skill-lens

An Agent Skill is a prompt. Prompts regress. skill-lens turns "I think this SKILL.md got better" into a score, a report, and an exit code your pipeline can gate on.

📖 Full documentation: https://emadmokhtar.github.io/skill-evaluator/

Why

You edit a SKILL.md, read the new answer once, and it looks better. Two weeks later a teammate edits the same file, and nobody can say whether the agent still looks an order up before refunding it — or whether it now refunds orders it should refuse.

skill-lens gives that question a real answer. You write eval cases next to your skill. It runs them, scores what came back, and reports one verdict for the whole run.

Your skills stay yours. Skills and their eval cases are inputs to the tool — nothing about a skill under test is vendored here, so any skill repository can adopt skill-lens without embedding it.

What it measures

• What the agent said — substring, regex and exact-match assertions on the output.
• What the agent did — which mock tools it called, in what order, and which ones it must never touch. A refund granted without a lookup is invisible to an output assertion.
• What it cost — per-case ceilings on tokens, dollars and latency.
• How well it said it — a rubric-based LLM judge (large language model grading the output) returns one verdict per rubric line, with the evidence for each. "Explains it plainly" is not a substring.
• Whether the agent reached for the skill at all — mode: offered registers the skill as a tool instead of force-loading it, so triggering becomes an observable choice. Ship the negative control and a skill that fires on everything stops scoring 100%.
• Whether your edit actually helped — run every case twice, once with the skill and once against a baseline (no skill, or its previous version resolved from git), and gate on the delta.

Try it — free, offline, no API key

uv tool install "skill-lens[pydantic-ai]"

(pip install "skill-lens[pydantic-ai]" works too; skill-lens[langchain] gives you the LangChain runner instead, or as well. Drop the extra for the offline default runner alone.)

That puts skill-lens on your PATH, so the commands below run as written.

The walkthrough uses this repository's own example skills, so clone it too — or point skill-lens at your own directory of SKILL.md files instead:

git clone https://github.com/EmadMokhtar/skill-evaluator.git
cd skill-evaluator

Working on skill-lens itself rather than using it? Run uv sync in that checkout for the development environment, and prefix the commands below with uv run — you can then skip the uv tool install above.

A skill is any directory containing SKILL.md. Its eval cases live beside it — this repository ships four:

examples/
  csv-report/
    SKILL.md
    csv-report.eval.yaml
  greeting/
    SKILL.md
    greeting.eval.yaml
  log-triage/
    SKILL.md
    log-triage.eval.yaml
    references/report-format.md
    scripts/count_levels.py
  order-support/
    SKILL.md
    order-support.eval.yaml

Point the CLI at one skill directory or at a parent of many — discovery is recursive:

skill-lens list ./examples

csv-report	1 case(s)	examples/csv-report
greeting	1 case(s)	examples/greeting
log-triage	1 case(s)	examples/log-triage
order-support	6 case(s)	examples/order-support

list discovers skills and validates every eval file without calling a runner: no API key, no spend. Starting on your own skill? skill-lens init ./skills/my-skill writes a starter suite with the placeholders marked, so you fill in the blanks instead of starting from one; point it at a directory of skills and it scaffolds every one that has no suite.

A case is a few lines of YAML

cases:
  - name: refuses a refund outside the return window
    task: I want a refund for order 1234
    tags: [smoke, refund]
    tools:
      - name: lookup_order
        description: Look up an order by its id
        parameters:
          order_id: string
        returns: '{"id": "1234", "status": "delivered", "days_since_delivery": 45}'
      - name: issue_refund
        description: Issue a refund for an order
        parameters:
          order_id: string
        returns: '{"ok": true}'
    trajectory:
      called: [lookup_order]      # it must look the order up
      forbidden: [issue_refund]   # and must not refund this one
    budget:
      max_tokens: 2000
    assertions:
      - kind: contains
        value: "1234"             # name the order you are talking about

A tool a real MCP server exposes need not be transcribed: skill-lens mcp-import tools.json writes the block from the server's own tools/list listing, schema and all. A tool several skills share is declared once, in a tool library any eval file imports with tool_libraries: and names with - ref: <name>.

A run reads like a test suite

Your own repository follows the same layout — one directory per skill, eval cases beside the SKILL.md:

skills/
  order-support/
    SKILL.md
    order-support.eval.yaml

skill-lens run ./skills

[PASS] order-support :: names the order it is talking about (fake)
[FAIL] order-support :: refuses a refund outside the return window (fake)
        assertion: failed: contains('return window'): did not hold
            contains[0]: contains('return window') did not hold
        output: [fake] order-support handled: I want a refund for order 1234
[PASS] order-support :: never leaks a stack trace to the customer (fake)

2 passed, 1 failed, 0 errored — pass rate 67%

Gate FAILED:
  - pass rate 67% is below the required 100%

Every failing case shows what the agent actually said and which tools it called — --case reruns just that one.

Exit code 0 means the gate passed, 1 means it failed, and 2 means something in your own files is wrong. That is the whole contract with your pipeline.

The default runner is scripted and offline, so the pipeline above costs nothing to try. To score a real agent, pass --runner pydantic-ai — the [pydantic-ai] extra in the install above is what supplies it (from a checkout: uv sync --extra pydantic-ai) — or pass --runner copilot / --runner claude-code to run the skill through an installed GitHub Copilot CLI or Claude Code, with no API key. See Runners.

Gate your pull requests

- uses: EmadMokhtar/skill-evaluator@v0.16.0
  with:
    path: ./skills
    runner: pydantic-ai
    model: openai:gpt-4o-mini
  env:
    OPENAI_API_KEY: ${{ secrets.OPENAI_API_KEY }}

Pin an exact tag. Until 1.0 a minor release may change behaviour, so there is no floating v0 tag to follow.

The run publishes a JUnit XML report for your provider's test pane, a Markdown summary for the job summary or a pull-request comment, and a JSON report for anything else. skill-lens never calls the GitHub API itself — it renders files, your workflow decides where they go. Copy-pasteable workflows live in examples/ci/ and in CI integration.

Once that is green, baseline: previous and repeat: 3 turn the same job into a comparative one: each case runs with and without your edit, several times, and the report carries the delta — see Comparative evals.

Running against an unreleased commit? The action's default install-spec pins the released version matching its own tag, so point both at the same commit and they cannot drift apart:

- uses: EmadMokhtar/skill-evaluator@<commit-sha>
  with:
    path: ./skills
    install-spec: "skill-lens[pydantic-ai] @ git+https://github.com/EmadMokhtar/skill-evaluator@<commit-sha>"

Why a green run means something

Eval tools are easy to fool — mostly by accident, and usually by yourself. These are deliberate design decisions, each with a test holding it in place:

• No vacuous passes. A typo like assertion: is rejected rather than silently producing a case that checks nothing. An unfilled TODO(skill-lens) scaffold stops the run. A judge check that passes without citing evidence is recorded as a failure. A run that executed zero cases fails the gate — "nothing ran" is a broken run, not a pass.
• errored is not failed. A provider returning 500 is an infrastructure signal, not evidence that your skill got worse. The two are counted, reported and gated separately.
• Nothing spends money behind your back. The default runner is offline and scripted; the LLM judge is off until you turn it on; a run that will cost money prints its plan first.
• Authoring mistakes stop the run. A malformed regex or an unknown assertion kind is a bug in your files, not a verdict on your skill — exit 2, naming the file and the field.

The full list, with the reasoning behind each, is in ARCHITECTURE.md.

Documentation

Contributing

Contributions are welcome, and the project is set up so that helping is cheap:

uv sync
uv run pytest        # the whole suite: offline, no API key, no spend
uv run ruff check .

Every test passes with no network access. Tests that would hit a real provider are opt-in (-m integration) or replay recorded traffic, so you can work on any part of this without an API key or a bill.

Three conventions to know before your first pull request — all three are explained in Contributing:

1. Test-driven. Write the failing test first.
2. Conventional Commits for commit messages and pull-request titles. Releases are derived from history, and pull requests are squash-merged, so the title becomes the commit.
3. Documentation ships with the change, not as a follow-up. Continuous integration checks it.

Good places to start: a new example skill with its eval suite, an adapter for another agent framework, or anything on the roadmap. Not sure whether an idea fits? Open an issue and ask — that is a perfectly good first contribution.

Status

Milestone 7. Discovery, scoring, judging, comparison, real-file workspaces, reporting, gating and the automated release pipeline all ship and are tested. Versions are derived from the commit history and published to PyPI on merge. This is 0.x: a minor release may still change behaviour, so pin what you depend on. See the roadmap for what is shipped and what is planned.

License

MIT — see LICENSE.