Fixtures

tests/conftest.py is deliberately small. Two things, both about building a project to test against.

write_project(...)

A helper, not a fixture — call it directly when you need a project with specific contents.

from tests.conftest import write_project

def test_something(tmp_path):
    project = write_project(tmp_path, ...)

It writes a config.yaml and a tests.yaml into the directory you give it and returns the path. Because it takes tmp_path, each test gets an isolated project and nothing leaks between them.

Reach for it when the test is about a config shape — an unusual weights block, a constraint that should disqualify, a scorer option.

simple_project fixture

@pytest.fixture()
def simple_project(tmp_path: Path) -> Path: ...

A ready-made minimal project. Request it as an argument when the test needs a project rather than a particular one:

def test_history_is_empty_before_any_run(simple_project):
    assert list_runs(simple_project.parent, simple_project.name) == []

Choosing between them

Building on the example projects

For behaviour that should hold against a real config, load one of the committed examples rather than inventing a new one:

config = load_config("projects/support_triage")

They are offline, deterministic, and already exercised by CI. This is what the back-compat tests do — the point is precisely that these are configs real people copied.

Never mutate them from a test. Copy into tmp_path first:

import shutil

def test_migration_upgrades_a_real_database(tmp_path):
    src = Path("projects/local_demo/results/arena.sqlite")
    db = tmp_path / "arena.sqlite"
    shutil.copy(src, db)
    ...

That database is a committed artefact from a real run, and it is the only realistic migration fixture in the repo.

Fixtures deliberately absent

tests/test_local_connector.py stands up a real local HTTP server for the cases that genuinely need one. Reuse that approach rather than adding a global fixture — it keeps the cost where it is needed.