In April 2026, Kosmo Che worked on stabilizing the evaluation pipeline for the xlang-ai/OSWorld repository, focusing on backend development with Python. Addressing a persistent issue with dynamic module caching, Kosmo implemented a solution that snapshots sys.modules before executing new modules and cleans up any additions afterward, ensuring clean task boundaries and preventing stale imports from affecting subsequent evaluations. This approach improved the reliability of regression testing and protected against flaky scores. Kosmo’s work involved close cross-team collaboration, multi-author commits, and a deep understanding of Python internals, demonstrating thoughtful engineering depth in maintaining robust evaluation infrastructure.