repair-missing-agency-reasoning — WorksGood live mirror

Metadata

Status	done
Assigned	`agent-733`
Model	`pi:zai:glm-5.2`
Created	2026-07-25T07:54:02.392495074+00:00
Started	2026-07-25T13:00:51.827831335+00:00
Completed	2026-07-25T13:45:56.345765830+00:00
Tags	`evaluation`, `migration`, `pi`
Tokens	4942699 in / 18966 out

Description

Incident

make-hashed-project is stranded in PendingEval with WG-EVAL-PIPELINE-REPAIR-EXHAUSTED. Its FLIP and evaluator plans were scaffolded before Pi reasoning became mandatory. The persisted calls contain exact Pi routes but no reasoning; Pi-only execution correctly fails with WG-PI-REASONING-MISSING. Lifecycle repair then replays the same invalid plan byte-for-byte until its bounded repair budget is exhausted, leaving a completed implementation operator-blocked.

Authoritative incident identity: source attempt 1, pipeline evalp-cf8647216311909aacbaa9d2; FLIP plan b3:d7cd2331644ef581ba7ab579def9113e7c8c5030c493b9bb6255e434a11e0a42; evaluator plan b3:f3642b022aabbb4816c3b118278e365356b5ce3d76b877dd72881a4e19ed7de2.

Requirements

Add a deterministic, bounded migration/recovery path for pre-Pi persisted agency plans whose route is exact pi:<provider>:<model> but whose reasoning is absent.
Never execute or silently accept a missing-reasoning plan. At an explicit migration boundary, resolve reasoning from the authoritative role/tier configuration, persist it into a newly hashed plan/generation, and record old/new plan identity plus provenance.
If exact Pi reasoning cannot be resolved, fail closed with an actionable operator diagnostic. Never synthesize Claude/Codex/Nex/native routes or an implicit thinking default.
Recover the existing source attempt without rerunning the completed source worker or changing its artifacts/commit. Re-arm only its failed FLIP/evaluator satellites under a coherent atomic lifecycle identity.
Preserve original plans and failures as immutable history. Stale verdicts from the old plan/pipeline may never score the repaired attempt; new evidence is consumed exactly once.
A daemon restart, retry, or concurrent coordinator tick cannot duplicate calls, exhaust repair on the same invalid bytes again, or create mixed reasoning across FLIP stages.
wg show/status must distinguish migration-required, migrated/rearmed, active evaluation, and true operator ambiguity.

Validation

Fixture matching make-hashed-project migrates route-preservingly to explicit effective reasoning and completes FLIP/evaluation without rerunning source work.
Missing reasoning with no authoritative role/tier value stays fail-closed and operator-actionable.
Non-Pi/malformed legacy plans remain rejected; no cross-system fallback.
Old and new plan hashes, reasoning, source attempt, pipeline/generation, producer run, and consumed verdict are auditable.
Restart/concurrent-tick regression proves bounded exactly-once migration and verdict consumption.
Existing retry-drift, durable evaluation, Pi-only model-plane, and low-score gate regressions pass; fmt and clippy pass.

## Incident
`make-hashed-project` is stranded in `PendingEval` with `WG-EVAL-PIPELINE-REPAIR-EXHAUSTED`. Its FLIP and evaluator plans were scaffolded before Pi reasoning became mandatory. The persisted calls contain exact Pi routes but no `reasoning`; Pi-only execution correctly fails with `WG-PI-REASONING-MISSING`. Lifecycle repair then replays the same invalid plan byte-for-byte until its bounded repair budget is exhausted, leaving a completed implementation operator-blocked.

Authoritative incident identity: source attempt 1, pipeline `evalp-cf8647216311909aacbaa9d2`; FLIP plan `b3:d7cd2331644ef581ba7ab579def9113e7c8c5030c493b9bb6255e434a11e0a42`; evaluator plan `b3:f3642b022aabbb4816c3b118278e365356b5ce3d76b877dd72881a4e19ed7de2`.

## Requirements
- Add a deterministic, bounded migration/recovery path for pre-Pi persisted agency plans whose route is exact `pi:<provider>:<model>` but whose reasoning is absent.
- Never execute or silently accept a missing-reasoning plan. At an explicit migration boundary, resolve reasoning from the authoritative role/tier configuration, persist it into a newly hashed plan/generation, and record old/new plan identity plus provenance.
- If exact Pi reasoning cannot be resolved, fail closed with an actionable operator diagnostic. Never synthesize Claude/Codex/Nex/native routes or an implicit thinking default.
- Recover the existing source attempt without rerunning the completed source worker or changing its artifacts/commit. Re-arm only its failed FLIP/evaluator satellites under a coherent atomic lifecycle identity.
- Preserve original plans and failures as immutable history. Stale verdicts from the old plan/pipeline may never score the repaired attempt; new evidence is consumed exactly once.
- A daemon restart, retry, or concurrent coordinator tick cannot duplicate calls, exhaust repair on the same invalid bytes again, or create mixed reasoning across FLIP stages.
- `wg show`/status must distinguish migration-required, migrated/rearmed, active evaluation, and true operator ambiguity.

## Validation
- [ ] Fixture matching `make-hashed-project` migrates route-preservingly to explicit effective reasoning and completes FLIP/evaluation without rerunning source work.
- [ ] Missing reasoning with no authoritative role/tier value stays fail-closed and operator-actionable.
- [ ] Non-Pi/malformed legacy plans remain rejected; no cross-system fallback.
- [ ] Old and new plan hashes, reasoning, source attempt, pipeline/generation, producer run, and consumed verdict are auditable.
- [ ] Restart/concurrent-tick regression proves bounded exactly-once migration and verdict consumption.
- [ ] Existing retry-drift, durable evaluation, Pi-only model-plane, and low-score gate regressions pass; fmt and clippy pass.

Depends on

1 non-public dependency hidden

Required by

done .flip-repair-missing-agency-reasoning

Log

2026-07-25T08:59:52.315513423+00:00 Prior attempt left an on-track core migration implementation in eval_lifecycle.rs and coordinator wiring. cargo check reaches one expected non-exhaustive status counter; continuing WIP rather than restarting. Direct implementation remains appropriate because lifecycle, coordinator, status, and tests share one tightly coupled identity model.
2026-07-25T09:00:56.057234390+00:00 Task unclaimed: agent 'agent-709' (PID 1169888) process exited
2026-07-25T09:00:56.064325816+00:00 Minted evaluation pipeline evalp-86838843737d63dfb9d3bd9e for source attempt 3 (coordinator dead-agent retry)
2026-07-25T09:01:57.181379803+00:00 Spawned by coordinator --executor pi --model openai-codex:gpt-5.6-sol --isolation required-worktree
2026-07-25T09:02:07.795236456+00:00 Starting implementation: inspect prior WIP and lifecycle/evaluation plan invariants before adding bounded migration.
2026-07-25T09:03:38.105918260+00:00 Task unclaimed: agent 'agent-712' (PID 1181235) process exited
2026-07-25T09:03:38.113144610+00:00 Minted evaluation pipeline evalp-3b862848b9bc8ae68098ea68 for source attempt 4 (coordinator dead-agent retry)
2026-07-25T09:04:38.542023327+00:00 Spawned by coordinator --executor pi --model openai-codex:gpt-5.6-sol --isolation required-worktree
2026-07-25T09:04:58.080911701+00:00 Starting implementation: inspect prior WIP and persisted evaluation lifecycle/plan structures before adding bounded Pi reasoning migration.
2026-07-25T09:05:08.197558028+00:00 Reconciliation: task recovered from orphaned state (was InProgress, agent: agent-714)
2026-07-25T09:05:08.197589179+00:00 Minted evaluation pipeline evalp-838ede778cda391a6d41fc64 for source attempt 5 (coordinator orphan reconciliation retry)
2026-07-25T09:05:09.112320793+00:00 Spawn failed (attempt 1/5): isolated worktree /home/bot/wg/.wg-worktrees/agent-708 is owned by live attempt agent-714 for task 'repair-missing-agency-reasoning' (status Working); no process launched. Recover by terminating/reaping that attempt or archive the worktree explicitly after inspection. exec_mode=default, executor=pi
2026-07-25T09:05:10.318437460+00:00 Spawned by coordinator --executor pi --model openai-codex:gpt-5.6-sol --isolation required-worktree
2026-07-25T09:05:23.987268564+00:00 Starting implementation: inspect prior WIP and agency lifecycle persistence/migration paths before adding bounded recovery.
2026-07-25T09:06:37.342073244+00:00 Task unclaimed: agent 'agent-715' (PID 1199305) process exited
2026-07-25T09:06:37.372269378+00:00 Minted evaluation pipeline evalp-ee03eda2cd083f7bb77676c0 for source attempt 6 (coordinator dead-agent retry)
2026-07-25T09:07:41.464114500+00:00 Spawned by coordinator --executor pi --model openai-codex:gpt-5.6-sol --isolation required-worktree
2026-07-25T09:07:54.158877421+00:00 Starting implementation: inspect prior WIP and persisted evaluation lifecycle before adding bounded Pi reasoning migration.
2026-07-25T09:09:31.068060518+00:00 Task unclaimed: agent 'agent-716' (PID 1207533) process exited
2026-07-25T09:09:31.075175264+00:00 Minted evaluation pipeline evalp-fac9ed6ed6cfe5a320d845bf for source attempt 7 (coordinator dead-agent retry)
2026-07-25T09:10:31.225521403+00:00 Spawned by coordinator --executor pi --model openai-codex:gpt-5.6-sol --isolation required-worktree
2026-07-25T09:10:55.480059377+00:00 Starting implementation: inspecting prior WIP, Pi agency plan persistence, lifecycle repair, and incident-shaped regressions.
2026-07-25T09:11:22.758707345+00:00 Task unclaimed: agent 'agent-718' (PID 1209321) process exited
2026-07-25T09:11:22.766812728+00:00 Minted evaluation pipeline evalp-a5d8719585701d22734231b7 for source attempt 8 (coordinator dead-agent retry)
2026-07-25T09:12:25.137247306+00:00 Spawned by coordinator --executor pi --model openai-codex:gpt-5.6-sol --isolation required-worktree
2026-07-25T09:12:40.655020071+00:00 Starting implementation: inspecting prior WIP and evaluation lifecycle migration paths
2026-07-25T09:13:00.250283143+00:00 Task unclaimed: agent 'agent-720' (PID 1210256) process exited
2026-07-25T09:13:00.258011046+00:00 Minted evaluation pipeline evalp-9ce9acf91152f156e833ad4f for source attempt 9 (coordinator dead-agent retry)
2026-07-25T09:14:35.400443455+00:00 Spawned by coordinator --executor pi --model openai-codex:gpt-5.6-sol --isolation required-worktree
2026-07-25T09:15:10.027842395+00:00 Task unclaimed: agent 'agent-722' (PID 1211128) process exited
2026-07-25T09:15:10.035211798+00:00 Minted evaluation pipeline evalp-48691ff532ff7c188a7a780e for source attempt 10 (coordinator dead-agent retry)
2026-07-25T09:22:38.540735065+00:00 Spawned by coordinator --executor pi --model openai-codex:gpt-5.6-sol --isolation required-worktree
2026-07-25T09:22:57.307846061+00:00 Reconciliation: task recovered from orphaned state (was InProgress, agent: agent-728)
2026-07-25T09:22:57.307874742+00:00 Minted evaluation pipeline evalp-09ca5e8a85caf36d940d6b96 for source attempt 11 (coordinator orphan reconciliation retry)
2026-07-25T09:22:57.849803858+00:00 Spawn failed (attempt 2/5): isolated worktree /home/bot/wg/.wg-worktrees/agent-708 is owned by live attempt agent-728 for task 'repair-missing-agency-reasoning' (status Working); no process launched. Recover by terminating/reaping that attempt or archive the worktree explicitly after inspection. exec_mode=default, executor=pi
2026-07-25T09:23:02.841466703+00:00 Spawn failed (attempt 3/5): isolated worktree /home/bot/wg/.wg-worktrees/agent-708 is owned by live attempt agent-728 for task 'repair-missing-agency-reasoning' (status Working); no process launched. Recover by terminating/reaping that attempt or archive the worktree explicitly after inspection. exec_mode=default, executor=pi
2026-07-25T09:23:03.710562925+00:00 Spawn failed (attempt 4/5): isolated worktree /home/bot/wg/.wg-worktrees/agent-708 is owned by live attempt agent-728 for task 'repair-missing-agency-reasoning' (status Working); no process launched. Recover by terminating/reaping that attempt or archive the worktree explicitly after inspection. exec_mode=default, executor=pi
2026-07-25T09:23:08.769884419+00:00 Spawn failed (attempt 5/5): isolated worktree /home/bot/wg/.wg-worktrees/agent-708 is owned by live attempt agent-728 for task 'repair-missing-agency-reasoning' (status Working); no process launched. Recover by terminating/reaping that attempt or archive the worktree explicitly after inspection. exec_mode=default, executor=pi
2026-07-25T09:23:08.769884419+00:00 Circuit breaker tripped: spawn failed 5 times. Last error: isolated worktree /home/bot/wg/.wg-worktrees/agent-708 is owned by live attempt agent-728 for task 'repair-missing-agency-reasoning' (status Working); no process launched. Recover by terminating/reaping that attempt or archive the worktree explicitly after inspection. exec_mode=default, executor=pi. Task marked incomplete for evaluator review.
2026-07-25T09:26:17.016640223+00:00 Task reset for retry from incomplete (attempt #11) — reason: ran on gpt-5.6-sol and came back incomplete; shifting to zai glm-5.2
2026-07-25T09:26:17.016692095+00:00 Minted evaluation pipeline evalp-b05788437643e80ed133121a for source attempt 12 (resume-in-place retry)
2026-07-25T09:26:20.028702982+00:00 Spawned by coordinator --executor pi --model zai:glm-5.2 --isolation required-worktree
2026-07-25T09:30:26.119034739+00:00 Resuming from prior WIP (10 retries). Prior agent left on-track core migration in eval_lifecycle.rs + coordinator wiring (542 insertions). Fixed the one expected non-exhaustive status counter in status.rs (added MigrationRequired/MigratedRearmed counters). cargo check now passes. Continuing: verify Config deps (ReasoningProvenance, resolve_reasoning_detail, parse_exact_pi_route, AgencyStage::role/default_tier) exist; then audit lifecycle_conflict + ensure tests exist.
2026-07-25T13:00:47.571560532+00:00 Task unclaimed: agent 'agent-729' (PID 1261609) process exited
2026-07-25T13:00:47.725955154+00:00 Minted evaluation pipeline evalp-aeb82ec6b118b4d170bfff0d for source attempt 13 (coordinator dead-agent retry)
2026-07-25T13:00:51.903069832+00:00 Spawned by coordinator --executor pi --model zai:glm-5.2 --isolation required-worktree
2026-07-25T13:07:09.733353129+00:00 Resuming on-track WIP (cargo check passes). Prior agent shipped complete migrate_missing_pi_reasoning boundary in eval_lifecycle.rs + coordinator wiring + status counters. Verified all Config deps exist (resolve_reasoning_detail, parse_exact_pi_route, ReasoningProvenance, AgencyStage::role/default_tier) and execution-time WG-PI-REASONING-MISSING defense exists in llm.rs:558. Adding the missing bounded-migration test coverage (no tests exist yet) + smoke scenario + report doc.
2026-07-25T13:44:20.693081257+00:00 Implemented + committed + pushed (b4a432d9 impl+tests, 3e19d6d3 report). NOTE: root cause of the 11 retries was the worktree-sweep bug (BUG-worktree-sweep-destroys-uncommitted-work.md) reaping this 0-commits-ahead branch every ~5s; committing immediately defeated it. Earlier confusion was self-inflicted: my bash commands used 'cd /home/bot/wg' which hit the MAIN repo while my edits went to the worktree — resolved by operating in the worktree consistently.
2026-07-25T13:44:29.561392573+00:00 Validated: migrate_missing_pi_reasoning boundary (eval_lifecycle.rs) + coordinator wiring + status counters. 7 new unit tests pass (route-preserving, exactly-once/restart idempotent, fail-closed no-reasoning, non-Pi rejected, health states, verdict consumed once). eval_lifecycle 27/27; status/coordinator bin 171/171; pi_sole_model_plane+pi_two_tier 30; failed_pending_eval 34; pending_eval_state 15; agency_pipeline 15; task_lifecycle+agency+coordinator_agent+e2e_smoke+triage_smoke 23. cargo fmt --check clean; cargo clippy 0 errors (2 pre-existing warnings in untouched call<'a>/load_graph lines). Report at docs/reports/repair-missing-agency-reasoning.md.
2026-07-25T13:45:56.345768+00:00 Task marked as done; scheduled evaluator is advisory evidence only (execution is not a quality pass)