skillxp
Skill invocation runner. Stage an Agent Skill on a real harness, invoke it headlessly, and report what actually reached the model, with transcript evidence.
Open Source MIT Licensed.
Why skillxp?
Skill authors publish for 25+ platforms that each load, present, and manage skills differently, and mostly invisibly. skillxp makes that behavior observable, from the session transcript rather than the model’s self-reporting.
Transcript evidence
Every claim about what a platform did with your skill traces to a session transcript, with line numbers. The model’s opinion of its own context is never the source.
Rates, never anecdotes
Loading behavior is model-dependent, so one run proves little. Repeat runs the same spec N times in fresh fixtures and reports how often each traced phrase reached each location.
Provenance classification
The trace report classifies each occurrence of a phrase, distinguishing harness-injected content from model-pulled tool results, model output, and echoes that prove nothing.
Sandboxed fixtures
Runs execute against an isolated home cloned per run, so your installed skills can’t leak into the experiment and the experiment can’t pollute your user scope.
Built on the stack
agentsummons invokes the harness, agentminutes parses the transcript, and skillxp adds the third layer of lore, covering how each harness discovers, activates, and records skills.
Observations, never verdicts
skillxp produces observations; graders consume them. The Agent Skill Implementation loading benchmark is the first consumer, with its checks and verdict logic in its own repo.