skill-evallisted

Run libi's agent-driven skill-eval scenarios. Use after editing a bundled skill, MCP wiring, or agent instructions to verify the inner libi agent still behaves correctly (e.g. picks gpt-image-2, keeps native audio). Heavy + token-costly — run manually, only the scenarios a change warrants.
Nagellabs/libi · ★ 2 · AI & Automation · score 72

Install: claude install-skill Nagellabs/libi

# Skill-Eval (behavioral regression tests for libi skills) Libi's primary user is the **inner agent**. This harness boots a hermetic `LIBI_TEST_MODE=1` libi, runs the inner agent against one `.md` scenario with production-exact skill/MCP wiring, checks deterministic trace invariants against fake-fal's recorded calls, and leaves the transcript for YOU to judge behavior. ## When to use After editing any of: a bundled skill (`mcp/skills/<name>/SKILL.md`), MCP tool surface/schemas, agent instructions (`mcp/templates/instructions.md`, `mcp/instructions.ts`), or `lib/mcp-config.ts` wiring. If you changed one of these and did NOT run the relevant scenario, the behavioral change is unverified. ## The loop 1. **Discover.** Open `skill-eval/INDEX.md`. Map your change to scenarios via the `covers` / `skills` columns (e.g. you edited `ai-asset-generation` → run every scenario whose `covers` includes a model id or `native-audio`). Heavy runs ⇒ confirm the chosen set with the developer before running; never run the whole library blindly. 2. **Run** each chosen scenario: `npm run skill:eval -- skill-eval/scenarios/<group>/<file>.md` (add `--agent claude-code` to override; `--keep` to retain the temp LIBI_HOME). 3. **Read the verdict.** The CLI prints per-run `HARD-PASS` / `FAIL` / `TIMEOUT` and a `JSON_SUMMARY` line. Hard invariants are mechanical — already decided. 4. **Judge behavior YOURSELF.** Open `<reportDir>/transcript.md` and check each `## Behavioral ex