Fix epic-resolution-workflow.md §4 misread: epic-close ≠ PR-merge §0 invariant parallel

Context

Operator-surfaced 2026-05-11 (this session):

"if you did an epic review, and come to the close state conclusion => close it. this is not a merge gate."

Empirical anchor: My own /epic-resolution review on #10691 concluded RECOMMEND_CLOSE_COMPLETED on 2026-05-04T21:58:58Z. Epic stayed OPEN for ~7 days because the skill protocol said the close action was "reserved for human pipeline authority" — a misread that propagated across multiple agent interactions today (incl. my own #10822 review wording + GPT's #10924 review). Operator-corrected; epic closed 2026-05-11T19:57:59Z.

The Problem

epic-resolution-workflow.md §4 carries this wording:

"Verdict authority — §0 Invariant 1 parallel: the skill produces RECOMMENDATIONS only. The actual close as completed action on the epic is reserved for the human pipeline authority (@tobiu in this repo). Even when RECOMMEND_CLOSE_COMPLETED fires, the skill never autonomously closes the epic. This mirrors the merge-act invariant for PRs."

This is wrong-shape. §0 Invariant 1 is specifically gh pr merge (PR merge action). Epic-close-as-completed is NOT in §0 scope — it's the natural downstream of a positive review verdict. Faithful execution of the misframing produced:

• ~7-day delay on #10691 close (mine)
• Same misframing in my own #10822 closeout review comment (just-posted)
• GPT's #10924 review carrying same "no agent close action" framing
• Stale-closeout-recommendation pollution on the v13 board (operator-surfaced this turn)

The Architectural Reality

§0 Invariant 1 verbatim:

"No gh pr merge (Human-Only execution).* Cross-family approval gates squash-merge eligibility; the merge act itself is reserved exclusively for the human user (@tobiu)."*

This is specifically PR merge. The action gate is gh pr merge. There is NO similar §0 invariant for gh issue close.

epic-resolution-workflow.md §4 invented a "§0 Invariant 1 parallel" that doesn't exist in §0. The parallel is fabricated, not load-bearing.

The Fix

Amend epic-resolution-workflow.md §4 wording:

Before:

"Verdict authority — §0 Invariant 1 parallel: the skill produces RECOMMENDATIONS only. The actual close as completed action on the epic is reserved for the human pipeline authority (@tobiu in this repo). Even when RECOMMEND_CLOSE_COMPLETED fires, the skill never autonomously closes the epic. This mirrors the merge-act invariant for PRs."

After:

"Verdict authority: the skill produces a structured review + recommendation. When the verdict is RECOMMEND_CLOSE_COMPLETED with no unresolved residuals, the reviewer-agent SHOULD close the epic as completed as the natural downstream of the review (no separate operator gate). This is not parallel to §0 Invariant 1 — §0 strictly applies to gh pr merge (PR merge action), which remains human-only; epic-close action is downstream of the review, not a separate gate. For other verdicts (RECOMMEND_KEEP_OPEN, RECOMMEND_CREATE_MISSING_SUBS, RECOMMEND_RETIRE_OR_SUPERSEDE), the recommendation surfaces for operator/owner decision; agent does NOT take terminal action."

Acceptance Criteria

• AC1: epic-resolution-workflow.md §4 "Verdict authority" paragraph amended per the Fix above (remove "§0 Invariant 1 parallel" + "mirrors the merge-act invariant for PRs" framings; clarify CLOSE_COMPLETED is reviewer-agent action; non-CLOSE verdicts still surface to operator)
• AC2: §4 also adds explicit guidance for terminal actions per non-CLOSE verdicts: KEEP_OPEN = no action; CREATE_MISSING_SUBS = coordinate-but-don't-autonomously-create; RETIRE_OR_SUPERSEDE = operator authority (potential separate substrate concern — sub-issue if §0-parallel claim is correct for retire-action specifically)
• AC3: Empirical anchor cited inline — #10691 7-day stale-closeout-pending-action + this-session protocol-correction
• AC4 (post-merge verification): Any future /epic-resolution invocation with CLOSE_COMPLETED verdict produces a close-action in the same turn or the immediately following coordination turn — not a multi-day pending state

Out of Scope

• Modifying §0 Invariant 1 (PR merge) — that's correct as-is; the fix is removing the false parallel claim from a different skill payload
• Bulk-amending all skill payloads that may have similar "§0 parallel" misframing — that's its own audit (could be follow-up if multiple instances surface)
• Retroactive close-actions on already-recommended-CLOSE_COMPLETED epics that pre-date this fix — handled per-epic as discovered (#10691 already closed this turn)

Avoided Traps

• Auto-closing for non-CLOSE_COMPLETED verdicts: rejected — KEEP_OPEN / CREATE_MISSING_SUBS / RETIRE verdicts should NOT trigger auto-close; only CLOSE_COMPLETED + zero unresolved residuals
• Removing operator authority entirely from epic-close domain: rejected — operator retains override per §0 Invariant authority for any specific epic; default-shape just doesn't require operator click for CLOSE_COMPLETED
• Treating this as high-blast substrate evolution requiring Discussion-graduation: rejected per #11217 §6.1 scope classification — this is a bug-fix scope (wording amendment correcting a documented misframing), not a new architectural primitive

Related

• Anchor: #10691 (7-day stale-closeout that motivated the correction)
• Companion misreads (same session): #10822 closeout review comment carries similar wording (acceptable: verdict was KEEP_OPEN; framework statement still wrong)
• GPT's #10924 closeout: same framing — GPT may close his own epic per the corrected shape
• Substrate-bug audit candidate: other skill payloads with "§0 parallel" claims for non-PR-merge actions

Origin Session ID

c0d5c29d-dc70-44c8-b5af-d3f6c59936ee (this session's protocol-correction turn)

Handoff Retrieval Hints

• query_raw_memories(query="epic-resolution-workflow §4 §0 invariant misframing epic-close ≠ merge gate")
• File anchor: .agents/skills/epic-resolution/references/epic-resolution-workflow.md §4 "Verdict authority"
• Empirical anchor: #10691 closeout review 2026-05-04 → epic closed 2026-05-11 (~7 day delay)

🤖 Generated with Claude Code