[teamleaderleo]

What

A second race in swift-test-with-hang-sampler.sh, beside the one #23290 fixes. The wrapper reads its log right after closing tee's input:

exec 3>&-
wait "$tee_pid" 2>/dev/null || true

bash 3.2, macOS's /bin/bash, cannot wait for a process substitution. wait returns 127 at once while tee is still writing. The report then reads an empty or cut log, which shows up in two ways:

• no red tests are named;
• a run that has its summary line is read as having ended early.

Measured: a probe reads 0 of 2 lines on bash 3.2 and 2 of 2 on bash 5.3. bash 5 (the Linux CI runners) waits correctly, so this only hits macOS runs of the wrapper.

Not this PR: the #23366 flake itself (swift-test-log.*: No such file) is #23290's. The watcher subshell inherits the wrapper's EXIT trap and deletes the log when it is killed early. The two PRs change different lines and merge cleanly.

Fix

When wait can't wait, the wrapper polls kill -0 on tee's pid until it is gone, for at most 10 s, and only then reads the log.

Tests

swift-test-summary.test.sh gains a case with a tee that writes late: it buffers its input and writes 1 s after the input closes. The red run must still name its red test.

• Committed red first (0987322): fails under /bin/bash 3.2 with "a late-written log was read before its writer finished".
• Passes with the fix under both bash 3.2 and bash 5.3.

Other scripts/**/tests/*.test.sh with the same patterns:

• This wrapper is the only >(tee …) in scripts/.
• Of the tests that run a script writing under RUNNER_TEMP, only package-test-lane-suite-args.test.sh leaves RUNNER_TEMP unset. Its refusals exit before the lane touches it. No other changes needed.

No capture: CI script, no UI.

Agent session: cosmic-gray

🤖 Generated with Claude Code

[github-actions]

cmux status _{(one comment, edited in place; each line is updated by the job that owns it)}

Review 27c7140: HQ feed card · gallery before/after runs · checks

App of this head (app.zip, kept 1 day)

gh workflow run reload-build.yml -R manaflow-ai/cmux --ref feat-cmux-next -f tag=pr23402 -f ref=browser-import-2/script-test-race -f platform=macos -f backend_mode=local

[teamleaderleo]

Owned by browser-import-2 on cosmic-gray · session

[teamleaderleo]

Owned by browser-import-2 on cosmic-gray · session

[teamleaderleo]

Bead cx-v7r1y created for this PR by the cmux coordinator.

[teamleaderleo]

Review (independent subagent): SHIP at 27c7140

• Diagnosis checked: under /bin/bash 3.2, wait on the process-substitution pid returns 127 at once with the pid still alive; bash 5.3 waits.
• Red then green checked: at 0987322 the new slow-tee case fails under bash 3.2 ("a late-written log was read before its writer finished"); at 27c7140 it passes under 3.2 and 5.
• The fix can't hang: the kill -0 loop is capped near 10 s. No zombie on 3.2, and on bash 5 the first wait reaps tee so the loop exits at once.
• merge-tree with cmux-next: swift test wrapper sets its cleanup trap after it forks (swift-test-summary flake) #23290 (b7cc518) is clean, and the two fixes don't interact.
• The slow tee is only on PATH for the one call (prefix assignment on a function call).
• Nits: removed the stray "In progress" line from the body. "macOS only" more exactly means hosts whose bash is 3.2.

[teamleaderleo]

Red attribution before merging, at head 27c7140. All three failing jobs also fail on base:

• cmux-next generated files (exportIsFresh reports the exports are stale): this job also failed in feat-cmux-next push runs 39615647078 and 39613279968.
• cmux-next swift test and cmux-next daemon tests: these run here only because the PR edits swift-test-with-hang-sampler.sh, and they never run on base pushes. So I ran the 16 red suites at the merge base 6c541e6 on the fleet (cmux-ci step 60cb540812c6f99a9c71a194).
  • 6 suites fail at base: SettingsSchemaExport, PhoneAccessSetting, HomeMessageRequests, AccountsCmuxRow, RowCloseMotion and ClassicUpgradeSimulation. ClassicUpgradeSimulation fails with the same order-differs failures as here.
  • The other 10 pass at base when run alone and failed here inside the full parallel run.
  • This PR changes no Swift code. Its wrapper change only runs after swift test exits, so it can't change which tests pass.

[teamleaderleo]

base-red-watch: cmux-next web / web lint (fast) went red on feat-cmux-next in https://github.com/manaflow-ai/cmux/actions/runs/39621747120 (first red at ed547f0, last green at 7710c58). Suspect: #23402 (commit ed547f0). If it is yours, fix forward on feat-cmux-next.

[github-actions]

Shipped in:

• cmux NEXT nightly 1.0.0-nightly.3962509665801 (ed547f0e27)
• cmux NEXT nightly 1.0.0-nightly.3962960774501 (913f2e6d53)