Overview / Motivation

Auto-filed by the workflow-fix-on-bug protocol from a workflow-fix prose note raised on task #1417 (emitting agent: clean-result-critic, round-1 verdict).

Goal

Extend scripts/verify_task_body.py check 20 (v4 conciseness caps): when the body carries a WARN-acknowledgment sentence, assert that EACH FIRED WARN class (Takeaways bullet-length / per-result 120-180 band / total-prose budget / caption length) is named in it; a fired-but-unnamed class degrades the acknowledgment (WARN naming the gap), so partial acknowledgments stop silently shipping.

Workflow gap

• Bug observed: #1417's body shipped a WARN-acknowledgment sentence covering two of the three fired check-20 WARN classes (per-result band + linked figures acknowledged; the 41-word Takeaways bullet-length WARN was not named). The verifier accepts ANY acknowledgment sentence without matching it against which classes actually fired; the clean-result-critic caught the gap manually at Lens 12.
• Why it is a workflow gap: the standing WARN-ship rule is "WARNs ship only when acknowledged in body" — enforcement of the acknowledgment's COVERAGE is prose-only today; the check is mechanizable (the verifier already knows exactly which sub-classes fired).
• Confidence (emitter): medium-high (critic marked it mechanizable: yes in spirit; the class list is finite and verifier-internal).
• verified-at-filing: grep -cE 'fired.*class.*acknowledg|ack.*class' scripts/verify_task_body.py → 1 hit, context READ (check 31's linked-figure docstring citing the standing acknowledge-in-body rule — does not implement class-matched assertion); git log --oneline --since='7 days ago' -- scripts/verify_task_body.py → 4 commits (checks 45/46/31 families), none touching WARN-ack coverage (2026-07-18).

Proposed change (candidate diff sketch — refine in planning)

• In check 20 (v4 conciseness), after computing the fired WARN sub-classes:
• ack = _find_warn_acknowledgment_sentence(body) # existing convention: a

• body sentence acknowledging verifier WARNs (e.g. "Verifier WARNs

• acknowledged: ...")

• if fired_classes and ack:

•   unnamed = [c for c in fired_classes if class_keyword[c] not in ack]
  

•   if unnamed:
  

•       warn(f"WARN-acknowledgment sentence does not name fired class(es):
  

•             {unnamed} — extend the acknowledgment or fix the WARN")
  

• Class keywords: bullet-length -> "bullet", per-result band -> "per-result"/
• "120", total budget -> "total prose"/"budget", caption -> "caption".
• WARN-tier only (never flips the overall verdict); v4-only, forward-only.

Scope / surfaces

• Primary target: scripts/verify_task_body.py
• Grep before editing: grep -rn 'acknowledg' scripts/verify_task_body.py .claude/skills/clean-results/SPEC.md .claude/agents/clean-result-critic.md — keep SPEC.md's WARN-ship rule wording + Lens 12's rubric in sync if the acknowledgment convention gets a mechanical shape; add a pin in tests/test_verify_task_body.py.

Constraints / invariants

• Workflow-surface only. scripts/workflow_lint.py --check-asks passes; ruff clean; existing verify_task_body tests stay green (forward-only: no new hard FAIL on grandfathered bodies).
• This session runs under EPM_WORKFLOW_FIX_SESSION=1 and carries a workflow_fix_target: Provenance line — recursion guard applies.

Provenance

• workflow_fix_target: scripts/verify_task_body.py
• fingerprint: 0673ceca0111

Verbatim surfaced prose (clean-result-critic, #1417 round 1): "a mechanizable extension for verify_task_body.py — assert each fired check-20 WARN class (bullet-length / per-result band / total budget / caption) is named in the body's WARN-acknowledgment sentence; this round shipped an acknowledgment covering two of three fired classes."