ouyangbo/hermes-agent
1
0
Fork 0
Code Issues Pull Requests Actions 11 Packages Projects Releases Wiki Activity

fix(macos): scrub gate env, refuse EACCES, normalize marker paths

Browse Source 
Review fixes from kokhlo's live-hardware review:

- The boot-gate probe now runs with PYTHONHOME / PYTHONPATH /
  PYTHONSTARTUP / __PYVENV_LAUNCHER__ scrubbed: an inherited
  PYTHONHOME=<venv> boots a staged copy that would otherwise die with
  "No module named 'encodings'", papering over the exact prefix
  failure the gate exists to catch.
- OSError is split by errno: ENOENT/ENOEXEC (fixtures, foreign-arch
  images) still skip; EACCES after our own chmod now refuses the
  install instead of silently accepting a broken copy.
- Marker writes and both marker comparisons go through os.path.realpath,
  so the managed-runtime layout (cpython-3.11-macos-* symlinked to
  cpython-3.11.15-macos-*) no longer reports stale on a fresh install.

Tests: +3 (env-scrub spy, EACCES refusal, symlinked-home state).
100 passed in the module + doctor neighborhoods.
This commit is contained in:
Zheqing Zeng
2026-08-26 17:22:48 -07:00
committed by kshitij
parent 1c92b95a12
commit 37bccf343e
2 changed files with 92 additions and 13 deletions
Download Patch File Download Diff File Expand all files Collapse all files
hermes_cli/macos_tcc_anchor.py
+32 -9
View File
@@ -38,6 +38,7 @@ uv-managed. Best-effort: never raises to callers.
from __future__ import annotations
import errno
import logging
import os
import platform
@@ -61,6 +62,13 @@ class _BootGateFailed(Exception):
"""Staged copy refused to boot; the live venv must stay untouched."""
def _marker_value(source_file: Path) -> str:
"""Canonical marker value: fully resolved so symlinked spellings of the
same store binary (``cpython-3.11-macos-*`` → ``cpython-3.11.15-macos-*``)
compare equal."""
return os.path.realpath(str(source_file))
def is_macos() -> bool:
return platform.system() == "Darwin"
@@ -231,19 +239,34 @@ def _materialize_aliases(
def _passes_boot_gate(staged: Path, venv_dir: Path) -> bool:
"""Launch *staged* and demand encodings + the venv prefix.
``OSError`` (fixture / wrong arch / exec-format) is treated as a skip —
we cannot verify, and we also cannot brick a host that cannot run the
binary. A real crash (dyld, encodings, wrong prefix) refuses the install.
The probe runs with ``PYTHONHOME``/``PYTHONPATH`` scrubbed: an inherited
``PYTHONHOME`` papers over exactly the prefix-resolution failure the gate
exists to catch. ``OSError`` is split by errno — ``ENOENT``/``ENOEXEC``
(fixture binaries, foreign-arch images) means the binary cannot run here
at all, so the symlinked venv was equally dead and installing cannot make
things worse: skip. Anything else (notably ``EACCES`` after our own
chmod) is a broken install about to go live: refuse.
"""
env = {
k: v
for k, v in os.environ.items()
if k not in ("PYTHONHOME", "PYTHONPATH", "PYTHONSTARTUP",
"__PYVENV_LAUNCHER__")
}
try:
proc = subprocess.run(
[str(staged), "-c", "import encodings, sys; print(sys.prefix)"],
capture_output=True,
text=True,
timeout=30,
env=env,
)
except OSError:
return True
except OSError as exc:
if exc.errno in (errno.ENOENT, errno.ENOEXEC):
logger.debug("boot gate skipped: cannot execute %s (%s)", staged, exc)
return True
logger.warning("boot gate: staged copy not executable: %s", exc)
return False
except subprocess.TimeoutExpired:
return False
if proc.returncode != 0:
@@ -282,7 +305,7 @@ def _install_anchor(venv_dir: Path, source_file: Path) -> None:
f"staged copy at {tmp_path} failed encodings/prefix probe"
)
os.replace(tmp_path, venv_py)
_anchor_marker(venv_bin).write_text(str(source_file), encoding="utf-8")
_anchor_marker(venv_bin).write_text(_marker_value(source_file), encoding="utf-8")
_materialize_aliases(venv_bin, venv_py, refresh=True)
except Exception:
try:
@@ -317,8 +340,8 @@ def ensure_tcc_anchor(project_root: Path | None = None) -> Path | None:
if not venv_py.is_symlink():
marker = _anchor_marker(venv_py.parent)
try:
if marker.is_file() and marker.read_text(encoding="utf-8").strip() == str(
source_file
if marker.is_file() and marker.read_text(encoding="utf-8").strip() == (
_marker_value(source_file)
):
_provision_libpython(venv_dir, source_file, refresh=False)
if _passes_boot_gate(venv_py, venv_dir):
@@ -361,7 +384,7 @@ def tcc_anchor_state(project_root: Path | None = None) -> tuple[str, str]:
if not venv_py.is_symlink():
marker = _anchor_marker(venv_py.parent)
source_file = _interpreter_file(source)
expected = str(source_file) if source_file is not None else source
expected = _marker_value(source_file) if source_file is not None else source
try:
if marker.is_file() and marker.read_text(encoding="utf-8").strip() == expected:
return "active", str(venv_py)
tests/hermes_cli/test_macos_tcc_anchor.py
+60 -4
View File
@@ -12,6 +12,7 @@ the existing macOS CI job, not against one-byte fixtures.
from __future__ import annotations
import errno
import os
import platform
import shutil
@@ -332,15 +333,47 @@ class TestBootGate:
assert not tcc._passes_boot_gate(tmp_path / "staged", tmp_path / "venv")
def test_skips_unexecutable_staging(self, tmp_path, monkeypatch):
# OSError (fixture binaries, exec-format on a foreign arch) means the
# binary cannot run here at all — the symlinked venv was equally dead,
# so installing cannot make things worse. Skip, don't refuse.
# ENOENT/ENOEXEC (fixture bad-shebang binaries, wrong-arch images)
# mean the binary cannot run here at all — the symlinked venv was
# equally dead, so installing cannot make things worse.
def noexec(*a, **k):
raise OSError("cannot execute")
raise OSError(errno.ENOENT, "No such file or directory")
monkeypatch.setattr(tcc.subprocess, "run", noexec)
assert tcc._passes_boot_gate(tmp_path / "staged", tmp_path / "venv")
def test_refuses_eacces_after_chmod(self, tmp_path, monkeypatch):
# EACCES after copy2+chmod is a broken install about to go live —
# refuse, never skip (review: kokhlo on #95605).
def denied(*a, **k):
raise PermissionError(errno.EACCES, "Permission denied")
monkeypatch.setattr(tcc.subprocess, "run", denied)
assert not tcc._passes_boot_gate(tmp_path / "staged", tmp_path / "venv")
def test_scrubs_pythonhome_pythonpath(self, tmp_path, monkeypatch):
# A staged copy that dies with `No module named 'encodings'` boots
# fine under PYTHONHOME=<venv> — the inherited var papers over the
# exact failure the gate exists to catch (review: kokhlo on #95605,
# verified live). The probe must run with a scrubbed environment.
captured = {}
def spy(*a, **k):
captured.update(k)
return self._proc(a[0], 0, f"{tmp_path / 'venv'}\n", "")
monkeypatch.setattr(tcc.subprocess, "run", spy)
monkeypatch.setenv("PYTHONHOME", "/poison")
monkeypatch.setenv("PYTHONPATH", "/poison")
venv = tmp_path / "venv"
venv.mkdir(exist_ok=True)
assert tcc._passes_boot_gate(tmp_path / "staged", venv)
env = captured.get("env")
assert env is not None, "gate must pass an explicit scrubbed env"
assert "PYTHONHOME" not in env
assert "PYTHONPATH" not in env
def test_accepts_matching_prefix(self, tmp_path, monkeypatch):
venv = tmp_path / "venv"
venv.mkdir()
@@ -352,6 +385,29 @@ class TestBootGate:
class TestTccAnchorState:
def test_state_active_through_unpatched_home_symlink(self, tmp_path, monkeypatch):
# The managed-runtime layout symlinks cpython-3.11-macos-* →
# cpython-3.11.15-macos-*, so pyvenv.cfg home and the marker record
# different spellings of the same binary. State must resolve both
# sides before comparing, or a fresh install reports stale
# (review: kokhlo on #95605, hit on a live venv).
_darwin(monkeypatch)
patched = _build_store(tmp_path, version="3.11.15")
versionless = patched.parent.parent / "cpython-3.11-macos-aarch64-none"
os.symlink(patched.parent, versionless)
home = versionless / "bin"
root = tmp_path / "checkout"
venv_bin = root / ".venv" / "bin"
venv_bin.mkdir(parents=True)
(root / ".venv" / "pyvenv.cfg").write_text(f"home = {home}\n")
os.symlink(home / "python3.11", venv_bin / "python")
tcc.ensure_tcc_anchor(root)
status, _ = tcc.tcc_anchor_state(root)
assert status == "active"
def test_state_missing_then_active(self, tmp_path, monkeypatch):
_darwin(monkeypatch)
store_bin = _build_store(tmp_path)