Toward Scalable Terminal Task Synthesis via Skill Graphs
Zhiyuan Fan, Tinghao Yu, Yuanjun Cai, Jiangtao Guan, Yun Yang, Dingxin Hu, Jiang Zhou, Xing Wu, Zhuo Han, Feng Zhang, Lilin Wang
2026-04-29
Summary
This paper introduces SkillSynth, a new system for automatically creating tasks for training computer programs that can use the command line, like a virtual assistant in a terminal.
What's the problem?
Training these 'terminal agents' is hard because you need a lot of examples of them successfully completing tasks, and getting those examples is time-consuming and expensive. Existing methods try to create lots of tasks, but they don't necessarily ensure the agent learns a *variety* of useful ways to solve problems, leading to agents that aren't very flexible or capable.
What's the solution?
SkillSynth solves this by building a 'skill graph' that maps out different command-line actions and how they connect to each other through common 'scenarios'. Think of it like a flowchart of possible actions. The system then randomly picks paths through this graph, representing real-world workflows, and turns them into actual tasks the agent can practice. This way, the tasks are designed to force the agent to learn a diverse set of skills.
Why it matters?
This is important because it makes it easier and more effective to train these terminal agents. The tasks created by SkillSynth were actually used to improve another agent called Hy3 Preview, showing that this approach really works and can lead to more powerful AI assistants that can handle complex tasks in a command-line environment.
Abstract
Terminal agents have demonstrated strong potential for autonomous command-line execution, yet their training remains constrained by the scarcity of high-quality and diverse execution trajectories. Existing approaches mitigate this bottleneck by synthesizing large-scale terminal task instances for trajectory sampling. However, they primarily focus on scaling the number of tasks while providing limited control over the diversity of execution trajectories that agents actually experience during training. In this paper, we present SkillSynth, an automated framework for terminal task synthesis built on a scenario-mediated skill graph. SkillSynth first constructs a large-scale skill graph, where scenarios serve as intermediate transition nodes that connect diverse command-line skills. It then samples paths from this graph as abstractions of real-world workflows, and uses a multi-agent harness to instantiate them into executable task instances. By grounding task synthesis in graph-sampled workflow paths, SkillSynth explicitly controls the diversity of minimal execution trajectories required to solve the synthesized tasks. Experiments on Terminal-Bench demonstrate the effectiveness of SkillSynth. Moreover, task instances synthesized by SkillSynth have been adopted to train Hy3 Preview, contributing to its enhanced agentic capabilities in terminal-based settings.