Behavioral test harness for Claude Code skills: run a skill in an isolated fixture (throwaway workdir + HOME) via headless Claude Code and assert on what it actually does — response text, transcript tool calls, and filesystem effects. Supports trap scenarios (a seeded defect the skill must catch), so a skill that only passes happy-path prompts isn't considered tested.
Highlights
- Nine assertion types (output / file / tool /
no_writes_outside/max_turns_used) with trap support skillprobe loginstores a token locally — no environment variables to manage; env vars still work for CI- A conversational wrapper skill: drive the whole flow from inside Claude Code
- Credentials are redacted from reports; each run is isolated (not a security sandbox — see the README)
- 112-test suite; CI on Linux/macOS/Windows × Node 20/22
See the README for the quick start and the isolation model.