Hugging Face publishes a guide to training open models inside Claude Code, Codex and other coding harnesses
Hugging Face's post-training team released an open framework and guide for reinforcement-learning a model inside unmodified coding agents, built on OpenEnv, the Harbor task framework and the TRL trainer.
Logos: Hugging Face (public domain) and PyTorch (BSD) via Wikimedia Commons
Hugging Face's post-training team has published "The ultimate guide to multi-harness RL," describing how to train a small open model with reinforcement learning inside the coding harnesses people actually use, such as Claude Code, Codex and OpenCode.
The guide says the harnesses run exactly as they ship. A capture proxy between the harness and the model records the tokens and probabilities needed for training, Harbor supplies containerized tasks and sandboxes, and TRL handles training, all connected through OpenEnv, a project from Meta's PyTorch team and Hugging Face.
Lewis Tunstall of the post-training team announced the guide on October 3, saying it offers "a recipe" for getting the best performance from any open model in a custom harness.