Terminal-Universe mines 37,300 training rooms from old agent runs

Every coding agent run leaves two artifacts behind: the transcript of commands issued and files edited, and the workspace those commands ran against. The transcript gets filed. The workspace gets deleted. A preprint submitted to arXiv on 3 September 2026 argues that this is backwards.

The paper, Terminal-Universe: Turning Agent Trajectories into Scalable Terminal Environments, is about what post-training for terminal agents actually consumes. A trajectory is one frozen demonstration. An environment can be queried again and again, yielding a different verifiable task each time and returning execution feedback on every pass. Trajectories have piled up at scale; executable environments have not, even though the tool-execution history inside a trajectory describes the environment it ran in closely enough to rebuild. See also: the visibility gap around what coding agents actually do.

So Terminal-Universe does not generate environments from scratch. It mines the ones that were discarded. Other work has taken the same recycling instinct to agent context: Blast Radius buries dead context and recovers it later.

Why a workspace outlives a demonstration

A frozen demonstration supports imitation. It teaches an agent what someone did once, and nothing more. A recovered workspace supports interrogation: the same files, the same failing test, queried differently on every pass. That reusability is what the paper claims scales, and it is what a transcript cannot offer on its own.

The scarcity the preprint names isn't about disk space. Environments are expensive to construct and easy to lose: a sandbox exists for the length of one run, while the log outlives it by years. What Terminal-Universe offers is closer to archival recovery than to synthesis. See also: what a terminal sandbox actually blocks.

Replaying file operations to rebuild the sandbox

The mechanism runs in two stages. First, Terminal-Universe replays the file operations recorded in a trajectory, restoring each file to the state it held before the agent touched it. What comes back is a partial workspace: the files the agent modified, and not the rest of the tree. A completion agent then fills in the missing files and dependencies the workspace needs in order to run.

Task generation starts only after that. Working from the reconstructed environment, the framework recovers the task the original trajectory was attempting, then synthesizes new ones. That last step is what turns a single recording into a corpus.

Two axes: across codebases, and across turns

Reconstruction alone would leave an archive of static, single-shot problems. So the paper scales its synthetic tasks along two axes. Breadth mines directional dependency relations between related environments and builds cross-workspace queries spanning several codebases, which the authors present as a stand-in for ordinary development work. Depth stretches a single-turn query into a multi-round session, with a user agent doling out feedback and refining requirements as the session continues.

What the three headline figures do and do not show

Run against public terminal agent trajectories, the pipeline produced 37,300 task-sufficient environments. Supervised fine-tuning of Qwen3.5-27B on that corpus yielded two reported improvements.

Every figure is self-reported, on a model the authors fine-tuned, measured against benchmarks they picked. The abstract names no baseline and describes no control run. It also leaves "task-sufficient" undefined, so nothing separates an environment that can host a task from one that merely executes. Claims that outrun their validation are a familiar shape in agent research: nMAS reported feature-generation results tested only on dummy patient records.

The multi-round figure is harder to read. EvoCode-Bench v2 MT@4 gets named but not described, and multi-round evaluation introduces a simulated user whose feedback steers each turn. The abstract never says whether that user agent is held fixed, whether the same model plays it, or how much of the 13.8-point gain belongs to the model rather than to the conversation handed to it.

Provenance, licenses, and benchmark contamination

The abstract says the source trajectories were public. It doesn't say which collections were used, what their licenses permit, or whether a rebuilt workspace can surface files that were never meant to leave their original sandbox.

Contamination is the sharper question. Environments recovered from public trajectories come from the same population of real development sessions that public benchmarks are built from, and the preprint never addresses whether any of the 37,300 overlap with the workspaces behind Terminal-Bench 2.1 or EvoCode-Bench v2. A corpus assembled by replaying recorded runs is unusually exposed to that, because its inputs are real sessions rather than synthetic examples.

None of this undercuts the recycling argument, which is simple and checkable: if a workspace can be recovered from the log, discarding it was optional. Fidelity is the open question. Reconstruction depends on how much of the file system a trajectory happened to record, and wherever the log runs thin, a completion agent has to guess. The abstract reports the pipeline's yield, not how closely the rebuilt rooms resemble the ones the agents actually worked in. Sandbox results and real-environment results have a way of diverging: Qwen-UI-Agent scored higher on real phones than in sandboxes.