Stop autonomous resurrection of historical tasks at recovery boundaries #337
Loading…
Add table
Add a link
Reference in a new issue
No description provided.
Delete branch "fix/stale-task-reactivation"
Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?
Operator issue
Old tasks reappear across agents after rollout/recovery. Read-only .103 Peter state had 43 historical independent admissions sharing one unknown child snapshot. Unknown ownership was allowed to authorize model-selected independent work. Elapsed legacy polls also became active again.
Repair
Validation
Full suite 1848 passed / one optional skip / zero failures. Typecheck/build/diff-check passed. Dedicated negative regressions reproduce old behavior and prove durable held queues/restart, scoped recovery, genuine user replay, explicit resume and stop/deadline/budget/save fences.
Rollout
Operator explicitly requested completed repair and rollout. Use immutable published tested main artifact and platform owners. Historical evidence is retained, not bulk cancelled/marked completed or removed. Live Peter is native Linux 192.168.1.103, NOT legacy Value/.122.
Final exact head
4175778: full suite 1849 passed, one optional skip, zero failures. Native decision-interface CI green. Includes terminal legacy-result fence as well as retained generated queues, one-time host enrollment, authenticated explicit resume and no poll-expiry reactivation. Operator requested completed fleet rollout, including live Peter .103. Existing ongoing exact child work preserved; historical generated inputs are retained but not executed.DELIVERED: 0.1.234 /
52ac9d4/ immutable sha256:b32a4948c9915c4aa930229d610c5eca05ac76c8cb6a7d824dc9827fc5761781 live and independently verified on all TEN observed active installations: Che, Randy, Minato, Kakashi, SaaStaq Moist, SaaStaq Odo, Shikamaru, live Peter native Linux .103, Akira and separate native macOS Karsten. Native process/pointer/API identities checked; all six K8s workloads Ready on exact digest. Installed compiled-code historical permit/queue smoke passed on all ten. Live stored-state invariants passed: no held task active, no held worker running, no held generated input in executable pending. Snapshot examples: Peter 215 held nonterminal tasks / 17 held workers; Akira 194 / 62. History/checkpoints retained, true ongoing exact-owner work preserved. Full integrated suite 1865 passed / one optional skip. Owner activation used supported checkpoint restart where idle unavailable, explicitly NOT represented as an idle lease; no SIGKILL/global cancellation/data cleanup. Detailed receipt ~/.cache/hayate-maintenance/stale-task-delivery/verified-fleet.json and project notes/stale-task-recovery-delivery-2026-11-11.md. Kubernetes Karsten intentionally remains replicas 0; absent legacy installations not recreated. Core #338 complements #337 with direct historical-worker automatic retry fence.