Terminal-Universe Turns Agent Trajectories into Reusable Terminal Environments

Why trajectories are not enough

Terminal-based coding agents are generating an increasing number of interaction trajectories. These records show how an agent used tools, edited files, and responded to a request, but a trajectory is usually a frozen example. It cannot by itself provide a fresh query, a new execution outcome, or another round of user feedback. Post-training, by contrast, benefits from environments that can be queried repeatedly and judged through execution.

Terminal-Universe, introduced by the Qwen team, addresses this gap by treating trajectories as evidence about the environments in which they were produced. Rather than creating every project and dependency from scratch, the framework reconstructs a usable workspace from the recorded tool history and then explores that workspace to produce additional training tasks.

How the framework works

• Replay file operations. The system follows file creations and modifications in a trajectory to recover the state of files before the agent changed them. This produces a partial workspace rather than a guaranteed complete project.
• Complete missing context. Because trajectories may omit untouched files, dependencies, or configuration details, a completion agent supplies the missing components needed to make the recovered workspace more executable.
• Recover and diversify tasks. The reconstructed environment can be used to restate the original intent, but it can also support new queries that exercise different parts of the same codebase.
• Scale breadth. Terminal-Universe mines directional dependency relations among related environments and creates cross-workspace tasks. Such tasks reflect development situations in which an engineer must coordinate multiple codebases rather than modify one isolated repository.
• Scale depth. A single request can be expanded into a multi-round session. A user agent introduces feedback and changing requirements, allowing the coding agent to revise its implementation over time.

Reported results and implications

Applied to public terminal-agent trajectories, Terminal-Universe generated 37.3k task-sufficient environments. Supervised fine-tuning of Qwen3.5-27B on this corpus reportedly improved single-round performance on Terminal-Bench 2.1 by 11.9 points and multi-round performance on EvoCode-Bench v2 MT@4 by 13.8 points. The supplied material does not include detailed ablations or evaluation settings, so these numbers should be read as the paper’s aggregate claims rather than a guarantee across every deployment scenario.

The broader contribution is a change in how trajectory data can be valued. A trajectory is not only a demonstration of an answer; its tool calls and file changes can serve as a blueprint for reconstructing an environment. One recorded interaction may therefore become a source of multiple verifiable tasks, improving the return on existing data.

This approach also highlights two capabilities that are often separated in coding-agent benchmarks. Breadth concerns coordination across repositories and dependencies. Depth concerns sustained interaction, clarification, and requirement refinement. Real software work needs both. At the same time, the method inherits risks from reconstruction and synthesis: an automatically completed dependency may not match the original project, and generated tasks may drift from the original intent or duplicate one another. Reliable validation and quality control will therefore be as important as scaling the number of recovered environments.

来源：Hugging Face Daily Papers