Code Debug Agent

An automated agent for debugging and fixing bugs in Python code using LLM (Qwen) and LangGraph.

Description

This project implements an automated code debugging system based on "Qwen3-0.6B" LLM. The agent analyzes buggy code from the HumanEvalPack dataset and generates fixed code. After generation, the code is tested in isolated Docker environment.

Project Structure

buggy_agent/
├── main.py              # Entry point
├── code_agent.py        # LangGraph agent for code fixing
├── code_model.py        # Qwen model wrapper
├── code_intepretor.py   # Docker runner for code execution
├── evaluation.py        # Evaluation system
├── prompts.py           # LLM prompts
└── requirements.txt     # Project dependencies

Install Dependencies

pip install -r requirements.txt

Run the agent

python main.py

Workflow

1. Load Data: Load HumanEvalPack dataset with buggy code
2. Prepare Prompt: Create prompt with task description, error type, and examples
3. Generate Fixed code: LLM analyzes code and generates fixed version
4. Testing: Fixed code runs with unit tests in Docker
5. Evaluation: Calculate Pass@1 metric and detailed statistics