Merge pull request #3159 from jgaribay01/fix/prepush-venv-pytest

fix(hooks): pre-push skipped every Python project that uses a virtualenv
This commit is contained in:
Affaan Mustafa
2026-09-18 21:10:20 -04:00
committed by GitHub
2 changed files with 400 additions and 4 deletions
+160 -4
View File
@@ -117,16 +117,172 @@ if [[ -f "go.mod" ]] && command -v go >/dev/null 2>&1; then
go test ./... || fail "go test failed"
fi
# Resolve how this project runs pytest, into PYTEST_CMD as an argv array.
#
# Looking only for `pytest` on PATH meant the hook skipped every project that keeps
# its tools in a virtualenv -- which is most of them -- and reported "pytest is not
# installed" while sitting next to a .venv with pytest in it. A gate that silently
# declines to gate is worse than no gate, because the skip line reads like a pass.
#
# An array rather than one string, because a virtualenv path may contain spaces:
# a scalar command splits `/home/me/my env/bin/python` into two paths that do not
# exist, and the hook then rejects the push for a reason that has nothing to do
# with the code being pushed.
#
# Echoes the command it will run, so the reason for a skip is always visible.
PYTEST_CMD=()
# Does this command actually run pytest? Accepting `--version` is not evidence --
# plenty of programs take it and exit 0 -- so the output has to name pytest. The
# version is captured rather than piped: under `set -o pipefail` a `| grep -q` can
# report the SIGPIPE of the program it just matched.
#
# Only ever called on a command this script composed itself. Probing an arbitrary
# operator-supplied command is not safe: a wrapper that ignores `--version` and
# execs pytest runs the entire suite during the probe, and is then rejected for
# not having printed a version.
is_pytest() {
local version
version="$("$@" --version 2>&1)" || return 1
grep -qiE 'pytest[[:space:]]+(version[[:space:]]+)?[0-9]' <<<"$version"
}
# Does the repository itself ship this interpreter?
#
# A virtualenv is never committed -- it is platform-specific binaries, and every
# Python project gitignores it. One that IS tracked is the repository handing this
# hook an executable and asking it to run. The hook is installed globally, so
# cloning a hostile repository and pushing it to your own fork would be enough,
# and on a machine with no pytest on PATH this arm is the only thing that would
# run at all. A developer's own venv is untracked, so nothing legitimate is lost.
#
# The path is resolved through symlinks before git is asked, because `git ls-files`
# reports paths as indexed and does not follow links. A repository that commits
# `.venv` as a symlink to `.` next to a tracked `bin/python` would otherwise be
# queried for `.venv/bin/python`, a path git has never heard of, and the answer
# would be "untracked". Measured: that shape ran the planted binary twice.
repo_ships_interpreter() {
local bindir real top
bindir="$(cd -P -- "$1" 2>/dev/null && pwd -P)" || return 1
[[ -n "$bindir" ]] || return 1
real="$bindir/python"
top="$(git rev-parse --show-toplevel 2>/dev/null)" || return 1
top="$(cd -P -- "$top" 2>/dev/null && pwd -P)" || return 1
[[ -n "$top" && "$real" == "$top/"* ]] || return 1
# `:(icase)` because git matches index pathspecs case-sensitively even where
# core.ignorecase is set, while the filesystem underneath does not. On macOS's
# APFS -- the platform this hook most often runs on -- a committed
# `.venv/bin/Python` is what `$venv/bin/python` opens and executes, but a
# case-sensitive query for the lowercase name finds nothing in the index and the
# guard waves it through. Measured: that spelling ran the planted binary twice.
git ls-files --error-unmatch -- ":(icase)${real#"$top"/}" >/dev/null 2>&1
}
# `-I` isolates the probe: without it Python puts the working directory first on
# sys.path, so a repository that commits a `pytest.py` in its root gets that file
# imported -- and executed -- by a check whose only job is to answer whether pytest
# exists. Measured: a committed pytest.py ran during the probe. Isolation does not
# hide a real pytest, which lives in the interpreter's own site-packages.
resolve_pytest() {
# `${VAR+set}` rather than `-n "${VAR:-}"`, so that a variable set to nothing is
# still an override: `ECC_PYTEST_CMD=` and `ECC_PYTEST_CMD=" "` now behave
# alike, where the first used to fall through to discovery and the second failed
# the push. Falling through is the wrong half of that pair -- an override that
# evaluated empty (a command substitution that found nothing, say) would silently
# run a different runner than the operator asked for, which is the substitution
# this resolver refuses to make anywhere else.
#
# Not `[[ -v ECC_PYTEST_CMD ]]`: that is bash 4.2, and a stock macOS `/bin/bash`
# is 3.2, where it is a syntax error rather than a false. This hook ships to
# whatever `env bash` finds.
if [[ -n "${ECC_PYTEST_CMD+set}" ]]; then
# Taken as given. This is a deliberate override, and the hook cannot inspect it
# without running it -- a wrapper script may ignore `--version` and run the
# suite, so probing costs a duplicate test run and then blocks the push anyway.
# Pointing this at something that is not pytest turns the gate off, and that is
# the operator's call to make, not a misconfiguration for the hook to second
# guess. Word-split, so the command names something on PATH or an interpreter
# whose path has no spaces; a venv with spaces is found by the loop below.
read -r -a PYTEST_CMD <<<"$ECC_PYTEST_CMD" || true
[[ ${#PYTEST_CMD[@]} -gt 0 ]] || fail "ECC_PYTEST_CMD is set but names no command.\
Point it at your test runner, or unset it to fall back to discovery."
return 0
fi
local venv
for venv in "${VIRTUAL_ENV:-}" .venv venv env; do
if [[ -n "$venv" && -x "$venv/bin/python" ]]; then
if repo_ships_interpreter "$venv/bin"; then
log "Ignoring $venv/bin/python: the repository ships it."
log " A committed virtualenv is an executable the repository controls, and"
log " this hook runs on every push in every repository."
continue
fi
if "$venv/bin/python" -I -c "import pytest" >/dev/null 2>&1; then
PYTEST_CMD=("$venv/bin/python" -m pytest)
return 0
fi
fi
done
if [[ -f "uv.lock" ]] && command -v uv >/dev/null 2>&1; then
if uv run --no-sync python -I -c "import pytest" >/dev/null 2>&1; then
PYTEST_CMD=(uv run --no-sync pytest)
return 0
fi
fi
if [[ -f "poetry.lock" ]] && command -v poetry >/dev/null 2>&1; then
if poetry run python -I -c "import pytest" >/dev/null 2>&1; then
PYTEST_CMD=(poetry run pytest)
return 0
fi
fi
# `command -v` proves only that a file of that name exists on PATH. This one the
# script composed itself, so confirming it costs a harmless `pytest --version`.
if command -v pytest >/dev/null 2>&1 && is_pytest pytest; then
PYTEST_CMD=(pytest)
return 0
fi
PYTEST_CMD=()
return 1
}
if [[ -f "pyproject.toml" || -f "requirements.txt" ]]; then
if command -v pytest >/dev/null 2>&1; then
if resolve_pytest; then
ran_any_check=1
log "Python project detected. Running: pytest -q"
pytest -q || fail "pytest failed"
log "Python project detected. Running: ${PYTEST_CMD[*]} -q"
if [[ -n "${ECC_PYTEST_CMD+set}" ]]; then
# resolve_pytest deliberately does not verify the override is pytest, because
# probing it can run the operator's suite. What this gate can honestly do
# about a stale override is refuse to be quiet about it: a bypass announced
# on every push is not the silent gate this resolver exists to prevent.
log " via ECC_PYTEST_CMD -- the hook runs what you pointed it at, and does"
log " not check that it is pytest. Unset it to gate on the real suite."
fi
pytest_status=0
"${PYTEST_CMD[@]}" -q || pytest_status=$?
case "$pytest_status" in
0) ;;
# pytest reserves 5 for NO_TESTS_COLLECTED, which is not a red suite. A
# pyproject.toml that only configures ruff or black is still a Python project
# by this hook's test, and blocking those pushes would make the gate something
# people switch off. Never silent, though: a bad rootdir, testpaths or a
# conftest that fails to import also collects nothing, and swallowing that is
# the same skip-reads-like-a-pass hole this resolver exists to close.
5)
log "pytest collected no tests (exit 5). Not gating this push."
log " If this repository is supposed to have tests, that is the bug:"
log " check rootdir, testpaths, and conftest.py import errors."
;;
# The code is in the message because 1 (tests failed) and 4 (usage error)
# need different responses, and "pytest failed" alone cannot tell them apart.
*) fail "pytest failed (exit $pytest_status)" ;;
esac
else
log "Python project detected but pytest is not installed. Skipping."
log "Python project detected but no pytest found (checked \$VIRTUAL_ENV, .venv,"
log " venv, env, uv, poetry, PATH). Set ECC_PYTEST_CMD to point at it."
fi
fi
if [[ "$ran_any_check" -eq 0 ]]; then
log "No supported checks found in this repository. Skipping."
else
+240
View File
@@ -306,6 +306,246 @@ if (
passed++;
else failed++;
function writeExecutable(filePath, body) {
fs.mkdirSync(path.dirname(filePath), { recursive: true });
fs.writeFileSync(filePath, body);
fs.chmodSync(filePath, 0o755);
}
// The Python arm of the hook, exercised without a real interpreter: the stubs
// record the argv they were handed, which is what the virtualenv-path regression
// is actually about.
function runHermeticPythonPrePush({
venvName = null,
venvExit = 0,
trackVenv = false,
trackedVenvBasename = 'python',
trackedSymlinkVenv = false,
pytestCmd = null,
overrideStub = false,
pathPytestVersionLine = null,
} = {}) {
const tempDir = createTempDir('codex-pre-push-py-');
const projectDir = path.join(tempDir, 'project');
const callsPath = path.join(tempDir, 'calls.txt');
fs.mkdirSync(projectDir);
fs.writeFileSync(path.join(projectDir, 'pyproject.toml'), '[project]\nname = "demo"\n');
const initialized = spawnSync('git', ['init', '--quiet'], { cwd: projectDir });
assert.strictEqual(initialized.status, 0, initialized.stderr?.toString());
// Every stub records the argv it was handed. That record is the assertion: it is
// how a test tells a preserved path from a split one, and a command that was run
// once from one the hook probed first.
const record = `printf '%s\\n' "$0|$*" >> "${toBashPath(callsPath)}"`;
// A tracked venv has to live inside the repository to be trackable at all, and is
// found by directory-name discovery rather than by VIRTUAL_ENV.
const venvDir = venvName === null ? null : path.join(trackVenv ? projectDir : tempDir, venvName);
const venvPython = venvDir === null
? null
: path.join(venvDir, 'bin', trackVenv ? trackedVenvBasename : 'python');
if (venvPython !== null) {
writeExecutable(venvPython, `#!/bin/sh\n${record}\ncase " $* " in *" -c "*) exit 0 ;; esac\nexit ${venvExit}\n`);
if (trackVenv) {
// Staged, not committed: `git ls-files` reads the index, so this is enough to
// make the file repository-controlled without needing a committer identity.
const added = spawnSync('git', ['add', '-f', '--', venvPython], { cwd: projectDir });
assert.strictEqual(added.status, 0, added.stderr?.toString());
}
}
// The shape that defeats a naive `git ls-files -- .venv/bin/python` check: the
// repository commits `.venv` as a symlink to its own root plus a tracked
// `bin/python`, so git is asked about a path it has never indexed.
if (trackedSymlinkVenv) {
writeExecutable(path.join(projectDir, 'bin', 'python'), `#!/bin/sh\n${record}\nexit 0\n`);
fs.symlinkSync('.', path.join(projectDir, '.venv'));
const added = spawnSync('git', ['add', '-f', '--', 'bin/python', '.venv'], { cwd: projectDir });
assert.strictEqual(added.status, 0, added.stderr?.toString());
}
// Deliberately does NOT special-case --version: an operator's wrapper would not
// either, and the recorded calls are what prove the hook never probed it.
const overrideStubPath = overrideStub ? path.join(tempDir, 'bin', 'wrapper') : null;
if (overrideStubPath !== null) {
writeExecutable(overrideStubPath, `#!/bin/sh\n${record}\nexit 0\n`);
}
const pathBin = pathPytestVersionLine === null ? null : path.join(tempDir, 'pathbin');
if (pathBin !== null) {
writeExecutable(
path.join(pathBin, 'pytest'),
`#!/bin/sh\nif [ "$1" = "--version" ]; then printf '%s\\n' '${pathPytestVersionLine}'; exit 0; fi\n${record}\nexit 0\n`,
);
}
const override = overrideStubPath === null ? pytestCmd : toBashPath(overrideStubPath);
// Built from nothing rather than from process.env. The hook reads VIRTUAL_ENV and
// ECC_PYTEST_CMD from the ambient environment, so a developer running this suite
// inside an activated virtualenv, or with ECC_PYTEST_CMD exported, would resolve a
// pytest the fixture never created. Omitted, not blanked: now that a variable set
// to nothing is itself an override, blanking it here would make every one of these
// tests take that branch.
const env = {
PATH: pathBin === null
? process.env.PATH
: `${toBashPath(pathBin)}${path.delimiter}${process.env.PATH}`,
HOME: process.env.HOME ?? '',
ECC_SKIP_GIT_HOOKS: '0',
ECC_SKIP_PREPUSH: '0',
MSYS_NO_PATHCONV: '1',
...(venvDir === null || trackVenv ? {} : { VIRTUAL_ENV: toBashPath(venvDir) }),
...(override === null ? {} : { ECC_PYTEST_CMD: override }),
};
const result = runBash(prePushHook, {
env,
cwd: projectDir,
preservePath: false,
input: Buffer.from('refs/heads/main 1111111111111111111111111111111111111111 refs/heads/main 0000000000000000000000000000000000000000\n'),
});
const calls = fs.existsSync(callsPath)
? fs.readFileSync(callsPath, 'utf8').trim().split(/\r?\n/).filter(Boolean)
: [];
cleanup(tempDir);
return { result, calls, venvPython, overrideStubPath };
}
if (
test('pre-push runs pytest from a virtualenv whose path contains spaces', () => {
const { result, calls, venvPython } = runHermeticPythonPrePush({ venvName: 'my venv' });
const python = toBashPath(venvPython);
assert.strictEqual(result.status, 0, `${result.stdout}\n${result.stderr}`);
assert.deepStrictEqual(calls, [
`${python}|-I -c import pytest`,
`${python}|-m pytest -q`,
], JSON.stringify({ calls, python, stdout: result.stdout, stderr: result.stderr }, null, 2));
})
)
passed++;
else failed++;
if (
test('pre-push refuses to run a virtualenv python that the repository tracks', () => {
const { result, calls } = runHermeticPythonPrePush({ venvName: '.venv', trackVenv: true });
assert.strictEqual(result.status, 0, `${result.stdout}\n${result.stderr}`);
assert.deepStrictEqual(calls, [], JSON.stringify(calls));
assert.match(result.stdout, /the repository ships it/);
})
)
passed++;
else failed++;
// A case-folded spelling, because macOS resolves `$venv/bin/python` to a committed
// `Python` while git matches index pathspecs case-sensitively. Skipped where the
// filesystem is case-sensitive and the two names cannot collide.
if (fs.existsSync(__filename.toUpperCase()) || fs.existsSync(__filename.toLowerCase())) {
if (
test('pre-push refuses a tracked interpreter committed under a folded case', () => {
const { result, calls } = runHermeticPythonPrePush({
venvName: '.venv',
trackVenv: true,
trackedVenvBasename: 'Python',
});
assert.strictEqual(result.status, 0, `${result.stdout}\n${result.stderr}`);
assert.deepStrictEqual(calls, [], JSON.stringify(calls));
assert.match(result.stdout, /the repository ships it/);
})
)
passed++;
else failed++;
}
if (
test('pre-push refuses a tracked interpreter reached through a committed symlink', () => {
const { result, calls } = runHermeticPythonPrePush({ trackedSymlinkVenv: true });
assert.strictEqual(result.status, 0, `${result.stdout}\n${result.stderr}`);
assert.deepStrictEqual(calls, [], JSON.stringify(calls));
assert.match(result.stdout, /the repository ships it/);
})
)
passed++;
else failed++;
if (
test('pre-push blocks the push when the resolved pytest fails', () => {
const { result } = runHermeticPythonPrePush({ venvName: 'venv-red', venvExit: 1 });
assert.notStrictEqual(result.status, 0, `${result.stdout}\n${result.stderr}`);
assert.match(result.stderr, /pytest failed \(exit 1\)/);
assert.doesNotMatch(result.stdout, /Verification checks passed/);
})
)
passed++;
else failed++;
if (
test('pre-push does not block when pytest collected no tests (exit 5)', () => {
const { result } = runHermeticPythonPrePush({ venvName: 'venv-empty', venvExit: 5 });
assert.strictEqual(result.status, 0, `${result.stdout}\n${result.stderr}`);
assert.match(result.stdout, /collected no tests \(exit 5\)/);
assert.match(result.stdout, /rootdir, testpaths, and conftest\.py/);
})
)
passed++;
else failed++;
if (
test('pre-push runs an ECC_PYTEST_CMD override exactly once, without probing it', () => {
const { result, calls, overrideStubPath } = runHermeticPythonPrePush({ overrideStub: true });
assert.strictEqual(result.status, 0, `${result.stdout}\n${result.stderr}`);
assert.deepStrictEqual(calls, [`${toBashPath(overrideStubPath)}|-q`], JSON.stringify(calls));
// The override is not verified to be pytest, so it must at least be loud.
assert.match(result.stdout, /via ECC_PYTEST_CMD/);
assert.match(result.stdout, /does\n?.*not check that it is pytest/s);
})
)
passed++;
else failed++;
// Both blank forms, because they used to disagree: an unquoted empty value fell
// through to discovery while whitespace failed the push. A venv is present so a
// fall-through would be visible as a pass rather than as an absence.
for (const [label, blank] of [['empty', ''], ['whitespace', ' ']]) {
if (
test(`pre-push fails closed when ECC_PYTEST_CMD is set to ${label}`, () => {
const { result, calls } = runHermeticPythonPrePush({
venvName: 'venv-blank',
pytestCmd: blank,
});
assert.notStrictEqual(result.status, 0, `${result.stdout}\n${result.stderr}`);
assert.match(result.stderr, /ECC_PYTEST_CMD is set but names no command/);
assert.deepStrictEqual(calls, [], JSON.stringify(calls));
})
)
passed++;
else failed++;
}
if (
test('pre-push rejects a PATH pytest that does not identify itself as pytest', () => {
const { result, calls } = runHermeticPythonPrePush({
pathPytestVersionLine: 'true (GNU coreutils) 9.0',
});
assert.strictEqual(result.status, 0, `${result.stdout}\n${result.stderr}`);
assert.match(result.stdout, /no pytest found/);
assert.deepStrictEqual(calls, []);
})
)
passed++;
else failed++;
if (
test('pre-push accepts a PATH pytest that reports a pytest version', () => {
const { result, calls } = runHermeticPythonPrePush({ pathPytestVersionLine: 'pytest 8.0.0' });
assert.strictEqual(result.status, 0, `${result.stdout}\n${result.stderr}`);
assert.strictEqual(calls.length, 1, JSON.stringify(calls));
assert.match(calls[0], /\|-q$/);
})
)
passed++;
else failed++;
if (
test('check-plugin-cache fails when the installed cache is missing manifest-referenced files', () => {
const homeDir = createTempDir('codex-plugin-cache-home-');