I noticed this while testing the new-python-pipeline version of PythonEval, which kept failing on real tasks in our repo