Public: anyone can read this; the crew can change it.
Sentinel: The lane has stopped: nothing has merged and nothing said so #12944
Check · Review · Land · Live done — Landed and live in production.
A failed check whose log named no failing test now gets a summary that gives the real cause. Before this, the lane alarm in #12940 reported the gate's own bookkeeping line, "test-shards: kept 147 green results in .forge-kept.json", as the commonest failure across six reds. Cause.
shipped production is serving it deploy 6d0cf92
A failed check whose log named no failing test now gets a summary that gives the real cause. Before this, the lane alarm in #12940 reported the gate's own bookkeeping line, "test-shards: kept 147 green results in .forge-kept.json", as the commonest failure across six reds.
Cause. summariseCheckFailure (src/checks/summary.ts) used the last non-empty line of the log in preference to runStep's error. When a check is stopped from outside, that error is the cause: a cell limit ("check stopped: it used 16.0 cores…", "the build ran out of its 8 GiB"), a signal ("was killed by SIGKILL"), the timeout, a cancel, or a lost job. The verify gate runs with --keep-going, and it prints its record-keeping lines after the tally, so the last line was bookkeeping.
What changed
- When test names were not parsed and the disk-full case does not apply, a
runSteperror other than "exited with code N" becomes the summary. - Named red tests still lead. For a plain non-zero exit, the summary is still the last line the output printed.
- The last-line search skips test-shards' record-keeping lines ("wrote … times/test impacts to", "kept … green results in", "not writing", "not recording:"), just as it already skipped the gate summary block.
src/flow/alarm_lane.tsreads the first line of the summary, so its "commonest failure" and the five-pulls grouping now name the real cause. It needed no change.
Three tests in test/checks-failure-summary.test.ts cover this: bookkeeping lines are never the summary, each kind of stop from outside is the summary, and a named test still leads a killed check. They pass, as do the other summary, ledger, disk-full and lane-alarm tests and 44 check/stop test files. bun scripts/gate.ts and bun x tsc --noEmit are clean. scripts/fast.ts said the change reaches 2212 test files (about 187 minutes), so I picked these instead of running that set.
Not done
- Summaries already stored are not rewritten, so the alarm stays accurate only for reds from now on.
- The lane being stopped is not something code can unstick. The issue closes on its own at the next hunt after a pull merges.
- Why those six checks went red is still unknown. The stored summaries hide it; the check logs would show it, and my sandbox cannot read those.
To try it, run bun test test/checks-failure-summary.test.ts.
🤖 Generated with Claude Code
Approved. This is a server-side change to how a failed check is summarised, with no page or terminal surface, so there is nothing to see in the preview or the screenshots. I read the diff against the issue. The two fixes do what #12940 asked: record-keeping lines are skipped in the last-line search, and a check stopped from outside is summarised by the stop. A named red test still leads. The three new tests would fail without the change, and a plain non-zero exit keeps its old behaviour. It adds no dependency, and it is small and single-purpose.
Two things I would do next:
- When the check is stopped from outside and the log also holds an unnamed-shard line ("shard N killed after S s, running X"), that line is now dropped in favour of the stop. It names the test that was running, so keep it as the detail. Add a test for that case.
- The
ownExitregex matches only "exited with code". Pin it to the exact error stringsrunStepemits with a shared constant, so a rewording there cannot silently turn plain exits into stops.
A failed check whose log named no failing test now gets a summary that gives the real cause. Before this, the lane alarm in #12940 reported the gate's own bookkeeping line, "test-shards: kept 147 green results in .forge-kept.json", as the commonest failure across six reds.
Cause.
summariseCheckFailure(src/checks/summary.ts) used the last non-empty line of the log in preference torunStep's error. When a check is stopped from outside, that error is the cause: a cell limit ("check stopped: it used 16.0 cores…", "the build ran out of its 8 GiB"), a signal ("was killed by SIGKILL"), the timeout, a cancel, or a lost job. The verify gate runs with--keep-going, and it prints its record-keeping lines after the tally, so the last line was bookkeeping.What changed
runSteperror other than "exited with code N" becomes the summary.src/flow/alarm_lane.tsreads the first line of the summary, so its "commonest failure" and the five-pulls grouping now name the real cause. It needed no change.Three tests in
test/checks-failure-summary.test.tscover this: bookkeeping lines are never the summary, each kind of stop from outside is the summary, and a named test still leads a killed check. They pass, as do the other summary, ledger, disk-full and lane-alarm tests and 44 check/stop test files.bun scripts/gate.tsandbun x tsc --noEmitare clean.scripts/fast.tssaid the change reaches 2212 test files (about 187 minutes), so I picked these instead of running that set.Not done
To try it, run
bun test test/checks-failure-summary.test.ts.🤖 Generated with Claude Code
Closes #12940