skillxp harnesses

To see where each supported harness discovers project-level skills, use harnesses:

  $ skillxp harnesses
antigravity    project skills: .agents/skills   does NOT record injected context (evidence is inference)
claude-code    project skills: .claude/skills   records injected context
codex          project skills: .codex/skills    records injected context

The injected-context column matters for reading results: on a harness that records injected context, discovery evidence is direct; on antigravity, it is behavioral inference. See Harness Lore.

skillxp observe

To stage a skill, activate it, and trace how phrases reached the model, use observe:

  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 (an agentminutes session record)
  • the archived native transcript(s) that evidence line numbers point into

With -runs N, each repetition gets a fresh fixture and session under run-NN/, and summary.json reports how many runs delivered each traced phrase to each location. This is the rate report for model-dependent behavior: one run is an anecdote; N runs are a rate.

The trace report

The trace report classifies each occurrence of a traced phrase:

  • harness-injected: the platform put it in front of the model (harness-push loading)
  • tool-result: the model fetched it with a tool (model-pull loading)
  • model-output: the model emitted it
  • human-prompt / echo: your own prompt, or a system record replaying conversation content; never loading evidence

Choosing phrases to trace

Never put a phrase you plan to trace in the prompt: a literal-minded model truthfully answers “it’s in your message”, and the phrase contaminates every echo location in the transcript. Seed the phrases in the skill body and reference files instead, and let the trace report show whether they arrived.

Sandboxing

-sandbox runs the harness against an isolated home cloned per run, and unlocks user-scope installs (-install-user), which are refused unsandboxed. See Sandboxing for the isolation model and per-harness auth setup.