Agents in this repository run tests in a cloud container whose Python setup varies from image to image. A small script, platform/ci/pyenv.py, finds an interpreter that can import the modules you name and installs any that are missing. The weak part was "the modules you name". Twice today a run asked for pytest or pytest numpy matplotlib, got an interpreter back, and then watched the suite fail on numpy or scikit-learn.
A requirements file would fix that, but only 5 of the 1,083 research directories have one. The code itself does say what it needs, though. So pyenv.py --for research/marine-hull/tests now parses each .py file and collects its top-level imports. Then it drops:
- the standard library;
- relative imports;
- imports inside a
try:block, since the code runs without them; - anything the repository holds itself. Many scripts reach a sibling project's module through
sys.path, and pip has never heard oflinefigorcore.
What is left, plus pytest, goes to the usual install path.
Two checks. For research/codesign-transfer it gives cma, matplotlib, numpy and scipy, the same list that project's README writes out by hand. For marine-hull's tests it picked /usr/bin/python3.13, which ran the suite: 150 passed.
It has one honest limit. It reads only the folder you point it at. If a project imports a sibling's module, it does not follow that module's imports. Point it at the tests/ folder when a project also holds data fetchers with heavy dependencies you don't need.