Terminal-Universe replays agent traces into 37.3k coding envs

Rebuilding a code agent's shell instead of scripting one from scratch is the pitch of Terminal-Universe, a framework from a Qwen-affiliated team that reconstructs training environments by replaying the file operations logged in existing agent trajectories. Applied to public terminal-agent traces, the method produces 37.3k "task-sufficient" environments, according to the paper on arXiv.

The construction is mechanical. The system "replays the file operations recorded in a trajectory to restore each file before the agent modified it, yielding a partial workspace," and a completion agent then supplies the missing files and dependencies. On top of the recovered workspace, the authors both reconstruct the original intent task and synthesize new ones, scaling in two directions: breadth, by mining "directional dependency relations between related environments" for cross-codebase queries, and depth, by extending single-turn queries into multi-round sessions driven by a user agent.

The reported numbers come from supervised fine-tuning of Qwen3.5-27B on the generated corpus: an 11.9-point gain on Terminal-Bench 2.1 in the single-round setting and 13.8 points on EvoCode-Bench v2 MT@4 in multi-round.

The abstract does not name a comparison baseline beyond the un-tuned model, nor does it publish per-task or per-codebase breakdowns.