Files
hermes-agent/tests/test_packaging_metadata.py
Siddharth Balyan c349eca823 fix(packaging): ship locales/ i18n catalogs in wheel, sdist, and Nix (#38383)
* fix(packaging): ship locales/ i18n catalogs in wheel, sdist, and Nix

locales/ is a bare data dir (no __init__.py), invisible to packages.find
and package-data. Sealed installs (pip wheel, Nix store venv) dropped it,
so gateway/CLI commands rendered raw i18n keys like
gateway.reset.header_default.

- pyproject: [tool.setuptools.data-files] locales = ["locales/*.yaml"] (wheel)
- MANIFEST.in: graft locales (sdist)
- agent/i18n._locales_dir: env override -> source -> sysconfig data scheme
- nix/hermes-agent.nix: copy locales into the store + set HERMES_BUNDLED_LOCALES
  as defense-in-depth. The wheel's data-files already materialize into the
  uv2nix venv, so resolution works with no env var; the override pins the
  store path against a future uv2nix change that could drop data-files.
- tests: metadata regression, wheel + sdist build-install smoke tests, and a
  bundled-locales flake check that verifies BOTH the wrapper override and the
  env-var-less data-files path. Smoke test wired into CI.

Closes #23943, #27632, #35374.
Supersedes #23966, #27716, #30261, #33841, #35429, #35494, #35735, #36697.

* test: cap locale e2e timeout, tighten catalog count guard

The two wheel/sdist e2e tests inherit the global --timeout=30 from
addopts; a cold-CI run (isolated build env + venv create + network pip
install) can plausibly exceed it. Add @pytest.mark.timeout(300) so they
don't ride the unit-test budget and flake intermittently.

Also assert the shipped catalog count equals len(SUPPORTED_LANGUAGES)
instead of a hardcoded >=16 floor, so the guard self-updates and trips
on a single dropped catalog (not just a fully-empty graft).
2026-06-03 12:00:27 -07:00

229 lines
10 KiB
Python

from pathlib import Path
import tomllib
import pytest
# setuptools is declared in the [dev] extra and is the build backend, but
# guard the import so a runner without it skips these packaging checks
# instead of erroring out collection for the whole shard (it used to be
# picked up ambiently from the CI image; newer ubuntu-latest images don't
# ship it in the test venv).
find_packages = pytest.importorskip("setuptools", exc_type=ImportError).find_packages
REPO_ROOT = Path(__file__).resolve().parents[1]
def _packages_find_include():
data = tomllib.loads((REPO_ROOT / "pyproject.toml").read_text(encoding="utf-8"))
return data["tool"]["setuptools"]["packages"]["find"]["include"]
def test_every_on_disk_subpackage_is_covered_by_packages_find():
"""Regression test for #34701 (and the bug class behind #34034 / #28149).
``[tool.setuptools.packages.find]`` ``include`` is hand-maintained. Every
top-level package is listed twice — bare (``hermes_cli``) for the package
itself and ``hermes_cli.*`` for its subpackages — EXCEPT when someone
forgets the wildcard. v0.15.x listed ``hermes_cli`` without ``hermes_cli.*``,
so the wheel shipped ``hermes_cli/*.py`` but dropped the ``dashboard_auth``
and ``proxy`` subpackages. The dashboard then died on every install with
``ModuleNotFoundError: No module named 'hermes_cli.dashboard_auth'``.
This drives setuptools' own discovery against the live tree: every package
that exists on disk and would be found by a permissive ``<name>.*`` scan
must also be found by the actual ``include`` list. A subpackage added under
any listed package without the matching wildcard fails here instead of in a
user's container.
"""
include = _packages_find_include()
# What the real include list actually selects.
selected = set(find_packages(where=str(REPO_ROOT), include=include))
# Top-level packages we ship (bare names in the include list, no wildcard).
top_level = sorted({name for name in include if "." not in name})
# For each shipped top-level package, every on-disk subpackage must be
# covered by the include list.
expected = set(
find_packages(
where=str(REPO_ROOT),
include=[pattern for name in top_level for pattern in (name, f"{name}.*")],
)
)
missing = sorted(expected - selected)
assert not missing, (
"These packages exist on disk but are dropped from the wheel because "
"[tool.setuptools.packages.find] include is missing a wildcard. Add the "
f"matching '<name>.*' entry in pyproject.toml: {missing}"
)
def test_faster_whisper_is_not_a_base_dependency():
data = tomllib.loads((REPO_ROOT / "pyproject.toml").read_text(encoding="utf-8"))
deps = data["project"]["dependencies"]
assert not any(dep.startswith("faster-whisper") for dep in deps)
voice_extra = data["project"]["optional-dependencies"]["voice"]
assert any(dep.startswith("faster-whisper") for dep in voice_extra)
def test_manifest_includes_bundled_skills():
manifest = (REPO_ROOT / "MANIFEST.in").read_text(encoding="utf-8")
assert "graft skills" in manifest
assert "graft optional-skills" in manifest
def test_bundled_plugin_manifests_ship_in_both_wheel_and_sdist():
"""Regression test for #34034 / #28149.
Plugin discovery (hermes_cli/plugins.py) registers each bundled plugin by
reading its ``plugin.yaml`` / ``plugin.yml`` manifest. Those manifests are
data files, not Python modules, so they only reach installed packages when
declared explicitly:
- wheel -> ``[tool.setuptools.package-data]`` ``plugins`` glob
- sdist -> ``MANIFEST.in`` (Homebrew and other downstream packagers build
from the sdist)
v0.15.0 declared neither, so the wheel shipped every adapter's Python code
but none of its manifests, and *every* gateway platform failed with
"No adapter available for <platform>". Both channels must cover manifests.
"""
# There must actually be manifests on disk for the globs to match.
on_disk = list((REPO_ROOT / "plugins").rglob("plugin.yaml")) + list(
(REPO_ROOT / "plugins").rglob("plugin.yml")
)
assert on_disk, "expected bundled plugin manifests under plugins/"
# Wheel channel: package-data must declare a glob that matches plugin
# manifests anywhere under the plugins package.
data = tomllib.loads((REPO_ROOT / "pyproject.toml").read_text(encoding="utf-8"))
plugins_pkg_data = data["tool"]["setuptools"]["package-data"].get("plugins", [])
assert any(
g.endswith("plugin.yaml") or g.endswith("plugin.yml")
for g in plugins_pkg_data
), "pyproject package-data 'plugins' must ship plugin.yaml/plugin.yml (wheel)"
# Sdist channel: MANIFEST.in must recursively include the manifests so
# downstream packagers building from the sdist also get them.
manifest = (REPO_ROOT / "MANIFEST.in").read_text(encoding="utf-8")
assert "recursive-include plugins" in manifest and "plugin.yaml" in manifest, (
"MANIFEST.in must recursive-include plugins plugin.yaml/plugin.yml (sdist)"
)
# Minimum non-vulnerable Starlette: CVE-2026-48710 ("BadHost") was fixed in
# 1.0.1. Anything below that lets a malformed Host header desync
# ``request.url.path`` from the dispatched ASGI path, bypassing path-based
# authz in middleware/endpoints that gate on ``request.url``. Starlette is a
# transitive dep (fastapi in [web]; sse-starlette/mcp in [mcp]/[computer-use]/
# [dev]) so we pin it directly in every extra that exposes a server surface and
# enforce the floor in both pyproject and the committed lockfile.
_STARLETTE_CVE_FLOOR = (1, 0, 1)
def _version_tuple(spec: str) -> tuple[int, ...]:
# "1.0.1" -> (1, 0, 1); tolerant of pre/post suffixes by truncating.
head = spec.split("+", 1)[0]
parts = []
for chunk in head.split("."):
digits = "".join(ch for ch in chunk if ch.isdigit())
if not digits:
break
parts.append(int(digits))
return tuple(parts)
def test_starlette_pinned_above_cve_2026_48710_floor_in_pyproject():
"""Every extra that declares Starlette must pin a patched (>=1.0.1) version.
Regression guard for #35067 / CVE-2026-48710. A future edit that drops the
pin (re-exposing the unbounded transitive ``starlette>=0.27`` from mcp /
``>=0.40.0`` from fastapi) or pins a pre-1.0.1 version fails here instead of
shipping a Host-header auth-bypass to dashboard / MCP-HTTP users.
"""
data = tomllib.loads((REPO_ROOT / "pyproject.toml").read_text(encoding="utf-8"))
extras = data["project"]["optional-dependencies"]
found = {}
for extra, specs in extras.items():
for spec in specs:
name = spec.split("==", 1)[0].split(">", 1)[0].split("<", 1)[0].split("[", 1)[0].strip()
if name.lower() == "starlette":
assert "==" in spec, f"[{extra}] must exact-pin starlette, got {spec!r}"
ver = spec.split("==", 1)[1].split(";", 1)[0].strip()
found[extra] = ver
# The four server-surface extras must each carry the direct pin.
for extra in ("web", "mcp", "computer-use", "dev"):
assert extra in found, (
f"[{extra}] no longer pins starlette directly — CVE-2026-48710 "
f"regression risk (mcp/fastapi pull it transitively with no upper bound)"
)
for extra, ver in found.items():
assert _version_tuple(ver) >= _STARLETTE_CVE_FLOOR, (
f"[{extra}] pins starlette=={ver}, below the CVE-2026-48710 fix "
f"floor {'.'.join(map(str, _STARLETTE_CVE_FLOOR))}"
)
def test_locked_starlette_is_not_vulnerable_to_cve_2026_48710():
"""The committed uv.lock must resolve starlette to a patched version.
pyproject pins protect the declared extras, but the lockfile is what
hash-verified installs (``uv sync --locked``) actually pull. Assert the
resolved version is >= the CVE-2026-48710 fix floor so a stale-lock
regression can't ship a vulnerable Starlette to users.
"""
lock = (REPO_ROOT / "uv.lock").read_text(encoding="utf-8")
versions = []
in_starlette = False
for line in lock.splitlines():
if line.startswith("[[package]]"):
in_starlette = False
elif line.strip() == 'name = "starlette"':
in_starlette = True
elif in_starlette and line.startswith("version = "):
versions.append(line.split("=", 1)[1].strip().strip('"'))
in_starlette = False
assert versions, "starlette not found in uv.lock"
for ver in versions:
assert _version_tuple(ver) >= _STARLETTE_CVE_FLOOR, (
f"uv.lock resolves starlette=={ver}, below the CVE-2026-48710 fix "
f"floor {'.'.join(map(str, _STARLETTE_CVE_FLOOR))} — regenerate the "
f"lockfile after bumping the pin"
)
def test_locale_catalogs_ship_in_both_wheel_and_sdist():
"""Regression test for #27632 / #35374 / #23943.
locales/ is a bare data directory (no __init__.py), so it is invisible to
packages.find and to package-data (which attaches to a package). It must be
declared as setuptools data-files (wheel) AND grafted in MANIFEST.in
(sdist). Without both, sealed installs drop the catalogs and gateway/CLI
commands surface raw i18n keys like `gateway.reset.header_default`.
"""
data = tomllib.loads((REPO_ROOT / "pyproject.toml").read_text(encoding="utf-8"))
data_files = data["tool"]["setuptools"].get("data-files", {})
assert data_files.get("locales") == ["locales/*.yaml"], (
"pyproject [tool.setuptools.data-files] must declare "
'locales = ["locales/*.yaml"] so the wheel ships i18n catalogs'
)
manifest = (REPO_ROOT / "MANIFEST.in").read_text(encoding="utf-8")
assert "graft locales" in manifest, (
"MANIFEST.in must `graft locales` so the sdist ships i18n catalogs"
)
# Every on-disk catalog has the .yaml extension the globs above match.
on_disk = list((REPO_ROOT / "locales").glob("*.yaml"))
assert on_disk, "expected locales/*.yaml catalogs on disk"