harshit078/code-review-agent1

Code Review Agent — OpenEnv Environment

A reinforcement learning environment where an AI agent learns to review Python code across three difficulty levels.

Real-World Problem

Millions of code reviews happen daily at tech companies. Junior developers miss syntax errors, security vulnerabilities, and style issues. This environment trains an AI agent to perform accurate, structured code reviews — reducing review time and improving code quality.

Environment Overview

How It Works

The agent interacts with the environment through 3 API endpoints:

• —POST /reset — loads a random code sample and returns it as observation
• —POST /step — accepts the agent's text review and returns a reward score
• —GET /state — returns current task info and history

Tasks

Task 1 — Easy: Syntax Error Detection

Agent must identify the exact error type (SyntaxError, IndentationError, etc.) and the line number where it occurs.

Reward tiers:

• —1.0 — correct error type + correct line number
• —0.6 — correct error type, wrong line number
• —0.3 — related keywords mentioned
• —0.0 — missed completely

Task 2 — Medium: Security Vulnerability Detection

Agent must identify security vulnerabilities — SQL injection, hardcoded credentials, eval injection, path traversal, command injection, and more.

Reward tiers:

• —1.0 — 4+ vulnerability keywords matched
• —0.75 — 3 keywords matched
• —0.5 — 2 keywords matched
• —0.25 — 1 keyword matched
• —0.0 — missed completely

Task 3 — Hard: Full Code Quality Review

Agent must cover all 4 review categories. Each category is worth 0.25.

Observation Space

{
  "code": "def login(user, pwd):\n    ...",
  "task_type": "easy",
  "difficulty": "easy",
  "language": "python",
  "instructions": "Review this code and identify all syntax errors..."
}

Action Space

Natural language text — the agent's code review response.

Reward Design

Partial scoring is used throughout — agents are never penalized with binary right/wrong. Every partially correct answer earns credit, which enables meaningful gradient signal for RL training.

Running Locally

# Install dependencies
pip install -r requirements.txt

# Start the environment server
uvicorn app:app --reload --port 8000

# Run the baseline agent
python inference.py

Running with Docker

docker build -t code-review-agent .
docker run -p 7860:7860 \
  -e API_BASE_URL=https://api.openai.com/v1 \
  -e MODEL_NAME=gpt-4o-mini \
  -e HF_TOKEN=your_key_here \
  code-review-agent

Environment Variables

Project Structure

code-review-agent/
├── code_samples/
│   ├── easy_samples.json       # 10 syntax error samples
│   ├── medium_samples.json     # 10 security vulnerability samples
│   └── hard_samples.json       # 5 multi-issue samples
├── graders/
│   ├── easy_grader.py          # Keyword + line number scoring
│   ├── medium_grader.py        # Vulnerability keyword scoring
│   └── hard_grader.py          # 4-category partial scoring
├── environment.py              # Core OpenEnv class
├── app.py                      # FastAPI server
├── inference.py                # Baseline AI agent
├── openenv.yaml                # OpenEnv spec config
├── Dockerfile                  # HF Spaces deployment
└── requirements.txt