Documentation
README
Test Agent Skills
Measure behavior without inheriting current chat, global memory, or unrelated Skills.
Ownership
- Test owns live activation, outcome, routing, and baseline evidence.
- Audit owns static structure and trigger-overlap review.
- Create owns Skill changes; Test never rewrites a failing Skill automatically.
Layers
| Layer | Installed Skills | Measures |
|---|---|---|
| Target-only | Target Skill only | Activation, outcome, baseline uplift |
| Pack | Candidate sibling Skills | Routing and false activation |
Use target-only for core outcomes. Use pack for adjacent and unrelated routing cases.
Cases
- Core positive: target should activate and produce required artifact or answer.
- Paraphrase: same intent, different wording.
- Adjacent: sibling should own request.
- Unrelated: no package Skill should activate.
- Assertions: inspect files, output, commands, or explicit events.
Default one trial. Add repetitions only for known nondeterminism or regression confidence.
Isolation
- Fresh session, temporary home, temporary workspace.
- Install only layer-required Skills.
- Minimal authentication/configuration; no user memory or chat history.
- Guard host-home reads and writes.
- Adapter owns harness-specific install and event parsing.
- Save raw events, final output, workspace, version, and Skill hash locally.
Verdicts
pass: activation and outcome assertions pass.fail: deterministic assertion fails.inconclusive: runner or evidence cannot support claim.- Baseline also passes: no demonstrated deterministic uplift.
- Exact isolated
SKILL.mdread is medium-confidence activation; explicit activation event is high confidence.
Fake runner tests validate adapter plumbing, not provider behavior. Never claim unrun results.
uv run --script <skill-dir>/scripts/run -- test skills/<name> --agent <agent> --cases cases/<name>.yaml --baseline
uv run --script <skill-dir>/scripts/run -- test skills/<name> --agent <agent> --cases cases/<name>.yaml --pack