[Mikola Lysenko (mikolalysenko)]

LLM Description written by Claude Code:claude-opus-5-5

Fixes #1638

Root cause

The CLI tells users to commit two state files, .socket/vendor/state.json (the vendor ledger) and .socket/manifest.json (the agent-mode manifest). It never asked git whether it would actually commit them. The only gitignore checks (npm_dir::gitignored) looked at the vendored artifact directories, and only in some backends. So a .gitignore rule like state.json, *.json or manifest.json silently dropped the file from the commit. The scan still exited 0, and every other checkout then broke:

• vendored: the lock wiring is there but the ledger is missing, so rollback hits vendor_ledger_missing and list shows no patches.
• agent: the blobs are there but the manifest is missing, so CI's socket-patch apply reports noManifest, exits 0, and the package installs unpatched.

Changes

1. Vendored runs refuse (exit 1, nothing written, dry run included). vendor::ledger_gitignore_refusal lives in vendor/state.rs, next to the ledger. It makes one git check-ignore call over .socket/vendor/ and the ledger and honours the index, so a ledger that is already tracked is never refused. It runs at the shared boundary vendor_records_reusing, which covers vendor and scan/get --mode vendored for every backend. It also runs at the start of the hosted→vendored eject, before any pin is restored.
   • A file rule that only matches the ledger gives vendor_ledger_gitignored, and the message names the rule.
   • A .socket/ or .socket/vendor/ directory rule is left to the per-package vendor_artifact_gitignored refusals when every ecosystem in the run checks that itself (npm family, NuGet, JVM), so that behaviour is unchanged. Otherwise (gem, PyPI, Cargo, Go, Composer) it is now a run-level vendor_artifact_gitignored. Before this PR, those projects lost the ledger and the artifacts silently.
2. Agent-mode runs warn (manifest_gitignored, exit code unchanged). Both the search/scan path (download_and_apply_patches_into) and get <uuid> (save_and_apply_patch) warn whenever the run leaves patches recorded in a manifest git would not commit. This is a warning and not a refusal, because running scan and apply in the same CI job without ever committing .socket/ is a legitimate workflow.
3. CLI_CONTRACT.md documents both codes.

Tests (red → green)

Each new test fails on main and passes here. The ones I checked by reverting just the fix are marked ✔.

• in_process_vendor::a_gitignored_ledger_refuses_before_any_write ✔ covers state.json, *.json + !package*.json and **/state.json, each with and without --dry-run: exit 1, error code, nothing written, lock unchanged.
• in_process_vendor::scan_vendor_gem_refuses_a_gitignored_ledger ✔ runs scan --mode vendored on gem (a non-npm backend) with a state.json rule, which gives vendor_ledger_gitignored, and with a .socket/ rule, which gives vendor_artifact_gitignored.
• in_process_vendor::hosted_to_vendor_conversion::hosted_then_vendor_with_a_gitignored_ledger_keeps_the_hosted_wiring ✔ covers the eject (no manifest) and the manifest-driven takeover. The hosted lock and .npmrc stay untouched.
• get::get_failure_paths::agent_record_warns_when_git_ignores_the_manifest ✔ covers manifest.json, *.json + !package.json and .socket/, which warn, and a re-included manifest, which doesn't.
• get::get_failure_paths::get_uuid_warns_when_git_ignores_the_manifest ✔ covers the get <uuid> save path.
• Controls: a_re_included_ledger_vendors, a_tracked_ledger_is_not_refused, agent_record_does_not_warn_for_a_tracked_manifest. The existing hosted_then_vendor_under_a_gitignored_socket_dir_keeps_the_hosted_wiring and vlt::vendor_vlt_refuses_gitignored_payload still pass.

Local: cargo fmt --all -- --check and cargo clippy --workspace --all-features -- -D warnings are clean. Tests: socket-patch-core lib has 6542 passing and 4 failing. All 4 are read-only-directory tests (copy_tree::relax_loop_must_not_traverse_symlinked_root, vlt_heal::an_unremovable_hidden_lock_keeps_every_store_entry, pypi_poetry::wire_write_failure_maps_error_and_leaves_lock_untouched, pypi_requirements::wire_failure_rolls_back_already_written_files). They fail only because this container runs as root, which ignores chmod, and none of their files are in this diff. socket-patch-cli lib, in_process_vendor and get all pass (924, 139 and 170 tests). A full cargo test --workspace run ran out of disk while linking, so CI runs the full matrix. No wrapper changes are needed: the npm/, pypi/ and gem/ wrappers only dispatch to the binary.

/code-review high

The review found 10 issues. I fixed 1–5, 7, 8 and 10 in 8caf8d6: tracked files were falsely reported as ignored (--no-index), the directory-rule gap for backends that don't check it, get <uuid> not warning, remedy text that couldn't work under an ignored .socket/, four git spawns per check, a misplaced doc comment, and the helpers living in the npm module. Not fixed:

• Add packageManager field to specify pnpm version #6, git failing to answer is treated as "not ignored". This matches npm_dir::gitignored's documented fail-open. The npm-family artifact check already warns vendor_artifact_gitignore_unchecked for that same broken-git case, so a second warning would only duplicate it.
• Add support for public patch API proxy #9, a small git_project_ignoring test helper is copied in two test binaries. The two binaries include different shared-helper modules (common/hermetic.rs vs common/mod.rs), so sharing it would mean wiring a new module into one of them just for a 10-line helper.

CI

1a25bb3 ports #1639's setup-uv 0.13.1 pin. uv 0.13.2 broke the Pipenv pylock e2e legs on every PR, and the pin no-ops once #1639 merges. Bugbot found no issues on 8caf8d6.

https://claude.ai/code/session_01QvLvEUfidP565pLJTJUJLK

---

Note

Medium Risk
Changes vendor exit behavior and adds git subprocess checks on common code paths; low security impact but can block vendoring in repos with broad *.json or state.json ignore rules until fixed.

Overview
Fixes #1638 by detecting .gitignore rules that would keep critical Socket state out of git before users commit broken trees.

Vendored workflows now fail early (vendor_ledger_gitignored, exit 1, no writes, including --dry-run) when .socket/vendor/state.json would not be committed. The check runs on normal vendor / scan --mode vendored paths and on hosted→vendored eject before any lock changes. For ecosystems that do not already enforce per-artifact ignores (e.g. gem), a run-level refusal also covers directory rules like .socket/ that would drop the whole vendor tree.

Agent mode (get, batch download/apply) emits a manifest_gitignored warning (exit unchanged) when .socket/manifest.json would be ignored—so CI that never commits .socket/ still works, but teams that do commit get a clear signal. Already-tracked ledger/manifest files are not flagged.

Git probing was extended so check-ignore can honor the index (tracked files) vs --no-index for new paths. CI pins uv 0.13.1 in Pipenv/pylock legs because 0.13.2 rejects Pipenv’s empty environments pylock.

^{Reviewed by Cursor Bugbot for commit 6094254. Configure here.}

---

Generated by Claude Code

[Mikola Lysenko (mikolalysenko)]

BugBot review

---

Generated by Claude Code

[cursor]

Stale Bugbot comment from a previous run.

[Mikola Lysenko (mikolalysenko)]

[agent] e2e (ubuntu-latest, e2e_vex_build, pipenv:: --ignored, 2026.8.0) failed on 8caf8d6. This PR does not cause it: uv 0.13.2 refuses the environments = [] pylock that Pipenv writes, and the same leg fails on other PRs (#1591). I ported #1639's setup-uv 0.13.1 pin in 1a25bb3. It no-ops once #1639 lands. This PR touches no Pipenv code.

---

Generated by Claude Code

[Mikola Lysenko (mikolalysenko)]

BugBot review

---

Generated by Claude Code

[cursor]

Stale Bugbot comment from a previous run.

[Mikola Lysenko (mikolalysenko)]

bugbot run

---

Generated by Claude Code

[cursor]

Stale Bugbot comment from a previous run.

[Mikola Lysenko (mikolalysenko)]

Labeled Ready for review (burn-down agent).

• Head: 1a25bb33. Mergeable, no conflicts; CHANGELOG.md not touched.
• CI: 54/54 checks green on the head (35 success, 19 skipped by path filters).
• Bugbot: reviewed 1a25bb33, no findings; 0 open review threads.
• Reviewer focus: the vendor-state change in socket-patch-core/src/vendor/state.rs / npm_dir.rs (gitignored ledger/manifest handling, Vendored scan exits 0 when .gitignore covers .socket/vendor/state.json, so the commit silently drops the ledger and every other checkout's rollback fails with vendor_ledger_missing #1638). The two workflow edits are Fix nightly Pipenv pylock legs broken by uv 0.13.2 #1639's uv 0.13.1 pin and no-op once Fix nightly Pipenv pylock legs broken by uv 0.13.2 #1639 merges.

---

Generated by Claude Code

[Mikola Lysenko (mikolalysenko)]

[final reviewer] I reviewed 1a25bb3 and found nothing that blocks merging. I pushed one small fix, 6094254: artifacts_gitignore_checked had been inserted between emit_eject_refusal and its doc comment (vendor.rs:1400), so I moved that doc comment back onto emit_eject_refusal. The change is comments only, and cargo clippy -p socket-patch-cli -- -D warnings is clean. I removed Ready for review until CI is green on the new head, and I'll post the approver brief at that head.

What I checked on 1a25bb3

• in_process_vendor: 139 passed. get: 170 passed. Core vendor::npm_dir / vendor::state: 42 passed.
• The new ledger-refusal and manifest-warning tests fail when crates/*/src is reverted to main, so they exercise the change.
• I probed git 2.43 check-ignore by hand: the path is echoed back as given, tracked ledgers report nothing, !-negations are filtered, and a git error (exit 128) fails open.

Non-blocking notes and questions for the approver

1. Existing projects already in the Vendored scan exits 0 when .gitignore covers .socket/vendor/state.json, so the commit silently drops the ledger and every other checkout's rollback fails with vendor_ledger_missing #1638 state (an untracked ledger plus a state.json or *.json rule) will now exit 1 on every vendored run until the rule is fixed. That is intended, but it may deserve a release-note line.
2. Once anything under .socket/vendor/ is tracked, git (without --no-index) no longer reports a .socket/ directory rule. In that case gem, PyPI, Cargo, Go and Composer runs still get no run-level vendor_artifact_gitignored. main behaves the same, so this is not a regression.
3. If git is broken on a non-npm backend, the ledger check is skipped without any *_unchecked warning.
4. The new tests call the real git init without isolating HOME or core.excludesFile, so an unusual global excludes file on a dev machine could affect them. CI is fine.

---

Generated by Claude Code

[Mikola Lysenko (mikolalysenko)]

bugbot run

---

Generated by Claude Code

[cursor]

✅ Bugbot reviewed your changes and found no new issues!

Comment @cursor review or bugbot run to trigger another review on this PR

^{Reviewed by Cursor Bugbot for commit 6094254. Configure here.}

[Mikola Lysenko (mikolalysenko)]

Labeled Ready for review again (burn-down agent) after the final reviewer's push.

• Head: 6094254. Mergeable, no conflicts; CHANGELOG.md not touched.
• CI: 55/55 checks green (36 success, 19 skipped). No main-wide failures.
• Bugbot: reviewed 6094254, no findings; no open review threads.
• Slack: not announced this run (no Slack send tool available); the next run will retry.

---

Generated by Claude Code

[Mikola Lysenko (mikolalysenko)]

Final review brief

What it does. Vendored runs (vendor, scan/get --mode vendored, and the hosted-to-vendored eject) now ask git check-ignore once, index honoured, before writing anything (dry run included). If a file rule such as *.json or state.json ignores .socket/vendor/state.json and git doesn't already track it, the run is refused with vendor_ledger_gitignored. A .socket/ or .socket/vendor/ directory rule refuses the run as vendor_artifact_gitignored when it includes a backend with no per-package check of its own (gem, PyPI, Cargo, Go, Composer). Agent-mode scan/get also warn with manifest_gitignored; this is a warning only and doesn't change the exit code.

Risk: low to medium. The checks fail open: no git, not a work tree, or a git error mean no refusal. They cost one rev-parse and one check-ignore per run. The medium part is a deliberate behaviour change, documented in CLI_CONTRACT.md. A gem/PyPI/Cargo/Go/Composer vendored run in a repo that ignores .socket/ used to exit 0 and now exits 1. A mixed run (npm plus gem) under that rule now fails as a whole instead of per package.

Look here

• socket-patch-core/src/vendor/state.rs:1028: ledger_gitignore_refusal, which decides between a directory rule and a ledger rule
• socket-patch-core/src/vendor/npm_dir.rs:475: ignoring_rule_lines gains a no_index switch (existing per-package callers keep --no-index)
• socket-patch-cli/src/commands/vendor.rs:2781: the refusal, placed after the empty-vendorable early return so no-op runs aren't refused
• socket-patch-cli/src/commands/vendor.rs:1404 / :1739: artifacts_gitignore_checked (npm, NuGet, Maven count as self-checking) and the eject pre-check
• socket-patch-cli/src/commands/agent_download.rs:1333:`` warn_if_manifest_gitignored

Verified

• cargo clippy --locked -p socket-patch-core -p socket-patch-cli --all-features -- -D warnings: clean.
• cargo test -p socket-patch-cli --test in_process_vendor: the new ledger, re-include, tracked-ledger, gem and eject tests pass, and so do the existing Vendored yarn classic exits 0 when .gitignore covers the vendored tarball (*.tgz, vendor/, .socket/), so the commit drops it and every fresh checkout's install fails #831 tests.
• cargo test -p socket-patch-cli --test get -- manifest: 16/16.
• cargo test -p socket-patch-core --lib gitignore: 9/9.
• The new tests assert codes that don't exist on main, so they would fail without the change.
• Checked check-ignore behaviour by hand on git 2.43.
• The uv-pin commit 1a25bb3 is identical to main (git diff origin/main -- .github/ is empty).
• CHANGELOG.md is untouched.
• CI 55/55 green on 6094254 (36 success, 19 skipped). Bugbot: success. 0 open threads. Mergeable.

Changes I made: none.

Open questions (non-blocking, follow-up material)

1. If any file under .socket/vendor/ is already tracked and the repo has a .socket/ rule, git doesn't report the directory, so the run is refused as vendor_ledger_gitignored and told to add !.socket/vendor/state.json, which can't work under an ignored parent. The refusal is right; only the code and hint are wrong, and the case is narrow.
2. A gem/PyPI/Cargo/Go/Composer project that already committed its ledger and artifacts and then adds .socket/ to .gitignore is still not caught: the probe honours the index and reports nothing. That matches main today, but commit 8caf8d6's message overstates the coverage.
3. Are you OK with the run-level failure for mixed npm+gem runs under a .socket/ rule (see Risk)?

Auto-merge (squash) is armed, so approving sends this straight to the merge queue.

---

Generated by Claude Code