RSIGame

RSIGame is a research framework for improving existing generated games. Agents explore a running project, identify a specific problem, edit code or content, then replay the game to verify the change. A monitor tracks the best-scoring checkpoint instead of assuming the latest edit is best.

Start with the official installation guide. It requires Python 3.11+, Linux or macOS, a separately installed and patched GameCraft-Bench checkout, the OpenGame-RSIGame repair-agent fork with Node 20+, model credentials and replay tools. The Godot workflow specifies Godot 4.6.2. Configure paths and keys, run rsigame check, then develop a task from a base-game corpus.

This is not a one-click game maker. Model and image-generation calls can cost money; setup and evaluation require technical work. The published study uses a fixed-budget, post-run monitor by default, while live monitoring can stop runs early and changes budget comparability. GameSlash has not installed or reproduced the system.

The project code is Apache-2.0; separately installed dependencies and underlying game assets need their own license checks. The TableArtifacts dataset provides base games, scored builds and judge records. Intermediate rounds and recordings are not included there, and the repository still marks the supervised corpus and model weights as pending release.

RSIGame — source and installation