop the chars/4 context estimate from demoting an operator-verified model_context_override behind an unconfirmed catalog "emergency" fallback in combo priority ordering (#13870) (#14046)
  • fix(quota): resolve Quota Sharing plan/limits from the pool's canonical primary connection in enforceQuotaShare and recordConsumption, not the serving connection — a multi-connection pool whose "Limite" wizard override only ever lands on the primary connection was writing/checking consumption under a different dimension key when a request was actually served via a non-primary pool member, so the dashboard's "consumed" amount never reflected real traffic (#13876) (#14042)
  • Fixed a security issue where a revoked, expired, or banned API key could still resolve an owner scope in getApiKeyRequestScope() and keep accessing its own /v1/files and /v1/batches records instead of being rejected with 401 (#13881). (#14024)
  • fix(security): scope /api/files and /api/batches management siblings to the caller's own API key (session stays instance-wide), closing a cross-tenant read that let an anonymous or foreign-key caller enumerate and download other tenants' files/batches (#13882) (#14027)
  • security(images): close a DNS-rebinding TOCTOU (#13883) at the three newer public-only image download sites — resolveImageSource and the NanoBanana result-URL conversion in imageGeneration.ts, and resolveUpscaleImageSource in imageUpscale/shared.ts. All three validated a caller-supplied URL's DNS answer as public but then let the download perform an independent, un-pinned second resolution at connect time, so a host that answered differently between the two lookups (public, then loopback/LAN) could reach an internal address; they now set pinDns: true (reusing the existing createPinnedFetch helper already used by embeddings and the vision/audio/video bridges), binding the connection to the exact validated address. (#14032)
  • fix(i18n): reviewer pass over the 1,865 pt-BR leaves retranslated in #13782 (172 corrections) via the new review-locale script. (#13885)
  • fix(db): getPricingForModel() now reads through the existing 30s TTL getCachedPricing() helper instead of rebuilding pricing from scratch (3 SELECTs + JSON.parse + merge) on every call, eliminating the multi-second event-loop stall /api/usage/history hit when calculateAggregateCost() invoked it once per GROUP BY row (up to 531 times per request) (#13891). Every known pricing writer already invalidates this cache via touchPricing(), so writes remain immediately visible. (#14040)
  • fix(cli): translate the CLI for every locale — 38 catalogs carried 1–24 of 830 keys and fell back to English; sync-ui-keys --catalog=cli fills them (52,000 strings) and a completeness gate keeps them full. (#13892)
  • fix(dashboard): the request-log detail view now shows an explicit "payload omitted — exceeded the call log size limit" notice for a pipeline/request/response section that was replaced by the size-limit marker (_omniroute_truncated / [omitted: call log artifact size limit exceeded]), instead of silently rendering the marker verbatim under a generic "Pipeline Error" title as if it were a real upstream error. Also fixed .env.example documenting stale CHAT_LOG_ARRAY_TAIL_ITEMS=128/CHAT_LOG_MAX_DEPTH=6 defaults that no longer match the code's actual 1000/20. (#14045)
  • fix(sse): the CCR protocol instruction (the system note teaching a model how to call omniroute_ccr_retrieve) is now injected for callers reaching OmniRoute's MCP server through a namespacing gateway (Claude Code / Docker MCP style mcp__<gateway>__<server>__omniroute_ccr_retrieve, or a dotted/slashed prefix) — callerSupportsCcrRetrieve() previously matched by exact string equality only, so a genuinely reachable but gateway-namespaced tool name was never recognized and the instruction (and CCR compression itself) was silently skipped (#13897). (#14028)
  • fix(tests): replace the flaky 250ms wall-clock ReDoS guard in sanitizeErrorMessage's property test with a deterministic cost-scaling check, so the test proves bounded-backtracking instead of failing on machine load (#13907) (#14039)
  • fix(providers): detectVisionInput now recognizes Lemonade Server's labels[] vision capability, so Lemonade vision models import with supportsVision set instead of being treated as text-only (#13918) (#14023)
  • fix(sse): deepseek provider registry now declares defaultContextLength: 1_000_000 and an explicit deepseek-flash model entry, so unlisted/new DeepSeek models (like the new DeepSeek V4.1 Flash) no longer fall back to the generic 128k context limit (#13922) (#14026)
  • fix(sse): opt-in OPENCODE_PARK_AND_RESUME flag (default off): after repeated transient 429s the opencode rotation parks the request with a heartbeat and replays one capped leg of up to 3 sequential accounts instead of fanning out the whole fleet; with the flag off every 429 rotates as before (#13924) — thanks @maxmad64bis
  • fix(guardrails): resolve nested combo-ref steps to their real leaf models when deciding vision-bridge behavior, so a pass-through combo pointing at an all-vision-capable inner combo skips the describe-and-replace path instead of stripping raw images (#13927) (#14038)
  • fix(docs): re-sync the 65 documentation mirror sets — 817 mirrors rewritten: the 14 core sources edited since their translation, the 322 mirrors that were still English copies, and the frontmatter the old extractor had leaked into the newer locales' bodies. The docs pipeline now retranslates only the ## sections whose text changed, and the drift gate is blocking. (#13940)
  • fix(combos): synchronize allowedProviders and allow invariant override when updating combos from dashboard (#13951) — thanks @fouadSalkini
  • fix(i18n): translate the 3,719 __MISSING__ markers (61 keys × 61 locales) that eight base PRs stamped into the catalogs on 2026-09-16, restoring the real-translation ratio gate on the release tip. (#13974)
  • fix(providers): OpenRouter model discovery honors the per-connection base URL override instead of always importing the global catalog, so a connection pointed at a regional endpoint (e.g. the EU in-region host) no longer advertises model ids that endpoint cannot serve (#14001) — thanks @tiangao88
  • fix(providers): Perplexity Web no longer collapses runs of spaces in non-streaming answers (the path tool mode always takes), which flattened code indentation in write_file arguments and plain code blocks; citation markers are still removed with single spacing left behind (#13968) (#14009) — thanks @costajohnt
  • fix(docker): copy the app into the runtime image with --chown=node:node instead of a second chown -R layer, so the image no longer stores the ~2 GB standalone build twice (#13990) (#14010) — thanks @costajohnt
  • fix(opencode): an OpenCode free-tier refusal no longer counts as a healthy response and no longer takes the account it landed on out of rotation for that model: the 403 is recorded on the connection instead of staying unclassified, it stops clearing the refused account's failure history, and the request comes back without a pointless hop across accounts that would all get the same verdict. Every sibling account returns the same answer to the same request, so one refusal per account would otherwise empty the pool and leave later requests answered "no active credentials" (#14011) — thanks @maxmad64bis
  • fix(opencode): keyless OpenCode models answer again instead of returning 403 — requests now carry a versioned OpenCode user-agent, canonical session and request ids derived from the existing conversation fingerprint, a streamed upstream request, and a tool list. The upstream inspects which tool names a request declares, so rather than pinning a list, OmniRoute reuses the one a request of the same conversation was last seen getting through: a title or a summary, which its client sends without tools, goes out with the list that client already declared. Two opt-outs (OPENCODE_FREE_TIER_REQUEST_CONTRACT, OPENCODE_FREE_TIER_PLACEHOLDER_TOOLS) (#14013) — thanks @maxmad64bis (with thanks to @AStupidBear for the identity-header work in #13937)
  • fix(build): drop the orphaned httpClientAbortGuard.mjs entries from the pack-artifact allowlists — the #13636 crash-guard wiring was removed, so no producer or consumer ships the file anymore (#14029) — thanks @maxmad64bis
  • feat(sse): multi-account rotation now spreads sends per network egress and eases off fleet-wide when throttled, opt-in via OPENCODE_EGRESS_THROTTLE_ENABLED=1 (#14290) — thanks @maxmad64bis
  • Fix Antigravity image generation not rotating to another account when the upstream returns an explicit quota-exhausted 429, so a second configured account with available quota is no longer stuck behind the first account's terminal quota error. (#9908) — thanks @Ardem2025
  • fix(sse): requestQueue.maxQueueDepth = 0 (the documented default, "disabled") once again means an unbounded account queue. #12911 redefined 0 inside accountSemaphore as "reject when busy" for its Codex WS leases, and because chatCore forwards maxQueueDepth straight into that option, every request that found its account slot busy under default settings was answered 429 Semaphore queue full (0) instead of waiting. The lease keeps its refuse-don't-wait behaviour through an explicit failFast option. (#14101)
  • fix(sse): the streaming OpenAI→Claude translator relays reasoning for legacy callers that never pass requestedThinking (undefined), matching the non-streaming path and the pre-#12905 contract; only an explicit opt-out (false) suppresses it, and only that case synthesizes the reasoning into a text block. (#14101)
  • fix(security): requestRejectedFailure.ts (#12864) sanitizes the upstream message at its own lastError writes instead of trusting the caller to have done so, and the public-boundary guard now covers the extracted module. (#14101)
  • fix(adobe-firefly): never spawn a real Chrome for CDP session warm under a unit-test runner, so tests stop leaking a browser process that holds an OS handle on its DATA_DIR profile directory (#13289) — thanks @anhtahaylove
  • fix(providers): declare Agnes CN chat models' live reasoning_effort vocabulary so the catalog and sanitizer stop inventing tiers the CN API rejects. Probes on api.agnes-ai.cn (2026-09-14) match the international endpoint: 2.0/2.5 accept none/low/medium/high/max, 3.0 also accepts minimal and xhigh; off/ultra clamp off the wire and Hermes' default xhigh clamps to max on 2.x. (#13399) — thanks @HouMinXi
  • fix(providers): declare Agnes chat models' live reasoning_effort vocabulary so catalog/builder/sanitizer stop inventing aliases the API 400s. 2.0/2.5 accept none/low/medium/high/max; 3.0 also accepts minimal and xhigh. off/ultra still clamp off the wire. (#13655) — thanks @HouMinXi
  • Fix Antigravity quota parsing treating an unreported remainingFraction as 0% remaining instead of unknown, which made a genuinely exhausted quota indistinguishable from one the upstream simply didn't report. (#7138) — thanks @Ardem2025
  • fix(api): accept blockedModels in the key permissions update schema so the deny-list half of per-key model policy is no longer silently stripped before it reaches the route (#13666) — thanks @fouadSalkini
  • fix(tests): bump the APIKEY_PROVIDERS tripwire count to 241 — Agnes AI China (#13399) added one apikey/regional entry, and the stale 240 was failing a unit-test shard on every open PR against release/v3.8.51 (#13905)
  • fix(ci): the advisory forgotten-sibling-tests step no longer fails "Fast Quality Gates" when a PR touches a hub module — the cross-product of consumers × candidate tests reached millions of rows and rendering them exceeded V8's maximum string length, so the throw hit main()'s catch and exited 1. The report now lists at most 200 rows per section (and 5 000 per array in the JSON artifact) while the header keeps the exact totals (#13889)
  • fix(compression): Caveman's leader_phrases rule (and any other anchored file-pack rule) works again. #12825 replaced the keyword prefilter for file-based packs with a test of the rule's own regex, but ran it against a lower-cased snapshot of the original text — so ^(?:i will|…) never matched once the leading Sure, was still there, and the rule was skipped before it could see the text pleasantries had already cleaned. The prefilter now sees the text as earlier rules left it. (#14164) — thanks @gonisulaimann / @prabhtheone / @xiaoyaner0201 / @xiechimon
  • Dashboard: renamed the Claude connection field helpers to claudeConnectionFieldValues.ts — the .ts/.tsx pair from #13074 differed only by casing, which breaks webpack on case-insensitive filesystems and made esbuild resolve the wrong module in the browser-bundle guard. (#14164) — thanks @gonisulaimann / @prabhtheone / @xiaoyaner0201 / @xiechimon
  • fix(oauth): stop posting a Claude refresh token that another in-process refresh already consumed. Re-check the rotation map and DB inside serializeRefresh (both Layer 1 and Layer 2), record Layer 2 rotations, re-read the connection uncached on invalid_grant, and keep the Claude refreshToken instead of nulling it into sticky no_refresh_token. (#13874) — thanks @ai-jeremi-esky
  • fix(oauth): stop nulling the Claude refresh token on the first unrecoverable refresh failure so CredentialHealth no longer gets stuck sticky-dead — the retry budget from #11414 can now actually spend its second attempt instead of finding an already-cleared token (#13183) (#13185) — thanks @RaviTharuma
  • fix(stream): restore Claude SSE passthrough tool_use names to the casing the client actually declared instead of "upgrading" them to the canonical Claude Code spelling (bash → Bash), which broke third-party Anthropic-format clients (pi/OpenCode on claude-format executors like devin-cli-agentic) while leaving the JSON path correct; the pre-existing Claude Code protection (upstream downcase restored to declared PascalCase, #7926) is preserved (#12855) — thanks @Neuron-Mr-White
  • fix(oauth): classify embedded invalid_grant in Cline token refresh error bodies so permanently consumed refresh tokens trigger re-authentication instead of indefinite transient retry loops, and add cline to ROTATION_LOCK_GROUP to serialize concurrent sibling refreshes (#13466) — thanks @fouadSalkini
  • Add a visible manual callback-entry action to the Codex loopback warning so remote users can paste the authorization result instead of setting up an SSH tunnel. (#9944) — thanks @Ardem2025
  • Fixed Codex executor forwarding client reasoning sub-fields (enabled, max_tokens, exclude) that the Codex Responses API rejects with HTTP 400, taking down every combo target with a deterministic client error. The reasoning object is now whitelisted to effort/summary, and enabled: false maps to effort none when no more specific effort was requested. (#13643) — thanks @HouMinXi
  • fix(combo): retry the same target once when a streaming response fails before any content reaches the client (streaming upstream error), including native-pinned Codex turns whose set retries stay disabled — previously the caller returned a 502 immediately instead of using the existing transient-retry loop (#13630) — thanks @anhtahaylove
  • Emergency fallback: the budget-exhaustion reroute targets nvidia/openai/gpt-oss-120b again, as documented in ENVIRONMENT.md and the NVIDIA hosted-model snapshot; #14006 had switched the provider to groq inside an unrelated MITM change, so operators without a Groq connection got the original 402 back. (#14164) — thanks @gonisulaimann / @prabhtheone / @xiaoyaner0201 / @xiechimon
  • fix(security): client-supplied image URLs (image_url / mask_url / message parts on image generation and upscale, chat image_url parts inlined by the vision bridge) and the NanoBanana result download now pin the public-only outbound guard with DNS validation, instead of inheriting the operator provider policy — block-metadata on a default install let a request body make the server fetch loopback/LAN URLs and forward the bytes upstream (GHSA-34rg-3pqj-35g9) (#13748)
  • fix(authz): classify the 14 remaining spawn-capable /api/cli-tools/* routes (all-statuses, status, detect and the claude/cline/codewhale/codex/crush/deepseek-tui/droid/kilo/openclaw/pi/smelt-settings writers) and the /api/skills/install + /api/skills/executions pair as LOCAL_ONLY — they reach child_process.spawn transitively (getCliRuntimeStatus() / detectAllTools() / the skills sandbox) but only sat behind Tier 3 MANAGEMENT auth, which requireLogin=false waives; loopback/LAN enforcement now runs before any auth check, matching their already-gated siblings (GHSA-35fw-cv32-2373 — thanks Parth Narula; GHSA-jx89-f37j-pq89 — thanks Aeon). Tunnel-served dashboards lose the CLI Tools status badges, the same trade-off already accepted for grok/forge/jcode/qwen. (#13745)
  • fix(security): redact Groq (gsk_…), xAI (xai-…) and every OpenAI-compatible sk-… key shape (DeepSeek 32-hex, Moonshot/Kimi, Together, …) in error bodies and the opt-in credential-masker guardrail — the catalog only knew the exact 48-char OpenAI form, so those keys passed through the guardrail verbatim and gsk_/xai- also reached public error responses (GHSA-r4q7-7f24-m29p) (#13744)
  • fix(grok-cli): a 429 "used all the included free usage … rolling 24-hour window" on Grok Build is quota exhaustion for that model, not a 30s rate-limit wait. Combo skips the drained grok-4.6 login and tries the next account (#13984) — thanks @stormsia
  • Rate limit: requestQueue.maxWaitMs=0 (the disable sentinel from #12902) no longer trips the #12715 queue-budget gate — every request on a protected connection was rejected with an immediate 503 queue budget instead of waiting without a queue deadline. Execution stays bounded by executionMaxWaitMs and the upstream fetch-start timeout. (#14164) — thanks @gonisulaimann / @prabhtheone / @xiaoyaner0201 / @xiechimon
  • fix(mcp): MCP audit treats a non-callable better-sqlite3 export (better-sqlite3 export is not a function) as a native load failure, falls back to node:sqlite, and caches a failed driver load so dashboard polls stop reprinting (a database file that does not exist yet is never cached, so the connection recovers once the app creates it). Docker now refuses to ship without better_sqlite3.node. Native-load classification lives in sqliteLoadError.ts so core.ts stays under its frozen line cap. The build bootstrap keeps a deliberately narrower classifier: a corrupt binding there must not be read as "no encrypted credentials", or a fresh STORAGE_ENCRYPTION_KEY would be generated over an existing encrypted database. (#13903) — thanks @HouMinXi
  • fix(api): the model-test skip for image/music/video-only models (#13376) returned a result with no httpStatus, and the /api/models/test route hands that field straight to NextResponse — so a skipped test reached the client as HTTP 200 carrying status: "error" in the body. It now answers 422: the request is valid, but that model's modality cannot be exercised by a chat test. Also clears the TS2741 that was failing API Route Typecheck on the release branch. (#13730)
  • fix(sse): detectMalformedNonStream no longer flags a Claude-format message as malformed when the content array only contains empty text blocks or the (empty response) sentinel and stop_reason is "length" — a legitimate truncation (e.g. ollama/qwen3:1.7b exhausting its reasoning budget before producing visible text), not a real 502-worthy empty response. (#12935) — thanks @jasminsehic
  • fix(compression): output styles and the caveman output mode now place their injected instruction in the top-level system field instead of a synthetic messages[0] entry for Anthropic-shaped requests, fixing the upstream 400 ("use the top-level 'system' parameter for the initial system prompt") (#12584) (#13383) — thanks @Xore
  • Add a shadow release-acceptance report next to release-green.json. It does not close #12732 and is not a Mergify required check. (#13701) — thanks @HouMinXi
  • fix(sse): chat admission PSI uses the container cgroup memory.pressure file instead of host-wide /proc/pressure/memory, so a swapping host no longer 503s an idle Docker/cgroup OmniRoute with resource_pressure; the host file remains the fallback when the cgroup sample is missing (#12562) — thanks @SCys
  • fix(responses): ensure full compliance with the OpenAI Responses API streaming schema for strict deserializers (e.g. OpenAI Responses SDK, Grok CLI / pager): - Include output: [], background: false, and error: null in the response.in_progress lifecycle event across both the Responses transformer and response translator. - Include status (in_progress or completed) on all emitted output items (message, reasoning, function_call, custom_tool_call) in response.output_item.added, response.output_item.done, and response.output[]. - Include sequence_number: 0 in in-band Responses stream error frames (OPENAI_RESPONSES_ERROR_FRAME and buildResponsesErrorDataLine) emitted after early keepalive streams commit. - Ensure input_tokens_details (with cached_tokens: 0) and output_tokens_details (with reasoning_tokens: 0) are always populated in response.usage even when upstreams (e.g. Gemini) omit reasoning or caching tokens. (#13956) — thanks @TheDemonTuan
  • responses: Responses-to-Chat fallback now skips replayed web_search_call metadata while preserving the paired function result, preventing deterministic HTTP 400 failures on follow-up turns routed to Chat Completions providers (#13304 (#14090) — thanks @anhtran-ai)
  • fix(security): bump the adm-zip override to ^0.6.1 — 0.6.0 followed a symlink already present inside the extraction root and could write outside it (GHSA-vwc7-r8mq-g2x9 / CVE-2026-76845); 0.6.1 walks every path component with lstat and refuses symlinks. Reached only through onnxruntime-node's install script, which unpacks the vendor's own binary — no request-path exposure. (#13737)
  • fix(ci): register the four unit tests whose mutant kills were not counting — daily-reset-tz-threading, noauth-model-lockout, free-badge-provider-gate and local-token-budget-429-skips-cooldown — in stryker.conf.json tap.testFiles. check-mutation-test-coverage --strict reported 6 missing coverings across 4 mutated modules (accountFallback, sse/services/auth, combo/comboPredicates, combo/rrState) on the release line itself with no PR diff involved, so the mutation-test-coverage gate was red on every open PR regardless of its contents; the gate now reports no drift (#13814) — thanks @abhisheksharma2411
  • fix(ci): drop a tap.testFiles entry naming a deleted test file, and guard the direction the existing drift check never covered — an entry left behind after its test is removed or renamed costs mutation coverage silently, because Stryker resolves the list into its sandbox without failing on a dangling path (#13814) — thanks @abhisheksharma2411
  • fix(providers): give the TinyCMS wasm-bindgen Node DOM stub a dedicated window with a Location-shaped object (never window = global without location), so Next.js SSR getLocationOrigin cannot crash every route after TinyCMS is used once (#13957) — thanks @aref-alapour
  • fix(analytics): resolve account email/name in Utilization Account Split chart and fix legend bleeding through tooltip (#13029) — thanks @ZaimMarzuki
  • Electron release: electron/package-lock.json regained the optional electron-builder-squirrel-windows subtree (13 entries) that npm ci had been refusing as out of sync — the Linux desktop leg died on it — and electron-release.yml gained a build_ref dispatch input so a release whose tag was cut with the broken lock can have its assets rebuilt from the repaired line (#13140) — thanks @ozeas
  • npm publish workflow: the CycloneDX SBOM is attached to the GitHub Release on workflow_dispatch publishes too (when a release for the tag exists), not only on the release event — v3.8.50 shipped through a staged dispatch and its release carried no SBOM until it was attached by hand from the run's sbom-npm artifact (#13140) — thanks @ozeas
  • fix(providers): strip <script>/<style> blocks from the Vertex model-docs HTML even when the end tag carries junk before the > (</script foo>, which the HTML spec still treats as a close). The old regexp required </script\s*>, so such a block survived; the generic tag-stripping pass then removed both tags and kept the script BODY, letting its text reach the table cells the context-window/token-limit parser reads (CodeQL js/bad-tag-filter, alert #1007). (#13936)
  • fix(vertex): prefer live model discovery for OAuth, Service Account, and service-account-bound authorization-key credentials while falling back cleanly to the Gemini-only Express catalog for standard API keys. Model Garden resources now route by publisher protocol: Claude and Mistral use their native rawPredict APIs, and Grok plus current or future open MaaS publishers use Vertex's OpenAI-compatible endpoint with normalized request IDs such as xai/grok-4.6. Distinguish intentional API-key catalog rejection from transient HTTP or network failures. (#12471) — thanks @JxnLexn
  • fix(db): remove the periodic wal_checkpoint(TRUNCATE) scheduler. Truncating the WAL rewrites the shared wal-index (storage.sqlite-shm) while other connections and in-flight statements still hold it mapped, which crashed long-running servers with SIGBUS roughly every six hours (#13973). WAL hygiene is unchanged: PASSIVE checkpoints still run every five minutes, now count busy contention and retry after 60s instead of silently logging, and a WAL above OMNIROUTE_WAL_GUARD_MAX_MB runs wal_checkpoint(RESTART) so the file stays bounded without rewriting the mapped index. The shutdown checkpoint still truncates. OMNIROUTE_WAL_TRUNCATE_INTERVAL_MS is ignored and logs a deprecation warning. (#14005) — thanks @HouMinXi
  • fix(antigravity): surface image quotas in Provider Limits (#12311) — thanks @Bl0ck154
  • fix(perplexity-web): preserve system contract on follow-up requests (#12443) — thanks @tanveer-arch
  • fix(translator): map Claude stop_sequences and stop to Gemini stopSequences (#12785) — thanks @Siva010
  • fix(cursor): preserve native Claude effort model IDs before executor dispatch (#12838) — thanks @Pllutonyy
  • fix(sse): guard empty tool_calls[] and strip tool_choice without tools (#12901) — thanks @initguru
  • fix(resilience): allow maxWaitMs=0 as disable sentinel for execution expiration (#12902) — thanks @initguru
  • fix(sse): demote mid-conversation system roles to user in claude-to-openai translation (#12908) — thanks @initguru
  • fix(soniox): pass client parameters through and surface speaker diarization (#12948) — thanks @amirrezakm
  • fix(tests): drain call-log saves before chat-pipeline DB resets (#12780) (#12966) — thanks @visheshgubrani
  • fix(providers): resolve unsupportedParams via model aliases so K3 stops 400ing on temperature (#13037) — thanks @patrykkopycinski
  • fix(db): gate the auto-cleanup VACUUM on reclaimable space, not row count (#13079) — thanks @hartmark
  • fix(logs): recover concatenated JSON objects in the Provider Event Stream viewer (#13115) — thanks @hartmark
  • fix(quota): apply the equal-split fallback in the pool usage snapshot (#13159) — thanks @datrixlab
  • fix(deps): declare remark-gfm as a direct dependency (#13162) — thanks @marioschoenert-code
  • fix(responses): count input tokens locally for Codex OAuth (#13167) — thanks @anhtran-ai
  • fix(sse): deepseek-web resilience — premature session close + malformed tool-call recovery (#13226) — thanks @VictorRP7
  • fix(memory): skip FTS rewrite on access-count updates (#13331) — thanks @HouMinXi
  • fix(security,resilience): block origin-IP header forwarding and treat 413 as retryable TPM (#13350) — thanks @themedexperiencesusa
  • fix(build): copy ioredis and bcryptjs into the standalone bundle (#13352) — thanks @sistemabritto
  • fix(embeddings): send stored API key on private-host embeddings nodes (#13398) — thanks @HouMinXi
  • fix(claude-web): add charset=utf-8 to Content-Type headers to fix Arabic/Persian UTF-8 mojibake (#13416) (#13419) — thanks @KooshaPari
  • fix(compression): track only the recursion path in isStrictlySerializable (#13154) (#13423) — thanks @KooshaPari
  • fix(network): skip Chrome TLS impersonation for Groq (#13445) — thanks @HouMinXi
  • fix(gemini): send thinkingLevel for 3.8 Flash so thoughts stop eating the output cap (#13463) — thanks @HouMinXi
  • fix(compression): log warnings for unreadable settings rows (#13522) — thanks @KooshaPari
  • fix(proxy): add combo scope to fail-closed proxy guard (#13551) — thanks @KooshaPari
  • fix(combos): stop Gemini thinking from failing dashboard combo tests (#13560) — thanks @HouMinXi
  • fix(sse): strip trailing assistant prefill on official Claude OAuth (#13572) — thanks @HouMinXi
  • fix(usage): detach completed request previews (#13623) — thanks @jbovard2016
  • fix(providers): clamp SenseNova DeepSeek V4 Flash effort to high (#13626) — thanks @HouMinXi
  • fix(routing): skip redundant parseAutoPrefix for recognized built-in auto variants (#13647) — thanks @hummern
  • fix(embeddings): log server-side when a provider can't be resolved (#13687) — thanks @hartmark
  • fix(codex): forward the caller client version upstream instead of a pinned default (#13708) — thanks @zeeshanhaque21
  • fix(gemini): a tool name starting with a digit no longer fails the request (#13738) — thanks @L4XB
  • fix(docker): smoke-test images before promoting latest (#13761) — thanks @lorenzozanee
  • fix(admission): measure request bodies with the active cost budget (#13762) — thanks @lorenzozanee
  • fix(providers): add missing modelsUrl for qwen-cloud-token-plan (#13764) — thanks @lorenzozanee
  • fix(providers): bypass web-search fallback for Antigravity target (#13447) (#13765) — thanks @lorenzozanee
  • fix(resilience): name collision keys, surface header drops, retry embeds, skip far-reset pings (#13766) — thanks @lorenzozanee
  • fix(compression): preserve case, tags and negations through llmlingua engine (#13768) — thanks @lorenzozanee
  • fix(sse): skip already-refused route per request on 429 (#13795) — thanks @maxmad64bis
  • fix(antigravity): canonicalize tiered flash quota (#13809) — thanks @domenicomassafra
  • fix(executors): strip invalid OpenCode stream options (#13819) — thanks @qinghuanandejiangshi
  • fix(providers): refresh uncloseai free-model roster after upstream rotation (#13825) — thanks @legas888Oleg
  • fix: match compatible-provider models owned by public prefix (#13831) — thanks @sahildaswani
  • fix(ci): drain three base reds blocking every PR — generated SKILL.md, stryker registry, env/docs contract (#13834)
  • fix(sre): redact secrets split across stream chunks (#13837) — thanks @pacocartones
  • fix(qoder): unwrap split SSE error envelopes (#13838) — thanks @pacocartones
  • fix(streaming): track progress across chunk boundaries (#13839) — thanks @pacocartones
  • fix(nlpcloud): restore chatbot endpoint coverage (#13845) — thanks @pacocartones
  • fix(translator): pair Gemini tool responses per turn to prevent cross-turn ID collision (#13848) — thanks @TheDemonTuan
  • fix: reclaim expired half-open probe lease (#13849) — thanks @luw2007
  • fix(providers): normalize non-function tools for allowlisted built-in OpenAI-format providers (#13855) — thanks @xiechimon
  • fix(auth): enforce blocked models in all-access mode (#13861) — thanks @fewensa
  • fix(codex): preserve native custom tools in Responses WebSocket requests (#13864) — thanks @trycohn
  • fix(docs): restore the env/docs contract broken by the #13679 vars (#13875)
  • fix(chatCore): only preserve tool_result blocks for Claude-native targets (#13972) — thanks @phs1997
  • fix(windows): hide supervised server console (#13992) — thanks @prabhtheone
  • fix(ci): document OMNIROUTE_STRIP_SYSTEM_PREAMBLE — the env/docs base red blocking every PR (#14022)
  • fix(providers): declare Agnes official thinking effort tiers (#13655) (#14063) — thanks @HouMinXi
  • fix(resilience): extend process crash guard to combo hedge cancels and upstream fetch failures (#13636) (#14064) — thanks @HouMinXi
  • fix(codex): whitelist reasoning object keys before the wire (#13643) (#14065) — thanks @HouMinXi
  • fix(sse): aggregate the findInsensitive collision warnings into one line per catalog build instead of one per colliding key (#12972) — thanks @IAMBOBJIM
  • Electron release workflow: the publish-npm job now grants actions: read to the reusable npm-publish.yml it calls (its publish job requests it), which is what made GitHub refuse the whole v3.8.50 run at startup and ship the release with zero desktop assets; a workflow_dispatch now builds the requested tag instead of the dispatching branch and can skip the npm leg (publish_npm=false) when only re-attaching assets (#11974)
  • fix(video): Honor poll_interval_ms and max_polls for Agnes and other video job providers so client-requested polling delays prevent upstream status-query rate limits. (#14114) — thanks @kaung-minkhant
  • fix(providers): A Google Cloud Code / Antigravity 403 VALIDATION_REQUIRED ("Verify your account to continue") no longer permanently bans the connection — the phrase was in ACCOUNT_DEACTIVATED_SIGNALS, so a single transient occurrence parked a healthy, fully-quota'd account as banned (permanent: true, 1-year cooldown, never auto-recovers). Measured on a live deployment: one Antigravity connection returned 33 of these 403s inside 10 minutes and stayed active, while a sibling holding 100 % of its quota on all 17 windows was banned by exactly one — the only difference being which attempt happened to be served. The phrase now lives in a new non-terminal ACCOUNT_VERIFICATION_REQUIRED_SIGNALS table (open-sse/services/errorClassifier.ts, home of the 403 branch that consumes it) and classifies as recoverable PROJECT_ROUTE_ERROR, which also un-blocks the cloud-code project-route recovery added for Gemini Code Assist (#868, #6452) — it was unreachable for this wording because accountDeactivated is evaluated first. Genuine Antigravity ban wording (this service has been disabled in this account for violation) still terminalizes, and a body carrying both still bans. (#14848) — thanks @adevwithpurpose
  • fix(release): clear release/v3.8.51 base-reds — the OpenCode free-tier observed-tools retry no longer re-sends the exact tool shape the upstream just refused, /v1/audio/transcriptions answers "Invalid model ID" again for unsafe model ids, the combined chatCore tool-metadata extraction no longer leaves the _toolNameMap side channel on the dispatch body, and the node_modules cache key covers every postinstall helper (#15112)
  • fix(context): combo members whose window resolves only to the generic 128000 catch-all no longer clamp the runtime combo context limit — the sibling-side counterpart of #10734 (#14931 (#15053) — thanks @skygunner)
  • fix(context): state the messages-only basis (and tools reserve) in the proactive-compression log lines so they read consistently next to the full-breakdown guard numbers (#14931 (#15055) — thanks @skygunner)
  • fix(release): the local merge-train now runs the four i18n contract gates on every combined tree before an --admin merge — i18n:check-keys (every en.json key in all 65 dashboard catalogs), i18n:check-keys:cli, i18n:check-ratio (real-translation ratio) and the docs drift gate (check-translation-drift.mjs) — so a PR that adds dashboard keys without translations or edits a documented page without its mirrors is ejected from the train instead of landing on the release tip. Three such landings reached release/v3.8.51 in 48 h (#13670, 1b2349de, 7f1b4a5e) although each PR's own CI was red on the same gates, because the train validated only typecheck, size/complexity and changelog integrity. (#14437)
  • fix(db): concurrent cache misses now share one in-flight load, so the nightly usage_history rollup no longer loads the full pricing table once per group after a pricing sync, which could run the server out of memory (#15101) — thanks @cryptiklemur
  • fix(translator): /v1/responses streaming no longer emits response.output_text.delta for a message item after its response.output_item.done (#13693) — upstreams that interleave plain text after a real tool_calls delta (deepseek-v4, Kimi-style empty-tool_calls payloads) used to strand those deltas on an already-closed item, which Codex CLI rejects as OutputTextDelta without active item and whose text silently vanished from response.completed.output; post-close text now opens a fresh message item at the next free output_index so every done item stays immutable and the tail text reaches the client in full. (#14153) — thanks @gonisulaimann
  • fix(db): stop syncManagedAvailableModelAliases from silently adopting and later pruning hand-created custom model aliases (#11836) (#14409)
  • fix(providers): stop sending Vertex Express API keys to generativelanguage.googleapis.com during model discovery (#12328) (#14383)
  • fix(pack): add BUILD_OMNIROUTE_BASE_PATH to APP_STAGING_ALLOWED_EXACT_PATHS so the prepublish staging prune no longer deletes the base-path sentinel before the tarball is packaged, which previously crashed the Docker container at boot (same failure class as the BUILD_SHA / #7065 incident) (#12388) — thanks @Egorich-print
  • fix(providers): Perplexity Web now declares tools/tool_choice/parallel_tool_calls as unsupported, so a direct request asking it for tool calling gets a clear 400 instead of silently failing (#12104, #12442) — thanks @tanveer-arch
  • fix(gemini): build the upstream URL from the payload-rule-rewritten body model and drop the deprecated thinkingBudget when a thinkingLevel is set (#12826) — custom-model aliases such as gemini-3.7-flash-high (rewritten by an override rule to gemini-3.7-flash) previously 404'd because the executor inserted the original alias into /models/{model}:generateContent while the body named the real id. BaseExecutor.buildUrlForBody() now derives the URL model from the prepared body (falling back to the executor model when the body carries none, so every other provider is unchanged), chatCore logs the rewrite and keys Gemini RPM/RPD accounting to the wire model, and a Gemini 3.x string thinkingLevel supplied by a request or payload rule replaces the deprecated numeric thinkingBudget instead of being sent alongside it. — thanks @adivekar-utexas
  • fix(responses): Preserve Gemini/Antigravity tool-name sanitizer aliases alongside namespace identities when extracting Responses API tool metadata, instead of dropping alias-only entries during collision recovery (#12839) — thanks @Pllutonyy
  • fix(gemini): strip additionalItems (the draft-07 spelling of the tuple-typed array keyword also fixed for prefixItems in #12540) from tool parameter schemas — it triggered the same hard upstream 400 (Unknown name "additionalItems"), breaking tool-enabled Claude Code requests to Gemini models whose schemas still used the older spelling (#12872) — thanks @sonirahul
  • fix(build): repair better-sqlite3 native prebuild resolution for standalone packaging — detect musl (Alpine) and Android/Termux runtimes and try a 4-candidate resolution order (bundled app build, bundled app prebuild, root build, root prebuild) instead of a single platform/arch match before falling back to npm rebuild (#12961) — thanks @raheemuddin786
  • fix(opencode): clamp oversized embedding string inputs to 20k chars before they reach the upstream body (behavioural fix — the clampEmbeddingStringInput helper existed but was never wired into buildUpstreamBody/CLOVA single-text dispatch, so over-limit batches still hit models that reject them), and preserve the GLM-5.3 vision fragment plus reasoning-presence sentinel for opencode/zen relay turns (#13004) — thanks @alvinveroy
  • resilience: credit-exhausted upstream errors (e.g. "You have exhausted all your credits") now classify as quota exhaustion even when the upstream answers with a non-429 status, so the provider breaker applies the long quota cooldown instead of a transient retry (#13041 — thanks @turbolego)
  • fix(sse): widen compression worker eligibility gate to accept structured-clone-safe undefined/Date/Map/Set/RegExp values, restoring worker offload for real requests (#13154) (#14391)
  • fix(auth): restore keyless local-first posture on GET /v1/models for pre-existing installs with no configured credentials (#13354) (#14415)
  • fix(dashboard): tokens-per-second now measures generation time (duration − TTFT) instead of wall-clock duration, and counting tokens_reasoning defensively when a provider excludes it from completion_tokens — thinking models (kimi k2.6/k2.7/k3) no longer look artificially slow in /dashboard/logs; TTFT is persisted per request and shown in the logs grid/detail and Playground TPS uses the same formula (#13373) — thanks @ahmedhosnypro
  • fix(providers): strip unsupported GLM thinking parameters and reject test-all requests when a provider has no active connections (#13487) — thanks @rafaeldrincon
  • fix(auth): scope credit-exhaustion lockout to the single failing model on passthrough providers (e.g. bai, *-compatible-* aggregators) instead of terminating the whole connection, so other models on the same key (including free ones) stay usable (#13548) — thanks @ggdayup
  • fix(resilience): Cap cached completed-request stream diagnostics at 64 chunks per stage (with a truncation marker) so long streams cannot bloat the completed-details bridge; byte budget and preview detachment come from the JON-562 cache (#13621). (#13867) — thanks @smshagor-dev
  • fix(sse): guard ChatAdmissionController's idle-queue TTL sweep so it never evicts a session key that still has a parked waiter, and keep #queueTimestamps in lockstep with the normal drain path instead of depending on the sweep alone (#13649) — thanks @khnker
  • fix(dashboard): respect iOS safe-area insets in standalone PWA mode so the header and mobile sidebar no longer slide under the status bar / home indicator (#13663) — thanks @isavage
  • fix(chat-admission): derive the heavyweight queue wait from requestQueue.maxWaitMs (RATE_LIMIT_MAX_WAIT_MS, 15 s) instead of a fixed 2 s — the single default heavyweight slot is held for the whole upstream turn, so a 2 s wait expired inside every occupancy it exists to bridge and a lone large-context session shed its own next request (reason:"queue_timeout", activeHeavy:1, waiting:0) with nothing to protect against; an explicit OMNIROUTE_CHAT_ADMISSION_QUEUE_MS still wins, 0 included, and .env.example no longer documents the variable twice with different defaults (#13648) (#13676 — thanks @pacocartones / @abhisheksharma2411)
  • fix(providers): devin-cli-agentic no longer forwards a bare <summary> envelope as the final answer (SWE-2 models emit it heavily) (#13691) (#14396)
  • fix(providers): route the OpenCode Muse Spark model family through the Responses API so newly released muse-spark-* variants do not require individual routing patches (#13725) — thanks @AStupidBear
  • fix(video): robustly extract video result URL from top-level fields, stringified JSON metadata, and arrays in Agnes and job-based video polling (#13726) (#13902) — thanks @HouMinXi
  • fix(sse): Use fresh Anthropic scoped-limit evidence to isolate an exhausted native Claude model while keeping global or unproven quota failures connection-wide (#13888) - thanks @riez
  • fix(api): log cache-hit search requests so cacheHitRate reflects real hits (#13928) (#14394)
  • fix(compression): ponytail and less-code output styles carry the safety carve-out the original port omitted, and the declared-but-unread OutputStyle.boundaries field is now honored (boundaries are appended after SHARED_BOUNDARIES, so no style loses the shared guarantee) (#13938) — thanks @woodsonl
  • fix(auth): non-chat callers now treat credential diagnostic sentinels (allExpired/allRateLimited/blockedByKeyPolicy/leaseConnectionMismatch) as unusable instead of falling through as real credentials (#13945) (#14414)
  • fix(sse): stop reordering demoted mid-array system messages in Claude->OpenAI strict-provider requests (#13948) (#14397)
  • fix(providers): use shell:true on win32 for zcode .cmd/.bat shim spawn (#13963) (#14422)
  • fix(dashboard): the OAuth token-expiry label in a provider's quota card now wraps onto multiple lines instead of being truncated (#13967) — thanks @trycohn
  • fix(db): phase-offset the model-sync interval (+45 min, period unchanged) from the cleanup 6h scheduler so they never fire in the same second, and add the missing conversation_turn_nodes.last_seen_at index used by the retention DELETE (#13973) (#14567)
  • fix(startup): the first model-catalog sync no longer fans out to every autoSync connection at once (112 concurrent catalog JSON parses were enough to OOM a 3 GiB V8 heap); the cycle now keeps at most 4 fetches in flight and waits 90s after boot so it does not overlap the 30s startup cleanup (#13975) (#14113) — thanks @HouMinXi
  • fix(mcp): read-only MCP tools now authenticate internal management API hops with the configured service token, including the advanced telemetry and cache helpers, while audit logging loads better-sqlite3 through the standalone-safe runtime loader instead of failing after webpack bundling (#13993) — thanks @dakaribeckerer-art
  • fix(catalog): resolve canonical provider aliases for combo prefix stripping and align max_input_tokens with context_length (#13994) — thanks @fouadSalkini
  • fix(auth): back the manage-scope authz-bypass policy with a globalThis-shared store so every standalone-build webpack chunk sees the same configured prefixes (#13995) (#14473)
  • fix(api): GET /api/logs/export now settles its HTTP response instead of hanging forever when the DB row source throws mid-stream — the row-iteration loop is wrapped in try/catch, the failure is logged, and the JSON document is closed out cleanly with additive emitted/error trailer fields so the client always gets a response (#13999). (#14439)
  • fix(guardrails): The Vision Bridge no longer whole-request reroutes an explicitly requested bare model id whose vision capability is unknown. kimi-for-coding has no MODEL_SPECS entry, so its capability resolved to null (unknown) and the guardrail treated unknown exactly like proven text-only, silently answering every image request with command-code/moonshotai/Kimi-K2.6 on a different provider connection. Unknown capability on a bare, non-combo id now keeps the requested model and describes the images instead, and a successful whole-request reroute is logged at warn with the original and substitute model ids. Provider-qualified ids, zero-vision combos, forced modes and auto/ are unchanged (#14003) (#14844) — thanks @aniruddhaadak80
  • fix(docs): name the combo hang-stop from comboPredicates.ts instead of citing COMBO_LOOP_SAFETY_TIMEOUT_MS as an env var (#14004) (#14054) — thanks @HouMinXi
  • fix(dashboard): un-gate the Dev Tools sidebar section (Playground, Translator, Search Tools) from Debug Mode so it is discoverable by default (#14021) (#14416)
  • fix(dashboard): Home degrades to default settings instead of returning 500 when the SQLite key_value table is corrupted; getSettings() keeps failing closed so auth gates still require login (#14060) (#14421)
  • fix(combo): a combo member that is missing from the live catalog is skipped as model_not_in_catalog instead of the generic availability bucket, and that reason is grouped in the ALL_TARGETS_SKIPPED diagnostics (#14069) — thanks @RaviTharuma
  • fix(mitm): resolve AgentBridge cert-trust check against the active cert model instead of hard-coded server.crt (#14070) (#14425)
  • fix(providers): declare official Claude thinking-effort tiers on relay/third-party registry entries (Bedrock, Kiro, AgentRouter, GitLab Duo, Anthropic-Messages relays) so suffixed combo ids fold using the same tier boundaries as first-party claude/anthropic (#14075) — thanks @HouMinXi
  • fix(dashboard): the Free Stack ($0) combo template inserted three Qoder model ids the registry no longer carries (if/kimi-k2-thinking, if/qwen3-coder-plus, if/deepseek-v3.2), so a combo created from it was born with dead targets and nothing failed until the user ran it — now if/kimi-k2.7-code, if/qwen3.7-plus and if/deepseek-v4-pro. Its description also still advertised Qwen, a provider in RUNTIME_RETIRED_PROVIDER_IDS that the preset never contained, and a Qoder count that was wrong; it now names what the preset actually routes across, in every locale. A new guard resolves every entry of both preset literals against the real registry (treating passthroughModels providers as always-resolving), which caught a second drift the report had only suspected: Paid Premium carried cu/claude-4.6-opus-high, and Cursor has no claude-4.6-opus-* at all — that generation is sonnet-only there — now cu/claude-opus-5-high (#14077) (#14150) — thanks @abhisheksharma2411
  • fix(providers): custom OpenAI-compatible node models keep their synced vision capability through the Vision Bridge guardrail (#14081) (#14428)
  • fix(gemini): normalize protobuf-flavored schema type spellings (dict, bool, int32, float, list, ...) to valid Gemini/Antigravity types and deduplicate repeated required array entries during tool-schema sanitization (#14083) — thanks @Ardem2025
  • fix(opencode-plugin): honor features.combos: false in the OpenCode config shim, and keep first-class cc/...-low catalog ids on openai-compatible /v1 instead of Anthropic Messages (#14088) — thanks @RaviTharuma
  • fix(guardrails): fail closed on unknown-root filesystem paths followed by ambiguous prose in sanitized error messages (#14110) (#14445)
  • test(images): add regression coverage for Antigravity quota-exhausted 429 credential rotation (#14112) — thanks @Ardem2025
  • fix(sse): stop leaking a foreign combo/pool account's Codex quota headers to the caller (#14116) (#14465)
  • fix(providers): Perplexity Web no longer strips array subscripts from code — citation cleanup now skips fenced code blocks, inline code spans and <tool> payloads, so print(arr[0], arr[12]) survives in both rendered answers and tool-call arguments (#14121) (#14122) — thanks @Zicocoder
  • startup: OMNIROUTE_DISABLE_BACKGROUND_SERVICES=1 now also skips the boot-time model-catalog warmup, so no OpenRouter catalog fetch runs when background services are disabled (#14076, #14147 — thanks @rafacpti23)
  • fix(opencode): requests the upstream refuses because of the tools attached to them, such as the short title-generation call, are now replayed once in the other shape and the shape that worked is remembered for that prompt (#14148) — thanks @maxmad64bis
  • fix(opencode): overlapping requests to OpenCode no longer share the target format and client session of whichever request started last; a slow JSON request could get the raw event stream back when another request finished first (#14149) — thanks @maxmad64bis
  • fix(auth): a request that fails upstream (for example with a 502) no longer makes the next request for the same no-auth model answer 401 No active credentials: the short model lock now surfaces like a real connection's cooldown, so the request waits it out when a retry wait is configured and otherwise gets a 429 with a Retry-After. A 401 or 403 from a no-auth provider, or a refusal that depends on the request itself, also stops starting a token refresh that can only fail, which cost several seconds per request and let repeated failures open the provider-wide refresh breaker for 30 minutes (#14151) — thanks @maxmad64bis
  • fix(api): add encrypted_function_args: [] plaintext marker to collaboration function_call items in the Responses translators (#14154) (#14447)
  • fix(opencode): append configured placeholder tools alongside client-supplied tools on free-tier requests and prioritize configured tools over un-scoped generic observations, resolving 403 FreeTierError when using external tool-calling clients (#14156) — thanks @adevwithpurpose
  • fix(guardrails): getBestVisionModel() no longer rescans the whole model catalog on every call when nothing is usable. Only a successful pick was cached, so an install that sends images with no vision-capable describer walked every provider in PROVIDER_MODELS and issued a credential lookup per candidate, on the request-serving event loop, for every image of every request. A null outcome is now remembered for 30 seconds, dropped as soon as a candidate becomes usable, and can be tuned or disabled with OMNIROUTE_VISION_BRIDGE_NEGATIVE_CACHE_MS (#14161) — thanks @cryptiklemur
  • fix(providers): The "SSRF Guard" toggle on the Feature Flags page now takes effect; before, only OUTBOUND_SSRF_GUARD_ENABLED in the environment was read, so switching the guard back on in the dashboard did not undo an env false. A dashboard OFF saved earlier now applies after upgrading (#14172) — thanks @datrixlab
  • fix(v1beta): Images, PDFs and audio sent as inlineData to /v1beta/models/{model}:generateContent, and parts sent next to a functionResponse (such as a file gemini-cli read), now reach the provider instead of being dropped (#14173) — thanks @datrixlab
  • fix(codex): Recover stale account cooldowns from fresh quota observations, add manual account cooldown release, and populate the Codex quota card from cached usage while removing its obsolete Spark tile (#14175). — thanks @JxnLexn
  • fix(combos): Editing auto-routing factor sliders now selects custom scoring, so an existing mode pack no longer silently discards reset-window and other manual weights. Preserve additional combo settings when editing routing controls. Auto reset affinity combines weekly (65%) and session (35%) urgency, while retaining explicit window selections. (#14176) — thanks @JxnLexn
  • fix(build): combos page resolves provider prefixes through the client-safe alias map instead of the server-only model service, fixing the production build (#14179) — thanks @maxmad64bis
  • fix(api): API-key usage, budget, token, request-rate, provider-account and combo-exhaustion errors now expose the earliest known reset as structured retry_after seconds and an ISO reset_at, with the matching Retry-After header, so clients can tell users exactly how long to wait instead of reporting an opaque unavailable error; non-Anthropic usage-limit responses now use 429, while Anthropic /v1/messages deliberately keeps 400 to avoid Claude Code misclassifying quota exhaustion as an authentication failure. (#14188) — thanks @fouadSalkini
  • fix(resilience): classify CLIProxyAPI model_cooldown 429 errors as quota_exhausted (#6342) (#14190) — thanks @sprintberlin
  • fix(sse): the opencode rotation head now reuses the last healthy member and skips recently refused members while the opt-in skip flag is on; with the flag off the order is exactly the plain rotation (#14202) — thanks @maxmad64bis
  • fix(db): the Bun SQLite adapter now returns undefined instead of null for a .get() with no matching row, matching better-sqlite3, node:sqlite and sql.js. Under Bun the exclusive-lease guard used to report every connection as leased, so connection tests returned LEASE_ACTIVE, usage refresh was deferred and model sync found no connections (#14203) — thanks @Ercaner1988
  • fix(sse): the OpenAI-to-Claude streaming translator now reads Responses-style usage names (input_tokens/output_tokens) in addition to prompt_tokens/completion_tokens, so upstreams reporting Responses-style usage no longer leave Claude clients with zeroed token counts; OpenAI naming wins when both are present (#14204) (#14205) — thanks @soroush5
  • fix(sse): parse deepseek-web double-pipe DSML invoke/parameter tool-call markup (#14208) (#14446)
  • fix(sse): retry empty translated streaming turns through the normal credential path (up to STREAM_RECOVERY.EMPTY_TURN_RETRY_MAX retries) instead of exposing an empty 200 or an empty-content 502, and a stream that drops before anything reaches the client takes the same retry path; a stream that answers and then stops producing without closing is replayed the same way once its stall budget runs out, instead of buffering forever; off by default behind FLUSH_EMPTY_RETRY_ENABLED (#14213) — thanks @maxmad64bis
  • fix(catalog): gate the Claude effort-variant catalog entries (-low/-medium/-high/-xhigh) behind OMNIROUTE_DISABLE_THINKING_LEVEL_VARIANTS, mirroring the existing synced-effort-variant gate — with the flag on, these synthetic model ids no longer leak into /v1/models for catalog-only clients. (#14214) — thanks @texastoland
  • fix(providers): retire gemini-business — Google retired the StreamGenerate endpoint it depended on and the replacement protocol cannot be validated without a live enterprise account (#14217) (#14467)
  • fix(proxy-logs): search proxy-log rows keep the HTTP status the provider actually returned (upstream_status, null when no response arrived), matching the chat writer (#14220) — thanks @maxmad64bis
  • fix(proxies): stop re-serving a pool member that just failed at the head of the pool: candidates are ordered by crossed short-memory health signals (opt-in, off by default), so a request lands on the first useful try instead of paying the failed head first (#14221) — thanks @maxmad64bis
  • fix(subscription): serve last known good nodes instead of emptying the pool on fetch or parse failure (#14222) — thanks @maxmad64bis
  • fix(sse): skipped-account diagnostics for multi-account rotation: accounts passed over while cooling down are now named (masked id + remaining time, one info line each per request) and their per-account state is readable from the resilience connections endpoint, so an operator can tell skipped accounts apart from unused ones; proxy log entries also carry the masked serving account and the request correlation id when the rotation attribution flag is on (off by default, no selection change) (#14223) — thanks @maxmad64bis
  • fix(providers): retire the suno music provider — studio-api.suno.ai is permanently suspended (503) and the live host requires a paid captcha-solving dependency OmniRoute declined to add (#14224) (#14468)
  • fix(rate-limit): keep per-connection rate limit overrides across restarts and JSON re-imports — startup reloads the overrides before building limiters and reconciles existing ones, and both re-import paths carry rate_limit_overrides_json in the INSERT OR REPLACE (preserving the stored value when the backup has none) (#14226 — thanks @maxmad64bis)
  • fix(providers): OpenCode Zen GPT-5.6 models (sol, terra, luna) route to the Responses API instead of 503ing on /chat/completions (#14230) — thanks @mhenke
  • fix(proxies): non-global proxy scopes now share one validation rule, so a missing scopeId is rejected the same way everywhere with the same messages as before. Behaviour change: blank scopeId values (whitespace only) on stored assignments are now rejected instead of accepted; clearing a non-global scope without scopeId is now rejected as well. (#14231) — thanks @maxmad64bis
  • fix(proxies): the health sweep now tells a refused destination apart from a dead proxy (a stalled handshake is observed separately from a frank connection failure, and a 451 or proven 403 carries its cause alongside the verdict without touching any proxy status), and an opt-in bounded pass re-tests a proxy the sweep saw refused before its set-aside expires — PROXY_HEALTH_RECOVERY_ENABLED=true (default off: no extra probe runs) (#14233) — thanks @maxmad64bis
  • fix(sse): a 403 without a quota signal now reports a neutral permission code while preserving the upstream message, so non-quota refusals no longer look like billing exhaustion. (#14234) — thanks @maxmad64bis
  • fix(health): the system health endpoint offers an opt-in deep check (?deep=1, off by default, authenticated callers only) that samples the completions surface with a minimal non-streaming request and caches the verdict briefly; only 502/503 answers raise the failover signal, every other failure keeps serving the existing payload unchanged. Wiring only: the probe target is not configured yet, so ?deep=1 stays inert until a dedicated settings change supplies it. (#14236) — thanks @maxmad64bis
  • fix(sse): the default early-stream keepalive threshold (2 000 ms) equalled the tightest client first-byte watchdog, so combo requests — which always take the default tier, since a combo arrives as a bare name with no provider prefix, and never resolve inside 2 s — had their keepalive byte land at 2 003 ms while the client aborted at 2 011–2 016 ms with zero bytes received (a 499 with nothing diagnosable in it). The default tier is now half the observed watchdog (1 000 ms) and the keepalive cadence is 1 500 ms; the deliberately long slow tier for browser-session and anonymous-fallback providers is unchanged (#14247) — thanks @ishan-parihar
  • fix(sse): normalizeOpenAICompatibleFinishReason() now maps the Claude stop_reason vocabulary onto its exact OpenAI equivalent (end_turn/stop_sequence/pause_turn→stop, tool_use→tool_calls, refusal→content_filter, model_context_window_exceeded→length). An Anthropic-style end_turn returned by an OpenAI-format combo leg used to reach the OpenAI wire raw, and a strict OpenAI client reads an unrecognized finish_reason as a provider fault — failing the whole turn with Provider finish_reason: end_turn after the text had streamed (#14247) — thanks @ishan-parihar
  • fix(api-keys): the combo 403 now names the combo and the two real remedies instead of only stating its refusal, since the previous wording read like a routing bug and made callers retry the same model or fall through a whole compaction cascade (#14247) — thanks @ishan-parihar
  • test(providers): pin the opencode 400 model-unavailable catalog rule at registry level (#14251) — thanks @maxmad64bis
  • fix(logging): keep the provider reply when a call-log artifact only overflows through its request body (#14253) — thanks @maxmad64bis
  • fix(resilience): treat Grok Build OAuth 403 content refusals as non-terminal request rejections (#14258) (#14259) — thanks @sprintberlin
  • fix(sse): the injected web_search fallback tool now uses the flat Anthropic shape ({ name, description, input_schema }) for Claude/Anthropic Messages targets, so strict Anthropic-compatible upstreams (e.g. a vLLM /v1/messages) no longer reject it with tools.0.name missing (#14272) — thanks @Nikita75699
  • fix(providers): register qwen-cloud-token-plan/qwen3.8-flash and deepseek-v4.1-flash as vision-capable so image requests through a combo can fall through to them instead of stopping on an unavailable vision-flagged leaf (#14273) — thanks @costajohnt
  • fix(antigravity): never persist or send the __REQUIRES_GCP_PROJECT__ manual-project sentinel as a projectId; a connection that already holds it now behaves as "no project" (typed 422, dropped by the selector when a healthy sibling exists) instead of failing every request with HTTP 400 and being preferred over healthy accounts (#14282) — thanks @bufftop25
  • fix(providers): retired cloud gemini-cli connections now fail closed before executor fallback or proxy overrides can route them through another backend (#14289) — thanks @peterbussch
  • fix(dashboard): fresh Docker installs no longer loop on the onboarding screen — a one-shot bootstrap token unblocks non-loopback local operators without treating the docker0 gateway as loopback (#14296) (#14472)
  • fix(providers): surface the real transport diagnosis (DNS/socket cause) instead of a bare "fetch failed" in provider validation errors (#14309) (#14463)
  • fix(sse): an OpenCode free-tier 403 (FreeTierError) is now request-scoped for combo routing — it no longer records model lockout or connection cooldown; the synthetic noauth path is paused for a short TTL after such a refusal so auto-combo stops re-picking it every turn; dashboard copy and FREE_TIERS.md document that the free tier can only be used from within OpenCode (#14313) (#14675) — thanks @Laksopan23
  • fix(usage): requests stuck past the pending age limit stay visible on the dashboard with a distinct marked state instead of disappearing, and still count toward pending totals (#14319) — thanks @maxmad64bis
  • fix(ci): raise the Docker :next/:latest publish heap budget to 12288 MB so the webpack production build no longer OOMs (#14325) (#14469)
  • fix(sse): give synthesized Responses API keepalive and failure frames a real sequence_number (#14330) (#14572)
  • fix(api): stamp anonymous /v1/files and /v1/batches writes with a synthetic anonymous owner so the same caller can read/delete what it wrote (#14332) (#14757)
  • Fix dashboard 'Remove completed' button to fully drain batches when the total exceeds the backend chunk limit. (#14347) — thanks @shubhayu-dev
  • fix(providers): KIE flux/kontext no longer returns a fake HTTP-200 success with zero images when the poll response uses the resultImageUrl(s) shape (#14335) (#14570)
  • fix(api): GET /v1/auto-combo/{channel}/candidates refuses a presented API key that failed validation instead of treating it as anonymous. Folding a revoked/expired/banned key to apiKeyId: null is fail-closed for owner-scoped records, but here null means "apply no per-key exclusions", so an invalid key was handed the instance-wide candidate pool (#14333, #14340). — thanks @sxh313
  • fix(sse): a client abort on a streamed chat completion now cancels the upstream request after headers arrive (start-timeout wrapper no longer drops the abort link, stream tracker cancels the upstream reader on disconnect, and the combo quality-check peek exits on abort), so local upstreams stop prefilling/decoding abandoned requests (#14342) — thanks @VIPKaiser
  • fix(dashboard): the version lookups can now read the payloads the real sources send. MAX_VERSION_RESPONSE_BYTES was 16 KiB while the registry per-version document is 29,609 bytes and GitHub's latest-release payload is 154,002, so both HTTP fallbacks cancelled the body mid-read, swallowed the error and returned nothing - /api/system/version answered "latest":"unavailable" forever, and installs without an npm binary had no working source at all (#14339) (#14345) — thanks @VIPKaiser / @sxh313
  • fix(models): the provider-detail model visibility controls work again — after #12172 scoped the dashboard's hide/unhide writes to modality: "chat", the flag is stored in hiddenModalities.chat instead of the legacy top-level isHidden, but the dashboard reader still inspected only the legacy flag. Chat-hidden models therefore kept rendering as visible: the eye toggle never flipped, "Hide all" left the {active}/{total} active count unchanged, the Hidden filter stayed empty and hidden models were still offered as Test-all targets, while /v1/models (which resolves visibility server-side) did drop them. The precedence rule now lives in one shared module (isHiddenForModality) consumed by both the DB layer and the dashboard, so the reader cannot drift from the writer again; legacy rows hidden before #12172 keep reading as hidden (#14349) — thanks @SCys
  • fix(opencode): Keep each request on its own member list so concurrent requests no longer replace each other's list (#14353) — thanks @maxmad64bis
  • fix(runtime): omniroute repair/runtime check now work on npm 11+, which hard-rejects the --allow-scripts=<pkg> CLI flag for project-scoped installs — the runtime's own package.json now declares allowScripts instead (retrofitted into existing installs automatically). Also fixed isBetterSqliteBinaryValid() only checking the build/Release/ layout and reporting valid: false for a genuinely working binary installed under prebuilds/<platform>-<arch>.node; it now shares the same platform/arch detection doctor already used, so the two commands can no longer disagree about the same install. npm install failures during repair also now surface the actual npm error instead of being silently swallowed (#14355) (#14541) — thanks @ogunlowoaamos / @geekyNads
  • fix(api,images): DELETE /v1/batches/delete-completed stops re-validating the API key the scope helper already gated and logs an honest audit reason; Antigravity image requests no longer force image_size: 1K when the caller omitted it; the image-upscale call log no longer crashes on sanitized error objects — findings of the omni-code-review battery (#14365)
  • fix(cost): orphan health-check repair no longer deletes the spend policy of the deployment-time environment key. repairOrphanDomainRows treated api_keys as the complete set of domain_budgets / domain_cost_history owners, but the env-key identity (OMNIROUTE_API_KEY / ROUTER_API_KEY) authenticates without ever being persisted there — every row it owned looked like a broken reference and was silently deleted on every startup and periodic pass. Since checkBudget() is fail-open when no budget row exists, the deletion lifted the spend ceiling instead of tightening it. Synthetic identities (@/shared/constants/apiKeyIdentities) are now excluded from the orphan count and delete; genuinely unowned rows are still removed (#14367). — thanks @valimwiliam2020-art
  • fix(opencode-plugin-v2): publish the gateway catalog through the stable provider contract instead of the removed beta one (#14370) — thanks @maxmad64bis
  • fix(providers): gemini-web carries the tool-call history into its next prompt. buildGeminiToolPrompt() sent only the system messages plus the last user turn, and because that adapter has no native message array the model never saw the call it had already answered or its result, so it re-issued the same tool call until the client's loop limit stopped it (#14368) (#14374) — thanks @formilw / @sxh313
  • fix(cli): Context exports omit credentials by default, resolve keychain secrets only with explicit opt-in, and atomically write private files without following destination symlinks. Provider management rejects ambiguous selectors, verifies mutation read-backs, keeps import data separate from management authentication, and redacts additional credential field spellings (#14379).
  • CLI configuration security: redact generated previews and existing profile credentials, reject credentials in preview URLs, validate requests, and use private atomic configuration/backup writes without weakening the container write guard. (#14381)
  • fix(opencode): preserve explicit native and Claude conversation IDs before request translation, keeping the existing canonical session shape and fingerprint fallback. (#14390)
  • fix(providers): resolve provider-prefixed upstream model ids from synced live catalogs (#14393) (#14765)
  • fix(cli): MCP-backed CLI commands now send the required Streamable HTTP headers and parse SSE responses, fixing HTTP 406 failures with streamable-http (#14435) (#14549) — thanks @Joaquinbrialva / @IbrahimKhan12
  • fix(combo): A request-level 400 (parameter/schema validation such as Extra inputs are not permitted, unknown field, malformed body or context overflow) no longer cools down the connection and rotates through every sibling account that would fail identically; account-scoped 400s (auth or rate-limit text) still cool down (#13757, #14438) — thanks @HouMinXi
  • fix(providers): no-auth providers that expose no remote model endpoint can finally sync their catalog — model-sync treats a local_catalog response without the intentional marker as a failed remote fetch and returns 502 before importing, so it never pinned a stale catalog over a real outage. But the no-auth response path never set that marker, and a no-auth provider whose registry entry declares no modelsUrl has the local catalog as its only possible discovery source. Those providers could therefore never persist a single model while /models kept answering 200 OK — a silent, permanent dead zone (observed on chipotle, cloudflare-playground, duckduckgo-web, felo-web and theoldllm: 58 models across 5 connections). The catalog is now marked intentional exactly when no modelsUrl exists; providers that declare one still report upstream, and a genuinely failed live fetch is still rejected as degraded. (#14448) — thanks @aldoeliacim
  • fix(call-logs): chat-path call_logs.id is a fresh UUID (not the dashboard traceId), so two attempts in one request cannot collide or overwrite each other's artifact; an explicit-id UNIQUE collision still retries with a generated id (#14451). Thanks @VIPKaiser for the production evidence (#14338): UNIQUE constraint failed: call_logs.id reports. (#14474) — thanks @HouMinXi
  • Release the heavyweight chat admission slot when a client disconnects without cancelling the SSE body. Previously activeHeavy only ever increased on abandoned streams and pinned at OMNIROUTE_CHAT_MAX_HEAVY_IN_FLIGHT, shedding every subsequent heavy request with 503 chat_admission_busy until the process was recreated. (#14456) (#14457) — thanks @xiaoyaner0201
  • fix(resilience): release the admission lease immediately when a client aborts while the chat handler is still pending, not only after it eventually settles (#14456) (#14566)
  • fix(mcp): The "MCP Enforce Scopes" toggle on the Feature Flags page now takes effect without a restart; before, OMNIROUTE_MCP_ENFORCE_SCOPES was read from the environment only, so the page could show scope enforcement as on while every MCP tool call ran unchecked. The catalog default now says false, which is what the gate has always done (#14459) — thanks @datrixlab
  • fix(translator): A Claude image sent as a URL (source: { type: "url" }) now reaches the provider: inside a tool_result on the OpenAI path, and everywhere on the direct Claude → Gemini path, where it is passed as fileData (#14460) — thanks @datrixlab
  • fix(search): Search connections that answer with "insufficient credits" / "out of credits" text (e.g. Exa's 400) now enter the quota cooldown instead of a short generic retry; the phrases are search-scoped so LLM chat fallback classification is unchanged (#14462) — thanks @HouMinXi
  • fix(opencode): free-tier requests refused on a small tool subset are retried once with previously accepted tool names appended (#14464) — thanks @maxmad64bis
  • fix(providers): Xiaomi MiMo Token Plan connections can select the China, Singapore, or Amsterdam cluster that issued the key, instead of always calling Singapore (#14476) (#14490) — thanks @sprintberlin
  • fix(models): search providers no longer leak their imported searchTypes (web, news, x) into /v1/models as chat models; the active synced-catalog projection excludes search provider IDs, and sync-models now skips them while cleaning previously imported search rows. (#14477) — thanks @wenzetan
  • fix(rate-limit): resolve zone-less reset timestamps from Z.AI and GLM 429 error bodies as Asia/Shanghai (+08:00) instead of naive UTC, avoiding an 8-hour artificial cooldown lockout (#14479) (#14542) — thanks @HouMinXi
  • fix(tests): make executor-web-cookie-sweep hermetic for tencent-aistudio-web, stop real upstream calls (#14480) (#14569)
  • fix(api): close /v1/files and /v1/batches authz residuals — manage-scope key demotion, session+key ownership bypass, and unscoped pagination cursor (#14481) (#14767)
  • fix(api): every /v1/files and /v1/batches handler now applies the caller's API-key policy (endpoint allowlist, schedule, usage cap, rate limit) like delete-completed and the other /v1 routes (#14481) (#14495)
  • fix(security): document accepted-risk extract-zip advisories GHSA-7pqw-9j4j-h8q3 / GHSA-jmr9-qjv8-65gv + add regression guard (#14482) (#14564)
  • fix(guardrails): flag authority/educational-framing jailbreak attempts in the injection guard (#14483) (#14563)
  • fix(cache): fold tools/tool_choice/response_format into the semantic-cache Layer-1 hash and cap caller-supplied TTL, closing a cross-request cache poisoning path (#14484) (#14565)
  • security(self-hosted): enforce enforceApiKeyPolicy (schedule/rate-limit/allowedModels/quota) on the self-hosted unified-entry divert — only when the divert is configured, so cloud requests still run the policy exactly once inside handleChat() — and compare the optional shared self-hosted API key with a constant-time comparison instead of !== (#14485) (#14571)
  • Security — POST /api/cli/connect (#14486): the well-known INITIAL_PASSWORD=CHANGEME placeholder can no longer be exchanged for an admin-scoped oma_ access token from off-loopback. #13679 added that gate to /api/auth/login, but this route verifies the same management password, mints a token on success and is not loopback-only — so a fresh install (where sync-env.mjs copies the placeholder from .env.example) handed admin to anyone who could reach the port. Loopback pairing and any rotated password are unaffected. (#14491)
  • fix(security): stop trusting a world-writable opencode pool-strain marker in /tmp (#14487) (#14568)
  • Build: open-sse/executors/auggie.ts compiles again — the spawn-options helper constrained its stdio generic to readonly string[], which matches no spawn() overload, and the broken overload cascaded into eight "possibly null" diagnostics on the child's streams. src/app/api/v1/combos/projectCombo.ts also imported three combo types from comboStructure.ts, which only uses them internally; they come from the combo type module. (#14511)
  • Responses passthrough (regression from #14252): the shared pre-executor strip removed _nativeCodexPassthrough and its siblings before the executor could read them, so every Responses-native Codex/xAI request was silently downgraded to the translated path and lost client fields (metadata) to the #2608 allowlist. Executor-consumed markers now survive to their reader and are still stripped at serialization, so #14252's leak fix is untouched. (#14511)
  • Compression preview (regression from #14529): the Settings-page "Effective pipeline" preview kept showing the configured lossy plan (e.g. rtk → caveman) while header-less requests had started running the safe session-dedup → lite pair, reopening #12063. The preview now goes through the same lossy-request policy as the runtime, so it shows what an ordinary request actually runs. (#14511)
  • CLIProxyAPI account health (regression from #14544): with only CLIPROXYAPI_MANAGEMENT_KEY set, the health probe targeted http://undefined:<port> and reported a correctly configured local CLIProxyAPI as unreachable (also the typecheck:core TS2322 on the release tip). It now falls back to the documented 127.0.0.1. (#14511)
  • fix(auth): parseStringList and parseAllowedModels keep string entries when the API-key row arrives already decoded as an array instead of JSON text, so a key's allowed models and allowed combos are no longer dropped before policy evaluation (#14500). parseAllowedConnections is covered by #14698. (#14704) — thanks @abhisheksharma2411
  • fix(deps): align onnxruntime-node to 1.30.0 (in lockstep with @huggingface/transformers 4.3.0) and set onnxruntime-node-install=skip in .npmrc, so a failed CUDA/binary download in the onnxruntime-node postinstall no longer makes npm silently drop the whole optional @huggingface/transformers chain and crash the transformers loader at startup (#14523) — thanks @soaoen
  • fix(providers): Grok Build model discovery reads reasoning tiers when /v1/models returns objects ({id, value, label, default}) instead of bare strings, and keeps xhigh when the payload names it. grok-4.7 is in the built-in seed with the same four tiers as grok-4.6. A request that names no effort still leaves with the build default (high) on grok-4.7, not only on grok-4.5 and grok-4.6. A payload that omits the tier list still falls back to low/medium/high, so grok-4.5 does not gain an xhigh it does not advertise. (#14525) — thanks @HouMinXi
  • fix(providers): wire xiaomi-mimo-token-plan into the usage dispatcher (#14543) (#14759)
  • fix(providers): expand IMAGE_PROVIDERS with missing OpenRouter image models so combo routing stops silently dropping them (#14545) (#14762)
  • fix(compression): the output-styles path honors the dashboard's Auto-Clarity Bypass toggle, so turning it off keeps the selected styles on security, irreversible-action, clarification, and order-sensitive turns (#14551) — thanks @woodsonl
  • fix(executors): auto-learn now recognizes Anthropic's "X is deprecated for this model" wording (previously only "Unsupported parameter: X"), so a retired sampling param on a model not covered by the static strip rules — e.g. temperature on claude-opus-5-5 — is stripped and retried instead of surfacing a 400 to the client (#14561) — thanks @patrykkopycinski
  • fix(providers): resolve untagged claude-* ids on GitHub/GHE Copilot to native targetFormat (#14575) (#14758)
  • fix(proxies): bulk proxy import now accepts only statuses from the shared registry, excluding dead (#14579). — thanks @maxmad64bis
  • test(sse): lock per-key ceiling 429s out of neighbor cooldown scope (#14581) — thanks @maxmad64bis
  • fix(providers): route grok-cli model ids that are not in the static seed (e.g. grok-4.7 before a model sync) through the Responses wire format. Grok Build only serves /v1/responses, and these ids previously fell back to a chat-completions body that failed with HTTP 400 (#14583) — thanks @VIPKaiser
  • fix(auto-combo): keep models whose catalog declares only non-chat surfaces (image, video, audio, embedding, rerank) out of every auto/* candidate pool. Tag x.ai grok-imagine-* during discovery so they hit the same filter. Fixes the long-running leak of grok-imagine-*/Wan2.2/suno/chirp-* into auto/coding and auto/smart (#14584) — thanks @VIPKaiser
  • fix(routing): preserve structured pre-content SSE error semantics so request-scoped invalid_request_error refusals remain sanitized 4xx outcomes, can fall back to a different model, and never poison account/model health or retry the same rejected request; transient server failures keep their existing retry and lockout behavior across priority, round-robin, and nested combo dispatch. (#14585) — thanks @fouadSalkini
  • fix(sse): the rotation wave no longer replays a proxy that already returned 429 within the same request (#14588) — thanks @maxmad64bis
  • fix(mitm): Respect downstream SSE backpressure and cap oversized inspector chunks to prevent avoidable memory growth (#14594). — thanks @Vaishnavi220506
  • fix(sse): Grok Build (grok-cli) no longer 422s on Codex CLI requests: Responses namespace tool groups are flattened into function tools on the same-format lane and their {namespace, name} identity is restored on returned function calls (#14596) — thanks @QuangBlue
  • fix(api): API keys with disableNonPublicModels now accept the synced reasoning-effort variants (<model>-<tier>) that /v1/models and the combo builder advertise, instead of silently skipping those combo targets (#14597) — thanks @QuangBlue
  • fix(sse): translate-mode streams from Responses-API upstreams (e.g. grok-cli) no longer collect reasoning, .done snapshots and tool-argument deltas into the assistant content used by call logs, semantic cache and memory extraction (#14598) — thanks @QuangBlue
  • fix(compression): make purify_history drop-count notice byte-stable to preserve prefix cache (#14600) (#14761)
  • fix(grok-cli): Grok Build no longer 400s on Codex CLI turns that carry the native web_search tool: the OpenAI-only external_web_access and search_context_size arguments are dropped before dispatch while the tool itself is kept (#14609) — thanks @QuangBlue
  • fix(grok-cli): Codex CLI turns after a tool call no longer fail on Grok Build with Could not decode the compaction blob: a content: null on replayed reasoning items is dropped before dispatch, leaving the encrypted blob untouched (#14615) — thanks @QuangBlue
  • fix(sse): the applied-proxy capture context is shared across duplicated proxyFetch module instances, so the proxy journal can tell a received provider status apart from a local failure no matter which server copy serves the dispatch. (#14620) — thanks @maxmad64bis
  • fix(rate-limit): a queued request no longer runs in the async context of the request that freed the slot; each scheduled job keeps the output, logging and attribution of the caller that queued it (#14621) — thanks @maxmad64bis
  • fix(sse): a transient parked stream no longer trips stream readiness — the guard honors the parked state instead of returning 502 before the replayed leg arrives (#14623) — thanks @maxmad64bis
  • fix(providers): reach the reactive reasoning_effort 400 clamp-and-retry chain from CommandCodeExecutor (#14629) (#14774)
  • fix(sse): /v1/responses error frames sent after the early keepalive committed HTTP 200 now carry status_code, error_type and retry_after_seconds, so clients can tell permanent from transient failures (#14632) — thanks @QuangBlue
  • fix(compression): session-dedup now counts repeated large memory blocks with literal scanning instead of compiling them as regular expressions, avoiding the Regular expression too large failure reported in #6467. (#14633) — thanks @Jehu
  • fix(codex): Codex 400s for a replayed reasoning item whose encrypted_content cannot be verified now reach clients as invalid_encrypted_content instead of bad_request, so they can resend without the reasoning item (#14636) — thanks @QuangBlue
  • fix(sse): the proxy journal keeps one row per request actually sent, so refused outlets stop looking like first-try successes — count requests with COUNT(DISTINCT correlation_id) (#14637) — thanks @maxmad64bis
  • fix(streaming): mid-stream continuation now resumes truncated Responses-API streams (text deltas replayed as an assistant input item behind the same setting); a short complete Responses turn ending in response.completed is never re-requested (#14639) — thanks @maxmad64bis
  • fix(capabilities): Custom Models "Vision capable" now applies to combo steps stored with the provider-alias prefix (cx/<id>), so image requests to those combos are no longer rejected with capability_mismatch or stripped by the Vision Bridge (#14642, closes #14641) — thanks @xiaoyaner0201
  • fix(grok-cli): Codex desktop requests no longer fail on Grok Build with invalid_client_tool_schema: root anyOf/oneOf tool schemas with $ref or nested-union branches are reshaped into a bare union of inline object branches (#14649) — thanks @QuangBlue
  • fix(grok-cli): Combo conversations can return to Grok Build after a turn served by another Responses provider: encrypted_content is dropped from replayed reasoning items without a Grok Build id instead of failing with Could not decrypt the provided encrypted_content (#14650) — thanks @QuangBlue
  • fix(codex): A combo turn that falls back from grok-cli to codex no longer fails with The encrypted content for item rs_… could not be verified: Grok Build reasoning items (rs_<uuid>, tco_…) are dropped from the replay before dispatch (#14653) — thanks @QuangBlue
  • fix(proxies): the 429 set-aside curve is now tunable via PROXY_QUOTA_429_BASE_MS (default 120000) and PROXY_QUOTA_429_MAX_MS (default 3600000), read once at startup and bounded to 1000..3600000 with a warning fallback — unset keeps the current doubling curve unchanged (#14655) — thanks @maxmad64bis
  • fix(auto-combo): STRICT zero-cost pool filtering now surfaces its exclusion diagnosis (excluded / no-hard-stop share) to the operator log without changing routing (#14656) — thanks @maxmad64bis
  • fix(proxies): a pool member sharing a recently refused egress address is served last for providers whose quota is bucketed by egress address — order only, nobody is ever excluded, and other providers keep the plain rotation (covers refused members only; recently served members are a follow-up). (#14657) — thanks @maxmad64bis
  • fix(sse): the opencode rotation may serve the next attempt from another pool member after a 429 from an egress-bucketed provider, instead of retrying the refused egress address (#14659) — thanks @maxmad64bis
  • fix(sse): optionally retry once when a stream stalls before usable output, behind a new opt-in flag that leaves the default path unchanged (#14669) — thanks @maxmad64bis
  • fix(sse): Chat→Responses requests no longer reach GitHub Copilot /responses with neither input nor previous_response_id, which it rejected with 400 One of "input" or "previous_response_id" or 'prompt' or 'conversation' must be provided. The translator now injects a placeholder user item when input would ship empty (system-only turns, empty messages, orphan-filtered tool results) and no continuity field is present, and the auto previous-response-id policy keeps previous_response_id when input is empty instead of stripping the only continuity field. An explicit strip mode still strips. (#14673) — thanks @Laksopan23
  • fix(sse): Judge stalled buffered turns by raw stream content and keep reading while reasoning is open (#14679) — thanks @maxmad64bis
  • fix(call-logs): record encrypted reasoning presence, duration and requested effort in call logs (#14680) — thanks @maxmad64bis
  • fix(sse): Report buffered-turn verdicts and content-watchdog stall evidence in logs (#14213) (#14681) — thanks @maxmad64bis
  • fix(resilience): only report cgroup_high pressure as critical on kernel throttle proof (#14687) — thanks @maxmad64bis
  • fix(proxies): skip recently failed proxies by default, opt out with PROXY_SKIP_RECENTLY_FAILED=false (#14688) — thanks @maxmad64bis
  • fix(sse): serve buffered turns at the first useful chunk instead of buffering the whole turn (#14691) — thanks @maxmad64bis
  • fix(command-code): route Responses-shaped bodies to /provider/v1/responses, where reasoning: {"effort":"none"} is honored. The chat endpoint's effort enum is low|medium|high|xhigh|max — no disable value — and silently drops a nested reasoning.effort, so reasoning-off was impossible for GPT-5.6 and DeepSeek models. Verified live 2026-09-24: /responses + effort none → reasoning_tokens: 0; /chat + reasoning:{effort:"none"} → 41 reasoning tokens. (#14692) — thanks @adivekar-utexas
  • fix(compression): The settings preview now runs the same downgrade and default-plan path a real request runs, so an active profile is not described as a pipeline the request will not use (#14700) — thanks @gonisulaimann
  • fix(routing): Sync tier classification serves the stored pricing snapshot instead of stale built-in prices (#14706) — thanks @maxmad64bis
  • fix(sse): Strict zero-cost tally separates proven-exhausted quota from unknown-state exclusions (#14707) — thanks @maxmad64bis
  • fix(sse): Persisted connection cooldown takes precedence over transient rate-limit signals (#14708) — thanks @maxmad64bis
  • fix(providers): synthesize finish_reason after claude-web holds a tool_use turn open (#14711) (#14770)
  • fix(routing): reasoning-routing rules that force max / ultra now keep GPT-6 Astra (and any other model in the Codex max/ultra alias sets) instead of dropping it from the target combo or rejecting it as a single target; cx/gpt-6-astra-max / -ultra are read as max / ultra requests, and the rules editor offers those tiers for it (#14720) — thanks @QuangBlue
  • fix(routing): bare gpt-6-astra ids are Codex-native like the gpt-5.6-* tiers: with Codex and OpenAI both active, bare gpt-6-astra now reaches the Codex subscription instead of OpenAI, and /v1/models lists the bare Astra ids (#14721) — thanks @QuangBlue
  • services: restrict resolvePortPid's lsof probe to TCP listeners only, preventing connected clients from being returned instead of the listening process (#14722). (#14838) — thanks @HouMinXi
  • fix(compression): the compression settings tab saves only the fields you change, so settings changed on another page or tab (such as Auto-Clarity on the caveman page) keep their values, and a failed save rolls the field back (#14723) — thanks @woodsonl
  • fix(sse): The opt-in empty-turn retry (FLUSH_EMPTY_RETRY_ENABLED) keeps the request's credentials on the connection that served the turn when a retry fails, prefers another connection inside the key's allowlist, and replays a leased or pinned connection instead of rotating it (#14729) — thanks @L4XB
  • fix(i18n): the docs translation pipeline now rejects model output that leaks <think> reasoning or English meta-prose, or that loses table rows or code fences. A rejected chunk is retried, and a doc that never validates fails instead of being written (#14731).
  • fix(grok-cli): Grok Build (grok-cli) no longer 422s on Codex requests that declare freeform custom tools (apply_patch, or lite mode's namespaced exec) or carry their tools in additional_tools input items: tools are merged and sent as function tools on the same-format lane, and returned calls come back as custom_tool_call items with streamed input deltas (#14738) — thanks @QuangBlue
  • fix(resilience): A 422 request error no longer cools down the provider connection: it goes through the same classifier as 400, so one rejected request body cannot push every other client off a single-account provider (#14739) — thanks @QuangBlue
  • fix(auth): Vision and audio bridges work again under REQUIRE_API_KEY without OMNIROUTE_API_KEY: the per-process self-loop bearer now authenticates (chat/audio only, shared between the proxy and route bundles), and the describe call no longer creates "CLI Auto-Key" DB keys (#14740) — thanks @QuangBlue
  • fix(providers): Grok Build grok-4.6 is declared vision capable, so images reach it directly instead of being replaced by a Vision Bridge description (#14741) — thanks @QuangBlue
  • fix(combo): bare combo targets without an explicit provider ID retain the provider inferred from their model; explicit provider IDs remain authoritative. (#14785) — thanks @lorenzozanee
  • quality: normalize symlinked worktree directories and report paths in new-code mode gates so macOS base violations match accurately (#14744). (#14849) — thanks @HouMinXi
  • fix(sse): forward the per-message effort betas alongside Claude Code's message-level output_config — without a gating beta Anthropic rejects the body field the passthrough kept (400 messages.N.output_config: Extra inputs are not permitted); the client-beta allowlist now carries Anthropic's documented mid-conversation-output-config-2026-07-01 (also derived from the body shape for clients with no anthropic-beta of their own) plus Claude Code's wire token per-turn-control-2026-07-01 (captured from [email protected]) (#14746, dup #14747) (#14828) — thanks @Laksopan23
  • fix(providers): NaraRouter model imports use the live model catalog while retaining the pinned catalog when discovery is unavailable. (#14826) — thanks @lorenzozanee
  • fix(sse): pool-served accounts without their own proxy use the egress key actually applied to the attempt for pacing and refusal memory, instead of the shared direct sentinel (#14750) — thanks @maxmad64bis
  • feat(sse): record per-request added wait with cause in call logs (#14756) — thanks @maxmad64bis
  • fix(compression): the cache TTL input on the compression settings tab offers 1 to 60 minutes, the range the settings route accepts, so its spinner no longer reaches values that fail to save (#14766) — thanks @woodsonl
  • fix(dashboard): the live dashboard WebSocket client now backs off 1s, 2s, 4s, 8s, 16s, then every 30s when the server rejects a session after the upgrade (UNAUTHORIZED, forbidden origin, server full), instead of re-dialing about twice per second forever. Only an accepted session (the server's welcome frame) resets the backoff, and events from a superseded socket no longer start a second reconnect loop. (#14768) — thanks @woodsonl
  • fix(sse): transport errors keep the underlying cause code next to the sanitized message (#14775) — thanks @maxmad64bis
  • fix(proxyHealth): a sweep marks a proxy failed when the same target answers through another proxy (#14776) — thanks @maxmad64bis
  • fix(providers): model and combo health checks now skip providers requiring web-session credentials instead of sending chat prompts (#14780). (#14818) — thanks @lorenzozanee
  • fix(i18n): the docs translation pipeline now writes .i18n-state.json atomically and refuses to start from an unreadable state, so parallel or interrupted runs no longer wipe it down to their own entries (#14781).
  • fix(api): Global npm installs are detected as npm again, so the dashboard's "Manual Update" works instead of showing "Not a git repository. Download source or use npm install -g.": the install-mode check read __dirname, which the Next.js server bundle freezes to "/ROOT/src/lib/system", and now falls back to the entry script (argv[1], or pm_exec_path under pm2) when it lives inside an installed omniroute package (#14784, closes #14782) — thanks @cfigueiroa
  • fix(sse): share hardened replay cancel with JSON-to-SSE path (#14790) — thanks @maxmad64bis
  • fix(sse): over-cap buffered reads return without awaiting the tee clone cancel (#14791) — thanks @maxmad64bis
  • fix(api): chat completions keepalive bytes merge into the request journal (#14792) — thanks @maxmad64bis
  • fix(sse): mid-stream recovery lines carry the requesting call id and stitched outcomes surface at info (#14793) — thanks @maxmad64bis
  • fix(db): proxy health stats separate transport health from upstream refusals, with connection tests counted apart (#14794) — thanks @maxmad64bis
  • fix(dashboard): proxyless account cards show the connection's effective egress instead of a misleading direct label (#14796) — thanks @maxmad64bis
  • fix(usage): remove finished pending request by id instead of oldest-first (#14797) — thanks @maxmad64bis
  • fix(api-keys): adding a token limit to an API key no longer replaces the key's existing limit for the same scope. A key now holds one limit per scope and reset interval, so daily, weekly and monthly limits coexist. Migration 196 rebuilds api_key_token_limits with that uniqueness while keeping every limit, token counter and reset log. POST /api/usage/token-limits updates by id when one is given and returns 409 if that would collide with another limit, or 404 for an unknown id (#14801). — thanks @fouadSalkini
  • fix(proxies): set aside egress after repeated transport failures with cross-egress success (#14802) — thanks @maxmad64bis
  • fix(proxies): local-core selector switches land on members without recent refusals, falling back to the least recently set-aside member (#14804) — thanks @maxmad64bis
  • fix(sse): a handler that never resolves no longer holds the early-keepalive slow path open forever: after a 33-minute last-resort deadline the wrapper emits the route error frame, closes the stream, aborts the handler, and warns once (#14808) — thanks @maxmad64bis
  • fix(sse): after a 429, a request makes one real call to an account held back by an earlier request (proxy set aside or account cooling down) instead of returning that 429 straight away (#14809) — thanks @maxmad64bis
  • Resilience (#14815): a request that fails while being translated (locally, before any upstream call — e.g. "Maximum call stack size exceeded" on one large image) no longer puts the account into cooldown or counts toward the provider circuit breaker. The same body fails identically on every account, so the error now goes straight back to the client instead of making a single-account install report "all accounts rate limited". (#14830)
  • fix(sse): a 5xx whose message merely mentions a quota (for example Cursor's empty-turn error) no longer parks the connection until the cached quota reset, which could be a billing-cycle end weeks away; it takes the normal capped cooldown, while real quota errors (4xx) keep parking until the reset (#14816) — thanks @QuangBlue
  • fix(providers): "Import from /models" on media-only providers (Soniox, ElevenLabs, Topaz, fal-ai, Adobe Firefly, …) now stores the models: the local catalog emits apiFormat values the provider-model schema accepts instead of audio / images, the import dialog reports rejected rows instead of a false success, and model sync no longer deletes the imported speech / transcription / image rows (its discovery is chat-only, so it never re-added them, and every sync cycle or restart wiped them). Soniox gains tts-rt-v2 (#14825) — thanks @QuangBlue
  • fix(sse): report the real error when a stream fails before its first SSE event instead of claiming a readiness timeout; the error class and routing stay unchanged (#14829) — thanks @QuangBlue
  • fix(mitm): a failed AgentBridge start now reverts the /etc/hosts entries it had already provisioned. Provisioning runs before the proxy is spawned, so a spawn failure threw past every teardown step and left each target hostname pointed at 127.0.0.1:<port> with nothing listening there — whatever else owns that port (Caddy/Vaultwarden on :443, for instance) then answered the SNI with a TLS internal_error alert, and the hostnames stayed broken machine-wide with no way back from the UI. The rollback is best-effort and never masks the original startup error (#14832) — thanks @adevwithpurpose
  • fix(responses): Codex CLI with the web_search fallback no longer fails with "stream closed before response.completed": a Responses-shaped fallback result is streamed as Responses SSE, and the client's own rejection of the internal search call is dropped from the next turn (#14837) — thanks @QuangBlue
  • fix(sse): the web_search / web_fetch fallback tool is no longer cut when a large tool list (for example Codex with several MCP servers) is capped to the provider limit, so the model can actually search (#14839) — thanks @QuangBlue
  • fix(compression): RTK no longer drops diagnostic lines at truncation because of three disagreeing severity vocabularies. The engine hard cap (index.ts::defaultPriorityPatterns), the filter priority set built from preserve.errorPatterns / preserve.summaryPatterns, and rawOutput.ts::isLikelyFailureOutput each carried its own word list, and the keep stage fed by rules.includePatterns truncated before any priority pattern was consulted — so a 2026-09-20 12:07:04 FATAL deployment aborted - rollback required line was classified as failure output worth retaining while the compressed body dropped it, and a bare TS2345 line was protected by the filter layer but not by the engine. They now derive from one shared vocabulary (engines/rtk/severityVocabulary.ts, 23 words), and diagnostic lines bypass the keep stage without populating a filter's EMPTY includePatterns (which would switch the stage on and newly drop every non-severity line). Measured through processRtkText at the live maxCharsPerResult on ~250-char lines: 0/55 builtin filters kept all eight severity words before, 55/55 after. Token impact is neutral — same output length, only which lines survive. Reported in #14648. (#14845) — thanks @yelsoft16
  • fix(providers): OpenAI GPT-6 requests map max_tokens to max_completion_tokens (#14870) — thanks @fidelix
  • fix(reasoning-routing): a force + none rule now emits an explicit reasoning_effort: "none" instead of only clearing reasoning params, so providers whose thinking defaults ON (e.g. DeepSeek V4 behind the native Responses API) actually stop thinking instead of failing later turns with The reasoning_text in the thinking mode must be passed back to the API; Command Code clamps the literal none to low, mirroring its existing minimal clamp (#14873) — thanks @TheRealAlexV
  • fix(usage): Non-streaming responses from every Claude-format provider now store the full prompt total, not only responses served under the claude/anthropic/vertex*/anthropic-compatible* ids (#14878). Notable affected providers: kimi-coding, deepseek, xiaomi-mimo, zai, github (Claude through /v1/messages), poe, agentrouter, bailian-coding-plan, and opencode-go/opencode-zen models that target the Claude format. Anthropic usage.input_tokens excludes prompt-cache reads and writes. Those bodies fell through to the Responses API branch of the usage extractor, which stored input_tokens alone next to the cache counters. The cost calculator then subtracted the cache from an input that never contained it, so fresh input tokens were billed at $0 and stored input did not match the streaming path. The extractor now recognises an Anthropic Messages body (type: "message") by its shape and stores input + cache read + cache creation, as the streaming path already did. Semantic-cache replays of OpenAI answers cached for Claude clients also recover their full prompt total. Providers that report no cache counters (e.g. devin-cli-agentic) keep them as N/A instead of 0. Responses API usage (whose input_tokens already includes the cache), OpenAI chat and Gemini usage are unchanged. — thanks @fouadSalkini
  • fix(db): databases created by the npm 3.8.50 package, which recorded migration 163 as model_capabilities, now run the release's 163_radar_feed_cache_generated_at on upgrade instead of skipping it, so radar_feed_cache.generated_at exists and the renumbering CRITICAL stops (#14881) — thanks @agusgonzaleznic
  • fix(proxies): refused-egress set-aside defaults to a 300s base with a 900s cap, and an egress that repeatedly waits out the response-headers window is set aside under a short 60s/600s curve after three settled overruns in five minutes (#14891) — thanks @maxmad64bis
  • fix(sse): recover from a reasoning_effort 4xx that names no accepted set — opt in per provider with OMNIROUTE_REASONING_EFFORT_PROBE_PROVIDERS and the reactive path probes one tier down, learning that single tier when the probe is answered so later requests clamp up front instead of 400ing on every attempt (#14895) — thanks @Moseyuh333
  • fix(auth): model: "auto" on a fresh install no longer skips every OpenCode Free target as availability (#14905). The auto combo pins keyless targets to the synthetic noauth connection and pin-fail-closed turns that into the allowlist ["noauth"], which the #9057 guards rejected. An allowlist that explicitly names noauth now resolves; keys restricted to real connections are still denied. — thanks @TrippyEngineer
  • fix(auth): back off an OAuth connection that keeps failing with 401 (still-valid or invalid-token class) instead of re-selecting it with cooldown 0 and refreshing the token on every request; the single-model cooldown retry no longer waits out an auth cooldown (#14917) — thanks @QuangBlue
  • fix(sse): record streamed ttft_ms as request start → first chunk with text, reasoning or a tool call; translated streams stored ttft = latency and others ~0 ms, and the Codex Responses WebSocket bridge timed turns from connection open (#14919) — thanks @QuangBlue
  • perf(db): cache prepared statements in the better-sqlite3 adapter (LRU 200, same as the node:sqlite adapter) instead of recompiling the call_logs/usage_history SQL on every request (#14920) — thanks @QuangBlue
  • fix(cli): the serve port preflight's bind probe now also tries 0.0.0.0, 127.0.0.1 and ::1, so on macOS it sees an instance already listening on the default host or on loopback instead of reporting the port free (#0000) (#14923) — thanks @peterbussch
  • fix(credential-health): the credential health sweep could probe connections every 5 minutes instead of the 60-minute default #12138 established. getConnIntervalMs kept a 300_000 default parameter, and docs/reference/ENVIRONMENT.md, open-sse/config/constants.ts and the scheduler header comment all still advertised 5 minutes. The per-connection fallback now reads the same constant as the global default, so the two cannot drift apart again. Operators seeing frequent probe traffic should check healthCheckInterval on the connection and the Resilience settings interval — 0 disables probing entirely. (#14924) — thanks @adivekar-utexas
  • fix(command-code): stop rewriting reasoning_effort xhigh to max. Command Code's own validator advertises low|medium|high|xhigh|max, so xhigh is a native and distinct tier there, and the provider-wide rule silently spent a different reasoning level than the caller asked for. It also overwrote the nested reasoning.effort carrier that the Responses path reads. The model-level ceilings are untouched: GLM 5.1+, DeepSeek V4+ and Kimi K3+ still map xhigh to max, as do the Ollama Cloud and OpenCode Go gateways, and OpenRouter still passes xhigh through unchanged. (#14929) — thanks @adivekar-utexas
  • fix(providers): Free-model detection reads each provider's own /models verdict through one shared predicate: input/output pricing and free tags count (Vercel AI Gateway, xKiro), explicit isFree/free: false vetoes a zero token price (Kilo, LLM Gateway), and a zero token price no longer marks models billed per request, image, second or minute as free (#14930) — thanks @aldoeliacim
  • fix(combo): a failed Last Known Good Provider write on the combo success path now warns with the combo, execution key, and provider instead of a bare error, so a missing pin can be attributed instead of silently leaving the retry on a stale assumption. (#14934) — thanks @maxmad64bis
  • fix(dashboard): the NoAuth egress label map covered only 4 of the 5 EffectiveEgressKind values, so own reached getEgressLabelKey and surfaced as a TS2345 dashboard typecheck regression. The map is now typed Record<EffectiveEgressKind, string>, so it is exhaustive by construction and a future kind cannot drift past it again, and it gains an ownProxy label. Also registers 7 covering unit tests in stryker.conf.json tap.testFiles so their mutant kills count, clearing the mutation-test-coverage --strict drift that was failing Fast Quality Gates. (#14939) — thanks @adivekar-utexas
  • fix(opencode): with the opt-in Responses stall guard on (OPENCODE_RESPONSES_STALL_ROTATION), a RESPONSES_FIRST_BYTE_TIMEOUT_MS configured above STREAM_READINESS_TIMEOUT_MS is now capped at the readiness bound, so a silent streamed Responses reply rotates once then fails fast instead of waiting out the larger window; windows at or below the bound, readiness disabled (0), and the flag off are unchanged. (#14940) — thanks @maxmad64bis
  • fix(providers): a passing or skipped connection test no longer re-enables a connection an operator switched off on purpose; toggling a connection off (single or bulk) records the intent in providerSpecificData.operatorDisabledAt, and only an explicit re-enable clears it (#14941) — thanks @shipsfromrio
  • fix(combo): a chaos combo (auto/chaos) now answers a /v1/chat/completions request that omits stream or sends stream: false with a JSON chat.completion (and a JSON 502 when the whole panel fails) instead of a text/event-stream body; the progressive SSE panel broadcast is kept for stream: true (#14942) — thanks @shipsfromrio
  • fix(opencode-plugin): force sync (and the auto-sync tick) now fetches /v1/models before touching any cache, so a transient failure such as the 10s abort keeps the last good in-memory catalog and disk snapshot instead of deleting them (#14943) — thanks @shipsfromrio
  • fix(dashboard): the Logs page Export button now requests the server maximum (50k rows) instead of the silent 10k default and warns when the export was still truncated; GET /api/logs/export mirrors the cap metadata in X-OmniRoute-Export-* response headers (#14946) — thanks @shipsfromrio
  • fix(combo): A combo whose every target is behind an OPEN circuit breaker now answers all_targets_cooling_down naming each provider/model: circuit_open (Ns) with Retry-After and the wait recovery action, instead of the generic ALL_TARGETS_SKIPPED 503 whose hint pointed at quota and top-ups (#14949) — thanks @shipsfromrio
  • fix(providers): the Add / Edit compatible-provider modals now show the validation reason (e.g. prefix: "openai" is a reserved provider prefix …) instead of the generic "Invalid request", and the user guide documents the custom OpenAI-compatible flow and the reserved-prefix rule (#14950) — thanks @shipsfromrio
  • fix(settings): Storage backup, export, restore and import surface the API error message instead of crashing the settings page ("Failed to load settings") or showing "Export failed: [object Object]" when the route answers with the structured { error: { code, message } } envelope (#14846) (#14951) — thanks @shipsfromrio
  • fix(docs): the self-host guide now downloads docker-compose.selfhost.yml and .env.selfhost.example from the v3.8.51 tag, says the flow starts at v3.8.51 (v3.8.50 users go to the docker run flow) and that the Docker flow ignores ~/.omniroute/.env; auto/chaos is documented as the parallel panel fan-out it is instead of fault injection; the combo editor gets the missing pipeline label and pipelineDesc description in every locale (#14953) — thanks @shipsfromrio
  • fix(translator): the OpenAI → Gemini request translator no longer injects a default thinkingBudget that is larger than the client's max_tokens / max_completion_tokens. Gemini counts thinking against maxOutputTokens, so a request with max_tokens: 50 and no reasoning knob went upstream with thinkingBudget: 24576 and the model could spend the whole cap thinking, returning an empty content. The injected (not client-requested) budget is now capped at half of the output cap, but not below the minimum Google documents for the 2.5 Pro (128) and 2.5 Flash-Lite (512) tiers; explicit reasoning_effort and thinking.budget_tokens budgets are unchanged. (#14954) — thanks @shipsfromrio
  • fix(sse): a non-streaming 502 built after the upstream already answered HTTP 200 with a body that translated into no usable output (upstream_empty_response, upstream_response_failed, upstream_fake_success, all typed upstream_response_error) is no longer replayed on the same account by the single-model pre-output transport retry. That retry exists for connection resets before any response; here the upstream call was already made, so the replay only paid for the same verdict twice. Genuine pre-output 502/503/504/507 transport failures keep their one same-account retry. (#14955) — thanks @shipsfromrio
  • fix(sse): Mark responses served by the budget-exhaustion emergency fallback with X-OmniRoute-Emergency-Fallback: from=<provider/model>; to=<provider/model> and log the served reroute, so clients can detect the provider swap directly instead of diffing X-OmniRoute-Provider against the request. (#14956) — thanks @shipsfromrio
  • fix(rate-limit): a request that gives up on the local rate-limit queue (queue-wait budget or client abort) now leaves the Bottleneck queue instead of staying there as dead work. Before, the abandoned job was still dispatched later and spent a reservoir token and a minTime slot before rejecting, so with a per-connection rateLimitOverrides.rpm live requests queued behind dead ones, timed out, and added more dead jobs: the connection's queue never recovered while traffic continued. Jobs already past the queue (waiting out minTime) are unchanged. (#14957) — thanks @shipsfromrio
  • fix(search): A successful web search now clears the connection's recorded error (stale failed test, elapsed 429 cooldown), so the dashboard stops showing a serving search connection as failed. (#14958) — thanks @shipsfromrio
  • fix(executors): a 429 whose retry hint (Retry-After header or Gemini google.rpc.RetryInfo) exceeds the intra-URL retry window is handed back after one upstream call instead of being retried twice on the same throttled account, cutting the 429 volume of a rate-limited account pool by up to 3x before rotation (#14959) — thanks @shipsfromrio
  • fix(claude): make the cc_version= build revision overridable. CLAUDE_CODE_CLIENT_VERSION could already be overridden to clear Anthropic's model gate, but the 3-character revision suffix stayed pinned, so overriding the version alone advertised a version.revision pair no real Claude Code binary emits (e.g. 2.1.283.1e2). CLAUDE_CODE_CLIENT_BUILD_REVISION now overrides it, and every wire consumer reads it live alongside the version. Documented in .env.example and docs/reference/ENVIRONMENT.md. (#14969) — thanks @adivekar-utexas
  • fix(providers): Soniox speech sends the language and voice fields its API requires (defaults en / Adrian; OpenAI stock voices fall back to the default) instead of failing every request with 400 (#14974) — thanks @QuangBlue
  • fix(auth): API keys with "Disable Non-Public Models" accept a published model addressed by provider alias (sx/tts-rt-v2) instead of answering 403; the alias stays as strict as the provider that serves it (provider-node prefixes, hidden flags and key-group denies on the canonical provider) (#14975) — thanks @QuangBlue
  • fix(compression): keep session-dedup from rewriting the current turn, so an agent that re-reads a file gets the file back instead of a [dedup:ref …] marker. (#14978) — thanks @QuangBlue
  • fix(sse): Include GPT-6 Astra, Sol, and Luna in the built-in Codex Fast/Flex model selection while preserving explicit request tiers and custom model lists (#14986) — thanks @xiaoyaner0201
  • fix(sse): Count persisted-cooldown bypasses served through the rate-limited allow-list so operators can see re-served targets (#15004) — thanks @maxmad64bis
  • fix(dashboard): Aggregate provider usage by connection family and remove the duplicate xai dashboard entry (#15005) — thanks @maxmad64bis
  • refactor(sse): isolate the slow-stream deadline from the framework Request lifecycle and propagate cancellation explicitly through streaming admission/provider cleanup (#15010) — thanks @mdigitalbh81
  • fix(i18n): zh-CN and zh-TW now use the canonical 提供者 for "provider" in two recent strings, clearing the glossary-consistency test (#15016).
  • fix(cursor): ranged reads bridged from Cursor's native read tool keep their offset/limit instead of reaching the client as whole-file reads, which Claude Code answered with "file unchanged" and left the model re-reading the same file (#15036). — thanks @QuangBlue
  • fix(compression): Lite no longer truncates the results of the latest tool calls, so OpenAI-format agents (Cursor IDE, Kilo/OpenCode) stop re-reading a file whose content was cut to 2000 characters every time (#15037). — thanks @QuangBlue
  • fix(cursor): turns no longer fail with 502 "cursor-agent frame decode failed (flag=2, size=2)" — the stream driver now handles the Connect end-of-stream frame (JSON, not protobuf), ending the turn cleanly or reporting its Connect error code (#15074). — thanks @QuangBlue
  • fix(build): the npm package now ships betterSqlitePrebuildTarget.mjs and peer-stamp.mjs, so installing a freshly built tarball no longer fails in postinstall with ERR_MODULE_NOT_FOUND, and a unit test fails when a shipped script imports a file the package leaves out (#15077) — thanks @maxmad64bis
  • fix(compression): the compression worker pool no longer leaks request bodies when worker spawn fails synchronously (Turbopack standalone builds): every queued job fails open immediately and the broken pool short-circuits later runs, the fail-open is now logged (rate-limited), and the LLMLingua worker entry is passed as a URL object instead of a file:// string (Node ≥ 21 rejects the latter) (#12755). — thanks @insoln
  • Add modelsUrl to Cheaper Inference provider registry for live discovery (#14347) — thanks @shubhayu-dev
  • fix(antigravity): send the finite JSON upload as a replayable fixed request body even when the upstream response is streamed, avoiding the one-shot ReadableStream/duplex upload path and keeping retries replay-safe under large Codex/Responses payloads. (#15033) — thanks @mdigitalbh81
  • fix(authz): the API_PORT bridge reports each remote client to the dashboard by its own address and drops the forwarding headers it sent, so a remote client is no longer treated as the host itself; a private-LAN proxy can keep its X-Forwarded-For with OMNIROUTE_TRUST_PROXY=private (#15047) — thanks @HouMinXi
  • fix(api): honor allowedCombos during combo candidate pre-dispatch checks (#12886 follow-up). When an API key had both allowedCombos and allowedModels restricted, comboTargetPassesKeyModelPolicy verified inner targets only against direct model restrictions, causing all targets of an explicitly authorized combo to be skipped with 503/429. Explicit non-auto combo permissions now override the key's direct allowedModels and blockedModels rules for inner targets, while direct model requests and auto/* routers retain their restrictions. Restricted /v1/models catalogs now also re-check synthesized effort, no-thinking, and synced-effort IDs against the API-key allow-list. (#14197) — thanks @fouadSalkini
  • Security (GHSA-7j4q-6gx6-pg77): an API key restricted to specific combos (allowedCombos) could still dispatch the built-in virtual routes auto/* and qtSd/* — they are not persisted combo rows, so the allow-list lookup treated them as "not a combo" and let them through, reaching every connected provider (including OAuth subscription accounts). Virtual routes are now matched against the combo allow-list like any combo: denied unless the key allows all combos (combo/*) or lists the route; allowAutoCombos: false still applies on top. Reported and fixed by @aldoeliacim. (#15072)
  • fix(sse): forward Claude Code's afk-mode-2026-01-31 auto-mode beta on the Claude OAuth path alongside dangerous-tool-use-2026-09-03, so gateway sessions stay eligible for server-side auto-mode classification; a stream regression test pins message_delta.delta.safeguard_results and tool_use ids passing through unchanged (#14694) — thanks @fouadSalkini
  • fix(executors): break circular dependency between BaseExecutor and usage/DB layers by inlining body parse in base.ts (#14824) — thanks @SCys
  • sse: stop every upstream failure from throwing out of handleChatCore — persistFailureUsage read providerResponse from outside its scope (regression from #14544), so all 13 failure paths raised ReferenceError: providerResponse is not defined instead of returning the upstream error and recording failure usage. (#14599)
  • fix(providers): stop treating Anthropic max_tokens as the context window during model discovery — that field is the output cap, so Claude Opus 5 was advertised as 128K instead of 1M. Discovery now reads max_input_tokens for the window and max_tokens/max_output_tokens for the output limit. (#14260) — thanks @HouMinXi
  • fix: Claude model discovery now follows Anthropic's has_more/last_id pagination and preserves published input/output token limits. Empty, malformed, failed or incomplete responses retain the previous catalog with an explicit fallback warning instead of replacing it with partial data. (#14827 — thanks @JxnLexn)
  • fix(sse): stop an intermittent empty Claude stream from emptying a whole combo — forward the auto mode safeguards field only when its paired dangerous-tool-use-2026-09-03 beta travels with it (Anthropic rejects the field alone with a 400), exempt OmniRoute's own synthesized empty-response 502 from the per-model lockout that was pulling healthy accounts out of rotation for 60s, and surface the real upstream streaming error instead of masking every failure as "Claude returned an empty response (no content block)" (#14314) — thanks @fouadSalkini
  • fix(api): POST /api/cloud/auth returns maskedApiKey and projectId only to keys with the manage / admin scope (both fields are omitted for other keys), lists only the connections a key is allowed to use, and shows at most a quarter of a short key (#15048) — thanks @HouMinXi
  • fix(providers): the codex-app-server registry factory now wires the app-server executor with websocketFn: getCodexAppServerWebsocketTransport(), the same transport the internal codex fallback path uses, so the executor no longer runs without a WebSocket transport when reached through the registry (#14277). Reasoning effort resolution also accepts reasoning_effort on the request body, and an explicit model reasoning suffix (e.g. gpt-5.5-high) now takes precedence over client-injected body defaults. (#14281) — thanks @gonisulaimann
  • fix(api): /v1/combos (and the VS Code import surfaces) now compute combo capabilities from the combo's actual resolved targets — combo-ref steps expand through resolveNestedComboTargets, the same resolver the routing runtime and /v1/models use, so a nested combo whose leaves are all vision-capable advertises capabilities.multimodal: true instead of a hardcoded false that contradicted /v1/models for the same combo (#14232). Unresolvable refs (dangling names, cycles) and callers that don't pass the collection keep the previous conservative false. (#14271) — thanks @gonisulaimann
  • fix(combos): align combo health probes with small streaming model tests and shared response parsing. Keep provider-aware test deadlines active through response consumption, show errors inline, and clarify that target health checks do not exercise the combo routing strategy. Existing runtime rate limits remain unchanged. (#14278) — thanks @JxnLexn
  • fix(auth): the environment passthrough API key and the model-sync internal token are compared in constant time, like the other secrets (#15050) — thanks @HouMinXi
  • fix(cursor): the Cursor agent transport now honours the configured proxy. It talks raw HTTP/2 through http2.connect(), which opens its own socket and therefore bypassed both the global fetch proxy patch and the undici dispatcher — a provider proxy set in the dashboard (and HTTPS_PROXY) was silently ignored and Cursor traffic egressed on the host IP. A CONNECT tunnel (TLS-wrapped for https:// proxies) or a SOCKS4/4a/5/5h tunnel is now opened first and handed to http2 via createConnection; proxy selection reuses resolveProxyForRequest(), so precedence, NO_PROXY and local-address bypasses match every other egress path. SOCKS setup now honours aborts, and abort/stream close releases the established tunnel. (#14735) — thanks @MikeTuev
  • fix(cursor): model discovery now resolves the connection's configured proxy before calling Cursor's AvailableModels, rather than sending the import request directly. (#14735) — thanks @MikeTuev
  • fix(cursor): an agent harness no longer loops on the same step. When Cursor asked to run one of its built-in tools (Read / Shell / Write), OmniRoute bridged it to the declared client tool but told Cursor the exec had been rejected — so the model believed its own tool never ran and retried the same step on the next turn (reported as an agent re-reading a missing file forever). Those execs are now held open and answered with a typed success carrying the client's actual output (ReadResult.success, ShellResult.success, WriteResult.success), the way Cursor's own CLI does. Grep, Ls, Write and Fetch built-ins also gained bridges, and the decoder now keeps their arguments (GrepArgs.pattern/path/glob, WriteArgs.file_text) instead of dropping them. (#14737) — thanks @MikeTuev
  • fix(cursor): complete the Cursor CLI exec replies and bridge compatible PI tools; preserve parallel tool results and return errors for unsupported requests. Forward model effort, cache usage and Responses cache-write metrics, expose selectable Grok effort variants, and allow longer bounded readiness for long agent histories. (#14737) — thanks @MikeTuev
  • fix(cursor): tool calling through the cursor provider works again. Any request whose model invoked a tool hung until CURSOR_STREAM_TIMEOUT_MS and returned 502 cursor-agent stream timed out; when it did not hang, the model reported the OmniRoute tool namespace as unavailable and fell back to Cursor's built-ins, so harnesses such as opencode stalled. The agent protocol is now aligned with the agent.v1 schema embedded in the Cursor Agent CLI: tools go on RequestContext.tools (field 7, not rules on field 2) and are advertised as an McpDescriptor through mcp_meta_tool_options, the blocking mcp_state_exec_args (field 36) handshake is answered, span_context (field 19) is no longer mistaken for the event variant, Responses-API (flat) tool definitions are read, and a Cursor built-in shell exec is bridged to the client's shell tool instead of failing closed on the timeouts Cursor stamps on every exec. (#14737) — thanks @MikeTuev
  • fix(auth): a password change now ends every dashboard session issued before it (the browser that changed it keeps a fresh one), and signing out revokes that session on the server instead of only clearing the cookie (#15063) — thanks @HouMinXi
  • Security — data directory permissions (GHSA-2pg2-xm9r-8544): the CLI created ~/.omniroute (or DATA_DIR) and wrote .env — which holds STORAGE_ENCRYPTION_KEY, the key to every stored credential — without an explicit mode, so under the common umask 002 they landed as 0775/0664 and any local user could read the key and the database. First runs now create the directory 0700 and write .env 0600 (also when migrating Electron's server.env), and every run that touches encrypted storage repairs existing installs: world access is removed from the directory and .env/server.env become 0600. Group bits on the directory are kept so deliberate group shares keep working; informational commands (--version, --help) never touch the directory. (#14712)
  • fix(docker): start the Bun standalone runtime via process.execPath in run-standalone.mjs instead of a literal "node", since a Bun image has no binary named node on PATH. (#12525) — thanks @rqzbeh
  • fix(scripts): the env/docs contract gate now actually runs its CLI entry when the checkout path contains characters that file URLs percent-encode (spaces, non-ASCII). The raw file://${argv[1]} comparison never matched import.meta.url in those checkouts, so npm run check:env-doc-sync exited 0 without scanning anything; the entry guard now encodes the path with pathToFileURL, matching the other gate scripts. (#14155) — thanks @gonisulaimann
  • Security (GHSA-mh4f-3xj9-4gc4): server.env — which holds the generated JWT_SECRET, STORAGE_ENCRYPTION_KEY and API_KEY_SECRET — was written by scripts/build/bootstrap-env.mjs (dev/start, Docker) and by the desktop app without an explicit mode, so under the usual umask it came out 0644 in a 0755 directory. Both writers now create the data dir 0700 and server.env 0600, and repair a file an earlier version left world-readable on every start (the file is only rewritten when a secret is missing). Follow-up to GHSA-2pg2-xm9r-8544; the package-root .env no longer receives generated secrets since #11436. Reported by @peterbussch. (#14994)
  • fix: Unit tests disable Adobe Firefly's durable browser refresh, preventing mocked image-generation and session tests from opening the user's Chrome profile. The same guard applies to individually run Firefly tests. (#14876 — thanks @JxnLexn)
  • Security (GHSA-9p9m-h9rj-rhhg, critical): pre-request middleware hooks ran in a vm context built from host objects, so hook code could reach the server realm — directly (this.constructor.constructor("return process")() returned the host process) and by polluting the server's Object.prototype, which a later worker_threads Worker inherits as env (NODE_OPTIONS=--require …). Hooks now run in an isolated realm created from a null-prototype object: the request context goes in and the result, mutated context and log lines come back only as JSON strings, and the host never awaits or calls anything from the realm. The hook contract is unchanged except that the non-JS globals URL / URLSearchParams are no longer provided inside hooks. Reported by @alienkeric. (#14916)
  • fix(authz): the IP allow/deny list judges the address of the connection unless a loopback, private-network or Cloudflare proxy fronts the request, and behind one it reads the client from what that proxy added rather than from what the client sent; a proxy on any other public address is named in the new OMNIROUTE_TRUSTED_PROXIES (#15038) — thanks @HouMinXi
  • fix(providers): advertise the Kimi Code CLI version 2.1.1. The pinned constant was still 0.26.0, so the X-Msh-Version and User-Agent headers OmniRoute sends to api.kimi.com trailed the real CLI by two major versions. KIMI_CLI_VERSION still overrides the pin per deployment. (#14842) — thanks @HouMinXi
  • fix(quota): Moonshot connections pointed at api.kimi.com/coding now fetch Kimi Coding /usages instead of falling back to Open Platform users/me/balance, and membership usages.limit_* used_ratio windows are mapped onto Code 7d/5h for dashboard and reset-aware scoring. (#14201) — thanks @HouMinXi
  • fix(livews): the fresh-install anonymous window on the dashboard WebSocket now also requires the upgrade's real TCP peer to be local. Forwarding headers (x-forwarded-for / x-real-ip) mark a reverse-proxy or tunnel hop and fail closed, matching isLoopbackRequest() on the HTTP routes, so a remote client reaching a loopback-bound pre-setup instance through a tunnel can no longer read the dashboard before setup completes (#14256). (#14275) — thanks @gonisulaimann
  • fix(network): raise the direct response-start timeout floor to 300s for local/self-hosted upstream targets (Ollama via host.docker.internal, LAN IPs) so VS Code Copilot and other agent clients no longer 504 at the 30s default; add docker-compose.ollama-local.yml and docs/guides/LOCAL_OLLAMA_DOCKER.md (#13988) — thanks @appscluster
  • fix(ci): validate-release-green parses ci.yml with the already-declared js-yaml dependency instead of the undeclared yaml package, so a main-green nightly crash no longer posts an empty verdict on the tracker. (#13955) — thanks @HouMinXi
  • fix(mcp): the memory and skill MCP tools now act as the authenticated caller instead of an apiKeyId written into the tool arguments, so an MCP client can no longer read or change another tenant's memories or enable and run another tenant's or the global skills (#15068) — thanks @HouMinXi
  • fix(media): image and video combo strategies now fan out to all targets concurrently and return the first healthy success, so a slow first target (AI Horde queue exhaustion, provider that only fails after its full timeout) no longer blocks every sibling behind it (#13852) — thanks @opensource-elearning
  • fix(budget): a spent dollar budget no longer disables flat-rate capacity. The budget is scoped by apiKeyId and was enforced at the api-key policy phase, before any provider was resolved — so once the allowance was gone, a subscription/coding-plan provider that costs nothing per call was rejected alongside the metered APIs, and a mixed combo could not fall back to it. The chat path now defers the metered budget to the candidate boundary (enforceApiKeyPolicy(..., { meteredBudget: "defer-to-candidate" })) and re-applies it per resolved provider, before a credential is selected: a candidate that draws on the allowance is refused, one that does not stays eligible and is chosen by the normal ranking. Nothing else moves — the budget can only remove a candidate, never nominate one. Unauthorized metered spend is still prevented (an unknown provider counts as metered, and a request with no eligible non-metered candidate fails closed without dispatching), the refusal is tagged BUDGET_EXCEEDED so the combo loop never cools a healthy account for it, and flat-rate traffic no longer draws the allowance down while remaining fully visible in cost analytics. Every other endpoint keeps enforcing the budget at the policy phase, unchanged. (#14348) — thanks @valimwiliam2020-art
  • fix(build): OMNIROUTE_BUILD_PROFILE=minimal builds again — the ninerouter and mitm/cert/install stubs now export every name real routes import (getInstalledVersion, getLatestVersion, update, uninstall, installCertResult), and the webpack stub replacement plugin resolves stub paths from the project root instead of the importing file's directory (#12806, #12807). (#12810) — thanks @alpha-code2019
  • fix(sse): a model name can no longer carry percent-encoded dot-segments, ?, # or a backslash into the upstream URL, which let a client key reach other endpoints or models of a provider such as Gemini or Vertex and slip past a per-key model allow-list pattern (#15067) — thanks @HouMinXi
  • Security — Obsidian integration (GHSA-474q-g63r-w4rr): the operator-supplied baseUrl now goes through the outbound URL guard — cloud-metadata / link-local targets (169.254.0.0/16, IMDS hostnames, IPv4-mapped spellings) are rejected at POST /api/settings/obsidian before any fetch or persistence, every Obsidian REST call uses safeOutboundFetch with block-metadata, and redirects are never followed. Loopback, LAN and Tailscale bases keep working. (#14322)
  • fix(authz): a reverse proxy on the same host that sets X-Forwarded-Proto, X-Forwarded-Host, Forwarded or Via but no address header is no longer taken for the host itself, and the nginx example in docs/security/CORS.md now sets the address headers too (#15060) — thanks @HouMinXi
  • fix(dashboard): Render model usage over time with explicit zero values for inactive models so stacked chart series remain continuous and accurate (#14225) (#14228) — thanks @yugui923
  • fix(network): the outbound URL guard recognises internal addresses written as trailing-dot names (localhost., metadata.google.internal.) and as IPv6 embeddings of an IPv4 address (NAT64, 6to4, IPv4-compatible), plus site-local and multicast IPv6, multicast IPv4, 192.0.0.0/24 and the Azure and Oracle metadata addresses (#15039) — thanks @HouMinXi
  • fix(proxy): the reachability fast-fail probe now uses port 80 for http proxies that omit one (#14157). The WHATWG URL parser drops the port when it equals the scheme default, so http://user:pass@host:80 reached defaultPortForScheme with an empty port and got probed on 8080 — the probe failed, the request was rejected with Proxy Fast-Fail, and the healthy connection went into cooldown. https (443) and socks5 (1080) defaults were already correct; only the http case was wrong. Proxies with explicit non-default ports are unaffected. (#14184) — thanks @gonisulaimann
  • Security (GHSA-jmq6-8j86-8xqj): the provider connection test (POST /api/providers/{id}/test, and the same test run by test-batch and by POST /api/providers for a new connection) could spawn sh -c 'command -v …' on the host to check the local cline / qoder CLI, although those routes are reachable remotely. The local CLI probe now runs only for loopback / private-LAN callers (the LOCAL_ONLY semantics) and for the server's own credential-health scheduler; remote callers get the upstream test without the local runtime diagnosis. Reported by @zer0d4y5. (#14995)
  • fix(sse): stop applying the outbound bound-and-replay egress policy to loopback self-requests — a slow internal model-sync/auto-discovery call is no longer cut at 30s and replayed for another 30s, so internal requests stop occupying inbound slots long enough to starve live traffic past Cloudflare's 120s proxy read timeout (HTTP 524) (#14311) — thanks @fouadSalkini
  • fix(shared): quota-combo sync now unions custom and synced-upstream models with the static provider registry, and quota-share (qtSd/) models resolve through the existing WS→HTTP 426 fallback instead of failing to resolve (#13156) — thanks @rafazafar
  • fix(sse): Adobe Firefly and UC persona reference images are downloaded with the public-only, DNS-pinned, size-capped fetch, so an image URL on a loopback, private-network or metadata address is refused (the reason is logged, not returned); downloads through the pinned fetch no longer stall on bodies larger than about 64 KB (#15045) — thanks @HouMinXi
  • fix(combo): OmniRoute's own local model-cooldown 429 no longer counts as provider quota exhaustion, so a plain transient 429 on one target keeps the remaining same-provider combo targets eligible again; #14190's CLIProxyAPI model_cooldown classification had started matching our own identical cooldown body (#1731) (#15116)
  • fix(db): importing the proxy-log DB module no longer hydrates proxy logs from SQLite at import time, so modules that only reach db/proxies (e.g. the compression pipeline) stop opening the database as a side effect (#15116)
  • fix(dashboard): stop treating /dashboard/resilience/connections as local-only so a reverse proxy no longer clears the session and redirects to login (#14478) — thanks @HouMinXi
  • fix(responses): a Chat Completions provider that stopped on the token limit or a content filter now surfaces as response.incomplete with incomplete_details (max_output_tokens / content_filter), matching the real Responses API contract, instead of silently reporting status: "completed" (#14806). A truncated, degenerating generation was previously indistinguishable from a genuine successful completion. Fixed in both Chat→Responses stream translators (translator/response/openai-responses.ts and transformer/responsesTransformer.ts); upstream mid-stream errors still take priority. — thanks @hartmark
  • fix(network): safeOutboundFetch with guard: "public-only" now resolves the host and refuses a name that points at a private, loopback or link-local address, and the HTTP skill connects to the address that was checked (#15064) — thanks @HouMinXi
  • fix(mcp): Restore MCP server startup from the published bundle — the quota healthy override from #14555 made quotaPreflight import quotaCache back through usage.ts, closing an ESM init cycle that deadlocked the esbuild bundle ("unsettled top-level await"); the override state now lives in a dependency-free quotaCacheState leaf, with the #14555 park/override behavior unchanged. Refs #14547. (#14820)
  • fix(mcp): move sanitizeTimingMs to the zero-import leaf open-sse/utils/timingMs.ts so src/lib/db/proxyLogs.ts no longer pulls upstreamStatusCapture → providerRequestLogging → usage/migrations (top-level await) into every DB module; the per-attempt timing import from #14892 deadlocked the esbuild MCP bundle again (#14820)
  • fix(sse): the upstream start-timeout wrapper now keeps its client-abort link only when the settled result still has a streaming response body, so the post-headers abort propagation from #14342 is preserved while attempts that resolve without a live body no longer leave an abort listener on the client signal (restores the #12406 no-leak guarantee) (#14820)
  • Security (GHSA-cqwr-7mqw-chr9): the MITM / Agent Bridge sudo helper wrote the operator's sudo password and the command's data to the same stdin. When sudo did not prompt (a NOPASSWD rule, or a credential still cached from an earlier sudo), sudo -S tee -a /etc/hosts appended the sudo password to the world-readable /etc/hosts. The password now goes only to sudo -S -v (validate, runs no command), and the elevated command runs as sudo -n <cmd> with only its own data on stdin; hosts that never cache credentials (timestamp_timeout=0) fall back to a single sudo -S call, where sudo is certain to consume the password line. Reported by @adevwithpurpose. (#14836)
  • fix(memory): assemble the optional @huggingface/transformers specifier at runtime with webpackIgnore, so Next does not fail /health when the optional package is absent. (#13865) — thanks @HouMinXi
  • fix(api): /v1beta Gemini requests with generationConfig.maxOutputTokens now return properly shaped Gemini replies (#14165). The ingress converts the body to OpenAI chat format before re-entering handleChat while the URL keeps its /v1beta path, and with no path branch detectFormat's max_tokens heuristic misread the converted body as claude — non-streaming callers got an anthropic-shaped body the route's OpenAI→Gemini converter silently dropped, and streaming callers got a 200 SSE response with zero bytes, looping the antigravity CLI and the Google GenAI SDK. detectFormatFromEndpoint now treats a /v1beta path as OpenAI chat unless the body still carries the raw Gemini contents envelope. Raw Gemini bodies (client-raw-request contexts) are unaffected. (#14185) — thanks @gonisulaimann
  • Electron release: electron/package-lock.json regained the optional electron-builder-squirrel-windows subtree (13 entries) that npm ci had been refusing as out of sync — the Linux desktop leg died on it — and electron-release.yml gained a build_ref dispatch input so a release whose tag was cut with the broken lock can have its assets rebuilt from the repaired line (#11982)
  • npm publish workflow: the CycloneDX SBOM is attached to the GitHub Release on workflow_dispatch publishes too (when a release for the tag exists), not only on the release event — v3.8.50 shipped through a staged dispatch and its release carried no SBOM until it was attached by hand from the run's sbom-npm artifact (#12020)
  • fix(authz): the CLIProxyAPI install, start, restart and stop routes under /api/version-manager/ are loopback-only like /api/services/cliproxy/*, and the install route now validates tool and version (#15041) — thanks @HouMinXi
  • fix(quota): keep informational Vertex spend and unknown or unlimited quota windows from falsely exhausting provider connections. Preserve unknown-window metadata across snapshot reloads and safely ignore legacy Vertex spend lockouts, while continuing to enforce reported quota exhaustion and upstream 429 cooldowns. (#14276) — thanks @JxnLexn
  • fix(sse): the Grok Web and web-tools tool-markup scanners no longer take quadratic or worse time on a long run of whitespace or unclosed tags, so one request cannot stall the server (#15066) — thanks @HouMinXi
  • Security (Hard Rule #12): the Responses WebSocket proxy (dist/responses-ws-proxy.mjs, shipped with the standalone server) sent raw exception text to the client in two places — the 500 body of a failed upgrade and the response.failed frame when the upstream connection failed. The latter could carry the configured upstream proxy URL including its credentials. Both now return a fixed message; the detail stays in the server log / request history. Flagged by CodeQL (js/stack-trace-exposure, js/xss-through-exception). (#15073)
  • fix(ws): the /v1/ws and Responses WebSocket relays report each remote client to the app by its own address and drop the forwarding headers it sent, so a remote WebSocket client is no longer treated as the host itself; a Cloudflare edge, or a private-network proxy with OMNIROUTE_TRUST_PROXY=private, keeps its X-Forwarded-For (#15040) — thanks @HouMinXi
  • fix(release): let the Electron workflow start again — grant actions:read to the npm leg (#11973)
  • fix(ci): stop a stalled Codecov upload from cancelling the Coverage job and the main run (main twin of #11972) (#11978)
  • fix(release): drop the build_ref input — a dispatch builds the ref it is dispatched on (#12032)
  • fix(ci): accept CVE-2025-68121 in the prebuilt tls-client .so, auto-close base-red issues, guard Scorecard on the default branch (main twin) (#12086)
  • fix(cli): setup-claude prints an absolute CLAUDE_CONFIG_DIR hint (#11525) (#12161) — thanks @ljluestc
  • fix(ci): port the release-green ESLint gate fix to main (base-red #12363) (#12618)
  • fix(sse): accept complete tool calls for minimum response content (#13047) — thanks @Sheshiyer
  • fix(sse): hoist leading text system messages on the Claude mid-conversation-system passthrough (#13072) — thanks @JasonBroderick
  • fix(kiro): strip orphaned tool_use with no matching tool_result (#13170) (#13174) — thanks @nahdd123uosackr
  • fix(dashboard): show full text on hover when feature-flag description is clamped (#13189) — thanks @retroamx
  • fix(kiro): fall back from branded gateway to CodeWhisperer on profileArn-required 400 (#13193) — thanks @nahdd123uosackr
  • fix(guardrails): resolve prefixed combo names in Vision Bridge (#13332) — thanks @AStupidBear
  • fix(mitm): capture all passthroughs to Traffic Inspector, not just chat-matched (#13920) — thanks @VIPKaiser
  • fix(deepseek-web): parse DeepSeek V4 native DSML markup and bare parameter tool calls (#13930) — thanks @EkoSaputro14
  • fix(usage): count Claude-format thinking blocks as observed reasoning (#13965) (#13969) — thanks @Iammilansoni
  • fix(reasoning): resolve 400 errors and capability mismatch for gpt-5.6-luna (#14059) — thanks @kang-heewon
  • fix(openai-responses): trim tool_call_id and function name so padded call ids stay paired (#14093) — thanks @pacocartones
  • fix(sse): allow a stream to fail from INITIALIZED (setup-time failure) (#14094) — thanks @pacocartones
  • fix(sse): guard STREAM_HISTORY_MAX parse so completed-stream history stays bounded (#14095) — thanks @pacocartones
  • fix(api): sanitize /v1/batches create error body (Hard Rule #12) (#14115) — thanks @xiechimon
  • fix(api): drop stale catalog content-length when re-serializing model responses (#14092) (#14120) — thanks @Mistertelecom
  • fix(translator): lift Responses tool-output images into a multimodal user message (#14123) — thanks @Mistertelecom
  • fix(providers): strip temperature/top_p for OpenAI gpt-5.x reasoning models (#14133) — thanks @zhiru
  • fix(combo): exempt per-model-quota 403 from connection exhaustion (#14136) (#14140) — thanks @rafacpti23
  • fix(combos): use provider node prefix alias in combo builder qualifiedModel (#14143) — thanks @rafacpti23
  • fix(auth): support OIDC IdP issuer with trailing slash (#14119) (#14144) — thanks @rafacpti23
  • fix(combos): avoid server-only model import in dashboard (#14207) — thanks @VardoKen
  • fix(sse): retry direct stream readiness timeouts (#14209) — thanks @lorenzozanee
  • fix(sse): clarify compressed context token accounting (#14210) — thanks @lorenzozanee
  • fix(quota): expire stale Antigravity 429 exhaustion (#14211) — thanks @lorenzozanee
  • fix(auggie): resolve spawn EINVAL when running the CLI shim on Windows (#14215) — thanks @andreyzagid-tech
  • fix(providers): separate Codex GPT-5.6 image catalog ids (#14216) — thanks @domenicomassafra
  • fix: accept ChatGPT cookie headers for clean-room sessions (#14219) — thanks @lorenzozanee
  • fix(providers): stop rewriting GA qwen3.8-max to preview on opencode-go (#14181) (#14242) — thanks @xiechimon
  • fix(sse): treat antigravity empty completions with a normal stop as valid 200s (#14160) (#14243) — thanks @xiechimon
  • fix(db): avoid live WAL truncate after incremental reclaim (#14244) — thanks @Bl0ck154
  • fix(egress): strip _omniroute* markers at shared pre-executor boundary (#14252) — thanks @alfred-rootson
  • fix(gemini): strip tilde-prefixed Standard Schema keys from tool parameters (#14279) — thanks @hartmark
  • fix(cleanup): prune compression_engine_breakdown on retention schedule and usage resets (#14268) (#14285) — thanks @gonisulaimann
  • fix(providers): stop condemning refresh-less Claude setup-tokens as expired/no_refresh_token (#14287) — thanks @aldoeliacim
  • fix(skills): expand nested string-shorthand property types in injected skill tool schemas (#14288) (#14291) — thanks @xiechimon
  • fix(cli): route the copilot device flow through the github backend key and the real poll route (#14300) — thanks @xiechimon
  • fix(cli): probe every bind address before respawning so a wildcard listener is not read as free (#14307) — thanks @smkzw
  • fix(sse): dynamically inject skills beta only when code_execution tool is present (#14200) (#14310) — thanks @HouMinXi
  • fix(dispatcher): stale keep-alive burst on local egress (#14315) — thanks @VIPKaiser
  • fix(deepseek): default to OpenAI Chat Completions instead of Responses API (#14316) — thanks @SCys
  • fix(dashboard): retain token usage for in-memory request rows (#14324) — thanks @woodsonl
  • fix(ci): drain the release/v3.8.51 base-reds — env-doc, generated skill, TS2677, ESLint (Refs #13866) (#14331)
  • fix(resilience): model-scoped provider errors must not invalidate the credential (#14336) — thanks @valimwiliam2020-art
  • fix(deps): bump next 16.3.3 -> 16.3.5 (two post-RCE patch releases) (#14344) — thanks @aref-alapour
  • fix(cli): run codex through the server callback flow and correct the kiro social routes (#14298) (#14350) — thanks @xiechimon
  • fix(catalog): preserve model cache on runtime state updates (#14354) — thanks @Petrusreno
  • fix(ci): preserve failed npm install exit after retries (#14369)
  • security(video): close response-cache transcript retention gaps (#14373)
  • fix(hooks): fail closed without shared worktree stashes (#14389)
  • fix(test): boot ecosystem through trusted-peer server (#14403)
  • fix(test): align Vitest collectors and lifecycle data loading (#14406)
  • fix: preserve explicit tier policy during async pricing resolution (#14411)
  • fix: repair rotted import paths and a drifted test fixture (#14412) — thanks @gonisulaimann
  • fix: keep Qdrant embedding options limited to configured providers (#14427)
  • perf(release): resolve fragment origins with one git history walk in release:reconcile (#14440)
  • fix(routing): bridge intent classifier vocabulary to task-fitness table keys (#14452) — thanks @alexdrgpy06
  • fix(types): preserve narrowed intelligent routing weight patch (#14503)
  • fix(video): omit transcript echoes from active request details (#14517)
  • fix(test-gate): reject missing golden snapshots without explicit generation (#14521)
  • security(video): keep capture-hook exceptions out of request logs (#14522)
  • security(video): omit echoed cues from retained provider failures (#14527)
  • fix(cli): bind-probe the serve preflight when port discovery tools are unavailable (#14538) — thanks @xiechimon
  • fix(domain): bound quota park windows and add healthy override (#14359) (#14555) — thanks @texastoland
  • fix(codex-ws): resolve the Responses WebSocket upstream proxy through the full per-connection cascade (#14556) — thanks @xiechimon
  • fix(combos): purge combo-scoped proxy assignments when a combo is deleted (#14553) (#14562) — thanks @xiechimon
  • fix(resilience): sample recovery in background driver when selfRestart is disabled (#14576) — thanks @rapha4lx
  • fix(integrity): remove dead helpers and repair stale references in core services (#14580) (#14586) — thanks @gonisulaimann
  • fix(providers): keep reasoning tiers when a codex catalog sends objects (#14593) — thanks @HouMinXi
  • fix(capabilities): resolve the MiMo v2.6 family as vision-capable (#14587) (#14595) — thanks @xiechimon
  • fix(cli): stop kills port listeners only; isolate stop test from real processes (#14605) — thanks @peridot-augustus
  • fix(nix): pin dev shell to nodejs_24 and add python3 (#14613) — thanks @marcothedood
  • fix(sse): match the Codex 0.156.1 inference headers (#14625) — thanks @HouMinXi
  • fix(sse): advertise the grok 1.0.41 client version (#14626) — thanks @HouMinXi
  • fix(sse): advertise the claude code 2.1.280 client version (#14627) — thanks @HouMinXi
  • fix(providers): wait for the Gemini Web composer instead of sleeping 3s (#14643) — thanks @RaviTharuma
  • fix(resilience): keep own-listener hops out of the public pressure shed (#14644) — thanks @RaviTharuma
  • fix(resilience): let operators ignore host-wide PSI pressure (#14645) — thanks @RaviTharuma
  • fix(providers): reuse one Chromium process for Gemini Web (#14646) — thanks @RaviTharuma
  • fix(combo): cap the SSE quality peek so a long stream cannot fill the heap (#14647) — thanks @RaviTharuma
  • fix(providers): keep gpt-6-astra effort suffixes on the Responses API (#14652) — thanks @RaviTharuma
  • fix(sse): match the Copilot CLI 1.0.88 inference headers (#14664) — thanks @HouMinXi
  • fix(sse): keep Bedrock streamed tool-call arguments (#14670) — thanks @milunmoghe
  • fix(dashboard): let the web-search route model be cleared again (#14674) — thanks @MikeTuev
  • fix(providers): declare thinking efforts for first-party Claude registries (#14732) — thanks @HouMinXi
  • fix(providers): correct first-party registry display-name typos (#14734) — thanks @HouMinXi
  • fix(sse): preserve namespaced Gemini custom tool round trips (#14751) — thanks @JxnLexn
  • fix(cli): prevent serve crash when port is free on macOS (#14812) — thanks @prabhtheone
  • fix(sse): stringify object-form streaming tool arguments (#14813) — thanks @lorenzozanee
  • fix(translator): normalize non-object tool arguments for Claude (#14814) — thanks @lorenzozanee
  • fix(chatgpt-web): capture why the first-party bridge module failed to load (#14819) — thanks @abhisheksharma2411
  • fix(providers): drop the unvalidated IMAGE_PROVIDERS.gemini entry from #14545 (#14861)
  • fix(sse): restore the reselectedProxy declaration dropped by the #14659 merge (#14865)
  • fix(quality): clear the base-reds left by the 2026-09-25 merge-batch wave (#14867)
  • fix(sse): build deadline wrap field-by-field so it survives Next's request proxy (#14886) — thanks @HouMinXi
  • fix(sse): treat Responses incomplete and cancelled as terminal states (#14894) — thanks @HouMinXi
  • fix(sse): accept a chat completion truncated at finish_reason length (#14899) — thanks @HouMinXi
  • fix(providers): declare muse-spark-1.2/1.3-contributor effort tiers on opencode-go (#14900) — thanks @Ch3fUlrich
  • fix(codex): retire gpt-5.3-codex-spark and codex-auto-review (#14903) — thanks @Neuron-Mr-White
  • fix(handlers): roll back credentials when an empty-turn retry is rejected (#14914) — thanks @HouMinXi
  • fix(mitm): wait for drain in the standalone proxy SSE loop (#14921) — thanks @HouMinXi
  • fix(api): prepare the auto-combo candidate pool once in GET /api/combos/auto (#14925) — thanks @developerjillur
  • fix(sse): log the synthesized 429 when quota parking skips a provider (#14928) — thanks @HouMinXi
  • fix(dashboard): expose extended reasoning target efforts (#14964) — thanks @lorenzozanee
  • fix(transforms): apply per-provider system transforms for non-Claude providers (#14965) — thanks @AlejandroAkbal
  • fix(sse): log a recovered Antigravity fetch failure as a warning (#14979) — thanks @HouMinXi
  • fix(compression): skip fuzzy dedup when the caller cannot retrieve CCR (#14983) — thanks @HouMinXi
  • fix(sse): map a truncated Claude completion to finish_reason length (#14987) — thanks @HouMinXi
  • fix(db): stop hot-path cooldown writes from busting the /v1/models catalog cache (#14998) — thanks @afonsoft
  • fix(providers): point the regolo entry at its /v1 endpoints (#14996) (#15000) — thanks @afonsoft
  • fix(resilience): escalate per-kind breaker cooldown on repeated open cycles (#14960) (#15001) — thanks @afonsoft
  • fix(auth): bound the per-connection refresh mutex and rotation lane (#14970) (#15002) — thanks @afonsoft
  • fix(build): ship compression worker externals in standalone builds (#15011) — thanks @Jehu
  • fix(pwa): drop a stale service worker left behind by a deploy (#15017) — thanks @HouMinXi
  • fix(providers): derive new-connection default name from live names (closes #15006) (#15018) — thanks @toor11
  • fix(combo): trust finish_reason stop over the reasoning-ratio heuristic (#15020) — thanks @HouMinXi
  • fix(oauth): only accept the Trae loopback callback for a login the dashboard started (#15042) — thanks @HouMinXi
  • fix(auth): decide default-password logins from the socket peer, not forwarding headers (#15043) — thanks @HouMinXi
  • fix(oauth): require management auth on the OAuth login and connection routes (#15044) — thanks @HouMinXi
  • fix(oauth): guard GHE Copilot device-flow requests to the caller's gheUrl (#15046) — thanks @HouMinXi
  • fix(release): drain the v3.8.51 base-reds after the 2026-09-28/29 merge wave (#15052)
  • fix(release): drain base-reds r6 of v3.8.51 — docs i18n drift, UI value drift, ai-attribution history (#15104)
  • fix(release): clear 9 unit base-reds on release/v3.8.51 (shards 3 and 4) (#15109)
  • fix(ci): clear 13 release/v3.8.51 unit base-reds (gpt-5 sampling strip, command-code none, contract drift) (#15111)
  • fix(sse): refuse traversal in the Codex /v1/responses subpath (#15119) — thanks @HouMinXi
  • fix(sse): scan Devin and PromptQL tool markup in linear time (#15120) — thanks @HouMinXi
  • fix(telegram): skip malformed percent escapes in initData instead of throwing (#15121) — thanks @HouMinXi
  • fix(deps): Make package-lock.json valid for npm 10 again (the Node 22 npm ci in Release acceptance had rejected it since #13661): scope the promptfoo > undici override to undici@7 and the rimraf > minimatch > brace-expansion override to minimatch@9, so @ai-sdk/provider-utils keeps its declared undici@^6 (6.29.0) and the minimatch@3 shared with the ESLint plugins keeps brace-expansion@^1 (1.1.21). npm 10, 11 and 12 now resolve the same tree. (#15122)

📝 Maintenance

  • docs(readme): reconcile live v3.8.50 provider, free-tier, CLI, routing, test, community, sponsor, acknowledgment, and SVG metrics with their audited source denominators, including a deduplicated OmniRoute-in-Action snapshot and distinct contributor rankings for merged pull requests, GitHub-attributed commits, and Git history (#11356). (#11434) — thanks @hartmark
  • docs: sync the canonical provider count 354 → 356 across README.md, AGENTS.md, llm.txt (+ 42 i18n mirrors), the four README SVG diagrams, docs/reference/PROVIDER_REFERENCE.md (regenerated) and the package.json description after Opper (#11629) and 1min.ai (#11631) boarded the catalog — closes the check:docs-counts-sync strict drifts that kept release/v3.8.51 red (#11449). (#11673) — thanks @jonlwheat2-gif
  • test(providers): update count-derived assertions after the v3.8.51 provider additions — APIKEY_PROVIDERS 233 → 235 (Opper #11629 + 1min.ai #11631), reserved-prefix REGISTRY walk 395 → 398, WEB_FETCH_PROVIDERS now includes nimble-search (#11620), and the provider translate-path golden snapshot regenerated (#11449). (#11674) — thanks @jonlwheat2-gif
  • test(sse): bump the hard-lease connection-query inventory for src/lib/tokenHealthCheck.ts to 2 — the verify-only web-cookie sweep added by #11495 split the single getProviderConnections call into oauth + cookie variants, which the frozen inventory had not tracked (#11495). (#11677) — thanks @jonlwheat2-gif / @oyi77
  • chore(release):