skillxp

Skill invocation runner: install an Agent Skill in a fresh fixture, invoke an agent harness headlessly, and report what actually reached the model, with transcript evidence.

Skill authors publish for 25+ platforms that each load, present, and manage skills differently, and mostly invisibly. skillxp makes that behavior observable: it stages a skill on a real harness and answers "what did the platform actually do with it" from the session transcript, not from the model's self-reporting. It builds on agentsummons (headless invocation) and agentminutes (transcript parsing), and adds the third layer of lore: how each harness discovers, activates, and records skills.

skillxp renders no verdicts. It produces observations; graders consume them. The first consumer is the agent-skill-implementation loading benchmark, whose checks and verdict logic live in that repo's benchmark-runner/.

Status

Early development. Supported harnesses (validated versions): Antigravity CLI 1.1.4, Claude Code 2.1.205, Codex CLI 0.144.6. Results reflect headless behavior, which may differ from interactive use.

Install

brew install agent-ecosystem/tap/skillxp
# or
npm install -g skillxp
# or
pip install skillxp
# or
go install github.com/agent-ecosystem/skillxp/cmd/skillxp@latest

As a Go library: go get github.com/agent-ecosystem/skillxp. The harnesses you observe must be installed and authenticated. On Windows, use pip or a release binary; the npm package temporarily has no Windows support (a registry naming issue is being worked out with npm).

Quick start

# Where does each harness discover project-level skills?
skillxp harnesses

# Stage a skill, activate it, and trace how two phrases
# reached the model
skillxp observe -harness claude-code -install ./my-skill \
  -prompt "Activate the my-skill skill and follow its instructions." \
  -activation -trace "PHRASE-IN-BODY-1234,PHRASE-IN-REFERENCE-5678" \
  -out out/

observe writes a bundle: observation.json (run metadata and the trace report), session.json (the normalized transcript), and the archived native transcript(s) that evidence line numbers point into. One rule about phrases: never put a phrase you plan to trace in the prompt, or it contaminates every echo location in the transcript.

As a library:

obs, err := observe.Observe(ctx, observe.Config{ArchiveDir: dir},
    agentsummons.ClaudeCode, observe.Spec{
        SkillDirs:  []string{"./my-skill"},
        Prompt:     "Activate the my-skill skill and follow its instructions.",
        Activation: true,
    })
if err != nil {
    return err
}
occs := trace.Phrase(obs.Session, "PHRASE-IN-BODY-1234", obs.Profile.EchoSubtypes)

Documentation

Full documentation is available at skillxp.dev:

• Quickstart: install skillxp and observe your first skill invocation
• Use Cases: cross-platform skill CI, benchmarks, regression watching, and skill iteration
• CLI: the harnesses and observe commands, the observation bundle, and the trace report's classifications
• Go Library: Observe, multi-turn sessions, repeat runs and rates, and the three packages
• Sandboxing: isolated homes per run, user-scope installs, and per-harness auth setup
• Harness Lore: the empirically established per-harness behavior the profiles encode

License

MIT.