write-tests

name	write-tests
description	Write or review unit tests for code in this repo. Use when writing new tests, reviewing existing tests, or adding test coverage.
argument-hint	[module or file to test]

Writing Unit Tests

Follow these conventions when writing or modifying tests in this codebase.

Core Principles

• Only write tests that provide real information. Avoid testing trivial things like object construction or obvious attribute access. Test behavior and computations.
• Don't check types in tests. Types are verified by the static type checker (ty). Avoid using isinstance, type(), or other type assertions in tests.
• Write doctests to explain function usage. Doctests are a quick way to show how to use a function and expected output; they also are quick smoke tests.

File and Function Structure

Tests live in tests/ within each package. Name files test_<module>.py. Use plain functions — never classes:

"""Tests for the Spam."""

import numpy as np
from numpy.testing import assert_almost_equal as aae
from pytest import approx, mark, param, raises

from spam import Spam

def test_classic_spam() -> None:
    """Test classic spam values."""
    classic_spam = Spam("classic")
    assert classic_spam.weight == approx(3.14159265)

• Add a module-level docstring to every test file.
• Give each test function a Google-style one-line docstring.

Float Comparisons

Use default tolerance thresholds unless there is a specific, documented reason to override them.

Use case	Tool	Import
Scalars and simple lists	approx	from pytest import approx
NumPy arrays (gradients, hessians)	aae	from numpy.testing import assert_almost_equal as aae
Exact/near-exact array equality	assert_allclose	from numpy.testing import assert_allclose

# Scalars
assert Spam("classic").weight == approx(3.14159265)

# Lists / tuples
assert Spam("jalapeno").scores == approx([1, 2, 3])

# NumPy arrays
aae(Spam("maple").distribution, [[0, 1, 2], [3, 4, 5], [6, 7, 8]])

# Only override thresholds when truly necessary, with a comment explaining why
assert Spam("bacon").weight == approx(3.14, abs=0.005)  # loose: GitHub Actions gives different results

# Exceptions
with raises(ValueError):
    Spam("classic").process(threshold=-1)

Parametrize Similar Tests

Use pytest.mark.parametrize to collapse tests that differ only in inputs/outputs Use param(..., marks=[mark.regression]) to apply markers to individual parametrized cases

@mark.parametrize(
    ("spam_type", "weight"),
    [
        ("classic", 3.14159265),
        ("maple", 4.1),
        param("low sodium", 2.9, marks=[mark.regression]),
    ],
)
def test_weights(spam_type: str, weight: float) -> None:
    """Test energy for multiple molecules."""
    spam = Spam(spam_type)
    assert spam.weight == approx(weight)

Regression Tests

Mark tests that run less regularly but protect against regressions with @mark.regression. These are excluded from the default run (-m 'not regression').

When a regression test is related to a GitHub issue, link it in the docstring:

@mark.regression
def test_low_sodium_spam() -> None:
    """They forgot to remove the sodium originally.

    https://example.com/flying_circus/spam/issues/1
    """
    ...

Regression tests verify:

• Previously-broken behavior is fixed and stays fixed
• A wider range of code paths (e.g., all solvents, all supported methods)
• Edge cases that are expensive to test on every CI run

Fixtures

Use pytest fixtures for any setup that is repeated across multiple tests. Place fixtures in conftest.py:

• Within a package: tests/conftest.py
• Shared across packages: not currently applicable; each package has its own conftest.py

Imports

Prefer importing selectively from pytest for readability

from pytest import approx, mark, param, raises

Standard alias for numpy array assertions:

from numpy.testing import assert_almost_equal as aae

What NOT to Test

• Type correctness (use ty check instead)
• That a function returns something (trivial)
• Exact exception message text (fragile)
• Behavior of external libraries
• Implementation details that should be free to change

Focus on observable behavior: computed values, raised exceptions for invalid inputs, and correct transformations.