Key Info

Coding harnesses including Claude Code, Codex, Hermes, Pi, and opencode can now be turned into RL environments without modifying the harnesses or training code, using an open-source capture proxy.

Highlights

  • The same model scored 62% under Mini-SWE-Agent but only 33% under Claude Code, showing harness choice matters
  • Training inside a real harness normally requires reimplementing it as an environment, so models often train in scaffolds nobody ships
  • The fix is a capture proxy that the harness thinks is a model API, speaking four formats (OpenAI Chat Completions, OpenAI Responses, Anthropic Messages, and more)
  • Works with any open model, any task set, fully open source