[PDF Skill] macOS: dependency install guidance should default to venv, not system `python -m pip`
What version of the Codex App are you using (From “About Codex” dialog)?
Version 260210.1703 (602)
What subscription do you have?
Business
What issue are you seeing?
Summary
On macOS, Codex sometimes suggests commands like python -m pip install ... that target system/global Python.
This often fails (permissions / externally-managed environment) and is not a reliable default.
The guidance should default to creating and using a virtual environment first.
Environment
- OS: macOS (reproducible across setups)
- Tooling: Codex CLI / agent workflow
- Typical failing pattern:
python -m pip install ...without an active venv
What steps can reproduce the bug?
Steps to Reproduce
- Ask Codex PDF skill to create a pdf.
- Codex suggests or runs
python -m pip install ...to install Python dependencies (example:reportlab,pdfplumber,pypdf). - Install attempts to write to system package directories.
- Command fails due to permission/external-management constraints.
What is the expected behavior?
Expected Behavior
Codex should prefer:
python3 -m venv <path>source <path>/bin/activatepython -m pip install -U pippython -m pip install <packages>
Optional: if uv is used, still require an active venv for package installation.
Additional information
Actual Behavior
Codex proposes system-level pip installs as a fallback path that takes longer and does not perform as well.
Why this matters
This is a common macOS failure mode and not environment-specific.
Defaulting to venv-based installs would prevent recurring setup failures and reduce user friction.
Proposed Fix
Update built-in install guidance (skills/prompts/docs) so Python dependency instructions always use isolated environments first, and avoid system Python installs by default.
2 Comments
Thanks for the suggestion. What you're seeing here is default model behavior, not something the Codex harness is doing. If you see model behavior that you'd like to report, please use the
/feedbackcommand to upload the session details. The team responsible for training our future models use that data for improvements.Our latest models are quite steerable, so you can provide your own guidance in the form of a prompt or an AGENTS.md file (either a global one that spans all of your projects or a per-project one).
Thanks @etraut-openai
I asked Codex App himself how I would make those instructions (using a
venv, etc.) survive our session, and he suggested patching he's own PDF Skill file/Users/ophers/.codex/skills/pdf/SKILL.md.Then I asked where I would report the issue and he both pointed here and provided all the texts.