Root cause: waitForImageViaWebSocket() only parsed the singular
update_content.message (object) / payload.message / data.message shapes
in the celsius WebSocket frames chatgpt.com uses to deliver async
image_gen results. Some chatgpt.com deployments deliver the completed
tool-role image_asset_pointer message inside update_content.messages[]
(a plural array of { message: {...} } wrappers) instead, which produced
zero candidates, so the listener idled out the timeout and the request
failed with the generic 'ChatGPT Web completed without returning image
markdown' 502 with no x_image_resolution_failed flag.
Fix: also read update_content.messages[] and push each wrapped message
into the same candidate pipeline used for the singular shape.
Regression test: tests/unit/chatgpt-web-async-image-ws-shapes-7357.test.ts
drives the real ChatGptWebExecutor.execute() end-to-end (real SSE
parsing, real pollForAsyncImage()/waitForImageViaWebSocket()), mocking
only the network edges (tlsFetchChatGpt + global WebSocket), and proves
the plural-array frame now resolves to image markdown instead of being
dropped.
Gates run: check-file-size (OK), check-complexity (OK, 2054 <= 2056
baseline), check-cognitive-complexity (OK, 889 <= 890 baseline),
typecheck:core (clean), eslint on changed files (clean), full
tests/unit/chatgpt-web.test.ts (89/89), chatgpt-web-image-silentdrop.test.ts,
chatgpt-web-tools-5240.test.ts, chatgpt-web-models-split.test.ts,
chatgpt-web-sha3-boringssl-5531.test.ts, chatgpt-web-handoff-resume.test.ts,
chatgpt-web-citations(-escape).test.ts all pass.
captureCurrentProviderRequest mirrors every Bedrock Converse request into the
pending-request log tracker right after openAIToBedrockConverse() builds it,
including the decoded image.source.bytes Uint8Array. sanitizePayloadPII() and
redactPayload() in src/lib/logPayloads.ts both gate their recursive walk on
Array.isArray(), which is false for typed arrays, so each image fell into the
generic-object branch and got enumerated one JS key per decoded byte (twice,
once per function) before any truncation bound applied. For 3x ~1MB images
this took ~4s of synchronous, event-loop-blocking work, matching the
reporter's "1-2 images OK, 3+ fails" threshold and their --stack-size
observation (data-width pressure, not call-depth).
Add an opaque-binary short-circuit (ArrayBuffer.isView) ahead of the
Array.isArray branch in both functions, returning a fixed-size placeholder
instead of recursing. Apply the same guard to cloneBoundedForLog() in
open-sse/utils/requestLogger.ts for defense-in-depth (same blind spot, only
accidentally safe today via its own key-count slice).
Regression test reproduces the exact reporter shape (3x 1MB images) through
the real openAIToBedrockConverse() converter and protectPayloadForLog(),
asserting completion well under the previous ~4s and that binary bytes are
never expanded into per-byte object keys.
resolveRequestedModel() only special-cased "auto" and the composer
"-fast" suffix; every pinned Claude/GPT id carrying an effort/reasoning
suffix (e.g. "claude-opus-4-8-high", "gpt-5.5-high") fell through and
was sent to cursor's server verbatim as model_id, with an empty
parameters array. Cursor has no route for the suffixed id -- it only
knows the base id plus an out-of-band ModelParameter -- so it accepted
the request but returned an empty turn.
Split the known effort suffixes (-low/-medium/-high/-xhigh/-max) off
the base id: Claude ids surface an {id:"effort", value} parameter,
GPT ids surface {id:"reasoning", value}, matching the real cursor-agent
client's wire format. encodeAgentRunRequest()'s ModelDetails fields
derive from the same resolved base id, so the #3714 pinned-model
ModelDetails envelope stays correct without further changes.
Updates the existing resolveRequestedModel test that locked in the
buggy verbatim pass-through, and the #3714 ModelDetails test to assert
against the base id. Adds a dedicated regression test file proving the
Claude/GPT split plus non-regression of the "-fast" toggle and
unsuffixed ids.
#7268: classifyProviderError() only inspected the response body for
model-unavailable wording on 400/403/404, so a 401 body like "Model X is
not supported" (free-tier/aggregator providers) fell through to a generic
UNAUTHORIZED classification. Because chatCore.ts only calls lockModel(...,
"model_not_found", ...) on the MODEL_NOT_FOUND branch, the broken model was
never locked out and auto-combo kept re-selecting it every request. Added a
shared containsModelUnavailableMessage() regex (bounded, ReDoS-safe) in
errorClassifier.ts, consulted by the 401 branch before falling back to
ACCOUNT_DEACTIVATED/UNAUTHORIZED, and reused by modelFamilyFallback.ts's
isModelUnavailableError() for the literal "<model> is not supported" phrasing.
#7387: applySessionStickiness() (combo-level session stickiness) only gated
a sticky pin's release on testStatus (credits_exhausted/banned/expired) and
rateLimitedUntil (#6692's fix). It never consulted isAccountQuotaExhausted()
(src/domain/quotaCache.ts) — the authoritative per-window (5h/weekly) quota
signal that src/sse/services/auth.ts and sessionAffinityPin.ts (the
provider-level pin) already gate on. A connection whose quota window was
depleted, but that hadn't yet received a hard failure severe enough to flip
testStatus/rateLimitedUntil, was re-promoted to position 0 on every request
regardless of routing strategy. Added isStickyConnectionQuotaExhausted(), a
dynamic-import seam (mirroring resolveConnectionHealth/resolveSaturation, no
new static edge from open-sse/ into src/domain/) with an injectable checker
for tests, gating the release condition alongside the existing checks.
Regression tests: tests/unit/repro-7268-401-model-not-supported-lockout.test.ts,
tests/unit/repro-7387-sticky-quota-exhausted.test.ts (both RED before, GREEN
after). Existing sticky/error-classifier suites re-run and stay green.
Closes#7268Closes#7387
- byProvider now resolves the internal provider id to its configured
display name via getProviderById() (fallback: raw id for providers
not in the static registry). Fixes the Usage page showing "codex"
instead of "OpenAI Codex".
- byModel's in-memory dedup key now uses the normalized model name
instead of the raw one, so the same logical model recorded under a
bare and a provider-prefixed spelling (e.g. "glm-5.2" vs
"z-ai/glm-5.2") merges into a single aggregated row instead of
appearing twice with the same displayed name.
- Introduces a local UsageRows type alias in route.ts to shrink the
repeated "as Array<Record<string, unknown>>" casts back under the
frozen file-size baseline once the file was touched.
* fix(dashboard): resolve costs page 500 from out-of-scope t() in TopListCard (#7272)
TopListCard (CostOverviewTab.tsx) referenced the bare identifier `t`
from an outer component's scope when rendering the zero-cost /
!hasCostData branch, throwing "ReferenceError: t is not defined" and
crashing /dashboard/costs?range=all&apiKeyIds=...&groupBy=model
whenever a filtered slice landed only $0-cost rows.
Extracted TopListCard into its own component file and threaded the
resolved legacyFreeLabel string in as a prop, mirroring the existing
CostBreakdownTable pattern in the same file. Also fixes a case of the
typecheck:core dashboard .tsx coverage gap tracked in #7033.
* test(dashboard): move TopListCard #7272 regression test to vitest UI project
The node:test unit runner cannot load TopListCard's import chain
(@/shared/components -> ProviderIcon -> @lobehub/icons ESM), which made
the "Impacted unit tests (TIA subset; blocking)" CI job red with
"Unexpected token 'export'" on tests/unit/costs-toplistcard-legacy-free-label-7272.test.ts.
Moved the regression test to tests/unit/ui/ and rewrote it against the
vitest UI project (test:vitest:ui, blocking in the test-vitest CI job),
which handles the ESM import chain natively. Verified the test still
fails with "ReferenceError: t is not defined" against the pre-#7272
TopListCard body and passes against the fixed component.
* fix(db): pre-init sql.js WASM ahead of any getDbInstance() consumer (#7288)
* fix(db): close sqljs preinit ordering gap without top-level await (#7288)
The previous fix added a top-level await barrier at the bottom of
src/lib/db/core.ts to guarantee sql.js pre-init before any consumer
reached getDbInstance(). That made core.ts an async ES module, which
broke esbuild's CJS require() bundling for every test file that does
require("../../src/lib/db/core.ts") (tsx's CJS require hook rejects
requiring a transitive dependency with a top-level await), and caused
unrelated tests running in the same node:test process to fail with
"Promise resolution is still pending but the event loop has already
resolved".
Move the fix to the real startup entrypoint instead: registerNodejs()
(src/instrumentation-node.ts) now awaits ensureDbReadyForBoot() before
ensureSecrets()/clearStaleCrashCooldowns()/getSettings()/initAuditLog(),
all of which reach getDbInstance() transitively. ensureDbInitialized()
is idempotent, so later getDbInstance() calls are free cache reads.
The driverFactory.ts error-surfacing improvements from the original
#7288 fix (logging swallowed sync-driver errors, surfacing the real
sql.js pre-init failure instead of the generic "not pre-initialized
yet" message) are unchanged.
Updated tests/unit/db-sqljs-preinit-ordering-gap-7288.test.ts to prove:
no top-level await in core.ts, the corrected call order in
registerNodejs() (source-order assertion), and the original
getDbInstance()-no-longer-throws-the-misleading-message behavior driven
via the same warm-up path ensureDbReadyForBoot() now guarantees ahead
of every other startup step.
* refactor(db): drop the orphaned preInitSqlJsIfSyncDriversUnavailable helper (#7288)
Moving the ordering guarantee to registerNodejs() left this exported
helper with zero production callers — its own docblock still claimed it
was 'Chamada no top level de core.ts', describing an architecture the
hotfix removed. It was also redundant: ensureDbInitialized() already does
tryOpenSync-then-preInitSqlJs on the real boot path (core.ts:1358).
Its two tests exercised the helper as a stand-in for the real warm-up
('Simulates the fixed ordering'), so they proved a simulation rather than
production behaviour. They now drive ensureDbReadyForBoot()/tryOpenSync()
directly. The ordering guard still fails against the unfixed
instrumentation-node.ts (verified) and the whole file is 4/4 green.
The live parts of the driverFactory change (logSwallowedDriverError,
getSqlJsPreInitError) are untouched — both have real callers.
validateResponseQuality() only recognized Claude SSE lifecycle events
(message_start/content_block_*/message_stop/message_delta.stop_reason).
An OpenAI-shape stream (choices[].delta) that emits some bytes (e.g. a
role-only delta) and then closes without ever carrying finish_reason
(and without a data: [DONE] sentinel) fell through to the generic
replay branch and was forwarded to the client as a success instead of
triggering combo failover.
Adds OpenAI-shape lifecycle tracking (hasChoicePayload/hasTerminalMarker)
parallel to the existing Claude tracking: when an OpenAI-shape chunk was
seen but the stream ends without finish_reason or [DONE], and no
recognized content was found, mark the response invalid so combo
failover retries a sibling target. Healthy OpenAI streams (finish_reason
present, or real content found) are unaffected — they exit the peek
loop before reaching this check, preserving the #3399/#3685
pass-through contract.
Regression test: tests/unit/combo-streaming-openai-no-finish-reason-7285.test.ts
Root cause: validateOpenAILikeProvider's chat-probe status handling
(src/lib/providers/validation/openaiFormat.ts) only special-cased
401/403, 404/405, and >=500 — every other status, including 429, fell
through to the unqualified return { valid: true, error: null }. For
permanently-throttled free tiers (e.g. opencode-zen, classified
'avoid' in freeTierCatalog.ts), the dashboard connection Test reported
green forever while real traffic hit 429 on every request, with no
signal to the user.
Fix mirrors the existing validateBedrockProvider 429 precedent in the
same file: keep valid:true (the key is accepted) but add a warning
field describing the rate limit, instead of an indistinguishable pass.
Regression test: tests/unit/issue-7284-connection-test-masks-429.test.ts
mocks a 404 /models probe followed by a 429 chat probe and asserts the
result now carries { valid: true, error: null, warning: <rate-limit string> }.
Gates run (all green):
- node --import tsx/esm --test tests/unit/issue-7284-connection-test-masks-429.test.ts
- node --import tsx/esm --test tests/unit/provider-validation-firepass-403.test.ts tests/unit/validation-format-validators-split.test.ts (existing tests of the touched area, unaffected)
- node scripts/check/check-file-size.mjs
- node scripts/check/check-complexity.mjs (2054/2056 baseline, unchanged)
- node scripts/check/check-cognitive-complexity.mjs (889/890 baseline, unchanged)
- npm run typecheck:core
- npx eslint --suppressions-location config/quality/eslint-suppressions.json <changed files>
Termux/Android's Node reports process.platform === 'android'.
playwright-core's serverRegistry.js throws 'Unsupported platform:
<platform>' from a top-level IIFE at require time, so merely
importing the playwright package crashed — no browser ever launched.
claudeTurnstileSolver.ts was the only playwright consumer in the
codebase with a static top-level import; every other call site
(browserPool.ts, inAppLoginService.ts) already lazy-loads it. That
static import is unconditionally reachable from the Next.js
instrumentation hook on every boot via open-sse/executors/index.ts,
so any unsupported platform crashed the whole server at startup
regardless of which provider was configured.
Fix: import type { Browser, Page } (erased at compile time) and move
the chromium binding to a lazy await import("playwright") inside
solveTurnstile(), matching the existing pattern.
chatCore.ts fed applyCompressionAsync's supportsVision option from
isVisionModelId() — the deliberately-conservative model-id fragment
heuristic in src/shared/constants/visionModels.ts — instead of the
authoritative getResolvedModelCapabilities().supportsVision used by every
other vision-aware path (e.g. the vision-bridge guardrail).
gpt-5.5 is registered with supportsVision:true in modelSpecs.ts, but the
fragment list has no gpt-5.x entry, so the heuristic wrongly returned
false. That false reached lite.ts's replaceImageUrls(), whose gate is
`supportsVision !== false`, silently stripping every image_url block
before the request ever reached the executor.
getResolvedModelCapabilities().supportsVision resolves to null (not
false) for genuinely unknown models, which the same !== false gate
already treats as preserve-by-default — matching the conservative
semantics used elsewhere and avoiding the #4071/#4012 class of bug
(blinding a model that can actually see).
Root cause: the Providers page model-name filter
(filterConfiguredProviderEntries) matched only against
getModelsByProviderId(...), the static curated model registry, never
the live/synced catalog for the connection. Aggregator providers
(openrouter, kilocode, theoldllm...) declare a single-entry static
placeholder (e.g. openrouter's {id:'auto',name:'Auto (Best Available)'}),
so searching for any real upstream model name could never match and the
whole provider silently disappeared from the list.
Fix: source the live/synced catalog (already persisted per-connection via
GET /api/synced-available-models, the same store the combo builder's model
picker already relies on) via a new useSyncedModelsByProvider hook, and
union it with the static registry inside the filter. An empty/never-synced
catalog falls back to the static-only match so already-correct static
providers are unaffected.
Regression test: tests/unit/provider-model-filter-live-catalog-7250.test.ts
reproduces the original bug (static-only match returns 0 results for a
real model name) and proves the fix (live catalog match returns 1),
plus non-regression coverage for the static-only fast path and unrelated
providers.
* fix(cli): Windows cert check/uninstall key off the real CA identity, not a hardcoded legacy host (#7275)
checkCertInstalledWindows()/uninstallCertWindows() queried the Windows Root
store by the literal legacy hostname daily-cloudcode-pa.googleapis.com
regardless of the certPath passed in (the check's param was even
underscore-prefixed/unused). It only worked because that hostname happens to
be the generated CA's own commonName today (generate.ts derives it from
ANTIGRAVITY_TARGET.hosts[0]) -- a coincidence, not a derivation, with no
shared symbol coupling the two. installCertWindows() was already correct.
Both functions now derive a SHA-1 thumbprint straight from the certPath file
via the new exported certutilThumbprint() helper (reusing the existing
getCertFingerprint() logic checkCertInstalledMac() already keys off), so the
Windows store lookup/delete always matches the real generated CA regardless
of any future rename/reorder of ANTIGRAVITY_TARGET.hosts in generate.ts. Same
anti-pattern class as #6338 (DNS side).
* test(cli): assert certutil argv precisely instead of scanning for the legacy host (#7275)
The two negative substring checks tripped CodeQL's
js/incomplete-url-substring-sanitization (high) by looking like URL
sanitization while actually asserting a certutil argv. Replaced with
assertions on the exact argv / extracted certId, which are strictly
stronger: pinning the value proves no other identity can be passed.
Both still fail against the unfixed install.ts (5/5 RED) and pass with
the fix (5/5 GREEN).
deepMergeFallback in src/i18n/request.ts only substituted the English
fallback value when a key was entirely undefined. Keys backfilled by
scripts/i18n/sync-ui-keys.mjs with the __MISSING__:<english> sentinel
exist on the target object, so they passed through untouched and were
rendered raw to the user (395 zh-TW keys, 337 pt-BR keys, systemic
across locales).
Now any target leaf that still carries the __MISSING__: prefix is
treated the same as an absent key, so the clean EN value wins.
Does not touch the underlying translation content (395/337 strings) -
that is a separate content workstream.
The PKCE callback server binds the SERVER's loopback (localhost:PORT). When
the operator drives the OAuth flow from a different machine (OmniRoute on a
remote host/VPS), the provider redirects the browser to the operator's OWN
localhost:PORT — the confirmation screen hangs forever with no explanation.
start-callback-server now inspects the request Host: on a non-loopback host it
returns { remoteHost, tunnelCommand, message } so the UI can show the
'ssh -L PORT:127.0.0.1:PORT' instruction (or steer to the paste/import flow)
instead of a silent hang. Loopback access is unaffected. The Host header is
spoofable, so this drives only a UI hint — never an auth decision.
Logic extracted to remoteOAuthHint.ts (keeps the god-route under its size
budget and makes it unit-testable). TDD: 4 tests covering loopback (no hint),
null host (fail-open), and remote host (correct tunnel command for both the
fixed 1455 and OS-assigned ports).
Closes#7523
Every non-streaming Codex chat request for a ChatGPT-account connection failed
instantly with [502]: Response body is already used (reset after 1m). The
streaming/playground path was unaffected.
Root cause: peekCodexSseTransientError (open-sse/executors/codex.ts) peeked the
SSE prefix with response.body.getReader(), then called reader.releaseLock() and
response.body.getReader() a SECOND time on the same already-disturbed body to
build the replacement stream. Re-acquiring a reader on a disturbed body throws
on undici ('Response body is already used'); chatCore's generic upstream-error
handling then stamped the TypeError with a default 60s cooldown, masking a pure
code defect as a rate limit (and tripping the codex circuit breaker).
Fix: keep the single reader already held; never touch response.body again.
TDD: a getReader spy that throws on the 2nd acquire reproduces the exact hazard
— 1 test RED against the release code, 2/2 GREEN with the fix; the replacement
body stays byte-identical to the upstream SSE. No regression across the codex
unit suite. Reproduced live on the VPS 2026-07-16.
POST /api/oauth/codex/import accepted a payload with an already-invalidated
refresh_token and persisted it as an 'active' connection that could never
work — the failure surfaced confusingly only on first real use, long after
the import looked successful.
Validate each record's refresh_token against OpenAI's OAuth endpoint before
persisting, reusing refreshCodexToken() (free exchange, no quota). On an
unrecoverable refresh error the record is rejected with a clear re-auth
message; a valid token imports as before, with any rotated tokens applied.
Bulk import still processes each record independently — one dead token no
longer blocks the valid ones.
Reproduced live 2026-07-16: a 2026-07-10 auth.json imported clean but its
refresh_token returned 401 refresh_token_invalidated. TDD: 3 tests RED against
the old route, 5/5 GREEN with the fix.
Closes#7522
The connection Test button always reported success for a Codex connection
backed by a ChatGPT account: the probe sent `gpt-5.3-codex`, a codex-only
model the ChatGPT-account backend rejects outright with a 400 — the same
status the probe treats as 'auth accepted, body invalid'. A bad token and a
good token both came back 400, so Test could never fail on a bad token.
Probe with `gpt-5.5` (confirmed served for ChatGPT-account sessions via live
VPS test 2026-07-16) instead; `input: []` still yields the intended 400 for a
good token, 401/403 for a bad one.
Live verification (VPS): gpt-5.3-codex, gpt-5.6-sol and the gpt-5*-codex ids all
return 'not supported when using Codex with a ChatGPT account'; gpt-5.5 and
gpt-5.6-terra answer normally on the same account.
Closes#7521
* fix(electron): materialize Turbopack hashed-module symlinks during packaging
Reconstructed onto release/v3.8.47 to drop unrelated main-drift (deps/electron/proxy
files belong to #6620, not this PR); keeps only the author's changes.
Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>
* fix(electron): actually enable materializeSymlinks on the electron standalone path
The option existed in assembleStandalone but no production callsite passed it,
so packaged builds still shipped absolute symlinks into the build machine's
worktree for Turbopack hashed externals (better-sqlite3-<hash>, sqlite-vec-<hash>)
— verified by dpkg -c on a freshly built .deb. One-line enablement on the
electron prepare path, which is exactly the surface #6724/#6594 report.
Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>
---------
Co-authored-by: huohua-dev <258873123+huohua-dev@users.noreply.github.com>
Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>
With --depth=1 the base ref is grafted, so 'git diff base...HEAD' resolves a
wrong merge-base for PR branches that recently merged the release branch. The
three-dot diff then attributes ALREADY-MERGED sibling PRs' changes to the PR
under test, producing false high-signal reds (deleted test files / weakened
asserts that exist in no ref reachable from the PR).
Observed live on #7329: the job blamed it for tests/unit/ui/provider-plan-config.test.tsx
(deleted by an unrelated merged PR) and for #7106's antigravity files. Local
reproduction with full history returns PASS for the same head.
The job's checkout is already fetch-depth: 0, so the full base fetch only
updates the ref — negligible cost.
* fix(dashboard): align onboarding welcome feature cards vertically
* fix(dashboard): align onboarding tier content
* chore: scope onboarding PR to UI fix
* i18n(pt-BR): add onboarding.tier.flowCaption + afterSetup keys
The two new tier keys added to en.json were missing from pt-BR.json, tripping
the i18n-pt-br no-drift test (#6695). Add their pt-BR translations.
Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>
---------
Co-authored-by: Diego Rodrigues de Sa e Souza <8016841+diegosouzapw@users.noreply.github.com>
* fix(dashboard): add vision-capability toggle for custom OpenAI-compatible models (port from 9router#1904)
detectVisionInput()/getCustomVisionCapabilityFields() already honoured an explicit
supportsVision flag on a custom-model record, but there was no way to set it: the
POST/PUT /api/provider-models Zod schema and updateCustomModel()/addCustomModel()
silently dropped the field, and the 'Custom Models' add/edit UI had no checkbox at
all. Self-hosted/local backends that don't self-report an image input modality
(OpenRouter-style architecture.input_modalities) therefore had no way to be flagged
vision-capable, so the vision tag never appeared and image inputs were rejected.
Reported-by: nguyenphi37 (https://github.com/decolua/9router/issues/1904)
* refactor(dashboard): extract providerCredentialText from providerPageHelpers to respect the file-size gate
providerPageHelpers.ts is a frozen god-file (cap 1053, split(\n).length metric)
and this PR's own +3 lines (the #1904 supportsVision field) pushed it to 1054,
failing check:file-size. Extract the cohesive providerText utility + the 4
web-session-credential label/hint/title helpers into a new leaf module
(providerCredentialText.ts), re-exported from providerPageHelpers.ts for
backward compatibility so all existing import sites keep working unchanged.
File now sits at 946 lines, well under the frozen cap.
* refactor(db): extract tri-state override helper to keep the complexity ratchet at baseline
The #1904 supportsVision override added a second copy of the "absent keeps /
null clears / else coerce" block already used by preserveOpenAIDeveloperRole,
pushing updateCustomModel to 84 lines and check:complexity to 2057 > 2056.
The file-size failure was masking this one: the gate exits on its first red,
so complexity never ran until providerPageHelpers was back under its cap.
Fold both blocks into applyTriStateBooleanOverride(). Behavior is unchanged —
updateCustomModel is back under max-lines-per-function and the global count
returns to the 2056 baseline (cognitive-complexity stays at 890).
* Use OpenAI chunks for early chat keepalives
* Update keepalive assertion to match chat completion chunk format
---------
Co-authored-by: Diego Rodrigues de Sa e Souza <8016841+diegosouzapw@users.noreply.github.com>
* fix(combo): detect empty content_block in streaming SSE peek (port from 9router#1382)
The bounded SSE peek in validateResponseQuality() treated ANY content_block_start/delta/stop event as proof of real output and stopped buffering immediately, without checking whether the block actually carried text/tool_use content. Some upstreams (reported: DeepSeek, GLM via claude→openai translation) can open and close a text content_block with empty text and no tool_use on tool-heavy requests — the gateway logged success and forwarded a client-visible empty completion, and combo routing never failed over to the next model.
Track real content separately from 'a content_block_* event was seen': a tool_use/redacted_thinking block start is self-evidently real signal, a text/thinking block start is not (real content only confirmed via a subsequent delta carrying non-empty text/thinking, or an input_json_delta streaming tool arguments). A completed lifecycle (message_start + message_delta/stop) that never produced real content now fails validateResponseQuality(), matching the existing content_filter empty-stream detection path (#3685).
Reported-by: heishen6 (https://github.com/decolua/9router/issues/1382)
* refactor(combo): extract SSE lifecycle applier to keep the complexity ratchets at baseline
The #1382 empty-content_block peek added a branchy switch inline in
parseAccumulatedSse, pushing check:complexity to 2057 > baseline 2056.
Move the switch to a module-level applySseLifecycleEvent() and hold the
four lifecycle booleans in a single SseLifecycleFlags object threaded
through it, so the closure no longer copies flags in and out per event.
The per-event predicates (content_block_start / content_block_delta /
message_delta) are split into small guard helpers, which keeps the
applier flat — cognitive complexity punishes nesting, and an earlier
switch-only extraction traded the cyclomatic ratchet for a cognitive
regression at 891 > 890.
Logic is unchanged; both ratchets are now green (complexity 2055,
cognitive-complexity 890) and the #1382 regression tests still pass.
A fusion combo fans every panel model out in parallel and buffers each
model's full response text in memory simultaneously. With the runtime heap
capped by Dockerfile's OMNIROUTE_MEMORY_MB (default 1024MB), a large panel
(reported: ~73 models via an 'auto' combo with strategy: fusion) with
sizable concurrent responses can exceed the heap ceiling and OOM-crash the
whole container instead of failing one request.
handleFusionChat now rejects panels above a configurable hard cap
(FUSION_DEFAULTS.maxPanel = 40, overridable per-combo via
fusionTuning.maxPanel) with a clean 400 before fan-out begins.
Reported-by: Phong Vu (@fontvu) (https://github.com/decolua/9router/issues/1905)
* fix(api): check Vercel SSO-protection PATCH response on relay deploy (port from 9router#1037)
The Vercel relay deploy route disabled Deployment Protection (SSO) by firing a PATCH request with .catch(() => {}) and never checking res.ok. When Vercel rejects or no-ops the PATCH (plan doesn't allow disabling protection, an under-scoped token, etc.), the relay was still saved and activated as a healthy proxy pool, and later requests routed through it failed with an undiagnosed 403 Access denied from Vercel's own deployment protection — indistinguishable from an upstream-provider rejection (e.g. Codex/ChatGPT edge-IP blocking).
Extract disableSsoProtection() to check the PATCH response and surface an ssoProtectionWarning in the deploy response when it fails, so the failure source can be diagnosed instead of silently masked.
Reported-by: Rico Aditya (@ricatix) (https://github.com/decolua/9router/issues/1037)
* refactor(api): extract vercel-deploy POST helpers to keep the cognitive-complexity ratchet at baseline
The SSO-protection check added to POST pushed its cognitive complexity from
15 to 21, regressing the cognitive-complexity ratchet (891 > baseline 890).
Extract two pure helpers with identical behavior:
- buildDeployErrorResponse(): the sanitized non-ok Vercel deploy response
- resolveSsoProtectionWarning(): the SSO PATCH check + warning string
POST now reads as a flat sequence of guards. No behavior change.
* fix(cli): remove MITM DNS spoof entries before killing server process (port from 9router#1809)
stopMitm() killed the spawned MITM server process first and only removed the /etc/hosts DNS-spoof entries afterward. During that window any client whose DNS still resolved a target host to 127.0.0.1 but whose MITM listener was already dead got connect ECONNREFUSED 127.0.0.1:443 — exactly the community-confirmed workaround (stop DNS before stopping the server) proves. Swap the two steps so DNS is always cleared first, mirroring the ordering already used by repairMitm() and handleExitCleanup().
Reported-by: dionisius95 (https://github.com/decolua/9router/issues/1809)
* refactor(mitm): extract repair planning out of manager to respect the file-size cap
The #1809 DNS-before-kill ordering fix pushed src/mitm/manager.ts to 813
lines, over the 800-line cap check:file-size enforces for non-frozen files.
Move the pure repair-planning pieces (collectManagedHosts, the RepairPlan
shape and its filesystem/cert/DNS sweep) into a sibling src/mitm/repair.ts.
The in-memory session bookkeeping repairMitm() owns — cached sudo password,
orphaned flag, PID file — deliberately stays in manager.ts, so the seam is
"plan the repair" vs "own the session".
manager.ts is now 731 lines; behavior is unchanged. The DNS-first ordering
fix and its regression guard (tests/unit/mitm-stop-dns-before-kill-1809.ts)
are untouched and still pass.
* fix(mitm): split stopMitm() DNS/kill steps to fix complexity ratchet regression
stopMitm()'s new DNS-before-kill ordering (#1809) pushed its cyclomatic
complexity to 18 (max 15), regressing the complexity ratchet from 2056 to
2057. Extract the DNS-removal step and the process-kill step (in-memory +
PID-file fallback) into two private helpers, mirroring the existing
performRepairSteps() extraction pattern in repair.ts. Behavior unchanged;
complexity back at 2056 (cognitive-complexity drops to 889, one under
baseline).
parseInnerCall only split arg segments on a newline between the arg name
and its value. Cursor's live Composer/Auto output has been observed using a
single space instead, so those segments were treated as one long
(space-containing) arg name with an empty value, silently no-opping
Write/tool calls for Composer/Auto models.
Fall back to splitting on the first whitespace boundary when no newline is
present in the segment.
Reported-by: way-art (https://github.com/decolua/9router/issues/1811)
* fix(cli): verify better-sqlite3 native binary is actually loadable (port from 9router#2493)
isBetterSqliteBinaryValid() only checked the .node file's magic bytes (ELF/Mach-O/PE header), never whether the binary was built for the ABI (NODE_MODULE_VERSION) of the Node runtime that loads it. A stale or foreign-ABI binary passed the check and then segfaulted the process on the first database call instead of triggering a rebuild via npmInstallRuntime(). The fix adds a real load probe (require() in a throwaway subprocess) after the magic-byte check, so an incompatible binary is now correctly reported as invalid and the runtime self-heal reinstalls it.
Reported-by: Manikandan (@mrprohack) (https://github.com/decolua/9router/issues/2493)
* chore(changelog): move #2493 entry to changelog.d fragment
Consistency with the repo's canonical changelog.d/fixes/ workflow (avoids
merge-storm re-conflicts from editing CHANGELOG.md directly).
* fix(executors): forward X-Session-ID/X-Title agent metadata headers (port from 9router#2413)
Custom agent clients (e.g. non-OpenCode providers) commonly send X-Session-ID and X-Title headers for upstream request tracking/attribution, but forwardOpencodeClientHeaders() only forwarded x-opencode-* keys plus User-Agent, silently dropping these for every client. Extends the existing case-insensitive allowlist forwarding path with x-session-id/x-title.
Reported-by: Atikur Rahman Chitholian (@chitholian) (https://github.com/decolua/9router/issues/2413)
* chore(changelog): move #2413 entry to changelog.d fragment
Consistency with the repo's canonical changelog.d/fixes/ workflow (avoids
merge-storm re-conflicts from editing CHANGELOG.md directly).
Root cause: validateOpenAICompatibleProvider's chat-completions probe fallback treated ANY 4xx other than 401/403/429/400 as a silent 'credentials valid' pass with no warning, so a bogus/non-standard model id (e.g. Featherless/OpenRouter vendor/model typos) went undetected at Check time. The first real request then hit the upstream 404 model_not_found and the per-model lockout, holding the model unavailable for the configured reset window with no prior indication anything was wrong.
User-visible effect: 'Check' now returns valid:true with an explicit warning (including the upstream error message when parseable) whenever the chat probe answers 404, so a bad model id is caught before it reaches production traffic and the lockout.
Reported-by: advane204f (https://github.com/decolua/9router/issues/2032)
Root cause: SmartCrusher's system-message guard only excluded role === "system", but Codex CLI (open-sse/executors/codex.ts) sends its instructions/tool-schema turn with role "developer" (the Responses-API equivalent of system used by newer models). Every other system-exclusion guard in this codebase also covers developer (roleNormalizer.ts, contextManager.ts, claudeUpstreamMessages.ts, etc.) except this one, so Headroom happily tabular-compacted JSON arrays embedded in the developer turn (e.g. an update_plan tool schema example), corrupting the instructions the model needs to call the plan tool and breaking Codex CLI plan mode.
Fix: extend the guard in crushMessages()/collectCompactableArrays() (smartcrusher.ts) to skip role === "developer" alongside role === "system".
Reported-by: SingCJ (https://github.com/decolua/9router/issues/2132)
* fix(codex): strip regex lookaround from tool schema patterns (port from 9router#1556)
Codex/OpenAI's Responses API rejects JSON Schema pattern fields using regex lookaround (e.g. ^(?=.*@).+$) with a 400 'regex lookaround is not supported' error. The existing numeric-field sanitizer (coerceSchemaNumericFields) was only wired into the translated-request path (openai-to-claude.ts), not the native codex/openai passthrough path (normalizeCodexTools in open-sse/executors/codex/tools.ts), so lookahead/lookbehind patterns reached upstream unmodified and broke tool calls for clients that emit them (e.g. IDE agent harnesses validating an email field).
Reported-by: evin (@evinjohnn) (https://github.com/decolua/9router/issues/1556)
* chore(changelog): move #1556 entry to changelog.d fragment
Consistency with the repo's canonical changelog.d/fixes/ workflow (avoids
merge-storm re-conflicts from editing CHANGELOG.md directly).
* refactor(codex): table-drive the regex-strip recursion to keep the complexity ratchet at baseline
The #1556 lookaround strip walked every sub-schema field with its own
copy-pasted if-block (properties / patternProperties / definitions / $defs,
then prefixItems / anyOf / oneOf / allOf), pushing stripUnsupportedRegexPatterns
past the cyclomatic threshold and check:complexity to 2057 > baseline 2056.
Collapse the eight near-identical blocks into two loops over the field-name
constants, with the object-map recursion factored into a helper. Same fields,
same traversal order, same behavior — complexity is back at baseline 2056 and
the #1556 regression tests still pass.
* fix(sse): recognize xiaomi-tokenplan mimo as a thinking-mode model (port from 9router#1321)
The reasoning_content injector already handles DeepSeek/Kimi/K2/MiniMax thinking-mode upstreams, echoing a placeholder reasoning_content on assistant turns that lack one. Its THINKING_MODEL_PATTERNS list omitted the xiaomi-tokenplan mimo family, so requests through xiaomi-tokenplan/mimo-v2.5-pro still hit upstream's 400 'reasoning_content in the thinking mode must be passed back to the API', making the model unusable in multi-turn conversations (e.g. Codex CLI). Add a /\bmimo\b/i pattern so mimo models get the same treatment.
Reported-by: z.wl (@xxue-z) (https://github.com/decolua/9router/issues/1321)
* docs(changelog): add fragment for #7098 mimo thinking-model fix
Cloudflare Workers AI free tier (10k Neurons/day, account-wide) returns
429 with body 'you have used up your daily free allocation of 10,000
neurons' which matched no QUOTA_PATTERNS keyword — falling through to
rate_limit (~60s cooldown) instead of quota_exhausted.
Two layers:
1. Provider-specific rule for 'cloudflare-ai' in providerRuleRegistry
(scope: connection — budget is account-wide, not per-model)
2. Defense-in-depth: /daily free allocation/i in classify429 QUOTA_PATTERNS
Tests: 11/11 pass (provider rule + classify429 paths covered).
Closes#6980
Co-authored-by: Diego Rodrigues de Sa e Souza <8016841+diegosouzapw@users.noreply.github.com>
* fix(6954,6953): preserve system role + strip empty-signature thinking blocks
#6954 — System turns misattributed as assistant (claude-to-openai.ts:352)
The ternary `msg.role === 'user' || msg.role === 'tool' ? 'user' : 'assistant'`
mapped any non-user/non-tool role (including 'system') to 'assistant'.
Mid-conversation system turns (Claude format) lost their role on
translation to OpenAI format, causing them to be treated as assistant
output. Fix: add explicit 'system' branch to the ternary.
#6953 — Empty-signature thinking blocks poison Anthropic leg (openai-to-claude.ts)
Non-Anthropic providers (codex/gpt-5.x) synthesize thinking blocks with
signature:''\. On replay, the old code fabricated a DEFAULT_THINKING_CLAUDE_SIGNATURE
to fill the empty signature — but Anthropic rejects foreign signatures with
HTTP 400, permanently degrading combo/blend routes to codex-only.
Fix: strip thinking blocks with empty/missing signatures and redacted_thinking
blocks with empty/missing data entirely. They carry no replayable value.
Tests: 8 new tests (4 per bug), all passing. Existing #5312 and #5945
regression tests still pass — no interference.
* fix(6953): strip only signature:"" thinking blocks, preserve undefined signature
CI caught a regression: translator-helper-branches test had a Claude-format
thinking block without signature field (undefined) that was being stripped
by the original fix. The fix was too aggressive — it stripped both
signature:"" (non-Anthropic synthesized) and signature: undefined
(legitimate Claude-format).
Correct behavior:
- signature === "" (empty string): strip — hallmark of codex/gpt-5.x block
- signature === undefined: preserve with DEFAULT_THINKING_CLAUDE_SIGNATURE fallback
- redacted_thinking data === "": strip
- redacted_thinking data === undefined: preserve with fallback
Added regression test for undefined-signature preservation.
---------
Co-authored-by: Diego Rodrigues de Sa e Souza <8016841+diegosouzapw@users.noreply.github.com>
Root cause: the STREAMING branch of AntigravityExecutor.executeOnce() had no
!response.ok check at all — it unconditionally wrapped the upstream response
body in a pass-through TransformStream, unlike the sibling non-streaming
branch which already built a sanitized error via buildAntigravityUpstreamError.
When Google's 403 error body was binary/non-UTF8 (observed: gzip-magic-byte
payloads), those raw bytes were forwarded verbatim, corrupting the
client-visible error message ('[ERROR] [403]: <control-byte garbage>').
Fix: add the same !response.ok guard to the streaming branch, routing through
buildAntigravityUpstreamError()/buildErrorBody() (hard rule #12) instead of
piping unknown bytes through as if they were an SSE stream.
Reported-by: Duongkhanhtool (https://github.com/decolua/9router/issues/2461)
The release-green workflow already reproduces the release-equivalent gate
on release/** and opens a tracking issue on HARD failures — but main had
no such watch. Under the parallel-cycle model main only receives merged
work at the release squash, so a gate/infra fix that landed only on the
release branch leaves main red the whole cycle, and repo-wide gates
(CodeQL alert count, ratchet baselines) turn EVERY PR into main red on a
check unrelated to its diff. v3.8.49 hit this 3× in one night.
Adds a dedicated main-green job (push to main + the same 3 crons +
dispatch) that checks out main literally (no resolver, no injection
surface), runs the same validate-release-green.mjs, and opens/updates a
'🔴 main branch not green' issue pointing at the companion-PR fix. Gates
the existing release-green job with an if: so a push to main doesn't
re-validate release and vice-versa; schedule/dispatch sweep both.
Detection backstop for the prevention rule in _shared/merge-gates.md §8.
The E2E mock HTTP server's 500 catch sent error.stack straight to the
response body, which CodeQL flags as js/stack-trace-exposure (medium).
It's test-only localhost code, but the repo-wide CodeQL ratchet counts
open alerts across all branches — so this one alert (baseline 0 → 1)
turned the Quality Ratchet red on EVERY open PR into both main and
release, masking whatever each PR actually changed.
Surface error.message instead: clears the alert, keeps a useful signal
for a failing mock route, and doesn't log to stderr (node:test native
runner corrupts its report stream on console output). The test only
asserts status 200, so the 500 body is not checked.
Introduced by the #7304 integration test added this cycle.
* test(dashboard): dedicated regression guard for #6815 density guarantee
Coverage for the #6815 multi-column density guarantee was only ever
asserted incidentally, by two other guards (#7072, #3520) that pinned the
literal sm:grid-cols-2 token. That coupling evaporated the coverage when
PR #7027 migrated the component to a container-driven auto-fit template
and the literal token was removed from both files.
Adds a dedicated guard that simulates, from the shipped className, how
many columns the per-group card grid renders at a wide container width
-- supporting both the breakpoint-ladder and auto-fit mechanisms this
component has shipped with -- and asserts >1 column, without asserting
any specific Tailwind token.
* chore(changelog): fragment for #7291 density guard
typescript-eslint pins a hard upper bound on its typescript peer
(8.64.0 → ">=4.8.4 <6.1.0"). A major TS bump violates it, so the failure is
not one check — it is the whole toolchain at once.
#7068 is the demonstration: dependabot grouped typescript ^6→^7 with six
harmless dev bumps (@types/node, eslint, fast-check, knip, prettier,
typescript-eslint) and turned Build, Lint, Quality Ratchet, Unit (6/8, 8/8),
Integration (1/2, 2/2) and dast-smoke red in a single PR. The six innocuous
updates were blocked by the one that could never pass.
Ignoring the major lets the rest of the group flow on its own. TS majors are a
toolchain migration and deserve their own PR and their own CI run — not a
weekly automated attempt that cannot succeed until typescript-eslint widens the
peer.
Refs #7068
* fix(stream): reconcile encrypted Codex reasoning visibility without mutating upstream item
Resolves the collision between two open PRs on ensureVisibleResponsesReasoningSummary:
#7095 (xz-dev) found that chat clients see nothing when Codex exposes reasoning
only as encrypted_content, and added a visible placeholder — but did so by
mutating item.summary in place. #7176 (JxnLexn) found that same mutation
corrupts the forwarded response item, discarding the encrypted_content shape
Codex needs for follow-up requests, and removed the mutation — but that also
silently dropped the placeholder, so chat clients went back to seeing nothing.
The mutation existed only so a later line could read the summary text back off
the same item. getVisibleResponsesReasoningSummaryText() computes that text
without touching the item, so:
- synthetic response.reasoning_summary_text.delta / .part.done events still
carry the placeholder for chat clients (#7095's goal), and
- the forwarded response.output_item.done payload keeps its original
encrypted_content intact with no fabricated summary field (#7176's goal).
Applied at both call sites #7095 identified: the native Responses passthrough
in stream.ts/passthroughTailProcessor.ts, and the Responses-to-Chat-Completions
translator in openai-responses.ts.
Closes#7095, closes#7176.
Co-authored-by: Xiangzhe <xiangzhedev@gmail.com>
Co-authored-by: Jan Leon <Jan.gaschler@gmail.com>
* test(stream): guard the encrypted-reasoning mutation via the completed backfill path
The output_item.done line is echoed verbatim on the wire, so a re-introduced
item.summary mutation does NOT surface in that event — verified by re-injecting
the mutation, which left the existing assertion green. The mutation does surface
in the response.completed snapshot, where the captured reasoning item is
re-serialized when upstream sends an empty output (store: false).
Adds that case, which fails as expected when the mutation is re-introduced,
making the #7176 half of the reconciliation an enforced regression guard rather
than an incidental property of the current code path.
Co-authored-by: Xiangzhe <xiangzhedev@gmail.com>
Co-authored-by: Jan Leon <Jan.gaschler@gmail.com>
---------
Co-authored-by: Xiangzhe <xiangzhedev@gmail.com>
Co-authored-by: Jan Leon <Jan.gaschler@gmail.com>
