---
name: tool-use-data-synthesis
description: Synthesize training data for ANY tool-use / agentic benchmark, in ANY repo. Given a handful of runnable reference tasks, produce thousands of "same-distribution, different-entity" tasks, roll out real trajectories, and emit SFT, RL data. Use this when: you have a tool-calling / MCP / agentic benchmark and want to expand a few dozen tasks into thousands of in-distribution training samples without changing the environment or tools; when synthetic data mysteriously LOWERS the benchmark score and you need to diagnose why; or when you must produce checkable ground-truth for agentic tasks and cannot trust an LLM to fabricate facts. The first decision is set by the benchmark's SCORING MECHANISM (rule-based vs LLM-judged), which picks one of two synthesis paradigms. The overriding principle throughout is DISTRIBUTION ALIGNMENT.