Game Harness Lab

Game Harness Lab is a local web and CLI workbench for testing game agents and improving their harnesses. The public implementation includes Snake, Xiangqi, Military Chess, Gomoku, Tetris, an original platformer, Overcooked-AI and Pokemon battles through poke-env. StartLux Decision selects gameplay actions with thinking disabled. GLM-5.3 prepares instructions and reflects on completed games, while GEPA proposes and evaluates changes. Each game uses a frozen harness; code enforces game rules, action validity and scoring. Candidates must pass separate paired checks before becoming approved exports. Experimental Tetris workflows can also vary goal guidance, limited memory and observation organization. Here, RSI describes a repeated feedback-and-evaluation loop for harness configurations. It does not train model weights or guarantee improvement. Published results are limited to their recorded games, seeds and budgets: the latest three-way Tetris study retained the same parent policy in all groups and found no improved version. Requires self-hosting, a compatible Decision service and a coach API key. Model inference costs and hardware are separate. The main project's license was not established; dependency licenses do not establish a license for the whole project. This listing was checked against public code and reports; no new games or model tests were run. Source and setup: https://github.com/dctongsheng/game-harness-lab Latest experiment: https://github.com/dctongsheng/game-harness-lab/blob/main/docs/tetris-components-experiment-v1.md