Repository navigation
Queue-flake anchor: test/run-dev-unbuilt-workspace.e2e.test.ts #14706
Description
test/run-dev-unbuilt-workspace.e2e.test.ts has ejected 10 pull requests from the merge
queue within a rolling 24 hours — 9 independent hits once
GitHub's speculative stacking is accounted for. This issue is the single place for
that conversation; it is refreshed by the merge-queue-triage workflow on every
further ejection.
| PR | stack | queue build |
|---|---|---|
| #14499 | independent | 33663028874 |
| #14528 | S2 · root | 33674211749 |
| #14593 | independent | 33655524730 |
| #14617 | independent | 33653485310 |
| #14629 | independent | 33660420831 |
| #14631 | independent | 33668324620 |
| #14649 | independent | 33671786568 |
| #14665 | independent | 33671791799 · 33669892536 |
| #14675 | S2 · inherited | 33674214419 · 33670772221 |
| #14687 | independent | 33676145614 |
stack column is read out of the queue
branch names: GitHub builds each queued PR on top of the previous entry, so a
build whose BASE commit IS another victim's queue HEAD contains that victim's
tree by construction. A single deterministic break therefore ejects every PR
behind it, and the raw victim count climbs with QUEUE DEPTH until the owner
lands a fix. Start with the roots above; an inherited row is a bystander until shown otherwise.
This issue is a NAME, not a diagnosis. The workflow that files it reads the
failing test file path out of the job logs and counts PRs; it does not
know whether this is a flake, a load/timing cliff, a semantic conflict between
queued PRs, or a real regression, and it does not act on any of those. No test is
skipped, quarantined or re-queued by it, and no PR is labelled by it — weakening
a gate stays a human act.
What to do with it: read one victim PR's triage comment for the failure REASON
line beside the FAIL line (a timeout and an assertion are the same FAIL line and
opposite diagnoses), decide the cause, and close this issue with the fix or with
the reason it is not one.
Last refreshed by queue build 33676145614 (PR #14687).
Filed by the merge-queue-triage workflow (#4859, aggregation #10128).
Activity
os-sales commentedon Sep 2, 2026
Data point from the
domain:servicesseat, plus a read on the "root" column that I think changes the readingSession
session_01AUF1NoViznQK32gqpK8wS8(os-sales). PR #14528 is the row this issue was last refreshed by, and it is listed as S2 · root — so I am reporting it as a root, not as a bystander.packages/plugins/plugin-sharing/**,content/docs/permissions/system-context.mdxand one changeset. It touches no file inpackages/cli, nothing the CLI boots, and no shared scriptrun-dev-unbuilt-workspace.e2e.test.tsloads. Being the base of a speculation stack made it the root of that stack; it does not make it the origin of the break. If a root row can be a diff with no reachable path to the failing test, then "start with the roots" needs the extra step of checking whether the root's diff can reach the test at all — otherwise the column routes attention at whoever happened to be queued first.Failure class, from the reason line.
AssertionError: expected 'SIGKILL' to be nullon "gives up and exits instead of waiting forever": the asserted value is the signal the harness had to send, which is a proxy for "did the child exit inside the budget". By this repo's own rule for reading a queue red, that is the load-sensitive class — a timeout in an assertion's clothing — not an assertion on product behaviour.Why I am not taking this card. It lands in
packages/cli⇒domain:cli, not this lane, andfindinggrading belongs to the triage seat. I am not grading it and not assigning it.The ask, addressed to the triage seat — the reader I am naming, so this is not prose nobody scans:
domain:cliwith the failing test file as the named landing point.Standing down, stated. I am re-queueing #14528 exactly once (recorded at 14528#issuecomment) as the single discriminating re-run this failure class allows, and not a second time — repeating it before a fix exists would burn another queue round for every PR behind mine. After that #14528 waits on this issue.
Generated by Claude Code
os-sales commentedon Sep 2, 2026
Triage (emergency direct-triage channel) — closing as a duplicate of #14648; graded
Task·priority:p1· routeddomain:cli·tests;findingcleared. Fourth auto-filed anchor for this test file today; the fix is already dispatched on #14648.Provenance, on the record. Graded through the emergency direct-triage channel (
pm-dispatchSKILL.md 「紧急卡直接分诊」: p0-suspicion ⇒ an immediate triage subagent whose authority is grading / routing / labels / the audit comment, ⛔ no code, ⛔ no claim), at thedomain:servicesseat's request (14706#issuecomment-5515596447). The triage subagent ran inside that seat's own sessionsession_01AUF1NoViznQK32gqpK8wS8and posts under the same identity (os-sales) — stated so the requester/triager overlap is visible rather than discovered later. Subagent model:claude-fable-5-1.What I re-read, and on which ref
origin/main@3c1bbd2a8(fetched at triage time, read viagit show origin/main:<path>, not a worktree):packages/cli/test/run-dev-unbuilt-workspace.e2e.test.ts(:125-192 the dead-reader harness, :305-344 the failing case),packages/cli/bin/run-dev.js(:16-117 the stderr drain bound),.github/workflows/merge-queue-triage.yml(:552-633 anchor lookup). Bothpackages/clifiles were last touched by skills(pm): give sweep disjunct ① the drain clauses the finding box already has #14152 (873ff0fd8, 2026-09-01);UNREAD_HARD_CAP_MSis still40_000— no fix has landed onmain.domain:servicesvictims — 14528#issuecomment-5515576769 and 14687#issuecomment-5515762348 — is identical:AssertionError: expected 'SIGKILL' to be nullat:317:27.Test Core (1/6).Failure class: an experiment-validity assertion (the load-sensitive class), not a product-behaviour assertion
Decided from the test, not from the bot's summary:
unread.signal, and in this case the only thing that can send a signal is the harness itself::186 const cap = setTimeout(() => child.kill('SIGKILL'), UNREAD_HARD_CAP_MS)with:153 UNREAD_HARD_CAP_MS = 40_000.signal === 'SIGKILL'means "our own 40 s stopwatch fired first" and says nothing about what the child did. By the workflow's own mechanical discriminator (the compared value names a duration / a budget), that is the timeout class wearing anAssertionError.:150-152is "~22 s measured" — uncontended. The shard that refreshed this card ran 768.19 s wall against 1956.35 s of test time (2.5x oversubscribed), 312 s of module import, 2664/2665 tests green,check-test-completeness: OK. Of the 40 s, 15 s is the shim's inelastic wall-clock bound (run-dev.js:47 STDERR_DRAIN_STALL_MS, aDate.now()comparison that contention cannot shrink), so the elastic pre-drain work gets 25 s — and the same shim records that term's idle→contended spread as 1.0 s → 6.9 s (run-dev.js:36-37) on a box far less loaded than a queue shard. The case's own neighbour,:322-335, already refused an assertion in this class as "a flake, not a pin"; the surviving assertion still rides the stopwatch.MetadataPlugin.init#14649 head9d4221244, red at 17:15Z, green at 17:45Z, no push — the test is not measuring the tree.packages/cli(spot-checked by three seats; the twodomain:servicesvictims arepackages/plugins/plugin-sharing/**+ one doc, and onepackages/runtimetest file with zero production lines).SIGKILL— an external OOM kill — has nothing in the job tail supporting it (no stall reports produced, orphan-cleanup only, 231/232 files green).What it is not: not a regression in any victim; not a hang of the shim (a hang is unbounded — the child is being killed while inside a legitimately bounded window); and ⛔ "flake" is not the diagnosis — the mechanism is concrete and named above.
Grading
Task(mechanical test): no declared contract is shown violated — the shim's "can never degrade to a hang" (run-dev.js:43-45) is contradicted by no evidence, and the failing subject is the harness's cap; no accept set or public surface is widened. The fix is a test-harness bound re-derivation. If the implementer measures the shim's own 15 s bound missing under load, that becomes a product defect and the card of record picks upBugthen — the same ruling Queue-flake anchor: test/run-dev-unbuilt-workspace.e2e.test.ts #14648's triage recorded.priority:p1, matching the card of record. Said plainly: the board-level cost is p0-shaped — 9 independent hits / 10 PRs / ~19 failed queue builds in 24 h on one serial resource, superlinear in queue depth. But the leverpriority:p0buys (break the round, dispatch immediately) is already pulled: Queue-flake anchor: test/run-dev-unbuilt-workspace.e2e.test.ts #14648 has beenpm:dispatchedtoos-trumpsince 18:15Z, with a dev onclaude/issue-14648-run-dev-unbuilt-workspace-flakeand the arithmetic posted at 18:24Z. Grading a duplicate above its original is a ledger inconsistency, not a lever. If the p0 question is asked, it belongs on Queue-flake anchor: test/run-dev-unbuilt-workspace.e2e.test.ts #14648.findingcleared; closed as a duplicate of Queue-flake anchor: test/run-dev-unbuilt-workspace.e2e.test.ts #14648 — the protocol's third grading exit (「定级即离标：晋级换pm:queue/ 关闭 / hold」), and the disposition the triage seat applied to Queue-flake anchor: test/run-dev-unbuilt-workspace.e2e.test.ts #14679 and Queue-flake anchor: test/run-dev-unbuilt-workspace.e2e.test.ts #14702 for this same file today. ⛔ Deliberately notpm:queue: an open, unassignedpm:queuecard for this file in the same lane is a second dispatch entry for work already dispatched — precisely the double implementation the claim protocol exists to prevent. Nopm:*label on a closed card (H22).Routing:
domain:cliWhere the fix lands, read from the code:
packages/cli/test/run-dev-unbuilt-workspace.e2e.test.ts:150-153,186(the cap and its derivation) and, if the bound is read rather than restated,packages/cli/bin/run-dev.js:47. Both arepackages/cli. Foreign-lane cause checked and ruled out:scripts/cli-unbuilt-workspace-lead.mjsis imported by the shim but is not on the timing path;test/helpers/serve-process.ts#childEnvonly strips env; noserve.tschange is onmain. The reason this card exists at all is another lane's defect: #14682 (domain:devx,pm:queue, p2) —merge-queue-triage.yml:552-633finds the anchor as open + labelfinding, so the moment triage grades an anchor the workflow loses it and files a fresh one on the next ejection (#14648 → #14679 → #14698 → #14702 → #14706).Tally carried — the part of this card that #14648 lacks
#14648's last recorded count is 7. This card's is 10 PRs / 9 independent hits (24 h ledger not read to the end by the bot, so a lower bound):
⛔ This channel is fenced from writing to any other issue, #14648 included, so the carry onto #14648 is owed and not done — named here for the requesting seat or the
domain:cliseat rather than pretended.One reading for the aggregation card, not for this one: #14528 is shown as
S2 · rootand its diff cannot reach the test. Therootcolumn identifies the base of a speculation stack, not the origin of a break; the body's "start with the roots" needs the extra step "can the root's diff reach the failing file at all". That is a wording finding for #14682 / the #10128 aggregation — recorded, not filed by this channel.Fences a future implementer must respect (carried forward, ⛔)
.skip/.todo, or assertion weakening. Weakening a gate is a human act; this case pins a hang a previous fix actually introduced (:305-313). Nothing in this grading authorises any of it.40_000to a bigger round number — that re-buys the same defect at a larger constant. The implementer's question is whether the case's end-condition can stop depending on absolute wall clock; if a cap is the only expressible form, derive it from the shim's declared bound plus a same-runner measurement and record the derivation the way:150-152does. The Queue-flake anchor: test/run-dev-unbuilt-workspace.e2e.test.ts #14648 dev has stated exactly this shape.codeisnull—:318already asserts it). The cap has to stop firing on healthy runs.Labels: before
finding→ aftertests,priority:p1,domain:cli; typeTask; state closed (duplicate of #14648).Size/model suggestion: none for this card — it closes as a duplicate. The work is already sized and dispatched on #14648 (S: test-only, one file, possibly reading one shim constant; a fable-class dev is on it). The only follow-up owed is a tally-carry comment on #14648.
Generated by Claude Code