Most training for terminal-based coding agents lacks realistic, re-executable environments: trajectories are frozen demonstrations without the underlying workspace needed to re-run or verify solutions. Terminal-Universe flips that problem by recovering the latent workspace from a trajectory's tool-execution history, completing missing context with a completion agent, and turning each recovered environment into many testable tasks and multi-round sessions.
Key Findings
- Reconstruct-by-replay: Replays chronological read/write/edit operations to restore the earliest-observed file states, producing a partial workspace that reveals the environment structure so many missing pieces can be inferred rather than invented—this preserves real-world grounding.
- LLM-based completion + verification: Uses a completion agent to fill missing files/dependencies and an agentic verifier to author pytest suites; only tasks whose solutions pass these automated tests are kept, yielding high-quality SFT data.
- Scale and diversity: Applied to public terminal trajectories, the pipeline produced 37.3k task-sufficient environments and 31.9k verified SFT demonstrations, and fine-tuning Qwen3.5-27B on this corpus increased single-round Terminal-Bench 2.1 by 11.9 points and multi-round EvoCode-Bench v2 MT@4 by 13.8 points.
- Breadth & depth expansions: Mines directional cross-workspace dependencies to synthesize multi-repo queries (breadth) and converts single-turn traces into multi-round sessions that simulate iterative user feedback (depth), better matching developer workflows.
Who it's for and tradeoffs
Great fit if you need scalable, execution-grounded training data for code/terminal agents and want verifiable SFT examples derived from real agent behavior. It helps teams that train or evaluate coding tutors, autonomous dev agents, or retrieval-augmented code solvers. Look elsewhere if your trajectories lack sufficient file-access traces (reconstruction fails), if strict provenance/IP is required (completion agents may introduce synthesized content), or when you need fully human-curated environments rather than LLM-completed ones. The approach trades fully manual environment curation for scalable, LLM-assisted reconstruction that can hallucinate absent details and depends on the quality of recorded traces.