[aitianshi]

Closes #11

What

A driver ran with PYTHONPATH set to the folder that holds the nullius package (run.APP). From a checkout that is app/; from the PyPI install it is the tool's whole site-packages, so every package there came before the driver venv's own, and LongMemEval's Python 3.9 driver imported the tool's Python 3.12 click and stopped at units with a SyntaxError. run.driver_path() now gives drivers a folder in the user cache whose one entry links to the nullius package, the same from a checkout and from a wheel; run.py and prove.py both use it. APP keeps its other uses (the docker backend's mount checks and the relay container).

Touches

the standard

Verified

• New app/tests/test_driver_path.py: the driver path holds only nullius, linked to the real package; a fresh interpreter with it imports nullius.standard.jsonl but not web (from the checkout) nor a package beside nullius in a stand-in site-packages (from a wheel).
• The kit and benchmark driver tests go through the new path.
• New test: while the link is made, the folder on a driver's PYTHONPATH holds nothing else (the link is written beside it and renamed in). It failed before that change, when the .part link was written inside the folder.
• From the installed wheel, both acceptance boxes. A wheel built from this branch and one from main, each uv tool install --python 3.12 into a temporary tool directory and profile; the tool's site-packages holds click 8.5.1. Then nullius standard run longmemeval keyword-memory --dataset longmemeval_oracle --limit 1 --provider custom --model test-model --judge-model test-model --from-run, with the endpoint pointed at a closed local port so no model is called:
  • main: units: the program ended without replying (SyntaxError: invalid syntax), the issue's failure.
  • this branch: stage=units state=start n=1 of=500, then the unit runs and ends error NetworkConnectionError at the closed endpoint; no driver log holds a SyntaxError.
• Backend suite python3 app/tests/parallel.py: 1596 tests, OK.

Risk

A driver that relied on importing something else from app/ or the tool's environment would now fail at import; the bundled drivers import only nullius and their own environment. The link lives in the cache, so clearing the cache only means it is made again. The docker relay still mounts APP and puts it on its own PYTHONPATH inside its container; that is not a driver and is unchanged here. Undone by a revert.

[aitianshi]

/judge

[github-actions]

🟡 PR judge · 6 of 7 yes · #34 ↔ #11

Merge risk	🔴 MAJOR	the judge's reading of the diff; app/nullius/standard/prove.py is in the path of every run
Priority	🔴 EXTREME	the deepest path is app/nullius/standard/prove.py

Question	Verdict
Changes do what the issue asks · 2 of 2 criteria	✅ yes
Issue and pull request describe the same change	✅ yes
Description matches the diff, no more, no less	✅ yes
Diff stays in scope	✅ yes
Tests cover the change, or Verified says how	✅ yes
Docs follow the change	–
Risk named in the description	✅ yes
Before and after shown	–

Acceptance criteria, 2 of 2 met

#	Criterion	Verdict
1	From uv tool install nullius-eval, nullius standard run longmemeval keyword-memory --limit 1 gets past units.	✅ met
2	A driver's PYTHONPATH gives it the nullius package and no other package of the tool's environment, both from the wh…	✅ met

Checks the script runs itself, 4 of 4 pass

No machine attribution trailer in the commits or the description	✓
New code under app/, docs/, .github/ or .claude/	✓
No Flask import under the backend package	✓
No key-shaped string in the diff	✓