gitreins-phantom-complete-judge-fail

Answer

Diagnosed and verified against gitreins==0.16.0 (installed from PyPI) in a throwaway repo. Full solution: /workspace/solution.md (487 lines).

Root cause

In gitreins/cli.py :: cmd_task_complete, the status is persisted before evaluation:

task = tm.complete(args.id, force=force)   # status: complete + completed_at written NOW
...
result = judge.evaluate_task(task, skip_tier2=skip_tier2)
_persist_result(workdir, task, result)
if not result.passed:
    sys.exit(1)                            # exits, never reopens

TaskManager._apply_status() also never clears completed_at. So an INCOMPLETE verdict leaves a self-contradictory store: tasks.yaml and the board say complete, while .gitreins/history/2026-10-10/2316ac64/verdict.json says passed: false. I reproduced this exactly.

Exact recovery

1. Collect evidence — find the graded verdict and read passed, commit, evaluated_at: .gitreins/history/<YYYY-MM-DD>/<run>/verdict.json, where <run> = sha256("<task_id>:<evaluated_at>")[:8]. Use <run> as the board --evidence-run-id.

2. Targeted per-id replace (not a full YAML re-dump). The full script is in the doc; the key point is it edits only the named task's block — status: → in_progress, drop completed_at:. Verified it preserves comments and unknown keys, whereas gitreins task start destroys them: ``` tasks: # foreman audit note: do NOT drop this line # preserved

3. id: second-task status: in_progress # reverted external_ticket: OPS-42 # board link # preserved ```

4. Re-dispatch in a fresh worktree: gitreins task worktree <id> (created branch gitreins/task/<id> and a worker brief), then verify with gitreins task get <id> / gitreins worktree list.

5. Cite on the board: bash boardctl update "$BOARD_ID" --status in_progress \ --evidence-run-id "$RUN" --commit-hash "$SHA" \ --note "Judge FAIL (INCOMPLETE): run=$RUN commit=${SHA:-<none>} evaluated_at=$EAT; reopened + re-dispatched"

Prevention

Reopen on failed evaluation (patch in doc): add TaskManager.reopen() that sets in_progress and clears completed_at, and call tm.reopen(args.id) in the if not result.passed: branch. I applied and tested this: a failing task complete now leaves the task open with no completion stamp.

Verified

• Phantom complete reproduced (status flipped, exit 1, passed: false).
• Targeted recovery leaves task in_progress, comment/unknown key intact.
• Full re-dump confirmed destructive (drops comment + external_ticket).
• Fresh worktree re-dispatch and boardctl evidence citation both confirmed working.
• Preventive patch confirmed to eliminate the phantom state.