Gemini Subagent Reviewer (GSR)

Welcome to the GSR project repository! This project implements an advanced multi-agent code review system powered by Gemini.

🧠 Multi-Agent Approach

Instead of relying on a single prompt to review code, GSR uses a swarm of specialized subagents to analyze pull requests from different perspectives. This approach yields higher quality, more focused findings.

Specialized Agents

• Architecture Agent: Focuses on system design, separation of concerns, and adherence to design patterns.
• Logic Agent: Looks for bugs, edge cases, and correctness issues in the code.
• Security Agent: Identifies potential security vulnerabilities (e.g., DoS, injection risks).
• TechDebt Agent: Flags anti-patterns, legacy code usage (e.g., indexOf vs includes), and maintainability issues.
• Testing Agent: Ensures adequate test coverage and identifies missing tests for new logic.

Orchestration & Deduplication

1. Orchestrator: Splits the incoming PR diff into chunks and distributes them to the subagents in parallel.
2. Deduplicator: After all subagents report their findings, a dedicated DeduplicatorAgent (running on gemini-3.1-pro-preview with a sequential lock) merges duplicate findings and selects the most actionable suggestions, reducing noise for the developer.

---

📊 Evaluation Framework

To ensure our agent prompts are effective and not regressing, we use a custom Evaluation Harness located in tools/eval.

How it Works

• Ablation Testing: The harness can run reviews with specific agents disabled to measure their impact.
• Comparison (Local vs Production): It runs reviews against both the local development server and the deployed production server simultaneously.
• Auto-Evaluation: An independent Gemini instance reviews the findings from both targets and generates a detailed comparison report on accuracy, actionability, and noise.
• Persistence: Results are automatically archived and uploaded to an S3-compatible object storage bucket (Cloudflare R2 in production).

---

📁 Project Structure

• adk/: The core application.
  • backend/: Node.js Express server that handles GitHub API interaction and orchestrates Gemini agents.
  • frontend/: Vanilla JS interface to request reviews and view streaming progress.
  • prompts/: Contains the system prompts for all agents (including system_prompts_v2).
• tools/eval/: The evaluation harness used to benchmark performance.

---

🚀 Getting Started

Please see the ADK README for detailed instructions on setting up the environment and running the servers locally.

---

🤖 GitHub Action

GSR can also run as a GitHub Action on your own pull requests — see ACTION.md for setup and configuration (choose between subagent and basic review modes).