Agent-Skills.md

Agent Skills: review

Review local changes, a PR, or an issue for quality, correctness, security. Use when user asks for review, wants feedback on changes, or says /review.

UncategorizedID: keenranger/dotfiles/review

Repository

keenranger/dotfiles
keenranger

Install this agent skill to your local

pnpm dlx add-skill https://github.com/keenranger/dotfiles/tree/HEAD/agent/skills/review

Skill Files

Browse the full folder contents for review.

Download Skill

Loading file tree…

agent/skills/review/SKILL.md

Skill Metadata

Name
review
Description
Review local changes, a PR, or an issue for quality, correctness, security. Use when user asks for review, wants feedback on changes, or says /review.

Review target: $ARGUMENTS

Route to one of three modes based on the target, dispatch the matching reviewers in parallel, then synthesize a single verdict.

1. Classify the target

  • Empty, file paths, "staged", or a commit range -> LOCAL changes
  • Issue URL, "issue N", or #N that resolves to an issue -> ISSUE
  • PR URL, "pr N", or #N that resolves to a PR -> PR

Resolve a bare #N with gh pr view N, falling back to gh issue view N. An explicit keyword in the arguments (local, pr, issue) overrides detection.

2. Dispatch reviewers (parallel)

  • LOCAL: git-diff-reviewer agent (Claude) + codex:codex-rescue agent (Codex/GPT-5.4)
  • PR: pr-issue-reviewer agent (fetches gh pr view + gh pr diff) + codex:codex-rescue agent given the gh pr diff N output as explicit input -- codex:codex-rescue reads local git state, so the PR branch may not be checked out; pass the fetched patch instead of relying on the workspace
  • ISSUE: pr-issue-reviewer agent (fetches gh issue view; instruct it to report an issue verdict -- ready to implement / needs clarification -- not its default PR merge verdict) + a general-purpose agent that investigates the codebase for feasibility
  • Complex or critical targets: add 1-2 general-purpose agents for additional perspectives

After all agents return, normalize each reviewer's severity labels into critical/major/minor (e.g. git-diff-reviewer's Suggestions, pr-issue-reviewer's Blockers/Important/Consider), then deduplicate and elevate issues flagged by multiple reviewers.

3. Focus areas (apply what fits the target)

Code (LOCAL / PR):

  • Code quality: readability, structure, patterns, technical debt
  • Correctness: logic errors, edge cases, error handling
  • Testing: coverage, test quality, missing scenarios
  • Security: vulnerabilities, input validation, sensitive data
  • Performance: bottlenecks, optimization opportunities
  • Breaking changes: API changes, backward compatibility
  • Scope: verify changes are limited to what was requested -- flag additions, refactors, or config changes beyond the stated purpose

Issue (ISSUE):

  • Clarity: is the problem and desired outcome unambiguous?
  • Reproducibility: steps, environment, expected vs actual behavior
  • Acceptance criteria: is "done" defined and verifiable?
  • Feasibility: grounded in the actual codebase, not assumptions
  • Risks and unknowns: what could make this harder than it looks?

4. Output

  • Executive summary (1-2 paragraphs)
  • Detailed findings (organized by severity: critical, major, minor)
  • Risk assessment (low/medium/high with rationale)
  • Verdict and recommended actions:
    • LOCAL: ready to commit / split into commits / fixes needed
    • PR: ready to merge / needs work
    • ISSUE: ready to implement (with a suggested approach) / needs clarification

Goal: Catch real issues while avoiding false positives.