fix(supervisor): prevent non-decisions from permanently blocking open work #369

Merged
karsten merged 2 commits from fix/validator-pending-liveness into main 2026-11-12 17:38:09 +00:00
Owner

Generic host lifecycle fix: missing/queued/interrupted/inconclusive independent assessment retains a durable read-only retry, not an explicit block. No task/provider-specific behavior. Historical current-enrollment false blocks with unresolved verdict are assessed again; held legacy, real semantic denial, stops, expiry, budget, pending worker effects and approvals stay binding. Host records deferral reason for queued input/interruptions.

Failing-before: 9 initial new regressions fail. Passing-after: 14 new tests, full suite 2007 passed, one existing skip; typecheck/build and diff check pass. Updated older expectations to durable pending/backoff with deterministic retry clocks. Tests cover skipped and invalidated decisions, JSON restart, exactly one follow-up after fresh admission, historical false-block repair, and genuine holds.

Generic host lifecycle fix: missing/queued/interrupted/inconclusive independent assessment retains a durable read-only retry, not an explicit block. No task/provider-specific behavior. Historical current-enrollment false blocks with unresolved verdict are assessed again; held legacy, real semantic denial, stops, expiry, budget, pending worker effects and approvals stay binding. Host records deferral reason for queued input/interruptions. Failing-before: 9 initial new regressions fail. Passing-after: 14 new tests, full suite 2007 passed, one existing skip; typecheck/build and diff check pass. Updated older expectations to durable pending/backoff with deterministic retry clocks. Tests cover skipped and invalidated decisions, JSON restart, exactly one follow-up after fresh admission, historical false-block repair, and genuine holds.
Author
Owner

Integrated concurrently merged #368 normally (no force push). Its conservative historical-migration and stop/dependency/wait guards are preserved. This PR adds the remaining non-decision lifecycle repair: inconclusive/startup reviews retain durable retry ownership rather than re-blocking forever, new independent verdict replaces stale unresolved receipt, host records queue/interruption deferrals, and every unresolved turn retains backoff including absent proposal/diagnostic. No provider or task-specific changes. Final merged-base verification at 5ddcbae: npm run typecheck and npm test PASS: 2018 passed, 1 existing skip, 0 failed. Build included. 14 added regressions plus #368 guard suite all pass. Prior CI-less branch rules checked; no bypass requested.

Integrated concurrently merged #368 normally (no force push). Its conservative historical-migration and stop/dependency/wait guards are preserved. This PR adds the remaining non-decision lifecycle repair: inconclusive/startup reviews retain durable retry ownership rather than re-blocking forever, new independent verdict replaces stale unresolved receipt, host records queue/interruption deferrals, and every unresolved turn retains backoff including absent proposal/diagnostic. No provider or task-specific changes. Final merged-base verification at 5ddcbae: npm run typecheck and npm test PASS: 2018 passed, 1 existing skip, 0 failed. Build included. 14 added regressions plus #368 guard suite all pass. Prior CI-less branch rules checked; no bypass requested.
karsten deleted branch fix/validator-pending-liveness 2026-11-12 17:38:10 +00:00
Author
Owner

ACTIVATED AND VERIFIED on both Shikamaru and Akira using the supported native source owner, one idempotent install request each. Exact selected and actually running Core revision 537ba6f6b8; authenticated API identity and published software provenance confirmed; installed dist/harness.js hash and non-decision recovery logic checked. Native status activated. No task ledgers or configurations edited; no business worker cancel/resume/replay.

Real-provider verification against the installed artifact and configured homelab/homelab/gpt-6.1-sol: queued clarification displaces the initial review; pending checkpoint persisted to disk and restored in a new Harness; no early inference before retry deadline; independent review admits a single bounded read; exactly one actual read returns the correct random existing-file marker; independent final review marks completed. 2 real supervisor reviews, 1 real working turn, 1 read, no business effects. This is a deliberately bounded sandbox reproduction, not a claim that every historical business task has completed.

Final suite on the integrated main base: 2018 passed, 1 existing skip, zero failures; typecheck/build pass. Evidence ~/.cache/hayate-maintenance/supervisor-pending-liveness/{summary,shikamaru,akira}.json, live-provider.log and test logs.

ACTIVATED AND VERIFIED on both Shikamaru and Akira using the supported native source owner, one idempotent install request each. Exact selected and actually running Core revision 537ba6f6b818231111b59ad17bcb06c4d3b8a8c0; authenticated API identity and published software provenance confirmed; installed dist/harness.js hash and non-decision recovery logic checked. Native status activated. No task ledgers or configurations edited; no business worker cancel/resume/replay. Real-provider verification against the installed artifact and configured homelab/homelab/gpt-6.1-sol: queued clarification displaces the initial review; pending checkpoint persisted to disk and restored in a new Harness; no early inference before retry deadline; independent review admits a single bounded read; exactly one actual read returns the correct random existing-file marker; independent final review marks completed. 2 real supervisor reviews, 1 real working turn, 1 read, no business effects. This is a deliberately bounded sandbox reproduction, not a claim that every historical business task has completed. Final suite on the integrated main base: 2018 passed, 1 existing skip, zero failures; typecheck/build pass. Evidence ~/.cache/hayate-maintenance/supervisor-pending-liveness/{summary,shikamaru,akira}.json, live-provider.log and test logs.
Sign in to join this conversation.
No reviewers
No labels
No milestone
No project
No assignees
1 participant
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set.

Reference
harness/core!369
No description provided.