Werewolf LLM Project

This project simulates the classic social deduction game "Werewolf" using LLM (Large Language Model) agents. Each player is controlled by an LLM (mocked by default), and the entire game process is recorded, including role assignment, speeches, voting, and deaths.

Project Structure

• main.py - Entry point, sets up and runs the game
• game_engine.py - Game engine and flow control
• player_agent.py - Player and LLM agent logic
• roles.py - Role definitions
• logger.py - Structured game logging
• llm_api.py - LLM API interface (自行修改base_url和api_key)
• requirements.txt - Dependencies
• README.md - Project documentation

How to Run

1. Make sure you have Python 3.7+ installed.
2. (Optional) Install dependencies:

   pip install -r requirements.txt

3. Run the game:

   python main.py

4. After the game finishes, check game_log.json for the full game log.

Customization & Extension

• To use a real LLM API, implement call_llm_api in llm_api.py with your provider (e.g., OpenAI, Qwen, etc.).
• You can adjust the number of players and role distribution in main.py.
• Extend roles and actions in roles.py and game_engine.py for more complex gameplay.
• The log structure is JSON and can be used for visualization or analysis.

License

MIT