CodeMidas is an agentic pipeline that converts the implemented functionality inside open-source codebases into executable reinforcement-learning environments, using only source code as input rather than issues or commit history. Agents explore each codebase to derive behavioral specifications, construct execution-grounded tests, and filter candidate tasks via execution checks and repeated rollouts, yielding 5,545 training tasks from 3,185 repositories across 23 languages. Training the MiMo-V2.5 model on these tasks with GRPO improves performance on issue repair, whole-program construction, and terminal-use benchmarks, and ablations show that scaling up high-quality tasks continues to improve results.
