Patch release
Fixes a false-positive path-traversal error in soup autopilot on
Windows + Python 3.9.
What's fixed
-
soup autopiloton Windows + Python 3.9 — the--dataand--outputpath guards were usingPath.resolve() + relative_to(), which on Windows 3.9 occasionally leaves 8.3 short names (C:\Users\RUNNER~1\...) on one side of the comparison but not the other, causingrelative_toto raiseValueErroreven when the paths point to the same location. Replaced withos.path.realpath+os.path.commonpathvia a new_is_under_cwdhelper. Fixes #22 path-handling class of issues on Windows. -
CI encoding on Windows — added
PYTHONUTF8=1andPYTHONIOENCODING=utf-8to the CI workflow env.trl.trainer.grpo_trainerhas been flaky onwindows-latestbecause some upstream files are read without an explicit encoding and Windows defaults tocp1252. Also added a defensive skip intests/test_bugfixes.py::TestGRPOCPUMinNewTokensas a belt-and-braces layer.
Upgrade
pip install --upgrade soup-cliNo breaking changes. All v0.25.0 users on macOS / Linux / Windows py3.11+ are unaffected. Windows + py3.9 soup autopilot users should upgrade.
Security
No security changes. The code_exec_reward sandbox status, MLX backend coverage, and documented design limitations from v0.25.0 are unchanged — see #21, #22, #23 for the follow-up roadmap.