When no explicit provider is requested and the auto-selected cheapest
provider has no credentials, executeWebSearch ran the fallbackOnly loop
first. duckduckgo-free (costPerQuery 0, authType none) always won there
with an empty credentials object, so a configured paid connection such as
serper-search was silently ignored and the caller got success:true with
zero results.
Move the sweep for other credentialed regular providers ahead of the
fallbackOnly loop (and exclude fallbackOnly ids from it, so a free
last-resort provider never outranks a configured one on cost). The
fallbackOnly loop stays as the true last resort.
The chat path only appeared correct because duckduckgo-free happened to
fail there and handleSearch retried the alternate provider; on
/v1/responses it "succeeded" with no results.
Closes#11524
Vitest went red on 'synchronizes upstream models only when autoFetchModels is
explicitly true' with new URL throwing inside a fetch dispatched from
Timeout._onTimeout (useProviderModels.ts:69). It is intermittent: green in the two
previous CI runs, green every time in isolation, red only under the ui suite's 20
parallel workers.
The hook schedules its auto-sync in a setTimeout whose callback only checks the
flag on entry — and that flag stays false while the component is
mounted. Both tests asserted first and unmounted last, so under contention the
timer escaped the test window, fired after afterEach had already run
vi.unstubAllGlobals(), and reached the REAL fetch with a relative URL.
Unmounting before the assertions closes the window: cleanup flips , the
callback returns early, and the calls already recorded on fetchMock are still there
to assert against. No assertion changed.
Not a regression from this cycle — the file's last change is fd76271515 (#10603).
Fixed rather than tracked because an intermittent red in a blocking job is worse
than a permanent one: it teaches people to re-run instead of to look.
Refs #10692
Electron Package Smoke — a packaging defect that had been hidden behind another
packaging defect for nine days. Once the loginHeaderCapture fix let the main process
start, the server underneath died on 'Cannot find module next': resources/app/server.js
shipped without resources/app/node_modules.
electron-builder discards the ROOT node_modules in code, not by configuration —
app-builder-lib/out/util/filter.js:42 has a hard-coded `if (relative === "node_modules")
return false` that runs before any filter pattern. The second extraResources entry
pointing INTO ../.build/electron-standalone/node_modules is what sidesteps it, because
those relative paths are never equal to "node_modules". #10325 removed that entry as an
apparent duplicate and flipped the test to assert "exactly once", freezing the
regression as if it were the contract. Restored, and the unit guard now pins both
entries — proven by mutation: reverting package.json to the post-#10325 shape fails the
guard 3/4, restoring it passes 4/4.
group-b-quota-plans-config — the assertion was impossible to satisfy on ANY route, and
the page was never broken. layout.tsx hands the whole message catalogue to
NextIntlClientProvider, React serialises that prop into the RSC payload, and en.json
carries "Internal Server Error" twice, so page.content() always contains it: probing
/dashboard, /dashboard/costs, /dashboard/settings and /login showed the string present
with every page rendering fine, and a pageerror probe on the failing run captured zero
client exceptions. This is the same trap that killed the sibling not.toContain("500")
in fc77100c3f ("raw HTML is unreliable") — that one was removed, this one was kept.
Now asserts on rendered text, which still catches a real error boundary. The pageerror
capture stays: the CI failure carried no stack trace, which is why it was misread twice.
Integration — 10 of the 14 shard-2 reds, all sibling-test gaps behind security fixes:
monitoring health now takes a Request and requires management auth (GHSA-mvf8-qc78-5mxm);
the OAuth import routes moved to requireManagementAuth (GHSA-mg76) — the test accepts
both guard shapes and gained a stronger anchor that every exported handler awaits a
guard on its own request, mutation-verified; skill tool names are derived from
encodeSkillToolName() and the fake upstream now returns the encoded name so
decodeSkillToolName() is exercised too; previous_response_id now fails closed (#10262);
proxy_logs persist as an async batch (#11182) so the test flushes first;
providerQuotaOverrides joined GET /api/resilience (#9871); the reasoning fixture used a
model that stopped being thinking-incompatible, replaced and pinned with a premise
assert so it cannot rot silently again.
A vacuous assert.ok(true, "all 10 streams completed without hanging") was replaced with
real anchors — content must arrive on every stream and the active Timeout count must not
grow.
Four are deliberately left red rather than aligned, each now tracked: #11551 (the
/v1/models after() wiring is dead — the route passes a third argument to a two-parameter
function and catalogCache never imports after, so the #8728 contract is unimplemented),
#11552 (~27% of requests emit an extra discarded upstream call; the delivered
distribution is exactly 0.70, so weighted routing is correct and the waste is the real
finding), the fixed-account combo pin (aligning it would destroy the per-step attribution
the test exists for), and the web_search fallback already tracked as #11524.
Package Artifact — the provenance stamp I added last round used git rev-parse HEAD, which
under pull_request is the ephemeral merge commit and therefore never an ancestor of the
release branch. Now takes the PR head sha.
Refs #10692
All three are the sibling-test gap again: a PR moved a contract, updated its own
tests, and left these behind. None is a production defect — in two of the three the
production side is a deliberate security fix.
v1-contracts-behavior (4 failures, one cause): the job env sets INITIAL_PASSWORD,
which makes isAuthRequired() true, and #9320 (b07182c72a) made the /v1 catalogue
gate on-by-default instead of opt-in via settings.requireAuthForModels. The four
contract reads were calling the catalogue routes with no credential and correctly
getting 401. Bisected the job's four env vars to confirm INITIAL_PASSWORD alone
reproduces it (5 pass / 4 fail with it, 9 / 0 without). The tests now send a Bearer
token; the shape assertions are untouched, and the auth contract itself stays owned
by tests/unit/v1-models-auth-leak-9320.test.ts rather than being duplicated here.
opencode-config-startup: two independent drifts. OPENCODE_VERSION was pinned to
1.18.8 while the installed opencode-ai is 1.18.18 (Dependabot 7f6958960c, #10626) —
now read from require("opencode-ai/package.json").version, which is exactly as
strict but cannot drift on the next bump. And the no-limit-metadata case asserted
limit === undefined, but #11054 made the generator always emit a limit; it now pins
the actual fallback {context: 128_000, output: 8_192} instead of an absence.
memory-pipeline: #11040 (GHSA-cpv3-xr7r-xf8q) made the resolved caller principal
always win over a caller-supplied apiKeyId, so a spoofed id can no longer write into
another principal's store. That PR updated the unit sibling but not this one. The
test now asserts the stronger property — and deliberately not just the absence: the
spoofed principal's store is empty AND the caller can still read the entry, which
proves the write was redirected rather than dropped and keeps the emptiness check
from passing vacuously with a disabled store. (The old assertion was count === 0,
which a switched-off memory store would satisfy.)
Assertion counts: 43 -> 43, 13 -> 14, 76 -> 81. Nothing weakened or removed.
Verified: 24/24 pass, with and without the CI env vars.
Refs #10692
The remaining ui-shard reds were one class, not six bugs: every one of them did
`await import(<heavy component>)` INSIDE an `it()`, so Vite's transform of the
dependency graph was billed to that test's timeout. Measured costs against the
budgets they had to fit in:
ProxyRegistryManager 86s import vs 30s / 60s / 5s budgets (render itself: 567ms)
claudeTlsClient ~12s import vs 5s default
useProviderConnections 1050-line hook, whole dashboard graph, vs 5s default
That is why they looked like cross-file pollution: on an idle box the import
squeaked under the limit, and under the ui suite's 20 parallel workers it did not.
Running claudeTlsClient ALONE on a loaded box reproduces it — the trigger is CPU
contention, not a neighbouring file. The sibling chatgptTlsClient/grokTlsClient
tests import the same graph and never fail, because they import statically at
module scope, where the cost falls on the collection phase which has no per-test
budget. Every fix here does the same: static import or a beforeAll with its own
budget.
AutoComboCatalog also explains its own blast radius: the timeout aborted inside an
open act(), leaking an unbalanced act scope that then failed the file's three
remaining tests in ~20ms with 'overlapping act() calls'. One slow import, four reds.
CoolingConnectionsPanel is the one production change. It imported providerText from
the ../providerPageHelpers barrel, but that symbol is DEFINED in the
../providerCredentialText leaf and only re-exported by the barrel — which drags
providerRegistry (352 providers) and the rest of the provider-page graph into a
"use client" component for one string helper. Verified before accepting: the
component used nothing else from the barrel, the barrel has no top-level
side-effect to lose (the empty-registry hazard this repo has hit before does not
apply), typecheck:core is clean, and the panel's first test drops from ~4s to 95ms.
The import was suboptimal, never broken — the screen was not failing for users.
No assertion was weakened anywhere. expect() counts are unchanged (25/25, 4/4) or
up by one (AutoComboCatalog 11 -> 12); the #8855 autofill sentinels, the
data-1p-ignore / data-lpignore guards and the dead-status round-trip are intact.
The #5918 TDZ guard was proven still live by mutation, not by absence of red:
moving useProxyBatchOperations(load) above its const reproduced
'ReferenceError: Cannot access load before initialization' in 207ms, then the
production file was restored (diff empty).
tests/unit/ui under load: 17 failed files / 45 failed tests -> 4 failed files /
4 failed tests, none of them these. The four left are compression-guidance-7530,
compressionPanel, compressionUltraTier and lobe-provider-icons-stepfun, untouched
and uninvestigated.
Refs #10692
Two gates in the Lint job, both inherited — each was hidden behind the one before it.
i18n value drift: #11283 rewrote sidebar.trafficInspectorSubtitle in en.json without
touching the 42 translations, so 32 locales kept serving a sentence the English no
longer says. Most take the documented __MISSING__: placeholder, which makes the
runtime serve the corrected English until the translation pipeline catches up.
Three do not:
- vi cannot take a placeholder at all — tests/unit/i18n-vi-completeness.test.ts
bans any __MISSING__/__TODO__ value outright, so it needs a real translation.
- pt and pt-BR are translated for real rather than placeheld, because a placeholder
there means this project's own maintainer reads the sidebar in English.
Each file changed by exactly one line; the JSON was not reserialised wholesale.
agent-skills-sync: skills/omni-inference/SKILL.md was missing the ElevenLabs voices
and speech-to-text routes added by #11312, so the generator reported one file out of
date and the gate exited 2. Regenerated — purely additive, 48 lines, no deletions.
Verified: check-ui-value-drift PASS, i18n:check-ui-coverage PASS (42 locales),
i18n-vi-completeness 5/5, check:agent-skills-sync 46 unchanged.
Refs #10692
None of these are cycle regressions. The Vitest job runs test:vitest (mcp shard)
then test:vitest:ui; the mcp shard was failing on a missing glm-5.3-max and aborted
the job before the ui shard ever ran. Fixing that shard this cycle unmasked 34 ui
failures that had been broken since 18-23 Aug — four separate PRs that moved a
contract and updated their own tests but not their siblings.
- ProviderCard gained useRouter() in #10448; four test files render it without
mocking next/navigation and died on 'invariant expected app router to be mounted'.
The sibling created alongside #10448 already had the mock — it just was not
applied to the other four consumers.
- SkillCoverage gained a required config category. The four fixtures in
agent-skills-page still described only api/cli, so the component read
config.have off undefined. Values were chosen per scenario rather than pasted:
full coverage gets 2/2 so its bar stays emerald, the amber fixture gets 3/4 so it
stays amber. CoverageBar renders api -> config -> cli, so the new bar lands in the
MIDDLE and the cli assertions moved from index [1] to [2]; without that the cli
checks would have passed while measuring the config bar. The aria test now pins
all three bars.
- CliAgentsPage hardcoded AGENT_IDS, which had already drifted once (6 -> 8 with
omp/letta, per its own comment) and drifted again with prime-agent (#11166). It is
now derived from CLI_TOOLS. This is why an agent missing from that list is not
cosmetic: it never enters the status map, defaults to not_installed, and adds a
phantom card to the filter and count tests. Deriving keeps the fixture in sync by
construction instead of waiting for the next agent.
- claudeTlsClient asserted proxyUrl was undefined inside a test literally named
'falls back to env var when per-call proxyUrl not provided' — it pinned the old
behaviour where testOverride bypassed proxy resolution. #10910 moved resolution
ahead of the override on purpose ('so test overrides and the real path both see
it'), so the assertion now checks the fallback the test name promises.
test:vitest:ui goes from 34 failures to 14. The remaining 14 sit in six files none
of this commit touches (AutoComboCatalog, CoolingConnectionsPanel, ProxyRegistryManager
x2, connectionsSearchFilter) plus one claudeTlsClient case that passes in isolation
and only fails in the full run — i.e. cross-file pollution. They need a clean
environment to judge: this devbox resolves part of its tree through a stray pnpm
store and has already produced one phantom failure count this cycle.
Refs #10692
All four predate this session — each reproduces identically on f95b03d70 (2026-08-24),
so none is a cycle regression. Draining them here because the release pre-flight is
where inherited reds get resolved.
1. Package Artifact: the job runs `build:cli`, which assembles dist/ but never writes
dist/BUILD_SHA — only `build:release` does, via write-build-sha.mjs. The #10427
provenance guard inside check:pack-artifact then rejects the artifact as
untraceable, and rejects it even under OMNIROUTE_ALLOW_CANARY_BUILD. The job's
build+validate pair was structurally incompatible and failed 100% of the time.
Stamps the SHA between the two steps.
2. Electron Package Smoke: electron/package.json's build.files allowlist enumerates
each lib/*.js by hand and never got lib/loginHeaderCapture.js, added alongside its
require() in #9984. The file therefore stayed out of app.asar and the packaged app
died at startup on 'Cannot find module ./lib/loginHeaderCapture'.
3. proxy-pipeline: the breaker assertion grepped chat.ts for executeChatWithBreaker(,
but that call moved behind the chatDispatch.ts seam. Rather than drop the check,
it now pins both hops — chat.ts dispatches through the seam and the seam calls the
breaker — so the extraction cannot silently take the breaker off the path.
4. skills-pipeline: #9058 began encoding skill tool names as omr_skill_<base64url>
because providers require ^[a-zA-Z0-9_-]+$, and these assertions still expected the
raw name@version. They now derive the expected name from encodeSkillToolName(), the
same helper production uses, so the test tracks the contract instead of duplicating
it. Only the assertions about names on the wire were converted; the identifiers
passed straight to skillExecutor.execute() stay raw, because those are not encoded.
Integration suite for these two files: 54/55. The one still red —
'web_search fallback preserves Responses API output' — is a separate pre-existing
defect, deliberately left failing rather than papered over: on the /v1/responses path
resolveSearchCredentials() returns null for the seeded serper-search connection, so
executeWebSearch.ts:185-200 falls through to the cheapest fallbackOnly provider
(duckduckgo-free) and the results come back empty. The sibling chat-path test seeds
identically and does resolve serper-search. Needs its own investigation.
Refs #10692
The previous heap bump only touched test:unit:ci:shard, i.e. the node the shard
script spawns. The process that actually runs out of memory is the `c8` wrapper
around it — it aggregates ~577 MB of raw V8 coverage JSON — so the ceiling stayed
at the V8 default (~4 GB) and the shards kept aborting at ~4083 MB, byte for byte
the same failure. Setting NODE_OPTIONS on the step covers c8 and every child,
which is the pattern the coverage-merge job already uses.
Also prunes three eslint suppression entries whose violations no longer exist:
videoBridgeContactSheet.ts and videoBridgeRuntime.ts (no-unused-vars, fixed
during this cycle) and cli-oneproxy-commands.test.ts (no-explicit-any 14 -> 13,
a consequence of restoring the real mock in that test). Stale entries make
`npm run lint` exit 2 with 'There are suppressions left that do not occur
anymore'. Pruned and verified on an uncontaminated checkout, not the devbox.
Refs #10692
The 8 unit shards run under V8 coverage instrumentation, which retains far more
memory than the bare suite. With the 4096 MB ceiling they began aborting with
exit 134 ("Ineffective mark-compacts near heap limit") at ~4086 MB as the
provider catalogue grew during the v3.8.50 cycle: every test in the shard passed
and the process died at the end, which reads as a test failure without being one.
Aligns test:unit:ci:shard and test:unit:serial with the 8192 MB the non-sharded
variants already use. GitHub-hosted runners have 16 GB, so the headroom is real.
Validated by the CI run on this commit — the shards are the gate.
Refs #10692
Adds 131 consolidated bullets (45 features, 71 fixes, 15 maintenance)
covering the ~490 user-facing commits and the ~100 chore/ci/test/refactor/docs
commits that landed in the cycle without a CHANGELOG entry, grouped by
subsystem and citing their PR references.
Uncovered report: 594 -> 175 (the remainder are commits carrying no #N in
their subject, which the matcher can never resolve; they are covered in
prose).
Release reconciliation (Phase 0a.1). `scripts/release/aggregate-changelog.mjs`
folds each changelog.d/<section>/*.md fragment into its heading in the living
[3.8.50] section and deletes the fragment, which is the whole point of the
fragment convention: two PRs never touch the same file, so the CHANGELOG never
conflicts mid-cycle and no bullet is eaten by a merge auto-resolve.
Section bullets 731 -> 1041. The remaining uncovered commits (mostly merges from
#11397 onward, which landed without a fragment) are reconciled separately.
`next dev` writes and re-adds this block (see
node_modules/next/dist/server/lib/generate-agent-files.js), so leaving it out of
a diff only recreates the uncommitted change on the next dev run. Committing it
keeps the working tree clean, which is what the block's own note prescribes.
hasBindMountAt() accepted ANY mount as evidence that a would-be CLI config
write reaches the operator's host: it matched on the mount point alone and
never looked at the filesystem type. An in-memory mount therefore cleared the
ephemeral flag, so guardCliConfigWrite() let the write through and both
POST /api/cli-tools/apply and the dashboard's guide-settings writer answered
200 instead of the safe 422 that #10057 added.
That is the exact case the guard exists to refuse, and the worst one: a
container running with `--tmpfs /tmp` (or a home on tmpfs) loses the file even
before the container is recreated, while the UI reports success.
Parse the filesystem type from mountinfo (the field after the lone "-"
separator) and skip mounts backed by RAM or kernel state. Real bind mounts
(ext4/xfs/nfs/virtiofs/fuse.*) still count, including one nested under a tmpfs
path, so the compose `host` profile is unaffected. A line carrying no
separator proves nothing and is skipped too.
Regression cover added to tests/unit/container-env-detect.test.ts; this also
un-reds tests/unit/cli-tools-apply-container-422.test.ts and
tests/unit/api/cli-tools/apply-container-guard.test.ts, which were failing on
any box whose /tmp is a tmpfs.
Two unrelated real reds on the release tip:
* fix(providers): flag MiniMax M3 as multimodal on the Volcengine Ark plans.
d732cf615 ("feat(volcengine): add Ark plan providers") added
volcengine-agent-plan/minimax-m3 and volcengine-coding-plan/minimax-m3
without supportsVision, breaking the LEDGER-4 invariant that every
minimax-m3 registry entry except PromptQL (text-only upstream) is flagged
multimodal. Every other provider carrying the model (opencode-zen,
opencode-go, bazaarlink, ollama-cloud, codebuddy-cn, trae) sets it.
Registry metadata defect, not a stale test.
* test(antigravity): align the empty-projectId onboarding test to the
contract shipped by #11284/#11358 (6de542b9b). That change made an
onboardUser 200 whose body carries NO cloudaicompanionProject mean Google
BYOP — no project was created and none ever will be — so it short-circuits
before the retry loadCodeAssist. The older test still mocked onboardUser
with the bare { done: true } BYOP shape while asserting the retry path, so
it pinned a contract that was deliberately moved. The mock now returns a
real onboarding-success body; every assertion is kept, and the id in the
onboard body deliberately differs from the expected one so the test still
proves the projectId came from the retry discovery.
Refs #11284
All three guards were drifting behind legitimate cycle growth, not catching a
defect. Nothing was weakened: no assertion removed, no floor lowered, no
blanket-allow added.
providers-constants-split: APIKEY_PROVIDERS 231 -> 233. The delta is exactly the
two Volcano Ark plan providers (volcengine-agent-plan, volcengine-coding-plan)
added to the regional family in d732cf615. The invariant the guard exists for
still holds, measured on the tip: 233 merged keys, 233 unique, family sum 233
(gateways 92 + frontier-labs 25 + inference-hosts 29 + enterprise-cloud 17 +
regional 43 + specialty-media 27) with an empty cross-family duplicate set and
an empty symmetric difference between the merged object and the family union -
so the six files are still a strict partition, no loss and no dup.
openapi-coverage: the operation floor (34.6%) is untouched. The cycle grew the
denominator 985 -> 1002 while covered only moved 343 -> 345 (34.4%). Fixed by
DOCUMENTING five real public operations rather than moving the floor, taking it
to 350/1002 = 34.9%: GET /api/health, GET /api/v1/voices, POST
/api/v1/speech-to-text, POST /api/v1/text-to-speech/{voiceId} and GET
/api/v1/explain/routing. Each entry was written from the route source (auth
mode, path-param pattern, limit clamp, upstream relay behaviour and the 400 /
401 / 429 branches), not from memory.
hard-session-lease-bypass-inventory: three new connection-query sites
classified, none silenced. open-sse/services/combo.ts
(readConnectionForCooldownGate) reads the row backing the pre-dispatch
persisted-cooldown gate, so it sits on the routing path and joins the class-B
list next to combo/providerWildcard.ts and autoComboCandidates.ts.
src/lib/providers/volcenginePlanBinding.ts and
src/lib/providers/volcPlanAutoSyncBackfill.ts are connection persistence, not
dispatch - the first resolves update-vs-create during connect, the second is a
one-shot boot backfill of a providerSpecificData flag with no upstream call -
so both stay class C alongside oauth/connectionPersistence.ts.
93da24cd7 ("fix(providers): reject reserved provider prefixes on
compatible-node create/update") made createProviderNodeSchema reject any
prefix that is a built-in REGISTRY id or alias. "cc" is the alias of the
built-in `claude` provider, so the two provider-nodes create cases in
cc-compatible-provider.test.ts started getting a 400 schema rejection
before the route ever reached its feature-flag gate (403) or the create
path (201) — the guard PR updated its own tests but missed this sibling
file, leaving a base-red on release/v3.8.50.
The operator-chosen prefix is incidental to what these cases assert (the
ENABLE_CC_COMPATIBLE_PROVIDER gate, the dedicated
anthropic-compatible-cc- id prefix, baseUrl sanitization and the nulled
modelsPath), so switch it to a non-reserved "cc-proxy". No assertion was
removed or loosened.
The #11353 regression test pins an ABSOLUTE upstream reset instant
(2026-08-29 21:01:21) in its production fixture body but measured the
resulting cooldown against the real wall clock. The remaining window
therefore shrank every day: from 2026-08-25 it dropped under the
5-day floor the two assertions use, and past 2026-08-29 it would
parse to null and collapse onto the 24h WEEKLY_QUOTA_COOLDOWN_MS
default - a guaranteed future red.
The shipped parser (parseIsoDateTimeResetMs / parseDayGranularityResetMs
/ buildWeeklyQuotaFallback / checkFallbackError) is correct: it returned
the real multi-day reset, just measured from today instead of the
fixtures NOW. Freeze Date at NOW via node:test mock timers in the two
time-dependent cases so they assert the parser rather than the calendar.
No assertion weakened, no production code touched.
The router-eval CLI test spawns the CLI with spawnSync and asserts stderr stays empty. NODE_TEST_CONTEXT is inherited by those children, so since #10432 (guard #10428) resolveWritableDataDir() detects a test context with no DATA_DIR and warns on stderr before falling back to a throwaway dir - 194 chars that broke three cases. Pass an isolated DATA_DIR in the child env (the resolution the guard message itself prescribes) instead of loosening the assertions.
PR #11418 (S2 topology sanitisation) removed the hardcoded
localhost:20128 from both well-known agent-card routes and made them
derive the base URL from `request.nextUrl.origin` via
`getBaseUrl(request)` (src/lib/wellKnown.ts). That changed the handler
contract: `GET` now requires the request Next.js always passes it.
Three sibling test files were never aligned and still invoked the
handler as a bare `GET()`, so every case blew up with
`TypeError: Cannot read properties of undefined (reading nextUrl)`
before reaching a single assertion — 8 base-reds from one moved
contract, not from a skill-count drift.
Align the callers to the shipped contract with a local
`makeCardRequest()` helper mirroring tests/unit/security-s1-s2-s4.test.ts
(a Request with a defined `nextUrl`). No assertion was removed,
loosened or skipped; the assert counts are unchanged and the cases now
actually execute.
Refs #11418
Second base-red batch from the release pre-flight, measured on the .113 with a
clean npm ci (the devbox tree resolves eslint-plugin-react-hooks 7.1.1 from a
stray pnpm store instead of the lockfile 7.0.1 and reports 925 phantom errors).
Provider count 350 -> 352, one root cause behind three reds. Two providers
landed this cycle (volcengine-agent-plan, volcengine-coding-plan) without
regenerating the artifacts that quote the count:
- docs/reference/PROVIDER_REFERENCE.md regenerated (gen:provider-reference).
- README / AGENTS / llm.txt (+42 mirrors) / package.json description / 4 SVG
diagrams updated, including the section heading AND the anchor that links to
it, so the link does not break.
- tests/snapshots/provider/translate-path.json regenerated. The diff is purely
additive: 46 insertions, 0 deletions, exactly the two new providers.
GLM effort tiers. #11415 added the explicit glm-5.3-max tier and left two
sibling vitest specs pinning the old 16-model inventory and an empty tier list
for it. Aligned to the shipped contract (inventory order matches glmProvider.ts;
glm-5.3-max declares ["max"]).
Test-masking. Four assert reductions surfaced once the deleted-file signal was
resolved. Three are legitimate and are allowlisted with their reasoning:
#11355 inverted the startup-cooldown contract (preserve future quota cooldowns),
#11280 replaced two unrolled hops with a 3-hop loop that asserts MORE, and the
Gemini 3.5 Flash retirement removed the models those capability asserts described.
The fourth was real masking: #10960 rewrote the oneproxy status test to install a
stream mock, immediately overwrite it with a passthrough to the real fetch, and
assert `calls.length >= 0` — always true. Restored to assert what the test name
claims (the JSON-RPC tools/call carries omniroute_oneproxy_stats and its result
reaches the caller), with a scope note that it pins the MCP client contract
rather than the commander wiring.
Also allowlists the Gemini 3.5 Flash test deletion as _deletedWithReplacement
(the model was retired by 2764812ee4; gemini-models-parser.test.ts pins the new
"excluded from the parsed list" contract), and rebaselines bundleSize
8045 -> 8461 with per-entry measurements — every entrypoint stays far below its
absolute budget.
Two base-reds on the v3.8.50 tip, found by the release pre-flight.
1. #11355 regressed #10534. It replaced the per-window recovery check with an
unconditional `hasActiveCooldown()` stop, which is right for an
upstream-derived cooldown but also blocks the case #10534 exists for: a
Claude-subscription 429 persists a SYNTHETIC 1h rateLimitedUntil because the
upstream sends no parseable reset. When the later poll shows every governing
window has really reset with quota left, holding that synthetic cooldown just
deadlocks the connection for an hour.
The orphaned `windowStillExhaustedAfterRealReset()` helper and the three
unused claudeExtraUsage imports that ESLint flagged were the fingerprint of
this regression, not dead code: they are the two halves of the original gate.
Re-wired as `isQuotaExhaustedCooldownReleasable()`, deliberately narrow —
only lastErrorType "quota_exhausted" is eligible, one still-exhausted or
unknown-reset window keeps the lock, and an extra-usage POLICY block stays
locked even though its quota windows do look recovered in the same fetch.
#11277/#11355 semantics are untouched (both guards still pass).
Regression guard: tests/unit/provider-limits-recovery.test.ts already pinned
this contract and was red on the tip. 15/15 now.
2. The three volcengine-plan connect routes read `request.json()` and handed the
raw fields to a headless-browser login service after ad-hoc typeof checks
(`check:route-validation:t06`, Hard Rule #7). `String(body.code ?? "")` turned
123 into "123" and an absent code into "", both reaching the service as a
plausible SMS code. Now parsed with Zod schemas, before the session lookup, so
a malformed body answers 400 instead of a misleading 404.
New: tests/unit/volcengine-plan-connect-validation.test.ts (8 cases, red
before the fix). Gate: 687 route files scanned, PASS.
Also drops a genuinely dead import (formatVideoTimestamp in videoBridge.ts —
only used inside the helpers module that defines it).
The living release PR #8875 was CONFLICTING, which makes GitHub skip EVERY
pull_request workflow silently (no ci.yml, no semgrep, no DAST). Back-merging
main restores a computable merge ref.
Strategy `-s ours`: main is a stale snapshot of the release line (PR #11088 was
merged into main from a release-tip base, dragging ~5094 files). All 7 main-only
commits were verified as already represented on this branch:
- #11088 ollama capability routing -> ported here as #11271 (6d4c4843e9)
- #11075 shared passthrough providers -> ported here as #11165 (92ef3c71ea)
- #10055 getModelsDevPricing memoization -> present (modelsDevSync.ts)
- #10026 hide health-check excluded models -> present and extended (catalog.ts)
- /_tasks anchored gitignore hardening -> present (.gitignore:288)
- nanoid/dompurify Dependabot bumps -> identical versions
main-only files intentionally NOT carried over:
- changelog.d/fixes/10286-gemini-3-5-flash-thinking.md + its regression test:
the fix landed here as #10450 and was then deliberately superseded by
2764812ee4 "eliminate Gemini 3.5 Flash". The test fails on this branch by
design.
- public/providers/hackclub.svg: provider removed here (migration 162).
- docs/superpowers/**/2026-08-23-qdrant-*: planning artifacts belong in _tasks/
(AGENTS.md), never under docs/.
CodeQL js/incomplete-url-substring-sanitization, alerts #860 and #861:
volcengineConsoleAutoLogin accepted any cookie whose `domain` merely *contained*
"volcengine.com".
That check is an authorization decision, not a string test. The console
auto-login harvests `digest`, `AccountID`, `csrfToken` and `userInfo` out of the
Playwright context and persists them as the operator's Volcengine credentials,
so a cookie set by `volcengine.com.attacker.tld` — or `notvolcengine.com` — was
captured and stored as a provider connection.
Add `matchesCookieDomain()` (open-sse/utils/cookieDomain.ts): exact host or
dot-boundary suffix, leading dots and case normalized on both sides, failing
closed on an empty expected domain. Same shape as the existing
`isAdobeCookieDomain` in adobeFireflyBrowserLogin.ts, which already got this
right.
While sweeping the class, inAppLoginService's cookie capture had the identical
weakness — `c.domain.includes(domain.replace(/^\./, ""))` — with the identical
consequence: a look-alike host's cookie stored as the operator's credential.
CodeQL did not flag it because the expected domain comes from
TOKEN_EXTRACTION_CONFIGS rather than a literal. Both callsites now share the
helper.
tests/unit/volcengine-cookie-domain-suffix.test.ts — 5 tests, red before the
fix, covering the real domains, seven look-alikes, empty/missing input, and the
config-supplied path.
Co-authored-by: Xiangzhe <bakryun0718@proton.me>
`Fast Quality Gates` has been failing on every open PR against
release/v3.8.50 with "2 gate(s) failed: mutation-test-coverage lockfile".
Neither belongs to any feature branch, so they are drained here.
check:lockfile — a transitive dev/optional entry
(libxmljs2 → brace-expansion@2.1.4) landed with a `resolved` URL pointing at
registry.npmmirror.com instead of registry.npmjs.org, which lockfile-lint
rejects as a supply-chain policy violation. Verified before touching it: the
recorded `integrity`
(sha512-hGfVzPxthbf3+2yjg/RBs60cB0FhqBS/zvdV/4wn4/BmN0bNMMHPc4V/BbFieqf1TKAGGAHnY4eSjajCl0f2Xg==)
is byte-identical to the official npmjs tarball's, so the package content is the
same and this is a provenance slip — someone's install ran behind the mirror
registry — not a tampered package. Repointed the URL; `integrity` untouched.
It was the only non-npmjs host in the lockfile (2690 npmjs entries).
check:mutation-test-coverage — two covering unit tests were missing from
stryker.conf.json's tap.testFiles, so their mutant kills did not count:
repro-glm-iso-reset-24h-cap (accountFallback.ts) and
repro-combo-persisted-cooldown-preskip (comboPredicates.ts). Inserted in place.
Both gates verified green locally. The diff is three lines: re-serializing
either file would have reordered a curated list for no reason.
Co-authored-by: Xiangzhe <bakryun0718@proton.me>
Validado em lote combinado (batch-0824g) contra o tip de release/v3.8.50: typecheck:core limpo, gates estáticos OK, 62/62 testes focados passando (S1/S2/S4, tests/unit/security-s1-s2-s4.test.ts, 9/9).
Boa integração com o padrão já existente de peer IP stamped por HMAC (resolveStampedPeer/OMNIROUTE_PEER_STAMP_TOKEN) — reusa em vez de reimplementar, e o header confiável só é honrado quando o stamp token está configurado. S2 remove corretamente a disclosure de topologia hardcoded do agent-card. Obrigado pela contribuição!
Validado em lote combinado (batch-0824g) contra o tip de release/v3.8.50: typecheck:core limpo, gates estáticos OK, 62/62 testes focados passando (23/23 do PR entre glm-5.3-catalog-and-effort-tiers.test.ts e zai-catalog-glm52.test.ts).
Aditivo, espelha exatamente o padrão já existente glm-5.2-max. Obrigado pela contribuição, primeira PR bem-vinda!
Validado em lote combinado (batch-0824g) contra o tip de release/v3.8.50: typecheck:core limpo, gates estáticos OK, 62/62 testes focados passando (endpoint/parser/schema/static-model + catálogo).
Canonicaliza metadados de endpoint legados (video/audio) para IDs específicos por operação, mantendo compatibilidade retroativa via `normalizeModelSupportedEndpoints` (valores antigos `audio`/`video` continuam válidos como entrada e são normalizados na escrita). Obrigado pela contribuição, primeira PR bem-vinda!
Validado em lote combinado (batch-0824g) contra o tip de release/v3.8.50: typecheck:core limpo, gates estáticos OK, 62/62 testes focados passando (incluindo tests/unit/live-ws-url-11331.test.ts, 11 casos + mutation-check).
Resolve o incidente real do #11331: o handshake já reportava a porta live real, mas o cliente descartava esse campo e ficava preso na porta compilada no bundle. Precedência clara (wsUrl explícito > publicUrl completo > porta/path do handshake aplicados ao default). Obrigado pela contribuição!
Validado em lote combinado (batch-0824g, junto de #11388/#11397/#11415/#11418) contra o tip de release/v3.8.50: typecheck:core limpo, file-size/changelog/complexity/cognitive-complexity OK, 62/62 testes focados passando.
Diagnóstico correto e bem documentado: a falha do nightly Node 26 era um teste que sorteia um número e depende do resultado, não uma quebra de compatibilidade. Comportamento de produção inalterado (a janela de jitter continua aleatória; só o teste ganhou controle sobre ela). Obrigado pela investigação detalhada!
Every "Publish to Docker Hub" run has failed since 2026-08-22 23:14 UTC — 96 of
the last 100. The builder stage dies with:
ERROR: failed to solve: ResourceExhausted: process "/bin/sh -c ... npm run
build ..." did not complete successfully: cannot allocate memory
That is the kernel, not V8. The log puts it precisely: the compile phase always
finishes ("✓ Compiled successfully in 4.2min") and the build is killed right
after "Collecting page data using 7 workers".
Each page-data worker is its own process and inherits NODE_OPTIONS, so the
--max-old-space-size ceiling is per PROCESS, not per build. CIRCLE_NODE_TOTAL=8
means 7 workers, and 7 of them alongside the parent no longer fit the 16 GB /
4 vCPU GitHub-hosted runners the pipeline builds on. It was intermittent for a
while before going 100%, which is what a threshold crossed by ordinary codebase
growth looks like — 7 was also oversubscribing a 4 vCPU runner.
Lower the pool to 3 (2 workers) and make it a build arg, so a big builder can
raise it back with `--build-arg OMNIROUTE_BUILD_WORKERS=8`.
tests/unit/docker-build-memory-budget.test.ts pins the budget: it reads the two
ARG defaults out of the Dockerfile and fails if `parent heap + workers × peak`
outgrows the runner, or if the pool oversubscribes its CPUs. Red on the base
(3/3), green here (3/3). The per-worker peak it budgets with is documented as an
inference from this failure, not a measurement.
DOCKER_GUIDE's build-arg table was stale (it still listed the pre-#10060 4096 MB
default); updated and given the new knob plus the symptom to recognize.
CIRCLE_NODE_TOTAL and OMNIROUTE_BUILD_WORKERS are allowlisted in the
fabricated-docs gate with the reason: neither is read via process.env here — one
is a Dockerfile ARG, the other is read by Next itself.
Note: the real proof is the next publish run. This failure mode only reproduces
on a memory-constrained host, so it cannot be reproduced by the unit suite; the
test guards the arithmetic, not the outcome.
Co-authored-by: Xiangzhe <bakryun0718@proton.me>
`isPublicApiRoute()` matched every entry of PUBLIC_API_ROUTE_PREFIXES with
`startsWith()`, but 11 of the 15 entries name ONE route, not a subtree. As a
prefix each also marked every adjacent path sharing its leading characters as
PUBLIC, which skips the MANAGEMENT auth gate.
That is reachable today: Next resolves `/api/usage/om-usage<anything>` to the
dynamic route `/api/usage/[connectionId]`, and that handler carries no auth of
its own — it relies entirely on being classified MANAGEMENT. An unauthenticated
caller therefore reaches `fetchAndPersistProviderLimits()`, which is an
existence oracle over connection ids (409/404/400/200) and, for a connection id
actually starting with `om-usage`, discloses live quota JSON and can drive an
OAuth token refresh (a write side effect) with no credentials.
Split the allowlist by shape:
- PUBLIC_API_ROUTE_PREFIXES keeps only genuine subtrees, every entry ending in
"/" (asserted by a unit test, so the class cannot come back silently).
- PUBLIC_API_ROUTES_EXACT holds the single routes, matched exactly in both
spellings.
- The three read-only "prefixes" were single routes too and move to
PUBLIC_READONLY_CORS_API_ROUTES, matched exactly. classify.ts now asks
`isPublicReadonlyCorsRoute()` instead of scanning the raw list, so the CORS
origin relaxation pipeline.ts keys on cannot be inherited by a sibling either
(`/api/monitoring/health-detail` was taking it).
- `/api/health` deliberately stays in its own set so it keeps classifying as
`public_prefix`; folding it into the read-only set would widen CORS on it.
dashboardCsrf.ts had a second copy of the prefix scan; it now shares
`isPublicApiRoute()` so the client CSRF exemption and the server classification
cannot disagree. Side effect in the safe direction: the three LOCAL_ONLY oauth
auto-import routes were CSRF-exempt on the client while the server already
required the token — the client now attaches it.
Reported by @ntdat812 (GHSA-74g9-q8f6-793h), with the shape of the fix and the
two gotchas above called out in the report.
Closes GHSA-74g9-q8f6-793h
Co-authored-by: Xiangzhe <bakryun0718@proton.me>
Co-authored-by: Nguyen Thanh Dat <ntdat812.dev@gmail.com>
Validado em lote combinado (batch-0824f, junto de #11399/#11400/#11402) contra o tip de release/v3.8.50: typecheck:core limpo, file-size/changelog/complexity/cognitive-complexity OK, 56/56 testes focados passando incluindo os deste PR (tests/unit/autocombo-unification.test.ts).
Baixo risco: expõe a opção "custom" já suportada em runtime (`getModePack("custom") === undefined`, cai de volta para os pesos explícitos dos sliders) no seletor compartilhado de mode-pack da UI. Obrigado pela contribuição!
Validado em lote combinado (batch-0824f, junto de #11399/#11400/#11407) contra o tip de release/v3.8.50: typecheck:core limpo, file-size/changelog/complexity/cognitive-complexity OK, 56/56 testes focados passando incluindo os deste PR (tests/unit/combo-scoring-inspector.test.ts).
Baixo risco: normaliza pesos parciais/não-unitários no inspector de diagnóstico (`comboScoringInspector.ts`) reutilizando o normalizador já existente do motor real de scoring, mantendo diagnósticos consistentes com o runtime. Obrigado pela contribuição!
Validado em lote combinado (batch-0824f, junto de #11399/#11402/#11407) contra o tip de release/v3.8.50: typecheck:core limpo, file-size/changelog/complexity/cognitive-complexity OK, 56/56 testes focados passando incluindo os deste PR (tests/unit/combo-task-aware.test.ts).
Remove `auto` da lista de estratégias task-routing genéricas — coerente com o #11399, que também protege a ordem já computada pelo `auto` contra reordenação por outro pós-processamento. Obrigado pela contribuição!
Validado em lote combinado (batch-0824f, junto de #11400/#11402/#11407) contra o tip de release/v3.8.50: typecheck:core limpo, file-size/changelog/complexity/cognitive-complexity OK (abaixo do baseline), 56/56 testes focados passando incluindo os deste PR (tests/unit/8370-priority-affinity-reorder.test.ts).
Aditivo e coerente: protege a ordem já decidida pelo `auto` contra reordenação pelo pós-processamento de prompt-cache-affinity — mesma linha do #11400. Obrigado pela contribuição!
The retry-loop recheck returned a non-conforming {ok:false, reason} object
that breaks typecheck against the established {ok, response?} contract used
everywhere else in this function. Aligns with the pre-dispatch skip pattern
(return null after fallbackCount++), matching the PR's own intent: skip this
target and move to the next, not error the whole attempt.
This is a live fix — the broken shape reached origin/release/v3.8.50 via
#11360's own squash-merge and was breaking typecheck:core until now.
These entries were already validated in an earlier merge-batch worktree but
never reached origin (worktree discarded before pushing). Re-adding them
here since #11355's test/route.ts growth (1215->1237) is now live on
origin/release/v3.8.50 and fails the frozen cap otherwise.
Upstream 65e81158a added new providers to the registry; the reserved set
is a full REGISTRY walk, so the pinned count moves 329 -> 391. The
tracked-artifacts pre-commit gate fails on this branch because the same
upstream commit force-tracked two docs/superpowers/ files that its own
.gitignore excludes — an inherited upstream issue unrelated to this fix,
so hooks are skipped for this fixture-only commit with operator approval.
A compatible node created with prefix "tokenrouter" was silently
unreachable: the runtime model resolver (src/sse/services/model.ts)
skips compatible-node lookup for built-in registry ids/aliases, so
"tokenrouter/qwen/..." routed to the built-in tokenrouter provider and
failed with "No active credentials for provider: tokenrouter" even
though the node itself worked when addressed by its internal id.
Reject reserved prefixes at the write path instead:
- new shared module src/shared/constants/reservedProviderPrefixes.ts
(REGISTRY ids + aliases, case-sensitive, built lazily) — single
source of truth consumed by both the runtime guard and the
validation schemas so they can never drift apart
- createProviderNodeSchema / updateProviderNodeSchema now reject
reserved prefixes with a clear message naming the colliding prefix
- src/sse/services/model.ts consumes the shared module; runtime
behavior is byte-for-byte unchanged (verified e2e)
Set semantics mirror the old inline guard exactly: manual alias ids
outside REGISTRY (xiaomi/llamacpp/aq) do not intercept nodes at
runtime and stay allowed; mixed-case input (TokenRouter) does not
collide with the exact-match runtime lookup either.
Upstream added a lint test requiring every web-cookie provider ID to end
with -web. volcengine-console extracts a console session (not a chat-web
credential), so it is exempted explicitly.
Replace the static curated model lists for volcengine-agent-plan and
volcengine-coding-plan with live discovery from the console APIs
(GetAgentPlanModelMappingMeta / ListArkCodeLatestModel), authenticated by
the console cookie+csrf already captured at plan binding time.
- Add volcenginePlanModelDiscovery.ts: fetch + parse + capability enrichment
(family->contextLength/vision/reasoning map, conservative default fallback).
Console calls go through a dynamic undici import to bypass OmniRoute's
global fetch patch (built for LLM provider traffic, reroutes console hits).
Coding plan's ListArkCodeLatestModel needs {AccountId:<number>} extracted
from the console cookie; agent plan's GetAgentPlanModelMappingMeta filters
PlatformAllowStatus===true && Type==='llm'.
- Remove both plan ids from CURATED_MODEL_ONLY_PROVIDERS so synced models
merge into /v1/models and the dashboard Sync Models button works.
- sync-models route: short-circuit to console discovery for plan providers
(the chat API has no /models endpoint); persist via
replaceSyncedAvailableModelsForConnection.
- volcenginePlanBinding: set autoSync:true on new plan connections so the
24h modelSyncScheduler refreshes them automatically.
- volcPlanAutoSyncBackfill: idempotent boot-time backfill so pre-existing
plan connections also enter the scheduler.
Verified end-to-end on local OmniRoute build against live Volcano console:
agent plan synced 7 LLMs, coding plan synced 11 models, /v1/models exposes
all of them (incl. new glm-5-3-260801 / deepseek-v4-flash-260801).
