Hosted and vendored uv wiring ignore `no-sources = true`: scan and vendor report success, `uv sync --locked` then fails, and a plain `uv sync` reinstalls the unpatched wheel

[mikolalysenko]

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

Summary

Hosted (scan --mode hosted) and vendored (vendor) wiring for uv projects both route the patched package through a [tool.uv.sources] entry, and they repoint uv.lock to match. When the project sets no-sources = true (in [tool.uv] of pyproject.toml or in a project uv.toml), uv ignores [tool.uv.sources] entirely. Neither writer checks for this. Both report success / exit 0 and leave a lock that uv considers outdated.

Impact

• Locked CI breaks right after wiring: uv lock --check and uv sync --locked fail ("The lockfile at uv.lock needs to be updated, but --locked was provided"). That's exit 1 on uv 0.8.17 / 0.12.22 and exit 2 on 0.5.31.
• The patch is silently lost: a plain uv sync (or uv lock) relocks six to registry = "https://pypi.org/simple". On a fresh checkout it installs the unpatched upstream wheel. Only uv sync --frozen installs the patch.
• No warning or refusal is emitted, so the user believes the project is patched.

Repro (Linux, main 61cfb9b, real uv 0.8.17)

cat > pyproject.toml <<'EOF'
[project]
name = "ns"
version = "0.1.0"
requires-python = ">=3.9"
dependencies = ["six==1.16.0", "idna==3.7"]

[tool.uv]
no-sources = true
EOF
uv lock && uv sync --no-install-project
# a free six 1.16.0 patch: .socket/manifest.json + blob for vendor, or the hosted routes (local mock patch API)
socket-patch vendor --json            # status success, applied 1; writes [tool.uv.sources] six = { path = ".socket/vendor/pypi/<uuid>/six-1.16.0-py2.py3-none-any.whl" }
#   or: socket-patch scan --mode hosted --json --yes --api-url $API ...   # status success, redirected 1; writes six = { url = "<hosted wheel>" }
uv lock --check                        # exit 1: lock needs to be updated
rm -rf .venv && uv sync --locked --no-install-project   # exit 1
rm -rf .venv && uv sync --frozen --no-install-project   # installs the patched wheel
rm -rf .venv && uv sync --no-install-project            # relocks six to the PyPI registry; installs UNPATCHED six

Control: the same project without no-sources = true passes uv lock --check and uv sync --locked, and a plain uv sync keeps the patch (hosted and vendored, same run). Putting no-sources = true in a project uv.toml instead gives the same failure (vendored, 0.8.17).

Expected vs actual

• Expected: docs/testing/uv-compatibility.md says the [tool.uv.sources] entry "carr[ies] the redirect (verified with --frozen, --locked, and ordinary installs)". When a project setting stops uv from honouring that entry, the writers should refuse or at least warn before writing anything, the way they already refuse other configurations they can't wire (redirect_uv_project_unsupported, pypi_uv_source_already_exists). That would leave the project untouched.
• Actual: both modes write the source plus the lock repoint and report success. The result fails --locked, and an ordinary install drops the patch.

OS × version (Linux; the check is config-driven, so it doesn't depend on the OS)

Mode	uv 0.5.31	uv 0.8.17	uv 0.12.22
vendored, [tool.uv] no-sources = true	❌	❌ (reproduced 2×)	❌
hosted, [tool.uv] no-sources = true	❌	❌ (reproduced 2×)	❌
vendored, uv.toml no-sources = true	–	❌	–
control (no no-sources), hosted and vendored	✅	✅	✅

Not bisected. git grep no-sources on main finds nothing in the crates or docs.

Suspect code

• crates/socket-patch-core/src/vendor/pypi_uv.rs:589: ensure_table(&mut doc, &["tool", "uv", "sources"]) is written unconditionally. Nothing reads tool.uv.no-sources or uv.toml.
• crates/socket-patch-core/src/utils/python_script.rs:142 (rewrite_sources, called from rewrite_project_metadata at :230): the same for the hosted path.

[mikolalysenko]

[agent] Triage: priority:p1 (uv). Confirmed on main: neither the hosted nor the vendored uv [tool.uv.sources] writer checks no-sources (no occurrence of no-sources in socket-patch-core). Distinct from open #545 (sources-table spelling); no open PR covers it.

---

Generated by Claude Code