FACET: Preserving Source Intent and Executable State in Terminal Task Synthesis

Summary

Training terminal agents requires scalable executable supervision, yet synthesizing high-quality terminal tasks is challenging. Each task couples an instruction, an initialized environment, a reference solution, and an executable verifier; inconsistent assumptions can make tasks unsolvable or mis-evaluated. Multi-stage synthesis may discard original goals, dependencies, state transitions, and procedural constraints. We present FACET (Fine-grained Agentic Construction of Executable Tasks), a framework addressing information preservation and cross-artifact consistency. FACET reconstructs related agent skills into coherent scenarios, then realizes and repairs the execution environment before generating final artifacts. The container state serves as shared grounding for instruction, solution, and verifier, with execution-based validation and targeted repair fixing errors without regenerating valid components. FACET produces complex tasks with dense executable checks, and successful trajectories provide effective supervision. Fine-tuning models at multiple scales improves Terminal-Bench 2.1 performance, and analyses support environment-grounded construction for task validity and solution-verifier alignment.