Open-source proxy turns coding-agent harnesses into reinforcement-learning environments

The project’s post says a capture proxy lets existing coding harnesses serve as reinforcement-learning environments without changes to the harnesses or training code. It reports 31% fewer tool calls on tasks the model already solved after adding a reward for using fewer calls, with reductions in every harness and about half as many calls under Codex. The post says the proxy, trainer, tasks, training code and seven trained models are open source.