PR Pilot — Autonomous PR Review

5 agents. One pipeline. Zero manual review.

PR Pilot autonomously reviews pull requests using a 5-agent chain — Reviewer analyzes, Fixer patches, Tester validates, Verifier gates, Escalator decides. Built for the Build with Gemini XPRIZE.

Agent Pipeline

Webhook → Reviewer → Fixer → Tester → Verifier → Escalator → Review Posted

Agent	Role	Latency
Reviewer	Analyzes PR diff for bugs, security, performance, style	~14s
Fixer	Generates minimal code patches for issues found	~29s
Tester	Detects test framework, generates tests for changed code	~16s
Verifier	Validates fixes and tests — no regressions	~17s
Escalator	Decides: auto-approve, request changes, or escalate	~18s

Pipeline total: ~94s | Cost: ~$0.04/review (Gemini 2.5 Flash)

Graduated Autonomy (The Safety Gate)

PR Pilot doesn't just approve or reject. It operates at three graduated levels — each escalating only when confidence demands it.

Level	Action	When
L1 · Comment	Issues found, fixes attached as review comments. No merge blocked.	Low confidence findings
L2 · Request Changes	Medium-confidence issues that should block merge. Fixes provided inline. PR blocked until author responds.	Missing validation, incomplete tests
L3 · Verified Auto-Approve	All agents agree: PR is clean, fixes verified, tests pass. Auto-approves and merges. No human in the loop.	Clean PRs, all gates passed

L1:  Reviewer: "line 42: unsafe shell exec"
     Fixer: "Use subprocess.run with list args"
     Escalator: COMMENT

L2:  Reviewer: "missing input validation"
     Fixer: "Add pydantic model for request body"
     Verifier: "Patch compiles, needs tests"
     Escalator: REQUEST_CHANGES

L3:  Reviewer: "No issues found"
     Tester: "Coverage maintained at 87%"
     Verifier: "All gates passed"
     Escalator: VERIFIED_AUTO_APPROVE

The safety gate is what separates autonomous review from automated rubber-stamping.

Architecture

FastAPI webhook → AgentChain orchestration → Gemini API → GitHub comment/post

• Cloud Run deployable — one gcloud run deploy command, zero-downtime, auto-scaling
• Full audit trail — every agent's reasoning, every fix generated, every decision logged to data/reviews/. Structured JSON for compliance and transparency.
• GitHub App integration — webhook receiver, PR API client, review comment posting
• Agent operations dashboard — real-time review status UI at /dashboard. Disabled by default; enable with DASHBOARD_TOKEN in your .env. See .env.example for details.

Quick Start

# Install
pip install -e ".[dev]"

# Configure
cp .env.example .env
# Fill in GEMINI_API_KEY, GITHUB_APP_ID, etc.

# Deploy to Cloud Run
gcloud run deploy pr-pilot \
  --source . \
  --region us-central1 \
  --allow-unauthenticated \
  --set-env-vars GEMINI_API_KEY=...,GITHUB_APP_ID=...

# Open a PR — PR Pilot reviews it
git checkout -b feat/new-feature
git push && gh pr create
# Seconds later: @pr-pilot-bot commented: "Reviewed. 2 suggestions. 0 issues."

XPRIZE Compliance

Built for the Build with Gemini XPRIZE. Every requirement met with documentation to match.

Requirement	Implementation
Gemini API	All 5 agents use Gemini 2.5 Flash via google-generativeai. Structured output support. Exponential retry.
Google Cloud Deploy	Cloud Run (serverless), Cloud Build (CI/CD), Artifact Registry (containers).
Stripe Integration	Stripe Checkout for premium features. Usage-based billing via Stripe Metered billing API. Webhook handling.
Code Quality	Type hints throughout. Structured logging (structlog). 85%+ test coverage. Pre-commit hooks.
Open Source	MIT licensed. Public repo with contribution guide, issue templates, and pull request template.

Requirements

• Python 3.11+
• Gemini API key
• GitHub App (for webhook events and PR API access)
• Google Cloud Run (optional — works locally too)

License

MIT — see LICENSE

---

Built for the Build with Gemini XPRIZE · Website