fix-no-cli — WorksGood live mirror

Metadata

Status	done
Assigned	`agent-853`
Created	2026-07-26T06:36:46.019286578+00:00
Started	2026-07-26T07:35:35.134282635+00:00
Completed	2026-07-26T08:02:45.077044230+00:00
Tags	`bug`, `evaluation`, `lifecycle`
Tokens	12113873 in / 41565 out

Description

Objective

Tasks keep landing in pending-eval / failed-pending-eval with evaluation_health.state = operator-required-ambiguity (eval-pipeline-repair-exhausted) and there is NO wg CLI to clear it. wg retry refuses these statuses; wg recover only catches pending-eval (not failed-pending-eval). The only remedy is risky graph.jsonl surgery (flip status + clear evaluation_lifecycle). This recurred for: remove-graph-depth-guard, make-hashed-project, make-wg-config, fix-chat-coordinator-2, fix-spawn-failures.

Work

Add a sanctioned CLI to resolve operator-required-ambiguity: e.g. wg retry (or a dedicated wg resolve-eval <task>) that clears evaluation_lifecycle/evaluation_health and resets the task to open for retry.
Make wg recover also catch failed-pending-eval (currently invalid as a filter status).
Ensure the eval reconciler stops re-pinning "historical PendingEval as a required gate" in a loop once the gate is unsatisfiable.

Validation

A pending-eval / failed-pending-eval / operator-required-ambiguity task can be retried via CLI with NO graph.jsonl edit.
wg recover accepts failed-pending-eval.
The reconciler does not loop-pin an unsatisfiable eval gate.

Reference

This session: 5 tasks stuck in pending-eval/failed-pending-eval, cleared only via graph surgery.

## Objective
Tasks keep landing in `pending-eval` / `failed-pending-eval` with `evaluation_health.state = operator-required-ambiguity` (eval-pipeline-repair-exhausted) and there is NO wg CLI to clear it. `wg retry` refuses these statuses; `wg recover` only catches `pending-eval` (not `failed-pending-eval`). The only remedy is risky graph.jsonl surgery (flip status + clear evaluation_lifecycle). This recurred for: remove-graph-depth-guard, make-hashed-project, make-wg-config, fix-chat-coordinator-2, fix-spawn-failures.

## Work
- Add a sanctioned CLI to resolve operator-required-ambiguity: e.g. `wg retry` (or a dedicated `wg resolve-eval <task>`) that clears evaluation_lifecycle/evaluation_health and resets the task to open for retry.
- Make `wg recover` also catch `failed-pending-eval` (currently invalid as a filter status).
- Ensure the eval reconciler stops re-pinning "historical PendingEval as a required gate" in a loop once the gate is unsatisfiable.

## Validation
- [ ] A pending-eval / failed-pending-eval / operator-required-ambiguity task can be retried via CLI with NO graph.jsonl edit.
- [ ] `wg recover` accepts failed-pending-eval.
- [ ] The reconciler does not loop-pin an unsatisfiable eval gate.

## Reference
This session: 5 tasks stuck in pending-eval/failed-pending-eval, cleared only via graph surgery.

Depends on

done .assign-fix-no-cli

Required by

✉ Messages 7 messages (7 unread)

#1user2026-07-26T07:09:02.436389042+00:00read

Live graph inventory: six source tasks are evaluation-held. Hard-stuck operator-required: make-wg-config (repair exhausted; .flip lacks evidence), fix-chat-coordinator-2 (FLIP on mismatched pipeline), resolve-prophage-source (failed-pending-eval; evaluator has no persisted route). Active but merely queued: smoke-probe, smoke-probe2, impl-maxagents-authority-fix; each reports active-evaluation with open FLIP/eval satellites. Please use these as validation cases and report a sanctioned recovery path.

#2user2026-07-26T07:09:12.587827232+00:00read

Additional CLI gap: `wg list --status pending-eval` and `--status failed-pending-eval` both reject those values as unknown, even though `wg list --json` emits those statuses. Please include this in the sanctioned operator recovery/visibility surface if in scope.

#3fix-no-cli2026-07-26T07:31:53.627595451+00:00read

Acknowledged live graph inventory — I will use hard-stuck operator-required and active queued cases to validate a sanctioned recovery path without graph surgery.

#4fix-no-cli2026-07-26T07:31:53.666595165+00:00read

Acknowledged status-filter gap — I will include pending-eval and failed-pending-eval in CLI list/recover visibility where status parsing is shared or directly in scope.

#5user2026-07-26T07:32:02.287645640+00:00delivered

Emergency recovery mode: profile is now Pi Codex Sol/high, eval gates are temporarily disabled, and only the three repair tasks are running. Resume preserved WIP, land/install the sanctioned pending-eval recovery CLI first. Do not re-enable evaluation settings or spawn/decompose.

#6user2026-07-26T07:32:20.603465050+00:00delivered

One authorized operator cleanup after the code fix: set only the effective evaluator reasoning to high via the supported role-specific config command, verify evaluator=pi:openai-codex:gpt-5.6-luna/high, and leave auto_evaluate/eval_gate_all false.

#7fix-no-cli2026-07-26T07:35:43.488331047+00:00read

Acknowledged queued requirements — I will preserve prior WIP, implement/install the sanctioned recovery and status visibility path first, validate against stuck-state fixtures without graph surgery, keep evaluation disabled, then set only evaluator reasoning to high and verify the effective route.

Log

2026-07-26T06:36:46.007770230+00:00 Task paused
2026-07-26T06:37:08.522372596+00:00 Task published
2026-07-26T06:51:10.753665725+00:00 Spawned by coordinator --executor pi --model zai:glm-5.2 --isolation required-worktree
2026-07-26T07:00:56.757978049+00:00 Starting fix-no-cli. Plan: (1) extend wg retry to accept PendingEval/FailedPendingEval (clears stuck eval gate via begin_source_attempt, no graph surgery); (2) add failed-pending-eval to wg recover parse_status + clear eval lifecycle on retry; (3) fix reconciler loop: migrate_missing_pi_reasoning must not re-run (oscillate diagnostic) once migration already ran for an attempt and repair is exhausted.
2026-07-26T07:15:20.671620079+00:00 Task unclaimed: agent 'agent-822' (PID 2806061) process exited
2026-07-26T07:15:22.128891770+00:00 Minted evaluation pipeline evalp-b041d47dfe2d7188f6c4f61a for source attempt 2 (coordinator dead-agent retry)
2026-07-26T07:18:20.142828427+00:00 Spawned by coordinator --executor pi --model zai:glm-5.2 --isolation required-worktree
2026-07-26T07:18:44.208999370+00:00 Reconciliation: task recovered from orphaned state (was InProgress, agent: agent-838)
2026-07-26T07:18:44.209012721+00:00 Minted evaluation pipeline evalp-413291868631c8de010690a6 for source attempt 3 (coordinator orphan reconciliation retry)
2026-07-26T07:27:35.977350013+00:00 Task paused
2026-07-26T07:31:35.557363547+00:00 Task resumed
2026-07-26T07:31:37.837945263+00:00 Spawned by coordinator --executor pi --model openai-codex:gpt-5.6-sol --isolation required-worktree
2026-07-26T07:31:53.708254943+00:00 Starting implementation; inspecting prior WIP and evaluation lifecycle/recovery/status parsing paths.
2026-07-26T07:35:20.810771508+00:00 Task unclaimed (was assigned to @agent-848)
2026-07-26T07:35:20.810814010+00:00 Minted evaluation pipeline evalp-b7b8c05b845930f574e4f217 for source attempt 4 (explicit unclaim redispatch)
2026-07-26T07:35:35.210722704+00:00 Spawned by coordinator --executor pi --model openai-codex:gpt-5.6-sol --isolation required-worktree
2026-07-26T07:35:51.220166829+00:00 Resuming preserved WIP; first inspecting branch state and existing implementation before edits. Will not decompose or re-enable evaluation settings per operator instruction.
2026-07-26T07:48:21.045341801+00:00 Targeted Rust tests now pass for retrying PendingEval, recovering FailedPendingEval, fresh recovery attempt minting, and list status filters. Reconciler regression is implemented against repeated unsatisfiable-gate ticks; adding installed-binary smoke flow next.
2026-07-26T07:59:32.592919806+00:00 Installed candidate globally with cargo install --path . --locked. Authorized cleanup completed: effective evaluator is pi:openai-codex:gpt-5.6-luna/high; agency.auto_evaluate=false and agency.eval_gate_all=false remain unchanged. Project/global profile pointers restored after applying the evaluator-only reasoning override.
2026-07-26T08:01:49.181790215+00:00 Validation: cargo build and cargo install pass; cargo fmt/check and cargo clippy pass (pre-existing warnings only); 34 retry tests pass; new recover/list/reconciler tests pass; installed-binary eval_operator_recovery_cli smoke passes. Full cargo test has unrelated pre-existing stale-profile/test-fixture failures; isolated first profile failure passes and task-owned tests are green.
2026-07-26T08:02:11.325057099+00:00 Committed and pushed: 153f53b6 — sanctioned eval-hold recovery CLI, list/recover status visibility, stable reconciler, docs, and owned smoke scenario.
2026-07-26T08:02:41.288255325+00:00 Final validation: installed binary accepts live [E] make-wg-config - Make wg config settings stick project-locally and complete the CLI surface (precedence + all knobs) [design, config, cli, daemon, ux] [E] impl-maxagents-authority-fix - Impl: fix the max_agents authority/reload-override bug (shared prerequisite) [coordinator, config, bugfix] [E] fix-wg-cleanup - Fix: wg cleanup nightly/orphaned error 'WG not initialized' in an initialized graph [bug, cli] [E] fix-dead-agents - Fix: dead agents hold worktree ownership for hours, blocking spawns (reaper gap) [bug, dispatcher, worktree] [E] fix-chat-coordinator-2 - Fix chat coordinator session directory registration [implementation, chat-control, regression] [E] fix-worktree-build - Fix: worktree build artifacts (target/) accumulate unbounded -> disk exhaustion [bug, worktree, disk] [E] impl-npm-publish-distribution - Impl: npm distribution via per-platform optionalDependencies (Shape A) [release, npm, distribution, ci] and ; owned smoke passes end-to-end. Effective evaluator remains pi:openai-codex:gpt-5.6-luna/high while auto_evaluate=false and eval_gate_all=false.
2026-07-26T08:02:45.077049180+00:00 Task marked as done; scheduled evaluator is advisory evidence only (execution is not a quality pass)