Skip to content

code-review.md

Latest commit

 

History

History
85 lines (69 loc) · 3.02 KB

code-review.md

File metadata and controls

85 lines (69 loc) · 3.02 KB
Raw

Code Review Cheatsheet

Overview

  • Automated PR reviews posted as inline comments on GitHub
  • Multi-agent fleet analyzes diff + full codebase context
  • Available for Teams and Enterprise (not with Zero Data Retention)
  • Findings do NOT approve or block PRs; existing review workflow stays intact
  • Average review time: ~20 minutes; average cost: $15-25 per review

Severity Levels

Marker Severity Meaning
🔴 Normal Bug that should be fixed before merging
🟡 Nit Minor issue, not blocking
🟣 Pre-existing Bug in codebase but not introduced by this PR

Setup (Admin Only)

  1. Go to claude.ai/admin-settings/claude-code
  2. Click Setup → install Claude GitHub App to your GitHub org
  3. Required permissions: Contents (R/W), Issues (R/W), Pull requests (R/W)
  4. Choose repositories and set Review Behavior per repo

Review Behavior Options

Mode When reviews run
Once after PR creation Single review when PR opens or is marked ready
After every push Review on each push (most reviews, highest cost)
Manual Only on @claude review comment

Triggering Reviews

@claude review
  • Post as a top-level PR comment (not inline on diff)
  • Put @claude review at the start of the comment
  • Requires owner, member, or collaborator access
  • PR must be open and not a draft
  • Works in any mode; also opts that PR into push-triggered reviews going forward

Customizing Reviews

CLAUDE.md

  • Claude reads CLAUDE.md at every directory level
  • CLAUDE.md violations introduced by the PR are flagged as nit-level findings
  • Bidirectional: if PR makes a CLAUDE.md statement outdated, Claude flags the docs too

REVIEW.md (review-only rules)

Add to repo root — not loaded for regular Claude Code sessions:

# Code Review Guidelines

## Always check
- New API endpoints have corresponding integration tests
- Database migrations are backward-compatible

## Style
- Prefer match statements over chained isinstance checks

## Skip
- Generated files under src/gen/
- Formatting-only changes in *.lock files

View Usage & Cost

Dashboard: claude.ai/analytics/code-review

  • PRs reviewed (daily count)
  • Weekly spend
  • Feedback (auto-resolved comments after developer fixes)
  • Per-repo PR counts and resolved comments

Set a monthly spend cap: claude.ai/admin-settings/usage

Pricing Notes

  • Billed by token usage; counted as extra usage (not against plan's included usage)
  • Costs appear on Anthropic bill regardless of Bedrock/Vertex usage
  • "After every push" mode multiplies cost by number of pushes

Gotchas

  • Not available for Zero Data Retention orgs
  • @claude review must be a top-level comment at the start of the line
  • If a review is already running, the request is queued
  • Default focus is correctness only; add REVIEW.md for style/coverage rules

Source: code-review.md