You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
v0.8.15 — prefer a locally installed runner over uvx/npx
The version gate no longer treats a YAML-quoted version as a different version.
A skill whose frontmatter said version: "1.1.0" was read as the literal string '"1.1.0"', which never equals the ledger's 1.1.0 — so the gate blocked a skill
that had not moved, and then wrote the quoted value back into the shared skill_versions.json. That second half was the worse one: the ledger is shared with tongbu-skills, which reads the same file, so a value one manager wrote could make
the other manager stop blocking that skill — a gate silently switched off by the
other gate's write. version_gate now strips the quotes with the same strip_yaml_quotes the frontmatter reader uses. Verified against the shared ledger:
390 skills allowed by both managers, identical sets.
Changed
Discovery, import, install and sync now prefer a locally installed module over uvx/npx. When a server's package is already importable in the interpreter running
mcptoon, the entry is written as python -m <module> instead of uvx <pkg>. uvx starts a
second interpreter and resolves the package into a fresh temp environment on every cold
run, so a host config full of uvx entries turns "open my agent" into "fetch N packages at
once" — the shape of a measured machine freeze on 2026-10-06, where a sync'd config with
three uvx rows and a uv cache building twenty temp environments landed on a box already
short on commit. The local form has no fetch, no download and no second interpreter. The
gate is two facts, not one: the module must resolve and expose a __main__
(python -m executes that), so an importable library with no entry point is never written
as a server. Servers whose package is not installed are unchanged — uvx is still the
fallback, and the behavior is preserved verbatim behind mcptoon config set runners allow-fetch.
Added
runners setting (prefer-installed | allow-fetch, default prefer-installed) — the
opt-out for the change above. mcptoon config set runners allow-fetch restores the old
fetch-on-first-run behavior for every surface at once.
Tests
tests/test_discover.py gains 7 tests over the local-runner resolver (installed module wins,
absent package keeps uvx, the opt-out wins, an importable module with no __main__ is not
runnable, an imported uvx row is rewritten on the way in), tests/test_sync.py gains 4
(the sync guard rewrites, keeps, and respects the opt-out), tests/test_installer_registry.py
gains 5 (the registry row's runner, including that trailing server args survive the swap),
and tests/test_exit_codes.py gains 4 (the new setting is validated and persisted like every
other). That is 20 tests for the runner change; the version-gate fix adds 3 more in tests/test_skills.py. Suite: 1946 passed + 2 skipped.