review-loop_skill

Overview

This skill performs iterative, multi-pass subagent code reviews with an automated fix–review loop. It runs three built-in review passes (Direct, Best Practice, Critical Think), suggests context-aware optional passes, applies safe fixes, and re-runs affected passes until the code is clean or iteration limits are reached. Use it for high-stakes changes that need multi-perspective verification before merge or deploy.

How this skill works

I first identify the review target (files, diffs, staged/unstaged changes, PRs, or a directory). Then I launch independent subagents for the three built-in passes in parallel and optionally for suggested passes triggered by code signals. Findings use a structured format (Pass, Severity, Location, Issue, Suggested Fix). Low/medium fixes are auto-applied, high-severity or complex fixes require user approval. After fixes I re-run only the passes affected and repeat until exit conditions are met.

When to use it

• Pre-merge or pre-deploy reviews for critical features or security-sensitive code
• When you want an iterative review → fix → re-review loop until issues are resolved
• Reviews that benefit from multiple perspectives (correctness, quality, risk)
• Large changes, multi-file diffs, or complex logic where single-pass review misses issues
• As a fallback subagent reviewer when another skill requests thorough verification

Best practices

• Run all three built-in passes for thorough coverage; enable suggested passes as needed
• Provide clear scope (file paths, PR number, or git diff) to avoid ambiguity
• Approve HIGH-severity fixes manually; let the tool auto-apply LOW/MEDIUM fixes
• Limit iterations to the configured maximum (default 3) and escalate stalls or ping-pong
• Accept suggested specialized passes (Testability, Performance, API Surface, etc.) when triggers are present

Example use cases

• Deep review of an authentication module before merging into main
• Run a 3-pass review on a production hotfix branch and auto-apply safe fixes
• Review a PR diff with suggested Performance and Testability passes enabled
• Verify staged changes locally using git diff --cached and iterate fixes
• Use as a fallback verification step when another agent delegates complex review