Repository navigation
v1.108.327 — a local model path is refused when the installed sentence-transformers would run its code
·
546 commits
to main
since this release
Security
- A local model path is refused when the installed
sentence-transformerswould run its code.
1.108.326 raised thesemanticextra's floor to 5.6.0 (GHSA-jhr6-gm9c-rqjv), and a floor is install
metadata:pip install -U jcodemunch-mcpwithout the extra named, or asentence-transformers
installed directly, keeps the older release, and nothing said so. The embedding call now checks the
release it imported. Whenembed_model(orJCODEMUNCH_EMBED_MODEL) is a path that exists and the
release is older than 5.6.0, is a pre-release of 5.6.0, or has a version that cannot be read, the call
refuses before the model is constructed and names the advisory and the upgrade command. Where the
reason appears:embed_reporeports it undererror_causes;search_symbols(semantic=true)returns
it as the error;search_symbols(fusion=true)still answers from its other channels and carries it
in a newsemantic_channel_errorfield, with the verdict's semantic channel readingunavailable.
That fusion exit used to catch every failure of its similarity channel and answeroff, the same as
a repository that was never embedded; any exception there is named now, not only this refusal, with
the message redacted and cut likeerror_causes. Two things follow. A fusion answer whose channel
failed is not cached, so the next call tries the channel again (a cached refusal would have been
replayed after the upgrade it asks for). And a fusion search that returns nothing while the channel
failed still readsabsent, as it did before: its lexical and identity passes score every candidate,
so zero rows is a fact about the corpus. The failed channel is labelled beside that verdict.
An embedded repository with no provider configured still readsoff; nothing was attempted there. A Hub
model name is not refused, because the bypass is in the local-directory path. A name that is a path
only after~or$VARexpansion is refused too, although the library would not load it from disk.
A fixed release behaves as before (LEDGER L-108). jdocmunch-mcp and jdatamunch-mcp have their own
call sites and do not have this check (L-109).
90 tools in the full surface; Python >=3.10. Published by the Release workflow with PyPI trusted publishing and sigstore-signed artifacts.