This paper addresses the scarcity of executable environments needed to train terminal-based coding agents by reconstructing usable environments directly from existing agent trajectories. Terminal-Universe replays the file operations recorded in a trajectory to restore each file and completes missing dependencies, turning recorded sessions into reusable workspaces. The framework synthesizes new tasks and extends interactions across multiple codebases and rounds of iterative user feedback, generating 37.3 thousand task-sufficient environments that improve model performance by 11.9-13.8 points on benchmark evaluations.
