Local Review

local-review

Overview

Run a disciplined local code review on git changes with no GitHub side effects. Use this skill when the user wants a /review-style assessment of local diffs.

Scope

• In scope: local diff review (working, staged, or branch compare), risk analysis, and action-focused findings.
• Out of scope: posting PR reviews to GitHub. Use the gh CLI or the harness's own PR-review command.

Workflow

1. Select review target.

• Working tree changes: --mode working (default)
• Staged changes only: --mode staged
• Branch changes against a base: --mode branch --base <ref> [--head <ref>]
• Deep review: add --deep to raise the default diff budget from 4000 to 8000 lines, or pass --max-diff-lines <n> explicitly.

2. Collect deterministic context. Resolve scripts/collect_review_context.sh relative to this skill directory, not the repository being reviewed. Do not look for the script in the target repo unless that repo deliberately vendors a copy.

bash <local-review-skill-dir>/scripts/collect_review_context.sh --mode working

or

bash <local-review-skill-dir>/scripts/collect_review_context.sh --mode staged

or

bash <local-review-skill-dir>/scripts/collect_review_context.sh --mode branch --base origin/main

If branch mode cannot resolve the default origin/main but local main exists, the helper falls back to main and prints a base note in the review target.

3. Analyze for:

• correctness and regressions
• security and data integrity
• performance and scalability
• maintainability and readability
• API or contract compatibility
• test coverage gaps

4. Produce response in the exact output contract below.

Output Contract

Always present sections in this order:

1. Findings

• Findings must come first.
• Order by severity: CRITICAL, HIGH, MEDIUM, LOW.
• Include file and line when available.
• Use this structure for each finding:
  • Severity: <level>
  • Location: <path:line>
  • Issue: <concise problem statement>
  • Risk: <why this matters>
  • Recommended fix: <specific fix>
  • Tests: <missing or required tests>

2. Open Questions / Assumptions

• List unknowns that affect confidence.

3. Change Summary

• Keep short. This is secondary to findings.

4. Residual Risks / Testing Gaps

• Explicitly call out what was not verified.

If there are no meaningful findings, state that explicitly and still include residual risks and test gaps.

Review Depth

• Quick review: small changes and low-risk files.
• Deep review: large diff, sensitive areas (auth, migrations, payments, infra), or user request.
• Trigger deep review automatically when diff is very large (for example, over 20 files or over 500 changed lines). If the context says the diff was truncated, rerun with --deep or a higher --max-diff-lines, then read high-risk touched files directly as needed.

Commands

# Default: review working tree changes
bash <local-review-skill-dir>/scripts/collect_review_context.sh

# Review only staged changes
bash <local-review-skill-dir>/scripts/collect_review_context.sh --mode staged

# Review branch changes vs base
bash <local-review-skill-dir>/scripts/collect_review_context.sh --mode branch --base origin/main

# Deep review with a larger default diff budget
bash <local-review-skill-dir>/scripts/collect_review_context.sh --mode branch --base origin/main --deep

Command Wrapper

If the harness supports command files, use commands/review.md as the canonical /review wrapper for this skill.

Boundaries

• Do not post reviews to GitHub or call the GitHub API
• Do not modify source code; this skill is read-only analysis
• Do not review files outside the git diff scope

Verification

• All findings reference specific files and line numbers from the diff
• Severity ratings are justified by stated risk
• Residual risks section is present even when no findings exist

Notes

• Prefer precise, evidence-backed findings over broad advice.
• Do not make GitHub review API calls in this skill.

Sibling skills

One of three review skills (diff / PR / multi-agent).

• verify-before-complete — orthogonal gate. Run before claiming a diff is "done"; this skill produces the findings, that one enforces the completion claim.
• error-handling-review — deliberately-invoked specialist lens for silent failures, swallowed errors, and unjustified fallbacks. Reach for it when this general pass flags error handling worth a deeper look.
• type-design-review — deliberately-invoked specialist lens rating a new or changed type's invariants and encapsulation. Reach for it when this pass surfaces a new type worth assessing.