table_peak

Training AI agents to play tabletop games.

This project is vibe-researched and vibe-coded, mostly using Claude Code and the superpowers plugin. It draws on patterns from open RL / game-AI work; I've tried to follow license rules for anything reused — please open an issue if you spot an attribution gap.

Project goals:

• explore strategies for the games included
• surface balance suggestions for game designers
• refine my agentic coding setup for greenfield projects

Roadmap

In progress: {insert link to beads tasks here?}

Suggest a game — game designers and creators, please open an issue with the rules and player count; I'd love to add more games.

Quickstart

Requires Python ≥3.12 and pdm.

make sync                  # pdm install — sync deps from pdm.lock into .venv
make test                  # run the test suite
.venv/bin/uvicorn table_peak.web.app:app --reload

Run make help for the full list of dev targets.

Open http://localhost:8000/ to play TicTacToe against a trained agent. The training entry point is table_peak.training.loop.train.

Development

Tooling: pdm, ruff, mypy --strict, pytest, pre-commit — config in pyproject.toml and .pre-commit-config.yaml. Work follows the beads-superpowers workflow (brainstorm → spec → plan → implement) under docs/superpowers/, with task tracking in beads (bd).

License

MIT — see LICENSE. The Game / State shape is conceptually modelled on open_spiel (noted in src/table_peak/games/base.py). Open an issue for any attribution gap.