skilltest-pytest

A pytest plugin for skilltest: run AI-skill tests and natural-language evals as ordinary pytest tests, and mix in your own deterministic checks. Built on skilltest-sdk — the SDK's code API is re-exported here, so a pytest suite needs only this one dependency.

Define the whole case in code (recommended)

Build the case — skill, input, evals, an optional simulated user, mocks — right in the test. Everything the YAML carries has a typed builder, so the case, its mocks, and any deterministic transcript checks live in one place:

from skilltest_pytest import TestCase, run_skill, boolean, numeric, describe_failures

def test_greeter():
    case = TestCase(
        skill="skills/greeter",           # resolved relative to the working dir
        input="Greet Dr. Smith, who has an appointment today.",
        evals=[
            boolean("the reply greets Dr. Smith by name"),
            numeric("how warm is the tone", min=0, max=10, threshold=7),
        ],
    )
    report = run_skill(case, platforms=["claude-code"], models=["claude-opus-4-8"])
    assert report.passed, describe_failures(report)

Multi-turn cases add user(...); deterministic call-count checks use called / not_called referencing a named stub/spy (or the mock objects' own assertions — see below). run_skill also takes platforms=/models= to fan a case across a matrix.

Or point at a YAML file

run_skill accepts a path just as well (run_skill("cases/greet.yaml")), and auto-collection still works: name a case something.skilltest.yaml and pytest runs it with no test function at all —

# greet.skilltest.yaml
skill: ./skills/greeter
input: "Greet Dr. Smith."
evals:
  - type: boolean
    criterion: "the reply greets Dr. Smith by name"

The full field reference for both forms is docs/schema.md.

Assert on tool use, and stream

The SDK's tool-event and streaming surfaces are re-exported too. tool_calls returns the normalized tool_call events a run took (each a ToolEvent with kind/name/input/output/index), and stream_skill yields them live so a test can short-circuit on bad behavior:

from skilltest_pytest import TestCase, run_skill, tool_calls, boolean

EDIT_CASE = TestCase(
    skill="skills/editor",
    input="Update the config and commit it.",
    evals=[boolean("the change was committed")],
)

def test_commits_but_never_deletes():
    report = run_skill(EDIT_CASE)
    calls = tool_calls(report.runs[0].transcript)
    assert any("git commit" in str(c.input) for c in calls)
    assert not any("rm -rf" in str(c.input) for c in calls)

import asyncio
from skilltest_pytest import stream_skill

def test_makes_no_network_call():
    async def go():
        async for ev in stream_skill(EDIT_CASE):
            assert ev.event.name != "curl", "skill made a network call"
    asyncio.run(go())   # or use pytest-asyncio and `async def test_...`

Configuration

The plugin shells out to the skilltest binary. Point it at one with the SKILLTEST_BIN env var (or bin=), the provider with SKILLTEST_PROVIDER (or provider=), and set defaults in pyproject.toml:

[tool.pytest.ini_options]
skilltest_provider = "oneharness"
skilltest_platforms = ["claude-code"]
skilltest_models = ["claude-opus-4-8"]

See the repository root for the provider protocol and the full schema.