Terminal-Universe: Turning Agent Trajectories into Scalable Terminal Environments
Terminal-Universe claims existing agent logs can be converted into reusable coding environments for post-training.
The paper says the framework replays file operations from terminal-agent trajectories to recover partial workspaces, then uses a completion agent to fill missing files and dependencies. It reconstructs the original task and also generates new tasks, including cross-workspace queries and multi-round sessions with simulated user feedback. Applied to public trajectories, it produced 37.3k task-sufficient environments. Fine-tuning Qwen3.5-27B on the corpus improved reported benchmark scores on Terminal-Bench 2.1 and EvoCode-Bench v2. ArXiv · AI/CL/LG's note
The paper says the framework replays file operations from terminal-agent trajectories to recover partial workspaces, then uses a completion agent to fill missing files and dependencies. It reconstructs the original task and also generates new tasks, including cross-workspace queries and multi-round sessions with simulated user feedback. Applied to public trajectories, it produced 37.3k task-sufficient environments. Fine-tuning Qwen3.5-27B on the corpus improved reported benchmark scores on Terminal-Bench 2.1 and EvoCode-Bench v2. ArXiv · AI/CL/LG's note
score 6