Repository navigation
PYTHONSAFEPATH still includes the current directory on Windows for script entrypoints #131484
Description
Bug report
Bug description:
We tried to support PYTHONSAFEPATH in coverage.py (pull request), and had a problem where Windows didn't behave the same as MacOS or Linux. I decided to make a small test repo for checking how it behaved natively in Python, and I think I've found a discrepancy.
The test repo is here: https://github.com/nedbat/safe-path-tests
The code prints a number of values from the sys module. Tox and GitHub Actions are used to run the code in a number of different ways on all three operating systems. Below are excerpts from the action runs, with links to the appropriate spots in the full runs.
When installing the code as an executable script entrypoint and run normally, the first entry in sys.path indicates the source of the code:
Windows (full action run):
py311: commands[21]> safepathtests
================================================================================
3.11.9 (tags/v3.11.9:de54cf5, Apr 2 2024, 10:12:12) [MSC v.1938 64 bit (AMD64)]
sys.argv = ['D:\\a\\safe-path-tests\\safe-path-tests\\.tox\\py311\\Scripts\\safepathtests.EXE']
os.getcwd() = 'D:\\a\\safe-path-tests\\safe-path-tests'
__file__ = 'D:\\a\\safe-path-tests\\safe-path-tests\\.tox\\py311\\Lib\\site-packages\\safepathtests\\work.py'
sys.path:
0: D:\a\safe-path-tests\safe-path-tests\.tox\py311\Scripts\safepathtests.EXE
1: C:\hostedtoolcache\windows\Python\3.11.9\x64\python311.zip
2: C:\hostedtoolcache\windows\Python\3.11.9\x64\DLLs
3: C:\hostedtoolcache\windows\Python\3.11.9\x64\Lib
4: C:\hostedtoolcache\windows\Python\3.11.9\x64
5: D:\a\safe-path-tests\safe-path-tests\.tox\py311
6: D:\a\safe-path-tests\safe-path-tests\.tox\py311\Lib\site-packages
sys.flags:
ignore_environment: 0
isolated: 0
safe_path: False
MacOS (full action run):
py311: commands[21]> safepathtests
================================================================================
3.11.9 (v3.11.9:de54cf5be3, Apr 2 2024, 07:12:50) [Clang 13.0.0 (clang-1300.0.29.30)]
sys.argv = ['/Users/runner/work/safe-path-tests/safe-path-tests/.tox/py311/bin/safepathtests']
os.getcwd() = '/Users/runner/work/safe-path-tests/safe-path-tests'
__file__ = '/Users/runner/work/safe-path-tests/safe-path-tests/.tox/py311/lib/python3.11/site-packages/safepathtests/work.py'
sys.path:
0: /Users/runner/work/safe-path-tests/safe-path-tests/.tox/py311/bin
1: /Library/Frameworks/Python.framework/Versions/3.11/lib/python311.zip
2: /Library/Frameworks/Python.framework/Versions/3.11/lib/python3.11
3: /Library/Frameworks/Python.framework/Versions/3.11/lib/python3.11/lib-dynload
4: /Users/runner/work/safe-path-tests/safe-path-tests/.tox/py311/lib/python3.11/site-packages
sys.flags:
ignore_environment: 0
isolated: 0
safe_path: False
But when run with PYTHONSAFEPATH=1, Windows still has the same first entry in sys.path while MacOS is missing it. I would expect it to be missing:
Windows (full action run):
py311: commands[22]> env PYTHONSAFEPATH=1 safepathtests
================================================================================
3.11.9 (tags/v3.11.9:de54cf5, Apr 2 2024, 10:12:12) [MSC v.1938 64 bit (AMD64)]
sys.argv = ['D:\\a\\safe-path-tests\\safe-path-tests\\.tox\\py311\\Scripts\\safepathtests']
os.getcwd() = 'D:\\a\\safe-path-tests\\safe-path-tests'
__file__ = 'D:\\a\\safe-path-tests\\safe-path-tests\\.tox\\py311\\Lib\\site-packages\\safepathtests\\work.py'
sys.path:
0: D:\a\safe-path-tests\safe-path-tests\.tox\py311\Scripts\safepathtests.exe
1: C:\hostedtoolcache\windows\Python\3.11.9\x64\python311.zip
2: C:\hostedtoolcache\windows\Python\3.11.9\x64\DLLs
3: C:\hostedtoolcache\windows\Python\3.11.9\x64\Lib
4: C:\hostedtoolcache\windows\Python\3.11.9\x64
5: D:\a\safe-path-tests\safe-path-tests\.tox\py311
6: D:\a\safe-path-tests\safe-path-tests\.tox\py311\Lib\site-packages
sys.flags:
ignore_environment: 0
isolated: 0
safe_path: True
MacOS (full action run):
py311: commands[22]> env PYTHONSAFEPATH=1 safepathtests
================================================================================
3.11.9 (v3.11.9:de54cf5be3, Apr 2 2024, 07:12:50) [Clang 13.0.0 (clang-1300.0.29.30)]
sys.argv = ['/Users/runner/work/safe-path-tests/safe-path-tests/.tox/py311/bin/safepathtests']
os.getcwd() = '/Users/runner/work/safe-path-tests/safe-path-tests'
__file__ = '/Users/runner/work/safe-path-tests/safe-path-tests/.tox/py311/lib/python3.11/site-packages/safepathtests/work.py'
sys.path:
0: /Library/Frameworks/Python.framework/Versions/3.11/lib/python311.zip
1: /Library/Frameworks/Python.framework/Versions/3.11/lib/python3.11
2: /Library/Frameworks/Python.framework/Versions/3.11/lib/python3.11/lib-dynload
3: /Users/runner/work/safe-path-tests/safe-path-tests/.tox/py311/lib/python3.11/site-packages
sys.flags:
ignore_environment: 0
isolated: 0
safe_path: True
The pattern continues in Python 3.12 and 3.13 as well. Linux behaves as MacOS does.
Is this correct behavior? Why does Windows keep the first sys.path entry when MacOS and Linux do not?
BTW: #123121 sounds similar, but I don't know if it's related.
Let me know if any of this is unclear, I'll try to explain. The test repo gets a bit intricate.
CPython versions tested on:
3.11, 3.12, 3.13
Operating systems tested on:
Linux, macOS, Windows
Activity
yhay81 commentedon Jul 29, 2026
The platform difference comes from how the entry point is executed, not from
Windows path handling: pip's Windows entry points are exe launchers with an appended
zip archive — the launcher starts a child Python process with its own
.exepath asthe script argument — so they take the zipapp code path, while POSIX entry points are
plain scripts. Reproduced with 3.14.6 and with main at 8b048eb (3.16.0a0):
Root cause: in
pymain_set_path0()(Modules/main.c), themain_importer_pathbranch(directory/zipfile targets) prepends unconditionally; only the non-importer branch
consults
config->safe_path:cpython/Modules/main.c
Lines 701 to 718 in 8b048eb
Skipping the insertion under
-Pwould not be a fix by itself:__main__is thenlooked up generically on
sys.path, so startup would fail — or import an unrelated__main__.So the question is whether keeping the explicitly named execution target on
sys.pathunder-Pis intended. The<script>docs already say a directory/zipfiletarget is prepended; the
-Pdocs just don't cover that case. If intended, I have ashort docs patch ready. If the behavior should change, it would need target-bound
loading of
__main__plus an import-precedence decision.(AI tools assisted this analysis; I reproduced the behavior and checked the cited
source myself.)