feat(repo-fleet-hygiene): apply fleet action plans behind one gate

[kyle-sexton]

Closes #2597

Summary

Adds repo-fleet-hygiene:apply, a sibling verb that consumes the existing fleet action-plan JSON behind one batch-wide confirmation gate. Audit stays read-only; execution is not a flag on audit-fleet.sh.

Fix

• New skills/apply/ with apply-plan.sh: --plan-file required, dry-run default, --apply + interactive confirmation or --yes
• Re-derives mutable OIDs / tips before destructive steps; fail-closed on drift
• Preserves plan ordering (merged local branches before worktree cleanup)
• Own allowed-tools surface so audit's Bash grant is not silently widened
• README/SKILL contract updated; fleet version bumped; CATALOG regenerated
• Tests for dry-run, confirmation stop, OID drift, ordering

Verification

bash plugins/repo-fleet-hygiene/skills/apply/scripts/apply-plan.test.sh
bash plugins/repo-fleet-hygiene/skills/audit/scripts/audit-fleet.test.sh

Related

Refs #2608 #2609 — action-plan prerequisite.
Refs #2605 — stranded-work delegation unblocked execution.

  

[cursor]

Bugbot is not enabled for your account, so this pull request was not reviewed.

Enable Bugbot in the Cursor dashboard to get automatic reviews on future PRs.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 715558fb9a

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[claude]

Claude finished @cursor[bot]'s task in 3m 37s —— View job

---

Claude Code Review

Todo

• Gather context (PR diff, CLAUDE.md, REVIEW.md)
• Run /review:code-review criteria against changed files
• Check whether the automated Codex findings were addressed
• Final summary

Summary

This PR ships in two commits. The second (b406289, "harden apply-plan validation and Bash 3.2") is a direct, targeted fix for all three issues the automated Codex review flagged on the first commit:

• P1 — unvalidated plan artifacts (apply-plan.sh): the parser now requires generated_by: repo-fleet-hygiene/audit, mode: read-only, that each action's canonical matches an audited: true repository entry, and that every action target resolves to a real findings[] entry of a matching kind (plugins/repo-fleet-hygiene/skills/apply/scripts/apply-plan.sh#L101-L229). Verified this is exercised by new tests apply-plan.test.sh §6 ("rejects plan without generated_by", "rejects action target without audit finding") — a hand-authored action list with no backing finding is now correctly rejected before any git command runs.
• P1 — Bash-4-only syntax: mapfile was replaced with a portable while read loop, and ${a,,}/${b,,} were replaced with tr '[:upper:]' '[:lower:]' (apply-plan.sh#L247-L252, #L370-L373). No remaining mapfile/,,}/readarray/declare -A usage in the script.
• P2 — empty TSV fields collapsing: the Python→Bash handoff now delimits fields with ASCII Unit Separator (\x1f) instead of tabs, so an intentionally-empty ref_name/expected_oid for prunable-worktree/missing-worktree no longer shifts columns (apply-plan.sh#L84-90, #L560-577). Covered by the new §7 test asserting prunable-worktree decodes to [prune-worktrees] rather than being misread as a skipped merged-worktree action.

I independently read the full apply-plan.sh (763 lines) and its test suite (511 lines) rather than just trusting the fix commit's message, and confirmed the fixes match what they claim. I was not able to execute the test suite myself in this environment (Bash execution outside gh pr * is not permitted for this review lane), so test-passing is based on static reading of the assertions, not an observed run — the PR description's own verification commands are author-claimed and unverified by me.

Docs (README.md, skills/audit/SKILL.md, skills/apply/SKILL.md, security-review.md, CHANGELOG.md) were updated consistently with the new :apply verb, and the allowed-tools-pairing.test.sh guard now covers the apply skill's own script grant so it can't silently widen.

No new correctness/maintainability findings beyond what's already fixed. One residual note, intentionally not raised as a finding here: apply-plan.sh's provenance checks are self-referential (they validate internal consistency of the JSON, not that it was produced by a real prior audit run against real repository state) — the actual backstop against a maliciously-crafted plan remains the printed ordered plan + explicit one-time confirmation gate. Since this repo carries a claude-security-review workflow, that class of finding belongs to the security lane per this org's REVIEW.md split, not this code-review lane.

[github-actions]

Claude has reviewed this PR 2 times. The lane skips further automatic reviews after 5; deleting this comment resets the count.

[claude]

Claude finished @cursor[bot]'s task in 1m 28s —— View job

---

Claude Code Review

Todo

• Gather context (PR diff, prior review comments, CLAUDE.md)
• Invoke /review:code-review and follow its instructions exactly
• Check whether the third commit (since the last lane review) introduces new findings
• Final summary

Summary

The lane already reviewed this PR at commit b406289 (see the comment above) and confirmed the three Codex-flagged P1/P2 issues (unvalidated plan artifacts, Bash-4-only syntax, TSV field collapsing) were fixed. HEAD has since moved to 3b51b43, "clear apply-plan CI hygiene and portability," which is pure lint/portability cleanup:

• oid_re in the embedded Python parser: \s+ → [ \t]+ (apply-plan.sh#L127). The heredoc it lives in is quoted (<<'PY'), so this was never a shell-quoting hazard; functionally a no-op since every evidence string producer in audit-fleet.sh emits headRefOid <oid> with a single literal space, never a newline.
• Three *) ;; no-op default cases added to case statements that previously had no default (worktree_path_for_branch, worktree_locked, refresh_worktree_cleanup) — shellcheck SC2249 silencing, no behavior change.
• # shellcheck disable=SC2310 comments on apply-plan.sh and apply-plan.test.sh for intentional status-in-conditional patterns — accurately scoped, not a blanket disable.
• Test file: dropped an unused OID_UNBOUND capture (add_merged_branch ... >/dev/null now discards output it never read) — correct dead-code removal, the call is still made for its git side effect.
• docs/SKILL-CHEAT-SHEET.md gains the /repo-fleet-hygiene:apply row; CHANGELOG.md drops a stray blank line for markdownlint.

I read this commit's diff in full against the surrounding code (not just the commit message) and found nothing that changes runtime behavior incorrectly — all four code hunks are lint/portability-only, and the regex narrowing is provably a no-op given how the only producer of evidence strings formats them.

No new correctness or maintainability findings on top of what the lane already reported at b406289. Nothing to post inline.

I was not able to execute apply-plan.test.sh or audit-fleet.test.sh myself in this environment (no test-runner tool available to this lane), so the PR description's verification commands remain author-claimed and unverified by me — same limitation noted in the prior review.