---

name: games-and-configs description: "Selects, validates, and adapts MuZero General games, MuZeroConfig fields, custom Game wrappers, and optional environment dependencies." disable-model-invocation: true metadata: disco-role: operating license: MIT

Games and Configs

Use this sub-skill when the task is to choose a built-in MuZero General game, inspect or modify MuZeroConfig, add a new game wrapper, validate observation/action contracts, troubleshoot optional game dependencies, or explain single-player/two-player game behavior.

Fast route

1. Read games-and-configs.md for the built-in game catalog, AbstractGame contract, custom game recipe, and optional dependency map.

2. Read configuration-reference.md before editing MuZeroConfig fields for game shape, self-play, network, training, replay buffer, reanalyse, GPU, or temperature behavior.

3. Read troubleshooting.md when imports fail, observations do not match observation_shape, legal actions are empty/illegal, rendering blocks, or optional dependencies are missing.

4. Discover available game modules without training:

   python sub-skills/games-and-configs/scripts/list_builtin_games.py --format table
   

5. Validate a selected or custom game wrapper without rendering/training:

   python sub-skills/games-and-configs/scripts/validate_game_module.py --module games.tictactoe --json
   

Scope and boundaries

This sub-skill owns:

• games.abstract_game.AbstractGame methods and expected Game wrapper behavior.
• Built-in game names, player counts, observation shapes, action spaces, network defaults, opponent defaults, and optional package needs.
• MuZeroConfig field groups and how they affect training/search/model construction.
• Custom game wrapper authoring and validation steps.
• Dependency handling for Gym classic-control, Box2D LunarLander, Atari/OpenCV/ALE, MiniGrid, and OpenSpiel.

Route elsewhere when the task is primarily about:

• Running training/testing, building CLI commands, Ray workers, or TensorBoard: training-and-cli.
• Debugging model tensor shapes, support transforms, MCTS action masks, or GameHistory stacking: models-and-mcts.
• Loading checkpoints or diagnosing learned dynamics from saved weights: checkpoints-and-diagnostics.

Safe defaults

• Use tictactoe, connect4, gomoku, twentyone, and cartpole for dependency-light validation in normal CPU environments.
• Skip render() and human_to_action() in automated checks; many wrappers call input().
• Treat lunarlander, atari, breakout, gridworld, and spiel as optional integrations until their specific dependencies are installed.
• Keep action spaces as contiguous integer indexes (list(range(n))) unless changing MuZero internals, because policy logits and one-hot dynamics use actions as tensor indexes.

Custom game checklist

Before training a custom game:

• For a new editable game, first stage the bundled source with python scripts/stage_muzero_source.py --dest <workdir>/muzero-general-source; then create a module importable as games.<name> inside that staged copy and pass --repo-root <workdir>/muzero-general-source to validators.
• Define MuZeroConfig with observation_shape, action_space, players, network, self-play/training/replay fields, and visit_softmax_temperature_fn.
• Define Game with reset(), step(action), legal_actions(), render(), and optional to_play(), close(), human_to_action(), expert_agent(), action_to_string().
• Ensure reset() and step() return observations exactly shaped like config.observation_shape.
• Ensure legal_actions() is non-empty before search and a subset of config.action_space.
• Run validate_game_module.py and then the model/MCTS smoke before training. Omit --repo-root for bundled built-ins; pass --repo-root only for staged custom-game copies.