Repository navigation
fix(testing): judge::stale treats a zombie pid as alive #6865
Description
judge::stale (plugins/testing/hooks/judge-lib.sh, around line 346) checks a pending marker's pid with kill -0, which succeeds for a zombie. When PID 1 never reaps orphans, as in a container started without --init, a killed job's marker looks alive, and the Stop hook waits out its whole 180 s deadline.
Evidence, from the investigation of #6719:
- In a container whose PID 1 was plain
sleep, the s7 case failed 3 times. The pending pid showed state=Z, ppid=1. - With
docker --init, the same case failed 0 times in 42 runs.
Possible fix: treat a pid whose /proc/<pid>/stat state is Z as dead, falling back to kill -0 where /proc is absent.
Priority: low. It costs a 180 s wait only under a non-reaping PID 1.
🤖 Generated with Claude Code
Activity
github-actions commentedon Oct 10, 2026
Thanks for the detailed report. I confirmed
judge::staleinplugins/testing/hooks/judge-lib.sh(line 346) relies onkill -0 "$pid", which does succeed for a zombie process, so under a non-reaping PID 1 (e.g. a container without--init) a killed job's marker can appear alive and the Stop hook waits out the full 180s deadline. The reproduction evidence (3/3 failures without--init, 0/42 with it) and the suggested fix (checking/proc/<pid>/statfor stateZ, falling back tokill -0where/procis unavailable) both look sound. No duplicates found. Marked low priority as you noted, since it only costs a wait under this specific condition.Automated triage. A maintainer reviews it.
kyle-sexton commentedon Oct 13, 2026
Owner decision (2026-10-13): approved the recommendation.
! kill -0 "$pid"succeeds (judge-lib.sh:369, checked again 2026-10-13). No PR addresses it.Basis: origin/main plugins/testing/hooks/judge-lib.sh:358-371
Basis: https://pubs.opengroup.org/onlinepubs/9799919799/functions/kill.html; https://man7.org/linux/man-pages/man5/proc_pid_stat.5.html
)and treat Z or X as dead; otherwise fall back to kill -0. Add a regression case to plugins/testing/hooks/test-judge.test.sh that fails before the fix, and bump the version and CHANGELOG.Next: an agent lane, via admission, opens one PR in plugins/testing; it changes a Stop-hook library, so a human merges it.
This was generated by AI.
kyle-sexton commentedon Oct 13, 2026
Owner batch-ratified this admission in an attended session on 2026-10-12: work-class: scoped. Rule: class C1-C3, no trust surface, the independent check confirmed the recorded decision, and no owner step remains.
kyle-sexton commentedon Oct 18, 2026
This was generated by AI during backlog shaping.
Closing as not planned. PR #7544 (merged 2026-10-16) moved the test judge off hooks and removed the test-judge Stop hook scripts this issue fixes;
plugins/testing/hooks/judge-lib.sh:1-4says no hook runs the library now.The owner approved this close in an attended backlog-shaping session on 2026-10-18 (question: "Close the 9 issues ...?"; answer: "Close all 9 (Recommended)").