Test selection

attest select lists the tests a change needs to run, so a pull request that touches one fixture does not run the whole suite. It reads the git diff and decides from file names alone: no model, no test execution, the same answer for the same diff.

uv run attest select --base origin/main

It diffs the merge base of <ref> and HEAD against the working tree, so committed changes, uncommitted changes and untracked files are all in. It prints one test path per line (sorted), the word all, or nothing when no test needs to run. Exit 2 means git or the ref failed; the error is on stderr.

--json prints the reasoning:

{"mode":"subset","tests":["tests/test_refund.py"],"reasons":[{"path":"tests/attest/fixtures/inc-1042.attest.json","class":"fixture"}]}

mode is subset or all (tests is empty for all). Each reason is a changed path and its class. --repo <dir> selects the directory whose tests to list; paths in the output are relative to it. The diff covers the whole git repository, so a lockfile or source change outside <dir> still selects all, while tests, fixtures and docs outside it are ignored.

Rules

Changed path	Class	Selection
pytest test_*.py, *_test.py; Vitest *.test.ts, *.test.tsx, *.test.js, *.test.mjs	test	that file, if it still exists
conftest.py	test	every test file below its directory
*.attest.json	fixture	the test files that reference the fixture (below)
*.md	ignored	nothing
uv.lock, poetry.lock, requirements*.txt, package-lock.json, pnpm-lock.yaml, yarn.lock, bun.lockb, bun.lock, pyproject.toml, package.json	all	all
attest.policy.yaml and other attest.*.yaml/yml/json/toml, world.yaml, *.world.yaml	all	all
anything else	unknown	all

A change to the Attest version is a change to a manifest or lockfile, so it selects all. Source files are unknown on purpose: the diff cannot say which tests exercise src/agent.py, so a source change runs everything. Selection only narrows a change that touches nothing but tests, fixtures and docs.

Which tests reference a fixture

A test references a fixture when its source holds the fixture’s capture name (the name field inside the fixture file) as a quoted string, as in attest_replay("inc-1042") or useReplay("inc-1042"), or a string that names the file (fixtures/inc-1042.attest.json). The match is on text, so it also finds names held in a constant defined in the same file, but not names built at runtime from pieces.

A fixture that a conftest.py references selects every test file below that conftest.py. A fixture that a helper module (any other .py, .ts, .tsx, .js or .mjs file that is not a test) references gives all, because the diff cannot say which tests import the helper. A fixture no test references gives all too: selection cannot tell what depends on it. So does a fixture that was deleted or that has no readable name.

Run the selection

tests=$(uv run attest select --base origin/main)
if [ "$tests" = "all" ]; then
  uv run pytest
elif [ -n "$tests" ]; then
  uv run pytest $tests
else
  echo "no tests selected"
fi

For Vitest, run npx vitest run $tests the same way. In GitHub Actions, fetch the base: actions/checkout with fetch-depth: 0, so the merge base exists.

attest init --ci github does not use selection; the generated workflow runs the full suite. Add the step above where a suite is slow enough to need it, and keep a full run on main so a selection error cannot hide a failure for long.