Port reviewer self-destruct + tight-PR norm to claude-gh / claude-jira / claude-local / .claude (post-#160)

[martin-conur]

Summary

PR #160 introduced two prompt-doc improvements in the claude-notion loadout only:

1. Reviewer self-destruct — the /reviewer worker, after posting its PR comment + radioing PM, runs task-done --remove-worktree as its final action (single-shot, self-cleaning) instead of idling with the tab open.
2. Tight-PR norm — reviewer frames every finding as fix-in-this-PR (no "defer to follow-up" for in-scope work); PM forwards all in-scope findings (blockers + nits) in one changes-requested round and grooms only genuinely out-of-scope work into a ticket.
3. Plus a PM-side note that Bash(task-reviewer *) should be allowlisted in the project's .claude/settings.json for hands-off dispatch, and a review-complete-with-findings tweak ("read its verdict on your next radio check, then forward").

#160's own body explicitly deferred porting these to the other loadouts. This ticket is that port.

Reference

The canonical diff to replicate is PR #160 (branch chore/reviewer-self-destruct-tight-pr), which edits three claude-notion files:

• claude-notion/commands/reviewer.md — steps 7 & 9 (idle → auto-destruct) + tight-PR norm paragraph after the verdict step.
• claude-notion/commands/pm.md — Bash(task-reviewer *) permission note, review-complete-with-findings tweak, tight-PR norm paragraph, and the "reviewer tab stays open" → "single-shot and self-cleaning" rewrite.
• claude-notion/steering/notion-workflow.example.md — two new sections: "Reviewer role (single-shot, self-cleaning)" and "Tight-PR norm — minimize deferrals".

This worktree is forked from #160's branch, so the claude-notion changes are already present here as the live reference — diff them against origin/main to see exactly what to mirror.

Scope — port to the 4 remaining loadouts (12 files)

Apply the equivalent changes, adapting tracker-specific wording (spec source, MCP/CLI references) — do NOT copy claude-notion's Notion-specific phrasing verbatim:

Loadout	reviewer prompt	pm prompt	workflow/steering doc
.claude (dogfood)	.claude/commands/reviewer.md	.claude/commands/pm.md	.claude/gh-workflow.md
claude-gh	claude-gh/commands/reviewer.md	claude-gh/commands/pm.md	claude-gh/steering/gh-workflow.example.md
claude-jira	claude-jira/commands/reviewer.md	claude-jira/commands/pm.md	claude-jira/steering/jira-workflow.example.md
claude-local	claude-local/commands/reviewer.md	claude-local/commands/pm.md	claude-local/steering/local-workflow.example.md

Tracker-specific adaptations:

• claude-gh / .claude — spec lives in a GitHub issue; use gh issue view. (.claude is the dogfood copy and tracks GitHub, hence gh-workflow.md.)
• claude-jira — spec is a Jira issue (key like PROJ-123) via the Atlassian MCP; keep the existing Jira phrasing.
• claude-local — spec is a local task file/slug; keep the existing local-file phrasing.
• The self-destruct mechanic (task-done --remove-worktree), the tight-PR norm, and the Bash(task-reviewer *) allowlist note are tracker-agnostic — port them faithfully.

Out of scope

• kiro- loadouts* — they have no commands/reviewer.md (different mechanism); claude-notion: reviewer self-destruct + tight-PR norm + PM reviewer perms #160 excluded them and so does this ticket.
• No bin/ script changes — this is prompt/doc-only.

Acceptance criteria

• All 12 files above carry the self-destruct + tight-PR norm + Bash(task-reviewer *) changes, with tracker-appropriate wording.
• Each loadout's reviewer prompt instructs task-done --remove-worktree as the final action (no "idle, tab stays open").
• No claude-notion regressions (already done in claude-notion: reviewer self-destruct + tight-PR norm + PM reviewer perms #160 — don't re-touch it) and no kiro changes.
• tools/check-drift.sh still passes (it covers bin/task-reviewer only, but run it to be safe).
• Existing test suites stay green.

Sequencing note

This PR is stacked on #160 (forked from chore/reviewer-self-destruct-tight-pr, targets main). Until #160 merges, this PR's diff will include #160's claude-notion changes; that resolves once #160 lands. Review/merge #160 first where possible.

[martin-conur]

PM: expanding scope to all 5 claude loadouts — #160 is superseded

Requested again 2026-09-15: "the reviewer worker should have in its body that after each
review it should destroy itself; PM should know this behaviour too." That is exactly this
ticket, so it moves to Ready and the scope grows by one loadout.

Why absorb claude-notion rather than merge #160 first

PR #160 has been open since 2026-06-20 — nearly three months. It is CONFLICTING
against current main, and main has moved a great deal underneath it: the #191 runbook
now lives in every steering template, #183 restructured those files with managed-region
markers, and #194 added a ci-guard section inside the managed region. Rebasing a
three-month-old prompt-doc branch through all of that, purely to then port it to four more
loadouts, is more work and more risk than implementing all five consistently in one pass.

So: implement all 5 claude loadouts here, including claude-notion. #160's diff stays
useful as the reference for what to say — its content is fully described in this ticket's
body — but it is not the thing to merge. I will close it as superseded when this lands.

Files — 15, not 12

Loadout	reviewer prompt	pm prompt	workflow/steering doc
.claude (dogfood)	.claude/commands/reviewer.md	.claude/commands/pm.md	.claude/gh-workflow.md
claude-gh	claude-gh/commands/reviewer.md	claude-gh/commands/pm.md	claude-gh/steering/gh-workflow.example.md
claude-jira	…	…	claude-jira/steering/jira-workflow.example.md
claude-local	…	…	claude-local/steering/local-workflow.example.md
claude-notion	…	…	claude-notion/steering/notion-workflow.example.md

Current state to change

All the reviewer prompts still say the opposite of the intent — each carries two mentions
of "do NOT auto-cleanup" / "the tab stays open". The PM prompts and workflow docs likewise
describe the reviewer tab as something the user cleans up by hand.

Constraints that did not exist when this was filed

• The steering docs now have managed-region markers (task-init: preserve user-authored sections in .claude/<loadout>-workflow.md across re-runs (managed-region merge) #183). Reviewer/PM guidance is
  template-owned, so it belongs inside the region — task-init must refresh it
  downstream. Do not put it below the end marker; that space is the user's.
• tests/radio_runbook.bats asserts the doc blocks are byte-identical within each
  loadout group (4 claude / 3 kiro). Your edits land in the claude group — keep all four
  in step, and note .claude/gh-workflow.md is the dogfood render of claude-gh's
  template, so those two must agree.
• Wording stays repo-generic per worker.md: pre-PR checklist (CHANGELOG, doc-beat, reuse, spec-comments, green gate) (#176) #177; adapt only tracker-specific phrasing (spec source,
  CLI vs MCP) as the table in the body sets out.

One judgment call to make and record

Self-destruct means the analysis scrolls away with the tab. The reviewer already posts its
full findings as a PR comment and radios PM a verdict, so nothing is lost — but say so
explicitly in the PM prompt, so a PM that wants to re-read a review knows to look at the PR
rather than hunting for a tab that cleaned itself up.