SEGym allows you to simulate patches for Python repos in isolated environments. You can use such an environment to let a solver (e.g. LLM) search for a patch for a given issue until the issue is resolved.
This project is not yet available on PyPI. To try it out, clone the repository and open the .devcontainer/devcontainer.json in VSCode. This will automatically set up a development environment with all necessary dependencies (including docker, git, poetry and the required Python packages).
Supply your own openai.Client compatible API.
FOR LMU: use openai_lmu.get_lmu_openai_client() to get a ready-to-use client.
Drawing strong similarities to the OpenAI gym, the SEGym class is the main entry point for the library. It allows you to create a new environment, reset it, and step through it.
No LLM generated content will modify local files, instead env starts up a docker container for every patch generation, ensuring that the host system is not affected by any potential bugs in the generated code.
For example usage, see demo.ipynb.