A working Qoder PAT was reported as "expired". The validator probes the Cosy endpoint
(api1.qoder.sh) — which IS the correct PAT path (the executor falls back to it after the
expected 401 from api.qoder.com). The bug was the verdict: isCosyAppError (added by #2860)
marked ANY Cosy 500 with "success":false as an auth failure, including a generic
{..."msgCode":500,"message":"Internal Server Error"} server fault — contradicting the
older #1391 "5xx = valid bypass" rule.
Narrow it: a Cosy 500 only marks the PAT invalid when the body carries an EXPLICIT auth
signal (unauthorized/forbidden/expired/token invalid/...); a generic Internal Server Error
falls back to valid-bypass. #2860's protection for genuine auth rejections is preserved.
Regression test: tests/unit/qoder-cli.test.ts — the two pre-existing generic-500 cases now
assert valid:true (they encoded the #3247 bug) + a new explicit-auth-signal case asserts
valid:false. 13/13 green.
A custom OpenAI-compatible image-edit provider received an empty `model`. In production
`globalThis.fetch` is patched with node_modules/undici's fetch, whose `FormData` class
differs from `globalThis.FormData`; passing a native FormData made undici serialize it as
the string "[object FormData]" (text/plain), dropping every field including `model`.
handleOpenAIImageEdit now assembles the multipart body as a Buffer with an explicit
boundary + Content-Type, which every fetch impl accepts verbatim.
Regression test: tests/unit/image-edits-multipart-3273.test.ts reproduces the exact prod
condition (routes through undici's fetch) — RED before (upstream got text/plain
[object FormData]), GREEN after. Existing image suites stay green (50/50).
Regression of #764. Claude Code → Groq (llama-3.3-70b-versatile) returned HTTP 400
because the model was treated as reasoning-capable: supportsReasoning() defaulted to true,
so applyThinkingBudget did not strip reasoning params, and the claude→openai translator
forwarded reasoning_effort (and re-injected it from output_config.effort) — which Groq
rejects on non-reasoning models.
- providerRegistry: mark llama-3.3-70b-versatile + llama-4-scout supportsReasoning:false
(gpt-oss / qwen3-32b keep reasoning — they accept reasoning_effort).
- stripThinkingConfig: also strip output_config.effort so the translator can't re-inject
reasoning_effort downstream.
Regression test: tests/unit/thinking-budget-groq-3258.test.ts (RED before, GREEN after);
existing thinking-budget suites stay green (45/45).
ds-web/deepseek-v4-pro emits tool calls wrapped as
<tool_call name="skill">{"name":"customize-opencode"}</tool_call> instead of the
canonical <tool>{json}</tool>. webTools.ts only matched <tool>...</tool>, so the block
was silently dropped (and when arguments were present, the surrounding tag leaked into
content). Add TOOL_CALL_TAG_RE to capture the JSON body — the real tool name comes from
the body, never the tag's name= attribute — and extend the early-exit + range stripping.
Regression test: tests/unit/web-tools-translation-3260.test.ts (RED before, GREEN after).
Existing web-tools suites stay green (26/26).
Bump 3.8.11 → 3.8.12 across package.json, lockfile, electron/, open-sse/, and
docs/reference/openapi.yaml; add the [3.8.12] cycle placeholder to the root
CHANGELOG and the 41 i18n mirrors. Integration branch for the v3.8.12 cycle —
fixes/features land here via per-issue PRs and it merges to main at release time.
CodeQL flagged 7 high alerts in code the cycle touched (the large release-PR diff
re-surfaces them). Resolved at the source — no dismissals:
- fix(images): resolveImageBaseUrl trimmed trailing slashes with `/\/+$/`, a
polynomial-ReDoS pattern (js/polynomial-redos) on the configured node base URL.
Replace it with a non-backtracking endsWith/slice loop.
- test(oauth): pin the Anthropic OAuth host with exact-equality asserts and a
parsed-hostname negative check instead of substring `.includes()`
(js/incomplete-url-substring-sanitization). The exact-equality assertions were
already present, so coverage is unchanged.
- test(images): drop the redundant `!includes("generativelanguage.googleapis.com")`
assert — the exact-equality assert on the resolved URL already guarantees it.
agy's gemini-3.1-pro-high/-low had no alias, so resolveAntigravityModelId sent the
speculative -high/-low suffix verbatim to upstream, which rejects it (400) for
gemini-3.x. Worse, the non-stream executor branch fed the 4xx response into the SSE
collector, returning a synthetic empty {object:chat.completion} envelope that masked
the error. Alias both to gemini-3.1-pro, and surface real upstream errors via
buildErrorBody for non-ok non-stream responses. + unit tests.
Image routes now resolve a requested model the same way across /v1/images/generations
and /v1/images/edits, via a shared resolver: built-in id -> custom provider prefix ->
bare combo/alias name (e.g. "image" -> its single image target). Previously a bare
combo name fell through to "Invalid image model".
/v1/images/edits gains two capabilities for custom OpenAI-compatible providers:
- multipart edit forwarding to the node's {base_url}/images/edits (was hard-rejected
unless chatgpt-web);
- JSON/data-URL edit input (images:[{image_url:"data:..."}]), converted to the same
fields the multipart reader produces (was "Invalid multipart body").
The chatgpt-web conversation-continuation edit flow is unchanged.
Residual of #3136: a local connection with an empty baseUrl still resolved
to this.config.baseUrl (OpenAI). Fall back to the provider's localDefault
(127.0.0.1:8080/v1) before the OpenAI default for the local-provider group.
Co-authored-by: tjengbudi <tjengbudi@users.noreply.github.com>
The image-generation handler read credentials.baseUrl (always undefined),
so custom OpenAI-compatible image providers fell back to the Gemini endpoint
(401). Resolve from providerSpecificData.baseUrl like the chat path, and
rewrite prefix/model to the internal node id before the exact-id lookup.
Co-authored-by: ngocquynh85 <ngocquynh85@users.noreply.github.com>
* chore(release): open v3.8.10 development cycle
Bump 3.8.9 → 3.8.10 across package.json, lockfile, electron, open-sse, and
docs/reference/openapi.yaml; add the [3.8.10] CHANGELOG section (root + 41 i18n
mirrors) as the integration target for the cycle. Entries land here as work
merges into release/v3.8.10; finalized by the release flow.
* fix(providers): resolve web provider alias collisions
Assign unique aliases to HuggingChat, Kimi Web, and Qwen Web so they no longer shadow primary providers or trigger startup warnings.
Add a unit test to enforce provider alias uniqueness and prevent future collisions. Also expand local ignore and VS Code exclude rules for agent, build, and worktree artifacts.
* fix(responses): normalize image_url parts across input paths (#3150)
Normalize image_url parts across all Responses input paths. Integrated into release/v3.8.10.
* fix(api-manager): preserve API key expiration local time (#3146)
Preserve API key expiration local time + clear button. Integrated into release/v3.8.10.
* Strip previous_response_id for stateless Responses upstreams (#3143)
Strip previous_response_id for stateless Responses upstreams (auto/strip/preserve). Integrated into release/v3.8.10.
* fix(opencode-plugin): map thinking cap to interleaved in model+combo (#3138)
Map caps.thinking to ModelV2.capabilities.interleaved for opencode-plugin. Integrated into release/v3.8.10.
* fix(providers): use synced models as fallback for all providers (#3148)
Use synced models as authoritative local catalog for all providers (+regression test). Integrated into release/v3.8.10.
* fix(qoder): bifurcate validation by token type — PAT→Cosy, regular API key→dashscope (#3149)
Bifurcate Qoder validation by token type (PAT→Cosy, regular→dashscope) +regression test. Integrated into release/v3.8.10.
* fix(antigravity): dynamic model resolution via MITM alias table (#3144)
Dynamic antigravity MITM model resolution in the executor (+bug fix +regression test; DB import dropped from client-reachable config). Integrated into release/v3.8.10.
* Feature/batch allow big (#3128)
Podman deployment options + larger upload body-size limits (+CONTAINER_HOST docs). Integrated into release/v3.8.10.
* fix(fireworks): preserve fully-qualified router/model IDs (#3133) (#3160)
Fireworks router IDs (accounts/fireworks/routers/...) were double-prefixed
with accounts/fireworks/models/ → upstream 404. Add optional
acceptedModelIdPrefixes to the registry entry and skip the prepend when the
model already starts with an accepted prefix.
Co-authored-by: KooshaPari <KooshaPari@users.noreply.github.com>
* fix(llama-cpp): route to configured local baseUrl instead of OpenAI (#3136) (#3161)
llama-cpp was missing from the local-provider group in buildUrl(), so it
fell through to the OpenAI baseUrl and returned an OpenAI 401. Add the
case to resolve the connection's providerSpecificData.baseUrl.
Co-authored-by: tjengbudi <tjengbudi@users.noreply.github.com>
* fix(t3-chat-web): parse cookies + convexSessionId from stored credential (#3007) (#3162)
The executor read credentials.cookies/convexSessionId, but the pipeline
only stores the pasted string under apiKey → t3.chat always 400'd. Parse
both values from apiKey (fallback accessToken), mirroring validation.ts.
Co-authored-by: minhtran162 <minhtran162@users.noreply.github.com>
* fix(minimax): stop capping MiniMax-M3 / M2.7 max_tokens at 8192 (#3141) (#3163)
MiniMax-M3 had no MODEL_SPECS entry and capitalized MiniMax-M2.7 missed
its lowercase spec (case-sensitive lookup) → both fell to the 8192 default
cap. Add the M3 spec (512K output), alias the capitalized ids, and make
getModelSpec lookups case-insensitive.
Co-authored-by: totaltube <totaltube@users.noreply.github.com>
* fix(github-copilot): discover model catalog live from api.githubcopilot.com (#3120, #3121) (#3164)
The github (Copilot) provider had a static hardcoded catalog with no
discovery source, so Import Models never refreshed (#3120) and advertised
non-entitled models that 400 on use (#3121). Add a live /models fetch with
fallback to the static list.
Co-authored-by: gabrielmoreira <gabrielmoreira@users.noreply.github.com>
* fix(combo): invalidate nested-combo cache on edits + log DATA_DIR (#3147) (#3165)
Editing a combo did not invalidate the 10s nested-combo expansion caches
(chat.ts getCombosCachedForChat + chatCore.ts getCombosCached; the exported
clearCombosCache was dead code), so a removed nested target/model could be
served as a phantom for up to 10s. Wire a shared monotonic combos-cache
version in readCache (bumped by invalidateDbCache("combos") on every combo
write); both cache layers treat a version mismatch as a miss.
Also log the resolved DATA_DIR/SQLITE_FILE absolute path at DB init so the
reporter's 'persists across restart + volume wipe' symptom (a multi-replica
Docker volume/DATA_DIR mismatch, not a routing bug) is diagnosable from logs.
Includes consolidated CHANGELOG entries for #3133/#3136/#3007/#3141/#3120/#3121.
Co-authored-by: ViFigueiredo <ViFigueiredo@users.noreply.github.com>
* fix(web-tools): parse bare JSON tool calls (#3157)
Parse bare JSON tool calls for deepseek-web (#2820) + fuzzy tool-name matching. Integrated into release/v3.8.10.
* fix(misc): minor fixes across reasoning cache, account fallback, binary manager (#3177)
Misc: ProviderProfile export, DeepSeek reasoning regex, binary guard. Integrated into release/v3.8.10.
* fix(kiro): minor OAuth social exchange tweaks (#3176)
Kiro social OAuth: optional targetProvider passthrough. Integrated into release/v3.8.10.
* deps: bump hono from 4.12.18 to 4.12.23 (#3179)
Bump hono to 4.12.23. Integrated into release/v3.8.10.
* fix(providerRegistry): update kilocode format and executor (#3166)
kilocode: openai format + default executor (matches kilo-gateway) + registry test. Integrated into release/v3.8.10.
* feat(metrics): cross-request TTFT and gap latency after tool calls (#3173)
Cross-request TTFT + gap-after-tool latency metrics (+test). Integrated into release/v3.8.10.
* feat(dashboard): provider stats API endpoint and dashboard page (#3175)
Provider stats dashboard + API (SQL moved to db module per Hard Rule #5, +test). Integrated into release/v3.8.10.
* fix(usage): sequential+spaced OAuth quota sync, reactive force-refresh, actionable 401 (#3156)
Sequential+spaced OAuth quota sync, reactive force-refresh on 401, actionable 401 in UI. Integrated into release/v3.8.10.
* fix(healthcheck): per-provider proactive-refresh skip list (rescue short-TTL OAuth) (#3159)
Per-provider proactive-refresh skip list (OMNIROUTE_HEALTHCHECK_SKIP_PROVIDERS) to rescue short-TTL OAuth. Integrated into release/v3.8.10.
* feat(quota): show OAuth token expiry on provider cards (small, blue, informative) (#3178)
Show OAuth token expiry on provider cards (small, blue, informative). Integrated into release/v3.8.10.
* fix(providers): empty refresh must not resurface just-cleared synced models (#3181)
Empty refresh must not resurface just-cleared synced models (fixes the release-blocking provider-models-route test). Integrated into release/v3.8.10.
* chore(release): v3.8.10 — 2026-06-04 (finalize CHANGELOG)
---------
Co-authored-by: Wilson <pedbookmed@gmail.com>
Co-authored-by: Xiangzhe <32761048+xz-dev@users.noreply.github.com>
Co-authored-by: Jan Leon <Jan.gaschler@gmail.com>
Co-authored-by: M.M <mr.maatoug@gmail.com>
Co-authored-by: Hernan Javier Ardila Sanchez <hjasgr@gmail.com>
Co-authored-by: Markus Hartung <mail@hartmark.se>
Co-authored-by: KooshaPari <KooshaPari@users.noreply.github.com>
Co-authored-by: tjengbudi <tjengbudi@users.noreply.github.com>
Co-authored-by: minhtran162 <minhtran162@users.noreply.github.com>
Co-authored-by: totaltube <totaltube@users.noreply.github.com>
Co-authored-by: gabrielmoreira <gabrielmoreira@users.noreply.github.com>
Co-authored-by: ViFigueiredo <ViFigueiredo@users.noreply.github.com>
Co-authored-by: PizzaV <103120356+pizzav-xyz@users.noreply.github.com>
Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com>
Co-authored-by: Nicolas Lorin <androw95220@gmail.com>
Kiro (AWS CodeWhisperer) tops out at Claude Opus 4.7 in the registry, but the latest Opus 4.8 is already served by the Claude Code provider and Kiro's executor passes the model id through to CodeWhisperer verbatim. Expose claude-opus-4.8 on the kiro provider so clients can select it via kiro/claude-opus-4.8.
- providerRegistry: add claude-opus-4.8 (1M context, 128k output) above 4.7
- pricing: add claude-opus-4.8 (and the previously-missing 4.7) to the kiro pricing block at Kiro's standard Opus rate so usage cost is non-zero
- tests: assert kiro exposes claude-opus-4.8 with matching context/output + pricing
Full CI surfaced real failures that local subsets missed (gh-merged PRs bypass
the hooks that run these gates):
- typecheck:core (Lint job): 3 now-unused @ts-expect-error in mcp-server/server.ts
(#3077 dynamic tool loops) → @ts-ignore (lenient, no TS2578).
- pack-artifact-policy.test.ts: build-reorg (#3124) renamed app/->dist/; the test
still asserted app/ paths + REQUIRED order (it sorts alphabetically).
- electron-packaging.test.ts: extraResources from .next/electron-standalone ->
.build/electron-standalone (#3124).
- glm-provider-model-import-route.test.ts: two GLM connections shared one apiKey,
so #3100 (#3023) dedup collapsed them → only one discovery fetch. Distinct keys.
Remaining CI flakes (batch expiration, ModelSync self-fetch) pass in isolation —
concurrency/port flakiness under --test-concurrency=4, not real failures.
* fix(sse): bound Antigravity short-retry 429 loop per endpoint
A persistent 429 on the short-retry branch (retryAfterMs ≤ 60s) looped
forever on the same endpoint because the branch did `urlIndex--; continue`
without checking the shared retry counter. Production log showed 77
consecutive 429s on one daily endpoint/account with zero fallback.
Gate the short-retry branch on `retryAttemptsByUrl[urlIndex] < MAX_AUTO_RETRIES`
(mirroring the already-bounded sibling), so a persistent 429 retries at most
3× per endpoint across all 3 base URLs then returns the 429 to the account-
fallback layer.
Regression test: 'bounds a persistent short-retry 429' in
tests/unit/executor-antigravity.test.ts — asserts 12 total attempts
(3 endpoints × 4) and a returned 429 with zero hang.
* fix(sse): lock Antigravity quota-exhausted account for full reset window
After the retry-loop bound, OmniRoute fell over to the next account but
re-selected the exhausted one first on every subsequent request (~60s wasted
per request). Root cause: the 429 body 'Individual quota reached. Contact
your administrator to enable overages. Resets in 164h27m24s.' was not
recognized as quota exhaustion, so the model was locked for only ~5s instead
of the real 6.8-day reset window.
Two detector fixes (mirrors Antigravity-Manager rate_limit.rs set_lockout_until):
1. classify429.ts — add QUOTA_PATTERNS: /individual quota reached/i,
/quota reached/i, /enable overages/i so looksLikeQuotaExhausted() fires.
2. accountFallback.ts — same patterns in classifyErrorText(); extend
parseRetryFromErrorText() to parse 'Resets? in XhYmZs' (reusing the
existing computeDurationMs helper) so the exact reset duration reaches
recordModelLockoutFailure as exactCooldownMs (uncapped, per user choice).
The lockout machinery already stores until = now + cooldownMs with no clamp,
bypasses getScaledCooldown when exactCooldownMs > 0, and keeps the longer of
existing/new, so the full 164h window flows through intact.
New patterns stay specific — plain 'too many requests'/'rate limit exceeded'
messages still classify as rate_limit.
Verified end-to-end against the real message:
- classify429 → quota_exhausted
- parseRetryFromErrorText → 592044000 ms (164h27m24s exactly)
- checkFallbackError → usedUpstreamRetryHint: true, cooldownMs: 592044000
* refactor(account-fallback): simplify error parsing and add cooldown safety
- Implement a 30-day cap on parsed retry durations to prevent indefinite account lockouts.
- Replace manual string matching with `looksLikeQuotaExhausted` for more robust quota detection.
- Streamline regex logic in `parseRetryFromErrorText` for better readability.
- Add unit tests for extreme cooldown values and free-tier exhaustion scenarios.
* fix(429): drop over-broad /quota reached/ pattern, keep specific matches
The bare /quota reached/ would also flag transient per-minute limits like
'request quota reached, retry in 60s' as quota_exhausted (multi-hour lock).
The Antigravity message is still caught by /individual quota reached/. Added a
regression assertion proving the transient case stays a rate_limit.
---------
Co-authored-by: diegosouzapw <diegosouza.pw@gmail.com>
* fix(tools): keep opaque object schemas open
* test: align opaque-object-schema expectations with additionalProperties:true
The opaque-schema fix intentionally injects additionalProperties:true on empty
object schemas (incl. the web_search passthrough shim and null/missing parameter
fallbacks). Update the pre-fix snapshot assertions to match the new behavior.
---------
Co-authored-by: diegosouzapw <diegosouza.pw@gmail.com>
* fix(sse): per-model 403 on passthrough providers locks the model, not the connection (#3027)
A per-model subscription 403 from a passthrough / per-model-quota provider
(e.g. ollama-cloud "this model requires a subscription, upgrade for access" on
deepseek-v4-pro) cooled down the ENTIRE connection instead of locking out only
the paid model, knocking out the free models on the same key and escalating an
exponential connection-wide backoff on repeats.
markAccountUnavailable's per-model lockout gate only covered 404/429/>=500, and
the only 403 -> model-lockout special case was hard-coded to grok-web. Generalize
it: a model-scoped 403 on an isPerModelQuotaProvider becomes a model lockout
(connection stays active). Terminal whole-key 403s (permanent/banned, account
deactivated, credits exhausted, project-route) keep their connection-level path.
Test-first: reproduction (paid model locked, connection active, free model still
eligible), regression guard (deactivated key still terminal -> banned, not
downgraded), and backoff guard (repeated 403s do not escalate connection backoff).
* feat(sse): persistent session + rolling-window memory for deepseek-web (#2942)
The DeepSeek web API takes only a single `prompt` string (no messages array) and
the executor created a fresh chat session per request, deleting it afterward — so
agentic multi-turn clients got per-turn amnesia.
Add two opt-in, per-connection settings (providerSpecificData), both defaulting to
the legacy behavior so plain-chat users are unaffected:
- historyWindow (number, default 0): when > 0, messagesToPrompt stitches the last N
non-system turns into a role-tagged transcript ("User:"/"Assistant:") so context
carries across turns within the single prompt string.
- persistSession (bool, default false): reuse one upstream chat session per userToken
(cached in the existing sessionCache) instead of creating/deleting one per request.
A reused session that fails is treated as stale (deleted in the DeepSeek UI): the
cache entry is dropped, a fresh session is created, and the completion is retried
once. Error paths invalidate the cached session so the next turn self-heals.
Test-first: pure prompt-builder window semantics (legacy/window/cap/empty) and
execute()-level session behavior via mocked fetch (fresh-per-request default, reuse,
stale-session fresh retry, history threaded into the prompt). Full existing
deepseek-web suite (35 tests) still green.
Note: chat.deepseek.com is an unofficial reverse-engineered surface; the live
round-trip can't be exercised in CI (no userToken/session). Treated as best-effort
per the issue; the prompt/session logic is unit-tested in isolation.
* feat(sse): tool-call translation for deepseek-web (#2820)
deepseek-web previously hard-failed any request carrying tools[] with a 400 (#2848),
so agentic clients could not use it for function calling at all. Add a bidirectional
translation layer (new open-sse/translator/webTools.ts, reusable by the other
web-cookie executors):
- Request: serializeToolsToPrompt() turns the OpenAI tools[] into a <tool>{...}</tool>
prompt contract, injected as a leading system message.
- Response: parseToolCallsFromText() extracts the model's <tool> blocks into OpenAI
tool_calls (arguments as a JSON string), strips them from content, and the executor
emits finish_reason "tool_calls" — for both non-stream and stream clients (the reply
is buffered since a tool block must be parsed whole).
A plain reply (no <tool> block) still streams normally with finish_reason "stop".
The superseded #2848 400-contract test is updated to assert the new translation.
Test-first: pure serializer/parser units + execute() round-trip via mocked fetch
(no-400, prompt serialization, non-stream tool_calls, stream tool_calls, plain reply).
Note: chat.deepseek.com is an unofficial reverse-engineered surface and the exact
<tool> emission depends on the model following the injected contract; the live
round-trip can't be exercised in CI. Best-effort per the issue; translation logic is
unit-tested in isolation.
* fix(sse): drop duplicate per-model 403 block — already in release via #3096
The release branch already scopes per-model 403 to model lockout (PR #3096,
commit 7042d562c) with the canonical !terminalStatus guard + exactCooldownMs
upstream hint. This PR's separate isTerminalOrRoute403 block shadowed it and
omitted the retry hint. Keep only the deepseek-web (#2942) + webTools (#2820)
changes here; the #3027 regression test is retained as coverage for #3096.
* feat(cursor): vision (image_url) input + tool-commit/output-constraint enhancements
Add image/vision input to the Cursor provider's agent.v1 endpoint, plus the
supporting prompt-engineering and resilience work developed alongside it.
Vision input
- Decode OpenAI `image_url` parts (base64 `data:` URIs and remote `http(s)` URLs)
and inline them as `SelectedContext.selected_images[]` — field numbers pinned
from the cursor-agent agent.v1 protobuf descriptor (SelectedImage.data oneof,
uuid, optional Dimension, mime_type). Cross-checked against composer-api's shape.
- New `resolveCursorImages` helper: SSRF-guarded remote fetches via the repo's
canonical `parseAndValidatePublicUrl` (always public-only for client URLs),
<=1 MiB per image (pre-decode + streaming cap), `image/*` enforced, max 12
images, sanitized `CursorImageError` (no stack/path leakage).
- `openai-to-cursor` translator now preserves `image_url` parts instead of
dropping them; executor `buildRequest` resolves images and attaches them to
the user turn. The no-image path is byte-identical to before (test-asserted).
Supporting cursor enhancements
- Tool-commit directive (raises composer-2.5 tool-call rate ~53% -> ~88%),
`tool_choice` none/required/specific handling, and output constraints
(`response_format` / `max_tokens` / `stop` surfaced as prompt instructions).
- `cursorSessionManager`: clear pending tool-call mappings on session close.
- `cursorVersionDetector`: export `FALLBACK_VERSION` as a single source of truth.
Tests & docs
- New unit suite for the image encoder + resolver (field layout, byte-identical
no-image path, SSRF / oversize / bad-base64 / too-many rejections, sanitized
error body), translator image-preservation tests, and live e2e tests
(base64 + remote URL, gated on `CURSOR_E2E_TOKEN`).
- Documented `CURSOR_TOOL_DIRECTIVE` and `CURSOR_IMAGE_FETCH_TIMEOUT_MS` in
`.env.example` and `docs/reference/ENVIRONMENT.md`.
* fix(cursor): address review — redirect SSRF, large-payload guard, stream OOM, case/NaN nits
Resolves the gemini-code-assist review on #3104:
- SSRF via redirect (critical): fetchImageBytes now uses redirect:"manual" and
re-validates every hop through parseAndValidatePublicUrl, so a public URL can't
30x-redirect to a private/link-local address. Bounded to 3 redirects.
- Large data URL (high): reject on raw payload length before the whitespace-strip
regex, so an oversized data URL can't burn CPU.
- Stream read (high): readCapped consumes the body as an async iterable (Node
Readable + Web Streams) or via getReader, capping mid-read; uncapped
arrayBuffer() is only a last resort.
- data: scheme (medium): match case-insensitively (RFC 2397) while preserving the
original payload.
- NaN timeouts (medium): CURSOR_IMAGE_FETCH_TIMEOUT_MS and CURSOR_STREAM_TIMEOUT_MS
fall back to defaults when the env value isn't a positive integer.
Adds tests: redirect-to-private blocked, redirect-to-public followed, too-many-
redirects rejected, uppercase DATA: accepted.
* fix(cursor): defend image fetch against DNS-rebinding SSRF
Address the @codex review on #3104: parseAndValidatePublicUrl only checks the
hostname string, so a public-looking host that (re)resolves to a private /
link-local / metadata IP would still be fetched. Each hop now resolves the host
via dns.lookup({all:true}) and rejects if ANY answer is private (isPrivateHost),
before connecting. IP literals are skipped (already validated by the URL guard).
This narrows but doesn't fully close the TOCTOU window vs fetch's own
resolution; a connection-time IP filter on the shared outbound guard would
close it for every caller. Adds unit tests for the IP gate and a mocked
DNS-rebinding case (public host -> 127.0.0.1, fetch never reached).
* fix: add AbortController timeout to fetchImageEndpoint
fetchImageEndpoint uses raw fetch() without timeout control.
Long-running image generation requests (~20-30s) are killed
by Next.js default timeout, producing upstream_error responses.
Replace fetch() with fetchWithTimeout() from shared utils,
defaulting to FETCH_TIMEOUT_MS (120s via OMNIROUTE_DEFAULT_FETCH_TIMEOUT_MS).
* fix: return 504 for image fetch timeout, add tests
Address review feedback from gemini-code-assist:
- Import FetchTimeoutError and catch it in fetchImageEndpoint
- Return 504 (Gateway Timeout) instead of 502 for timeout/AbortError
- Add 3 focused unit tests for timeout, non-timeout, and success paths
Clarifies timeout semantics: timeout is a gateway timeout (504), not
a bad gateway (502). Non-timeout fetch errors remain 502.
---------
Co-authored-by: mgarmash <mgarmash@37bytes.com>
synthesizeOpenAiSseFromJson combined role+content+reasoning_content in one delta, which the openai→openai translator re-split, duplicating reasoning_content across chunks. Now emits role, reasoning_content, content and tool_calls as separate sequential deltas (reasoning before content) — the shape a real reasoning model streams — so the translator passes them through with no duplication. Tests assert each field appears exactly once and reasoning precedes content.
#3089: reasoning openai-compatible upstreams that ignore stream:true and return application/json produced STREAM_EARLY_EOF because readiness only scans SSE data: frames. chatCore now detects a non-SSE JSON upstream body on the streaming path and synthesizes an equivalent OpenAI SSE stream (new synthesizeOpenAiSseFromJson util), preserving content + reasoning_content. #2952: semantic-cache hits returned application/json regardless of stream flag, so streaming clients lost reasoning_content; stream requests now SSE-wrap the cached completion via the same helper. Unit tests for the converter (4).
grok-web only sent the 'sso' cookie; Grok's anti-bot now rejects (403 code 7) requests missing the paired 'sso-rw' write cookie. Adds buildGrokCookieHeader() which emits sso plus sso-rw when the pasted blob carries it (never a phantom sso-rw from a bare value), used by both the executor and the connection validator. Updates the grok-web authHint. Tests: 6 new buildGrokCookieHeader cases; grok-web 62/62 unchanged.
Add xiaomi-mimo to the prompt-caching provider allowlist so Claude Code (via cc-switch) cache_control breakpoints are preserved instead of stripped by the OpenAI-format translator. Restores cache hits that worked when calling Xiaomi directly.
Bump 3.8.8 → 3.8.9 across package.json, lockfile, electron, open-sse, and
docs/reference/openapi.yaml; add the [3.8.9] CHANGELOG section (root + 40 i18n
mirrors) as the integration target for the cycle. Entries land here as work
merges into release/v3.8.9; finalized by the release flow.
Add a global OMNIROUTE_CODEX_WS_ENABLED kill-switch for Codex WebSocket transport, defaulting to enabled if feature flag lookup fails.
Track Xiaomi MiMo monthly token usage from OmniRoute usage history and expose it through the usage fetcher provider list.
Support MiniMax Coding Plan quota responses that expose remaining usage as
percentages instead of request counts, including the text quota surfaced under
the `general` model.
Also document Codex CLI WebSocket configuration, clarifying that bare ChatGPT
model ids must be used instead of `codex/`-prefixed ids.
Quality Gate was failing on New Code with 1 hotspot, 1 vulnerability and
4 reliability bugs. Resolved each at the source (no SonarCloud mark-as-safe):
Reliability (4 bugs → rating back to A):
- usage.ts: drop duplicate `case "opencode-go"` (S1862) — the 2nd case was
dead (first match wins) and would have called the wrong usage fetcher.
Also de-duplicate the same id in USAGE_FETCHER_PROVIDERS (mirrors the switch;
prevented a double quota-fetcher registration).
- ApiManagerPageClient.tsx: a dangling `.find(...)?.timestamp || null` expression
(S905) discarded the better matcher — `lastUsed` silently lost the
name-fallback for legacy logs without apiKeyId. Assign the complete matcher.
- chat.ts / auth.ts: `Promise` used in a boolean conditional (S6544). Both are
intentional (memoized promise / mutex default), not a forgotten await —
made the intent explicit via `!== null` and `??` (behaviour identical).
Security (vulnerability → rating back to A):
- nvidia-startswith-diag.ts: neutralize CR/LF before logging env-derived
values (S5145 log injection) in the ad-hoc diagnostic script.
Security Hotspot (reviewed → 0 open):
- pluginWorker.ts uses `vm.runInContext` to run plugin code — that IS the
worker's purpose, inside a hardened vm sandbox (createContext, require
allow-list of just `crypto`, 10s timeout); not eval/new Function. Suppress
S1523 scoped only to pluginWorker.ts in sonar-project.properties, with the
same documented-justification pattern as the existing hotspot suppressions.
