This is an AI agent system that can automatically play Tetris. The system uses AI models to control the game, achieving automated gameplay.

The system includes multiple AI iterator files that can be used to control Tetris with different AI models. These are the core components of the system:

Below is a visual comparison of different AI models playing Tetris:

Claude Sonnet 3.7	Gemini Flash 2.0	OpenAI o3-mini	Qwen 2.5 VL 72B Instruct
Claude's vision-based analysis	Gemini's approach to Tetris	OpenAI's decision making	Qwen's reasoning

Note: Some GIFs may be placeholders until recordings for all models are available.

This file specifically uses Claude models from Anthropic to play Tetris:

Claude AI playing Tetris through vision-based analysis

• Supported Models:

  • Claude 3.7 Sonnet (default)
  • Other Claude models can be specified with the --model parameter

• Key Features:

  • Designed specifically for Claude models
  • Direct integration with Anthropic's API
  • Includes simulation capabilities for testing
  • Provides detailed logging of Claude's responses
  • Can capture real screenshots or use simulated boards

• Usage:

  # Use default Claude model
  python tetris_claude_iterator.py
  
  # Specify a different Claude model
  python tetris_claude_iterator.py --model claude-3-opus-20240229
  
  # Use simulation mode with complex board
  python tetris_claude_iterator.py --complex
  
  # Use real screenshots instead of simulation
  python tetris_claude_iterator.py --no-simulate

This is the main, unified iterator that supports multiple AI models through different providers:

• Supported Models:

  • OpenAI o3-mini (direct API) Tested
  • Qwen2.5 VL 72B Instruct (via 302.ai API) Tested
  • Gemini 2.0 Flash (default, via OpenRouter)
  • Gemini Pro 2.0 Experimental (via OpenRouter)
  • Qwen2.5 VL 72B Instruct (via OpenRouter)
  • OpenAI o3-mini-high (via OpenRouter)
  • OpenAI GPT-4 Vision (direct API)
  • Qwen VL Plus (via DashScope API)
  • Qwen VL Max (via DashScope API)

• Key Features:

  • Supports multiple API providers: OpenRouter, OpenAI, DashScope, and 302.ai
  • Automatically selects appropriate output directories based on model
  • Provides fallback mechanisms for API failures
  • Includes simulation mode for testing without a real game
  • Can capture real screenshots or use simulated boards

• Usage:

  # Use Gemini Flash (default)
  python tetris_ai_iterator.py
  
  # Use Gemini Pro Experimental (Tested)
  python tetris_ai_iterator.py --use-pro-exp
  
  # Use Qwen2.5 VL 72B Instruct via 302.ai (Tested)
  python tetris_ai_iterator.py --use-302-ai
  
  # Use OpenAI o3-mini directly (Tested)
  python tetris_ai_iterator.py --use-o3-direct
  
  # Use Qwen model via OpenRouter
  python tetris_ai_iterator.py --use-qwen
  
  # Use OpenAI o3-mini via OpenRouter
  python tetris_ai_iterator.py --use-o3
  
  # Use OpenAI GPT-4 Vision directly
  python tetris_ai_iterator.py --use-gpt4-vision
  
  # Use Qwen VL Plus via DashScope
  python tetris_ai_iterator.py --use-dashscope
  
  # Use Qwen VL Max via DashScope
  python tetris_ai_iterator.py --use-dashscope-max
  
  # Use simulation mode with complex board
  python tetris_ai_iterator.py --complex
  
  # Use real screenshots instead of simulation
  python tetris_ai_iterator.py --no-simulate

This file is designed to use Google's Gemini models via OpenRouter:

• Supported Models:

  • Gemini 2.0 Flash (default)
  • Gemini Pro 2.0 Experimental

• Key Features:

  • Optimized for Gemini models
  • Uses OpenRouter to access Gemini
  • Includes simulation capabilities
  • Provides detailed logging of responses
  • Supports different output directories for different models

• Usage:

  # Use default Gemini model
  python tetris_gemini_iterator.py
  
  # Use Gemini Pro Experimental
  python tetris_gemini_iterator.py --use-pro-exp
  
  # Use simulation mode with complex board
  python tetris_gemini_iterator.py --complex
  
  # Use real screenshots instead of simulation
  python tetris_gemini_iterator.py --no-simulate

The AI iterators follow this general process:

1. Capture Game State: Either take a screenshot of the actual Tetris game or generate a simulated board
2. Send to AI Model: Send the image to the selected AI model (Claude, Gemini, etc.)
3. Get Move Suggestions: The AI analyzes the board and suggests moves as Python code
4. Execute Moves: The system executes the suggested moves using PyAutoGUI
5. Wait for Next Iteration: Wait for user input (space key) to continue to the next move

The system can be used with a real Tetris game or in simulation mode for testing and development.

During gameplay:

• Press q to stop all threads and the game
• Press A in the game window to toggle between AI/player control
• Press ESC in the game window to exit the game
• Press R in the game window to restart the game (when game is over)

1. Game window doesn't display

   • Make sure your system supports Pygame graphical interface
   • Check if all dependencies are installed: pip install -r requirements.txt

2. Module not found error

   • Make sure you're running the program in the correct directory (GamingAgent)
   • Ensure you've activated the correct conda environment: conda activate game_cua

3. Game window cannot be automatically detected

   • Start with manual window parameters: python run_tetris.py --manual_window

4. API key errors

   • Ensure you have the correct API keys in your .env file:
     • ANTHROPIC_API_KEY for Claude models
     • OPENROUTER_API_KEY for OpenRouter models
     • OPENAI_API_KEY for direct OpenAI models
     • DASHSCOPE_API_KEY for DashScope models
     • THREEZEROTWO_API_KEY for 302.ai models