skillZs
★ LIVE SKILL TAGS ★
>>> LIVE SKILLS INDEX <<<
* OPEN SOURCE *
NO LOGIN, NO TRACKING
※ REAL INSTALL DATA ※
← back to all skills
lwlee2608/agent-skills116 installs

review-loop

Use only when the user explicitly asks to review and fix in a loop until clean. Not for plain reviews (use review-code); never invoke proactively.

How do I install this agent skill?

npx skills add https://github.com/lwlee2608/agent-skills --skill review-loop
view source ↗

Is this agent skill safe to install?

  • Gen Agent Trust Hubpass

    The 'review-loop' skill automates an iterative process of reviewing and fixing code until it meets quality standards. It leverages local build tools (like make) and the GitHub CLI to manage development workflows. While the skill interacts with repository scripts and external issue data, its operations are strictly confined to the intended primary purpose of developer-led code maintenance.

  • Socketpass

    No alerts

  • Snykfail

    Risk: CRITICAL · 3 issues

What does this agent skill do?

Review Loop

Review, fix, re-review until a round finds nothing worth fixing. Fixes are new code — they need their own review.

Before round 1

If the target is empty or bare pr and the branch has no PR, ask whether to push and open one. Yes: push (branch first if on the default branch), open it with gh-create-pr, review it. No: review diff, no push.

Rules

  1. Review in a fresh subagent every round. Invoke review-code with <target> (default diff). If review-code is missing, spawn a subagent to review for correctness, security, resource, and performance defects, rating each finding Yes / Judgment call / No for worth fixing. Never review your own fixes inline.

  2. Pass -2o only if the user passed it, and only to round 1. Without it, no round gets a second opinion. Round 1 reviews the full change; later rounds review small fix diffs, where another agent adds cost but little coverage. Its Final Call reads code before any fixes exist, so it does not break rule 1.

  3. Use diff from round 2 for every target, including paths. GitHub does not see unpushed fixes, re-reviewing the whole codebase wastes context, and path-scoped reviews can miss fixes in related files. For pr <number>, check out the PR first.

  4. Fix every Yes; fix a Judgment call only if Trivial or Small. Skip No. When a round has a Final Call, act on it instead: Fix → fix, Optional → fix only if Trivial or Small, Skip → skip. The reviewer has not seen this session: skip a finding that contradicts a decision you can quote from the user, plan, or issue, never your own choice. Leave it open if Critical or High. Do not re-fix or re-argue a skipped item when a later round raises it again.

  5. Run the repo's checks after each round's fixes (make build / make test / make lint, else native commands). Fix breakage before the next review.

  6. Stop when a round has nothing new to fix. Cap at 5 rounds; round 5 is review-only, even if it finds actionable issues. Do not edit after that review — hand the open findings to the user. If a fixed finding comes back in any round, stop fixing and report the open findings. When the loop stops, check every checklist item in the linked issue against the code; find it from the target PR or the current branch's PR (gh pr view --json closingIssuesReferences), whatever the target; an item not done is an open finding unless a stated reason drops it.

  7. Commit each fix unless told otherwise. --commit each (default) — one commit per fixed finding. --commit end — one commit after the loop stops. Never push, except under --merge (rule 8). On the default branch, branch before the first commit — otherwise diff anchors at HEAD and committed fixes drop out of the next review.

  8. With --merge, merge only a clean loop. Merge only if no findings are open and checks pass; otherwise skip it and say why. If the branch has no PR or merge-pr is missing, say so and stop without a push. Otherwise push, then merge its PR by reading and following merge-pr's SKILL.md. The Skill tool refuses it because it is user-only; the flag is the user's go-ahead.

  9. Report at the end: rounds run and why it stopped, fixes (path:line — what changed), skipped findings with a reason each, open findings, check results, and the merge result under --merge.

Add the canonical catalog link to the repository README so users can inspect current installs and available audits. The publishing guide covers the complete discovery path.

<a href="https://skillzs.dev/skills/lwlee2608/agent-skills/review-loop">View review-loop on skillZs</a>