Repository navigation
🤖 tests: report-decision hold (real host) suite flakes in Test / Unit (6/6), dequeues merge-queue PRs #4458
Description
Summary
A test in the report-decision hold for queued follow-ups (real host) suite (src/node/services/taskService.reportDecisionHold.test.ts:86) fails intermittently in CI. On 2026-09-24 it failed in about half of the Test / Unit (6/6) runs, which appears to have started with #4308. It removed perf PRs from the merge queue twice that day (the React Compiler coverage stack, #4425–#4429 / #4457). Each time a re-run passed.
Impact
Merge-queue removals mean a new CI run and a re-enqueue, roughly 20+ minutes per occurrence, for any PR in the queue.
Next steps
- Reproduce with
bun test src/node/services/taskService.reportDecisionHold.test.ts --rerun-each=20, and under load (e.g. alongsidemake typecheck). Record the failing test name and assertion. - Look for timing-based coordination in the real-host harness (sleeps or grace timers) and replace it with explicit completion signals, per AGENTS.md.
Filed by the perf owner as a merge-queue health issue; this is not a perf regression. A related queue flake is #4435 (WorkflowRunner attemptDisposition).
Generated with xum • Model: anthropic:claude-opus-5-5 • Thinking: high
Activity
linear-code commentedon Sep 24, 2026
MUX-206
ThomasK33 commentedon Sep 26, 2026
Closing as fixed by #4641 (tracked as #4558). That PR found the cause of the
taskService.reportDecisionHoldreal-host flakes: the tests asserted before the second backend's fsynced config write had finished. They now await that write.Evidence: no failed merge-group run since #4641 merged (2026-09-26 02:32Z). Reopen if the suite flakes again, with the failing test name.
Generated with
xum• Model:anthropic:claude-opus-5-5• Thinking:high