Autonomous Code Refactoring Agent (Local Llama 3.1)

An autonomous AI agent that performs static code analysis and refactoring on your local codebase. Built with Python, LangChain, and Ollama, it runs completely offline using the Llama 3.1 Large Language Model.

Features

• Autonomous Agentic Loop: Uses a ReAct (Reason + Act) loop to read files, plan refactors, and execute changes.
• RAG Memory: Ingests the entire codebase into a local Vector Database (ChromaDB) to understand context across multiple files.
• Safety Guards: Includes a syntax validation tool (ast module) that prevents the agent from saving broken Python code.
• Privacy-First: Runs 100% locally on your machine—no code leaves your computer.

Tech Stack

• Core Logic: Python 3.9+
• Orchestration: LangChain
• LLM: Llama 3.1 (8B) via Ollama
• Vector DB: ChromaDB
• CLI: Typer

Installation

1. Clone the repo

   git clone https://github.com/VS251/code-refactor-agent.git
   cd code-refactor-agent

2. Install Dependencies

   python3 -m venv venv
   source venv/bin/activate
   pip install -r requirements.txt

3. Setup Ollama

   • Download Ollama
   • Pull the model: ollama pull llama3.1
   • Start the server: ollama serve

Usage

1. Ingest your codebase (Build Memory)

python ingest.py

2. Run the Agent

python main.py --file your_script.py --instruction "Refactor this to use a class structure and add type hints."

Architecture

• Ingestion: Scans .py files -> Embeds using all-MiniLM-L6-v2 -> Stores in ChromaDB.
• Retrieval: Agent queries ChromaDB for context ("How is this function used elsewhere?").
• Action: Agent uses read_file -> validate_syntax -> write_file tools to safely modify code.