PR Review Agent Base

Summary of: deepresearch/dev-workflow/pr-review-agent-base.md

Key Points

• Core architecture: GitHub Actions workflow triggers Claude API on PR events
• Diff processing: Fetch PR diff via GitHub API, focus only on changed files
• Prompt structure: Static context (CLAUDE.md) + dynamic diff + output schema
• Output format: Structured JSON with findings array and summary object
• Comment posting: Use GitHub Review API for inline comments with position mapping
• Confidence filtering: Only post findings with confidence >= 0.8
• Error handling: Skip oversized PRs, handle API failures gracefully
• Security: Never expose API keys, use GitHub Secrets

Critical Insights

1. Diff-aware is essential - Reviewing entire codebase creates noise; focus on changes only
2. Position mapping is complex - Converting file lines to diff positions requires careful algorithm
3. Start simple, iterate - Begin with single-pass review, add multi-pass filtering based on false positive rates

Quick Reference

Component	Purpose	Implementation
Trigger	Start review	GitHub Actions on PR events
Diff fetch	Get changes	GitHub API (accept: diff media type)
Prompt	Instruct Claude	Static context + diff + schema
Output	Parse results	JSON with findings array
Comments	Post feedback	GitHub Review API
Filter	Reduce noise	Confidence threshold >= 0.8

Workflow Steps

1. PR opened/updated triggers GitHub Action
2. Fetch PR diff and changed files list
3. Check size limits (skip if too large)
4. Build prompt with CLAUDE.md + diff
5. Call Claude API
6. Parse JSON response
7. Filter by confidence threshold
8. Post review comments via GitHub API

Minimum Viable Implementation

on:
  pull_request:
    types: [opened, synchronize]
permissions:
  contents: read
  pull-requests: write