Scan from a hatchling member of a uv workspace rewrites the member's pyproject.toml as a lockless Hatch project, so the root uv.lock goes stale, uv sync --frozen installs the unpatched release and vendored vex attests not_affected #1138

[agent] Found by the scheduled uv bug-hunt routine (ledger #310).

Summary

In a uv workspace, uv.lock lives at the workspace root and governs every member. When socket-patch scan runs with --cwd (or the shell) in a member directory whose pyproject.toml uses the hatchling build backend, the run never sees the ancestor uv.lock. detect_pypi_flavor falls through to step 8, "hatchling build backend → hatch", and both modes rewrite the member as a lockless Hatch project:

• vendored writes six @ {root:uri}/.socket/vendor/pypi/<uuid>/six-1.16.0-…whl#sha256=… into the member's dependencies, puts the wheel under packages/a/.socket/, and adds [tool.hatch.metadata] allow-direct-references = true. It exits 0 success.
• hosted writes six @ http://<patch-server>/…/six-1.16.0-…whl#sha256=… into the same place and exits 0 success.

The root uv.lock is never touched, so it no longer matches the workspace:

• uv lock --check and uv sync --locked fail at the root.
• uv sync --frozen, which is what CI and Docker builds use, installs the unpatched six from PyPI.
• The vendored run's vex, from the member, attests not_affected, although the frozen install has unpatched bytes.
• A plain uv sync relocks, and vendored mode then writes an absolute, machine-specific path (source = { path = "/home/<user>/…/packages/a/.socket/vendor/…whl" }) into the committed uv.lock. That lock breaks on every other checkout.

This is a realistic layout. uv init --package scaffolds hatchling members on uv < 0.8 (verified on 0.5.31; 0.12.23 scaffolds uv_build), so every workspace created with those uv versions has hatchling members. It is the uv-workspace counterpart of the governing-root refusals for npm / yarn / Bun (#884), pnpm (#590), vlt (#942) and cargo (#417), and the uv-side sibling of #505 (Hatch's own workspaces).

Repro (Linux, uv 0.12.23 or 0.5.31)

mkdir -p ws/packages/a/src/a && cd ws && touch packages/a/src/a/__init__.py
cat > pyproject.toml <<'EOF'
[project]
name = "root"
version = "0.1.0"
requires-python = ">=3.9"
dependencies = ["a"]
[tool.uv.workspace]
members = ["packages/*"]
[tool.uv.sources]
a = { workspace = true }
EOF
cat > packages/a/pyproject.toml <<'EOF'
[project]
name = "a"
version = "0.1.0"
requires-python = ">=3.9"
dependencies = ["six==1.16.0"]
[build-system]
requires = ["hatchling"]
build-backend = "hatchling.build"
EOF
uv lock && uv sync
(cd packages/a && socket-patch scan --mode vendored --yes --json)   # or --mode hosted; exit 0, "success"
git diff packages/a/pyproject.toml                                   # six @ {root:uri}/.socket/vendor/... (or the hosted URL)
uv lock --check; echo $?                                             # 1: the lockfile needs to be updated
rm -rf .venv && uv sync --frozen && .venv/bin/python -c 'import six; print(getattr(six, "SOCKET_PATCHED", 0))'   # 0: unpatched
(cd packages/a && socket-patch vex --product pkg:pypi/root@0.1.0 --output vex.json)  # vendored: six not_affected

I served patch data from a local mock of the patch API (six@1.16.0, which appends a marker to six.py), the same mock as earlier uv issues.

Expected vs actual

• Expected: fail closed and name the directory to run from, as the governing-root pre-check does for every other workspace family ("a run whose --cwd is a workspace member … reads the member's directory only, while the package manager installs from a lock in an ancestor directory", hosted/governing_root.rs module doc). Running from the workspace root is already refused, by design, with pypi_uv_workspace_unsupported / redirect_uv_project_unsupported. The member path should not quietly succeed where the root path refuses. A wired package also must not be attested not_affected while a frozen install of the project's lock installs the unpatched release.
• Actual: exit 0 success in both modes, a stale root uv.lock, unpatched frozen installs, and a not_affected VEX (vendored).

Matrix (Linux; real uv lock / sync)

rollback from the member restores the member's pyproject.toml byte-identically in both modes, so the damage is reversible once someone notices it.

Not bisected. The hatch routing predates the last release, and the ledger's Sept 30 note ("scan from a member dir falls through, 0 patches") was taken with a member that had no hatchling backend. No probe branch: this is path logic with no OS-specific branch.

Suspect code

• crates/socket-patch-core/src/vendor/pypi.rs:269 (detect_pypi_flavor) only looks at project_root. Its step 8 (return Ok((PypiFlavor::Hatch, …)), line 428) claims a hatchling member without checking for an ancestor pyproject.toml whose [tool.uv.workspace] members cover this directory, or an ancestor uv.lock.
• crates/socket-patch-core/src/hosted/governing_root.rs:72 (refusal) has cargo and npm-family arms but no PyPI / uv-workspace arm, so the hosted Hatch rewrite isn't pre-empted either.