mirror of
https://github.com/hanzoai/forge.git
synced 2026-10-30 15:00:39 +00:00
zooqueen
3c93d455c1
Not re-queued did not mean still ours. The sweep can hand a dropped job back and a second runner claim it before the first runner's report lands, and the fallback write settled whatever job it found: a job actively running another task went to Failure, the emitter skipped everything downstream of it, and those jobs stayed skipped when the real build succeeded. On a release pipeline that is a publish that quietly never happens. The write now carries the same task_id guard the other two do. The two sweeps also stopped sharing a conclusion. Zombie selects on a runner that has gone quiet, which is what a lost task looks like; endless selects on start time alone, and its runner is alive and has been reporting for hours. That job is wedged, not lost, so re-queueing it only spends hours more of a shared runner reaching the same wedge. Each sweep now says which one it is holding. The cap warning is driven by reaching the cap rather than by an in-memory count that a lost update leaves raised, so it no longer reports failing a job it did not fail. Covers the re-claim race, the wedged job, the log, and the drop count starting over on a rerun — the last so the whitelist clone that gives a new attempt a fresh budget cannot quietly become a copy of everything. Assisted-by: blue:claude-opus-5