reeval-python

Python authoring SDK for reeval. Write *.eval.py files, declare evals with the same options as TypeScript, and run them through reeval's language runner (over its stdin and stdout, runner protocol 3). Use the reeval-python version that matches your reeval: each says which runner protocol it speaks, and refuses another.

Import as reeval_python (from reeval_python import evaluate). The PyPI distribution is reeval-python (the name reeval is already taken on PyPI by another project).

With reeval installed, default discovery includes evals/**/*.eval.py. When your project has pyproject.toml or uv.lock and uv is on PATH, reeval starts the runner with uv run --project <root> python -m reeval_python. Otherwise it uses <root>/.venv/bin/python when present, then python3 (python on Windows). Set REEVAL_PYTHON to override the interpreter. The standalone reeval binary works the same way and needs only uv besides; install the reeval-python release made for your reeval, since a runner on another protocol stops the run.

Install

uv add reeval-python
# or: pip install reeval-python

Then uv run reeval run. reeval-python's reeval command runs the reeval binary of its own version: the first run downloads it from the GitHub release, checks it against the sha256 this release recorded, and caches it in <cache>/reeval/bin/<version>/.

Monorepo development:

cd packages/reeval-python
uv sync --dev
uv run pytest
uv run ruff check .
uv run pyright

Declare trees are checked against packages/reeval/test/conformance/declare/ goldens. From the reeval package, test/python-sdk.test.ts runs the same suite in CI.

Writing evals

from reeval_python import SyncAttemptContext, evaluate, expect


@evaluate("lands", {"runs": 3, "timeout": 300})
def lands(c: SyncAttemptContext) -> None:
    c.send("Create greet.py that prints hi")
    assert c.exec("test -f greet.py").ok
    expect(c.judge("Does greet.py print hi?")).to_pass()

Plain assert is first-class, as in pytest: each one in an eval file is a check in the run record, named by its test as written (assert c.exec("test -f greet.py").ok is assert c.exec("test -f greet.py").ok). A pass is recorded when it runs. A failure ends the body, as in Python, and is recorded with its message and both sides of a comparison as the reason (greet.py exists: assert 1 == 2). An AssertionError the body catches isn't a failed check, so a retry loop can assert; one the escaping error carries (each failure in a TaskGroup's ExceptionGroup, or one an error was raised from or while handling) is. Use typed results (ExecResult.ok, Verdict.passed, Turn.text, …) and c.file(path) for quick checks. Keep expect for reeval matchers (to_have_file, to_succeed, soft checks, labels) — Playwright-style.

Matchers that take a pattern (to_have_file, to_have_run, to_have_searched, to_have_fetched, to_have_called_tool, to_have_edited, to_have_changed) take a str or a compiled re.Pattern, str or bytes, with any flags. Python matches it with pattern.search, so anything re accepts works (lookbehind, re.VERBOSE, named groups). The record shows the pattern's repr.

import re

expect(c.turns[0]).to_have_run(re.compile(r"^(npm|pnpm) t(est)?$"))

Use snake_case on the context (read_file, write_file, file) and on matchers (to_succeed, to_pass, to_equal, to_contain, …). Matcher failures raise ReevalAssertionError.

Options are typed (EvalOptions, and JudgeOptions for c.judge(…)), so pyright refuses a misspelt key such as runz. They take reeval's keys in snake_case at any depth (min_lift, delay_ms); the keys of env, headers and the other maps of your own keep their spelling. Every duration is seconds or a timedelta: {"timeout": 90}, docker(start_timeout=timedelta(minutes=2)), expect.poll(…, timeout=10). docker(), fake_harness(), fake_judge() and plugin.* return values to pass as options, not data to edit.

def and async def bodies

A def body gets a blocking context: c.exec(…) runs the command and returns its ExecResult, and c.read_file(…) returns the contents. Use expect from reeval_python, whose matchers check at once.

An async def body gets the async context: every c.* call is a coroutine, and expect comes from reeval_python.aio, whose matchers you await. A call or matcher the body never awaits never runs, so it fails the attempt, and so does a task the body started and left running when it returns: await each one (asyncio.gather, a TaskGroup). Each expect refuses to run in the other kind of body, and a body that yields is refused where it's declared.

Annotate c (SyncAttemptContext or AttemptContext) and pyright checks the body too: a missing await, the wrong context for the kind of body, a generator or a returned value. An unannotated body is checked only when it runs.

from reeval_python import AttemptContext, evaluate
from reeval_python.aio import expect


@evaluate("lands, async")
async def lands(c: AttemptContext) -> None:
    await c.send("Create greet.py that prints hi")
    await expect(await c.exec("python greet.py")).to_succeed()

c.exec(cmd, retries=2, interval=2) runs a command that exits non-zero again, up to 2 more times, 2 seconds apart (interval is seconds or a timedelta, one second by default), and returns the last try. Use it for a check that can fail on a transient, such as a test that calls a model: expect(c.exec("pytest -q tests/hidden", retries=2)).to_succeed(). Any non-zero exit retries, and the record and output say passed on try 2 of 3.

expect.poll retries a matcher until it passes, every interval until timeout (one second by default). Hand it a function that returns one c.read_file(…) or c.exec(…), and reeval reads or runs it again on each try; it works the same in both kinds of body. Check the output with the matcher, not in the function: lambda: c.exec("cat x").stdout is refused.

expect.poll(lambda: c.read_file("ready.txt"), timeout=10).to_be("ready\n")
expect.poll(c, "the server wrote its log").to_have_file("server.log")

Names of evaluate_each cases

evaluate_each(cases, name) names each eval from name: %# is the case's index, %s the case as str() writes it, and $field a dict case's field. Each placeholder is filled once, so a value that contains $ or % stays as it is, and a $field the case doesn't have is left as written. Two evals with one name in a file are refused where they're declared.

@evaluate_each([{"lang": "en"}, {"lang": "fr"}], "greets in $lang (#%#)")
def greets(c: SyncAttemptContext, case: dict[str, str]) -> None: ...

Examples

• Feature smoke: packages/reeval/test/features/python-sdk/hello/ (uv project with a project dependency)
• Plan case: packages/reeval/test/plan/python/
• Contributor guide: .agents/skills/add-a-language/SKILL.md in the reeval repo