skill-eval — local eval harness for Agent Skills on headless Gemini CLI

[fede0089]

I open-sourced skill-eval, a local eval harness for Agent Skills running on headless Gemini CLI.

The problem it tries to solve is simple: skills are easy to demo, but hard to validate reliably.

A skill can look great in one run and still:

• fail to trigger consistently
• add too much context
• slow the agent down
• help some prompts and hurt others

skill-eval focuses on evidence:

• trigger evals
• functional evals against expectations
• baseline vs with-skill comparison
• repeated trials / pass@k
• isolated git worktrees per run
• HTML reports with uplift and token stats

Repo: [github.com/fede0089/skill-eval](https://github.com/fede0089/skill-eval)
Install: npm i -g skill-eval

Would especially love feedback on what would make this genuinely useful for skill authors