`vendor -g` and `vendor --revert -g` still rewire the current project: on vlt, `--revert -g` silently unpatches a vendored project (the #446 fix skipped vendor.rs)

[mikolalysenko]

[agent] Found by the scheduled vlt bug-hunt routine (ledger #307).

Summary

PR #446 (fixing #436 / #445) made every -g / --global-prefix run leave the cwd project's hosted pins and vendored wiring alone, but only in get, scan, apply, rollback and remove. The standalone vendor command was not changed, and it ignores global scope completely:

• vendor --revert -g run inside a vendored project reverts the project's vendoring: vlt-lock.json goes back to the upstream registry entry and .socket/vendor/ is deleted. It exits 0 and the global copy is untouched. The next vlt ci installs pristine left-pad, so the project is silently unpatched. This is the rollback -g and remove <purl> -g also unwind the current project's hosted pins and vendored wiring; on vlt they delete node_modules/left-pad too #445 failure again, reached through vendor instead of rollback.
• vendor -g run inside a project whose .socket/manifest.json holds a record (for example one written by get -g, which records the global patch in the cwd manifest) vendors into the project. It creates .socket/vendor/npm/<uuid>/…, rewires vlt-lock.json to a file~.socket+vendor+… node, and does a hosted→vendored takeover (vendor_takeover_reverted_redirect) when the project was hosted. It exits 0.
• SOCKET_GLOBAL=1 / SOCKET_GLOBAL_PREFIX behave the same as the flags.

Impact

A user who patched a global tool (get -g) and then runs vendor --revert -g or vendor -g from inside their project gets their project's lockfile rewritten, with exit 0 and no warning. With --revert -g the project is unpatched on the next frozen install. This is the same class of damage as #445, which was rated p1.

Expected (CLI_CONTRACT.md)

Global scope never touches the project's state (v5.0). A --global/--global-prefix run that starts inside a project acts on the global installs only. The --cwd project's hosted pins and vendor ledger are not its target …

scan and get with --mode vendored under -g are a usage error (exit 2: "global installs have no project lockfile … to wire vendored artifacts into"). vendor -g should be refused the same way, or be a no-op on the project. vendor --revert -g must not revert the project's ledger entries.

Repro (Linux, vlt 1.3.3; a local mock registry plus patch API, as in the ledger)

G=$PWD/gprefix; npm install -g --prefix $G --registry $REG left-pad@1.3.0
mkdir proj && cd proj
echo '{"name":"app","version":"1.0.0","dependencies":{"left-pad":"1.3.0"}}' > package.json
echo '{"config":{"registries":{"npm":"'$REG'"}}}' > vlt.json
vlt install
socket-patch scan --mode vendored --yes --json          # rc 0, project vendored
cp vlt-lock.json lock.vendored
socket-patch get pkg:npm/left-pad@1.3.0 -g --global-prefix $G/lib/node_modules --yes --json   # rc 0, global patched
socket-patch vendor --revert -g --global-prefix $G/lib/node_modules --yes --json             # rc 0
cmp vlt-lock.json lock.vendored   # differs: back to the registry entry; .socket/vendor gone
rm -rf node_modules && vlt ci --allow-scripts :scripts
node -p "require('left-pad')"     # pristine

The vendor -g variant: skip the scan --mode vendored step, then run get … -g followed by vendor -g --global-prefix …. vlt-lock.json gains "file~_d left-pad": "prod file:./.socket/vendor/npm/<uuid>/left-pad-1.3.0/node_modules/left-pad …" and the next vlt ci installs the vendored copy into the project.

Actual vs expected

Command (inside the project, global copy patched)	Expected	Actual
vendor --revert -g (flag or SOCKET_GLOBAL=1)	project's vendoring kept	project's vendoring reverted, rc 0, vlt ci → pristine
vendor -g	exit 2 usage error (as scan -g --mode vendored), or no project change	vendors into .socket/vendor, rewires vlt-lock.json, hosted→vendored takeover, rc 0

OS × version

OS	vlt 1.0.10	vlt 1.2.0	vlt 1.3.3	npm control (package-lock)
Linux	repro (vendor -g and --revert -g)	repro	repro (2/2 runs, flag and env)	repro (vendor --revert -g unwinds the npm project too)
macOS / Windows	untested (no probe branch this run)	untested	untested	—

The logic doesn't depend on the package manager: it reproduces against npm too. I'm filing it under vlt because that's where I found it, as with #445.

First bad

Release 4.0.0 predates vlt support. On main, vendor has never consulted global scope. #446 (551c362) fixed the sibling commands but left vendor.rs untouched, so this is a gap in that fix rather than a regression. Tested on main 61cfb9b.

Suspect code

• crates/socket-patch-cli/src/commands/vendor.rs:665: only the manifest-less eject path checks is_global(). The manifest-driven vendor path (falling through to the vendoring at ~:810) and run_revert (vendor.rs:3119) never consult crate::commands::project_state_in_scope (commands/mod.rs:44) or global_mode_conflict.

[mikolalysenko]

[agent] Triaged as priority:p1. The reporter confirmed the npm (package-lock) control also reproduces, so this is npm-family and not specific to vlt. I checked main (61cfb9b): vendor.rs never calls project_state_in_scope / global_mode_conflict on the manifest-driven vendor path or in run_revert. Only the manifest-less eject path checks is_global(). It isn't a duplicate. It's a gap left by #446, which fixed the sibling commands (#436, #445). No open PR covers it.

---

Generated by Claude Code

[mikolalysenko]

[agent] Claiming this issue (shared root cause: the vendor command never consults global scope (project_state_in_scope), so -g runs rewire or revert the cwd project). Branch: agent/fix-vendor-global-scope. Claim-ID: 2026-10-01T20:21:18Z-874386

---

Generated by Claude Code

[mikolalysenko]

[agent] Draft PR: #499

---

Generated by Claude Code