Root cause

The kanban janitor is a second phantom-done path. explicit_completion_evidence() (the completed_at_present branch) returned a CLOSE decision purely because completed_at was set on a non-terminal task — with no requirement of real result/summary/comment evidence:

if task.completed_at is not None and task.status not in TERMINAL_STATUSES:
    evidence = task.result or comments[-1].body if comments else task.result
    return "completed_at_present", shorten(evidence or "task has completed_at but non-terminal status")

build_scan() collected it and apply_closes() raw-UPDATEd status='done', stamping a tautological placeholder ("task has completed_at but non-terminal status") as the task's only "evidence". A fresh ready/triage/in_progress task that merely had completed_at stamped by a brief worker claim got rubber-stamped to done.

Live proof — t_5dbfc384

Task t_5dbfc384 carries a janitor_completed event closing it ~13 min after creation:

janitor_completed | {"reason": "completed_at_present", "previous_status": "in_progress",
                     "evidence": "task has completed_at but non-terminal status",
                     "source": "kanban-janitor"}

The completed event right before it had result_len: 0 — no real evidence. It had to be manually reopened and is now correctly in_review with a substantive result.

Why this is the 2nd path

This is the same pathology PR #1 fixed for complete_task() (CompletionEvidenceError). But the janitor bypasses complete_task entirely with a raw UPDATE status='done', so the PR #1 guard never fired here. This PR is the complement: it closes the janitor path.

Previously unversioned

The live janitor was a loose script at ~/.hermes/scripts/kanban_janitor.py, not in any git repo. This PR brings it under version control at scripts/kanban_janitor.py (verbatim copy of the live script as the starting point) and applies the fix to the repo copy.

The fix (scripts/kanban_janitor.py)

• Add has_completion_evidence(task, comments) — True iff non-empty result (after strip) OR any comment with a non-empty stripped body. A bare completed_at timestamp is NOT evidence. Mirrors kanban_db._has_completion_evidence; the janitor uses its own Task/Comment dataclasses so it needs a local copy of the predicate.
• explicit_completion_evidence() — the completed_at_present close decision is now gated on has_completion_evidence(...). Evidence-free completed_at tasks fall through (return None → not closed). Comment-pattern / trusted-operator checks still run normally.
• build_scan() — new phantom_completed_at anomaly category (capped at 50): non-terminal, non-running tasks with completed_at set but no evidence.
• render_markdown() — new "Phantom completed_at" triage section so operators see the anomaly instead of it being silently rubber-stamped.
• apply_closes() unchanged — once explicit_completion_evidence stops emitting evidence-free closes, apply_closes naturally never closes them.

Tests (tests/scripts/test_kanban_janitor_completion_evidence.py)

10 tests, all passing:

• has_completion_evidence: result-only True, comment-only True, whitespace-only False, nothing False.
• explicit_completion_evidence: bare completed_at no-evidence → None; with result → ("completed_at_present", ...); with comment → ("completed_at_present", ...); whitespace-only → None.
• build_scan-level: a fixture DB with a bare-completed_at no-evidence task → it appears in phantom_completed_at and NOT in close_decisions; a real-evidence task → in close_decisions, not flagged.

Verification

• python3 -m py_compile scripts/kanban_janitor.py → OK
• pytest tests/scripts/test_kanban_janitor_completion_evidence.py → 10 passed
• Scan-only run (no --apply) against a read-only snapshot of the live kanban → 0 evidence-free completed_at_present closes, applied: 0.
• Scan against a synthetic DB with an injected t_5dbfc384-style phantom task → task flagged in phantom_completed_at, absent from close_decisions, rendered in the markdown triage section. Pre-fix it would have been auto-closed.

Operator deploy step (Mac-gated)

After merge, sync the live janitor from the repo copy — deployment is an operator step, the live ~/.hermes/scripts/ file was intentionally left untouched by this PR:

cp <repo>/scripts/kanban_janitor.py ~/.hermes/scripts/kanban_janitor.py

🤖 Generated with Claude Code

Summary by CodeRabbit

• New Features

  • Added a kanban janitor tool for managing Hermes kanban boards.
  • Detects stale items, redundant tasks, and phantom completion records.
  • Auto-completes tasks with proper evidence.
  • Generates JSON and Markdown reports with optional external copy.
  • Creates database backups before applying changes.

• Tests

  • Added test coverage for completion evidence detection logic.