sanjuhs/open-env-meta-final-hackathon ? reverse-engineered prompt

Build me a project called CADForge that helps train an AI to write real CadQuery CAD code instead of just pretty meshes.

I want the app to take a design request, generate a complete Python file, then actually run the CAD build and score it based on whether it works, exports cleanly, stays editable, and matches the request. If the code is broken, it should clearly show the errors and give negative reward, then let the model try again and improve from the feedback. Please include the environment, the training notebook, and the scripts needed to run SFT and GRPO style training, plus a simple way to inspect logs and reward results.

Make it easy to rerun locally and in a hosted space, and include the main docs so someone can understand the workflow without digging too much. If you need current package or API details, look them up online while you build it.