Multi-agent AI code review GitHub Action powered by RocketRide. Three independent AI agents (Claude, GPT, Gemini) review your pull requests in parallel, each posting under its own GitHub App identity.
# .github/workflows/review.yml
name: RocketRide PR Review
on:
pull_request:
types: [opened, synchronize]
branches: [main]
issue_comment:
types: [created]
pull_request_review_comment:
types: [created]
jobs:
review:
runs-on: ubuntu-latest
steps:
- uses: rocketride-org/rocketride-reviewer@v1
with:
openai_api_key: ${{ secrets.OPENAI_API_KEY }}
anthropic_api_key: ${{ secrets.ANTHROPIC_API_KEY }}
google_api_key: ${{ secrets.GOOGLE_API_KEY }}
claude_app_id: ${{ secrets.CLAUDE_APP_ID }}
claude_app_private_key: ${{ secrets.CLAUDE_APP_PRIVATE_KEY }}
gpt_app_id: ${{ secrets.GPT_APP_ID }}
gpt_app_private_key: ${{ secrets.GPT_APP_PRIVATE_KEY }}
gemini_app_id: ${{ secrets.GEMINI_APP_ID }}
gemini_app_private_key: ${{ secrets.GEMINI_APP_PRIVATE_KEY }}That's it. Open a PR targeting main and all three agents will post reviews within minutes.
- A PR is opened or updated targeting your main branch
- Three AI reviewers run in parallel via RocketRide's pipeline engine
- A Claude-based aggregator deduplicates overlapping comments across agents
- Each agent posts its review under its own GitHub App identity (
claude-reviewer[bot],gpt-reviewer[bot],gemini-reviewer[bot]) - If no critical or high severity issues are found, all three agents auto-approve the PR
- Reply to any agent's comment (in PR comments or inline review comments) and only that agent responds -- scoped, conversational follow-up with full thread context
| Agent | Model | Role |
|---|---|---|
claude-reviewer[bot] |
Claude Sonnet | Independent code reviewer |
gpt-reviewer[bot] |
GPT-4o | Independent code reviewer |
gemini-reviewer[bot] |
Gemini 2.0 Flash | Independent code reviewer |
| Aggregator | Claude Sonnet | Deduplicates overlapping comments |
Full Review — Triggered on PR open/update. All three agents review the diff in parallel. The aggregator removes duplicate findings (same file, same line range, same intent) before posting.
Conversation Reply — Triggered when a developer replies to an agent's comment via issue_comment or pull_request_review_comment events. Only the originating agent runs, receiving the full comment thread and surrounding file context. The other two agents stay silent.
| Condition | Result |
|---|---|
| Zero critical/high findings across all agents | All three agents submit Approve |
| Any critical/high finding | Flagging agent(s) submit Request Changes, others submit Comment |
Create an optional .rocketride-review.yml in your repository root:
# Review scope: "full" (include file context) or "diff" (diff only)
review_context: full
# Target branch (default: main)
target_branch: main
# Severity threshold for auto-approval: "critical" or "high"
approval_threshold: high
# Add patterns to the default ignore list
ignore_patterns_extra:
- "migrations/**"
- "*.sql"
# Or replace the defaults entirely
# ignore_patterns_override:
# - "*.lock"
# - "dist/**"
# Large PR handling
max_chunk_lines: 500 # Max lines per diff chunk
chunk_overlap_lines: 20 # Overlap between segments
max_files: 50 # Skip review above this many changed files
max_total_lines: 5000 # Skip review above this many changed linesLock files, minified assets, generated code, build output, images, and fonts are ignored by default. See the PRD for the full list.
You need three LLM API keys and three GitHub App credential pairs (one per reviewer agent):
| Secret | Description |
|---|---|
OPENAI_API_KEY |
OpenAI API key for GPT reviewer |
ANTHROPIC_API_KEY |
Anthropic API key for Claude reviewer and aggregator |
GOOGLE_API_KEY |
Google AI API key for Gemini reviewer |
CLAUDE_APP_ID |
GitHub App ID for claude-reviewer[bot] |
CLAUDE_APP_PRIVATE_KEY |
Private key for claude-reviewer[bot] |
GPT_APP_ID |
GitHub App ID for gpt-reviewer[bot] |
GPT_APP_PRIVATE_KEY |
Private key for gpt-reviewer[bot] |
GEMINI_APP_ID |
GitHub App ID for gemini-reviewer[bot] |
GEMINI_APP_PRIVATE_KEY |
Private key for gemini-reviewer[bot] |
Each GitHub App needs these permissions:
- Pull requests: Read & Write
- Issues: Read & Write
- Contents: Read
See the Setup Guide for step-by-step instructions on creating the GitHub Apps.
| Input | Required | Default | Description |
|---|---|---|---|
openai_api_key |
Yes | — | OpenAI API key |
anthropic_api_key |
Yes | — | Anthropic API key |
google_api_key |
Yes | — | Google AI API key |
claude_app_id |
Yes | — | Claude reviewer GitHub App ID |
claude_app_private_key |
Yes | — | Claude reviewer GitHub App private key |
gpt_app_id |
Yes | — | GPT reviewer GitHub App ID |
gpt_app_private_key |
Yes | — | GPT reviewer GitHub App private key |
gemini_app_id |
Yes | — | Gemini reviewer GitHub App ID |
gemini_app_private_key |
Yes | — | Gemini reviewer GitHub App private key |
target_branch |
No | main |
Only review PRs targeting this branch |
review_context |
No | full |
full or diff |
config_path |
No | .rocketride-review.yml |
Path to config file |
# Install dependencies
pip install -r requirements-dev.txt
# Run tests
pytest tests/ -v
# Run tests with coverage
pytest tests/ -v --cov=src --cov-report=term-missing
# Lint
ruff check src/ tests/
# Format check
black --check src/ tests/
# Type check
mypy src/src/
├── main.py # Entry point: event detection, gating, orchestration
├── github_client.py # GitHub API wrapper (diffs, comments, reviews)
├── engine.py # RocketRide engine lifecycle (Docker)
├── pipeline.py # Pipeline execution (full review + conversation)
├── aggregator.py # Comment deduplication across agents
├── reviewer.py # Review posting per GitHub App identity
├── chunker.py # Large PR diff chunking and line remapping
├── config.py # Configuration and constants
├── models.py # Pydantic models (ReviewComment, AgentReview, etc.)
├── filters.py # File ignore pattern matching
├── retry.py # Retry with exponential backoff for transient errors
└── errors.py # Exception hierarchy