A CvC (Claude vs Claude) policy for CoGames. Clone this repo, improve the policy, and submit to compete.

# Clone
git clone https://github.com/Metta-AI/cogamer-policy-cvc.git
cd cogamer-policy-cvc

# Install
pip install -e ".[llm]"

# Play a game
softmax cogames play -m machina_1 -p class=cvc_policy.cogamer_policy.CvCPolicy --render=gui

# Evaluate
softmax cogames eval -m machina_1 -p class=cvc_policy.cogamer_policy.CvCPolicy -e 10 --format json

# Submit
cogames upload -p class=cvc_policy.cogamer_policy.CvCPolicy -n my-policy --setup-script setup_policy.py

The policy is a program table with 32 programs operating on GameState:

• Query programs — read game state (HP, position, inventory, junctions)
• Action programs — movement via A* pathfinding
• Decision programs — compose queries + actions (role selection, mining, combat)
• LLM program — periodic Claude calls for strategic analysis

See docs/architecture.md for details.

src/cvc_policy/          # Policy implementation
  cogamer_policy.py      # CvCPolicy entry point
  programs.py            # Program table (32 programs)
  game_state.py          # Observation processing + state
  agent/                 # Engine: roles, targeting, navigation, etc.
docs/                    # Architecture and strategy reference
tests/                   # Unit tests
setup_policy.py          # Setup script for cogames upload