skilltest-sdk

The Python SDK for the skilltest CLI. A thin, typed wrapper and nothing else: it runs the CLI as a subprocess and parses the stable --format json contract into Pydantic models. Test-framework integrations build on it — use skilltest-pytest if you want pytest collection; use this package directly from any other Python code.

Define the whole case — skill, input, evals, an optional simulated user, mocks — in code and pass it straight to run_skill. Everything a case YAML carries has a typed builder, so the case and its checks live in one place:

from skilltest_sdk import TestCase, run_skill, boolean, numeric, describe_failures, assistant_text

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)
assert report.passed, describe_failures(report)
# Mix in deterministic checks on the transcript:
assert "Dr. Smith" in assistant_text(report.runs[0].transcript)

Add user(persona=..., done_when=...) for a multi-turn case, and mocks=[...] of stub/spy/deny/rewrite (a called / not_called eval takes the mock object itself, or a name= you gave it). Validate a skill definition with validate_skill("skills/greeter").

Or point at existing YAML cases

run_skill also takes a path to a test-case YAML file — or a directory of them — so a suite that keeps cases as data works unchanged, and the pytest plugin auto-discovers *.skilltest.yaml files with no code at all:

report = run_skill("cases/greet.yaml")   # or run_skill("cases/") for the tree

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

Tool events

Each assistant turn carries the normalized tool events the skill took (shell commands, file edits, tool uses), lifted from oneharness's --events. Assert on what the skill did with tool_calls (the tool_call events across a transcript, in order); each ToolEvent has kind, name, input, output, index:

from skilltest_sdk import TestCase, run_skill, tool_calls, boolean

case = TestCase(
    skill="skills/editor",
    input="Update the config and commit it.",
    evals=[boolean("the change was committed")],
)
report = run_skill(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)   # never destructive

Streaming (opt-in)

stream_skill takes the same case (or path) and returns a SkillStream you iterate with async for to receive each event live, and break to short-circuit — closing the stream tears the harness down, so a bad turn is cut off instead of paid for in full. .report holds the final report once the stream runs to completion:

from skilltest_sdk import stream_skill

async def guard(case):
    stream = stream_skill(case)
    async for ev in stream:  # ev: StreamEvent — .case/.platform/.model/.turn/.event
        if ev.event.name == "bash" and "rm -rf" in str(ev.event.input):
            break
    return stream.report

The skilltest binary is resolved from the bin= argument, the SKILLTEST_BIN env var, or PATH; a provider override comes from provider= or SKILLTEST_PROVIDER. A failing eval is reported (report.passed is false), not raised; bad input raises SkilltestUsageError (CLI exit 2) and provider problems raise SkilltestProviderError (exit 3).

The Pydantic models are generated from the golden schemas in schemas/ — themselves generated from the CLI's own types — via just gen-contract, and a drift gate in CI fails if anything is stale, so the models cannot diverge from the binary. That covers both directions: the report models the SDK parses (_report.py) and the case models the builders construct (_case.py).