# Harness Selftest

> Fixture skill. Exists only so the eval harness has an artifact to bind its committed replay fixtures to, and so CI grades a real prompt set instead of only asserting that an empty fixture set fails. Triggered by the four cases in evals/prompts.json and by nothing else. NOT FOR any real task: it does nothing, and no agent should ever load it outside tests/skill_evals/.

- Skill: `broomva/harness-selftest` (Agent Skill)
- Install (CLI): `npx skillmds@latest add broomva/harness-selftest`
- Raw SKILL.md: https://api.skillmd.com/api/skills/broomva/harness-selftest/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: AI & ML
- Author: broomva (https://skillmd.com/u/broomva)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/broomva/harness-selftest

---


# harness-selftest

This file is the artifact under test for `tests/skill_evals/fixtures/harness-selftest`.

Every committed fixture in `../cases/` carries the SHA-256 of **this file** and of
the `description:` above. Editing either one invalidates all of them — which is the
point: it is the mutation CI performs on every PR to prove the gate can go RED.

If you edited this file on purpose, regenerate the fixtures:

```bash
python3 tests/skill_evals/fixtures/harness-selftest/generate.py
```