Relationships
↔ sibling #4568#4597 fan-out: a halt during a --from retry can name a lane as both failed and held, and the run-record judge refuses the save
Opened by stack72 · 11/10/2026· Shipped 11/10/2026
Found by the fan-out gates model test extended to multi-step lanes (swamp-club#4562).
Scenario (fleet harness): 2 workers, offline: fail, rollout batches [1, 1, 100%], maxFailures 0, approve [canary], lane steps prepare then apply (apply depends on prepare succeeding). w-1 fails prepare.
- Run: the canary w-1 fails prepare, apply is skipped; one failure is over the budget, so the rollout halts and w-2 is held. Run failed.
- Disconnect w-2.
- Retry with --from apply. w-1 keeps its failed prepare and reopens apply; w-2 reopens.
- w-2 is offline, so it fails before start. That is over the budget, so the rollout halts before w-1 starts, and w-1 is held.
- The halt is written with failed = [w-1, w-2] and held = [w-1]: in execution_service.ts the halt's failed list is every batch lane whose records include a failure (laneFailed), which counts w-1 by its kept failed prepare even though it never started in this attempt.
- judgeRunRecord refuses the save: its fan-out halt of job patch names a lane more than once. Under a datastore that judges saves, the run record cannot be written.
Only a lane of two or more steps can hold a kept failed step next to a reopened one, so the single-step model never reached this.
Expected: a lane the halt holds is not also counted as failed in that halt (or the failed list counts only lanes that ran in this attempt plus failures before start), so the halt names each lane once and the save is accepted.
Shipped
Click a lifecycle step above to view its details.
Sign in to post a ripple.