[Mikola Lysenko (mikolalysenko)]

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

Fixes #1931

What changes for users

In a Ruby project whose Bundler config sets disable_shared_gems to a false value ("false", "no", "0" or empty) in .bundle/config, the environment or the global config, Bundler installs and loads gems from the system gem homes, exactly as it does for path.system: true. socket-patch now treats it the same way:

• A leftover vendor/bundle is no longer crawled and no longer switches off the gem env homes, so apply patches the copy Bundler actually loads.
• A path recorded in the same config file as that flag is ignored too (local or global).

Before this change, apply could report success after patching a copy Bundler never loads.

Root cause

Bundler's Settings#path tier rule (the first of local, env and global that sets path, path.system or disable_shared_gems decides, and a truthy path.system or a falsy disable_shared_gems means system gems) was written twice in crawlers/ruby_crawler.rs. bundler_sets_explicit_path followed Bundler. Store discovery's copy (bundler_path_tier + config_path_tier) read system from path.system only and never saw the env BUNDLE_DISABLE_SHARED_GEMS.

The fix

• BundlerPathSettings::decides_path / uses_system_gems and BundlerPathTiers::deciding form the one tier resolver. Store discovery, the global-path shadowing check and bundler_sets_explicit_path all read through it.

• Removed: BundlerPathTier, config_path_tier, bundler_path_tier, app_config_sets_path, global_path_config_unless_env_path_settings and ambient_path_system_env. The env tier's flags reach discovery as BundlerEnvPathFlags instead of pre-filtering the global config file.

• parse_bundle_config_path drops a recorded path when the same file sets a falsy disable_shared_gems, as it already did for path.system: true.

• Side effects, both consistent with Bundler: an env BUNDLE_PATH__SYSTEM or BUNDLE_DISABLE_SHARED_GEMS no longer hides the whole global config file from discovery (only its path is shadowed), so a global BUNDLE_GEMFILE is still honored; and verification_only_gem_paths now passes the ambient env flags into discovery.

Tests (red → green)

Issue case	Test	Red without fix
#1931 local / global disable_shared_gems falsy + leftover vendor/bundle	falsy_disable_shared_gems_drops_leftover_vendor_bundle	yes (local false: [".../vendor/bundle/ruby/3.3.0/gems"])
#1931 env BUNDLE_DISABLE_SHARED_GEMS=false	env_falsy_disable_shared_gems_drops_leftover_vendor_bundle	yes (mutation: resolver without the disable_shared_gems arm)
same-tier recorded path	falsy_disable_shared_gems_drops_same_tier_path	yes (local: [".../vendor/x/ruby/3.3.0/gems"])
controls (truthy flag, shadowed falsy flag)	truthy_or_shadowed_disable_shared_gems_keeps_vendor_bundle	n/a (passes before and after)
env flags shadow a global path (replaces the removed helper's unit test)	env_path_flags_shadow_the_global_config_path	n/a

The mutation check (dropping the disable_shared_gems arm from uses_system_gems) also fails the existing bundler_sets_explicit_path_follows_settings_tiers, which shows discovery and the stale-install/vex guard now share one rule.

Local runs:

• cargo fmt --all -- --check: ok
• cargo clippy --workspace --all-features -- -D warnings: ok
• cargo test --workspace --all-features --no-fail-fast: 236 targets pass. Four targets fail only in this sandbox and none touch Ruby code: socket-patch-core --lib (4 tests), in_process_redirect (3), repair (2), e2e_bun_lockb (2). The first three inject write failures with read-only dirs, which the container's root user ignores. e2e_bun_lockb needs the live patch API (see Fix bun.lockb nightly e2e reaching the live blob API #1905) and depends on the local bun 1.4.4 build. Draft CI on ba9a49b was green.
• Ruby crawler unit tests: 112 pass. crawler_ruby_e2e: 30 pass. e2e_redirect_gem_stale_install: 42 pass. e2e_redirect_gem_build -- --ignored (Bundler 4.0.18): 37 pass. e2e_vendor_gem_build -- --ignored: 11 pass.
• No changes to the npm/pypi/gem wrappers: this is crawler logic only.

Code review findings not fixed here

/code-review high ran on the diff. The real findings are fixed in fe20764 (same-tier path, single resolver call, stale comments). Three were left as they are:

• Env BUNDLE_PATH root when a local tier decides without a path. Bundler ignores the env path there, but discovery probes both roots on purpose: bundle_roots_probe_in_bundler_precedence_order and refused_config_root_is_verification_only assert it. Narrowing that is a design change, so it's left for a maintainer to decide.
• Config files read several times per discovery. This is an existing pattern; Move Bundler settings parsing out of ruby_crawler into formats/gem/settings.rs, and read .bundle/config keys with one parser (two drifted today) #1935 (move Bundler settings into formats/gem/settings.rs with one snapshot) is where it belongs.
• BundlerEnvPathFlags beside BundlerPathSettings. It keeps the env BUNDLE_PATH as a raw OsStr path (non-UTF-8 safe) with a single source. Move Bundler settings parsing out of ruby_crawler into formats/gem/settings.rs, and read .bundle/config keys with one parser (two drifted today) #1935's BundlerEnv::ambient() snapshot can fold it in.

Follow-ups

• Move Bundler settings parsing out of ruby_crawler into formats/gem/settings.rs, and read .bundle/config keys with one parser (two drifted today) #1935 moves this resolver into formats/gem/settings.rs.

---

Generated by Claude Code

[Mikola Lysenko (mikolalysenko)]

BugBot review

---

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 fe20764. Configure here.}

[Mikola Lysenko (mikolalysenko)]

[agent] hosted-e2e (and the ci-ok gate that depends on it) is red on fe20764. It failed again on one re-run. The cause is outside this PR: the suite runs against live production (patch.socket.dev), and production changed after main's last green run at 05:42 UTC today. This PR only changes how the Ruby crawler finds Bundler install roots. The hosted redirect path these tests check doesn't touch that code.

• bun_hosted_install_proof: GET https://patch.socket.dev/patch/npm/minimist/1.2.2/…/minimist-1.2.2.tgz returns 403. That is a production grant or data problem, and no change in this repo can fix it.
• gem_bundler_hosted_install_proof: production now also serves a patch for actionpack 6.0.3 (grant 78b877c4…, patch f86d5caa-9eb0-4984-9ec7-87d5b0c83ed6). The fixture pulls actionpack in transitively, so the rewritten Gemfile now gets a Socket source block for actionpack before the activestorage one. wired_gem_registry_uuid (tests/e2e_hosted_production.rs:714) reads the first block, so it reports the actionpack UUID and fails the GEM_UUIDS any-of check. The test-side fix is to match the block that wraps gem "activestorage". That belongs in its own ci-janitor / agent/ci-* PR. No such fix exists yet.

All other checks are green, and Bugbot's review of fe20764 found no issues. A maintainer needs to look at the bun 403 on the production side.

---

Generated by Claude Code

[Mikola Lysenko (mikolalysenko)]

[agent] Update: hosted-e2e passed on a later re-run (19:12 UTC), so every check on fe20764 is now green. The production-side problems from my earlier comment cleared on their own. The PR is waiting on human review.

---

Generated by Claude Code

[Mikola Lysenko (mikolalysenko)]

[burn-down agent] Ready for review at fe20764.

• CI: 82/82 check runs on fe20764 green (success/skipped/neutral); no main-wide failures.
• Bugbot: reviewed fe20764, no findings; no open review threads.
• Mergeable, no conflicts with main; no CHANGELOG.md change.

---

Generated by Claude Code

[Mikola Lysenko (mikolalysenko)]

Final review brief

What it does: Fixes #1931. Store discovery had its own copy of Bundler's Settings#path tier rule. That copy counted only a truthy path.system as "system gems" and ignored disable_shared_gems. So BUNDLE_DISABLE_SHARED_GEMS=false next to a leftover vendor/bundle made apply patch a copy Bundler never loads. The PR replaces the copies with one resolver (BundlerPathSettings / BundlerPathTiers::deciding), used by discovery, the global-path shadow check and bundler_sets_explicit_path.

Risk: low. The change is one file of crawler logic. I checked it line by line against Bundler 4.0.18's settings.rb:

• Tiers are local/env/global; the first tier that sets path, path.system or disable_shared_gems decides.
• system = path_system || disable_shared_gems == false.
• Deployment applies only when no tier decides.
• to_bool matches bundler_truthy.

Look here:

• ruby_crawler.rs:460-472:`` discovery builds the tiers and calls .deciding().
• ruby_crawler.rs:518-521:``` the global path now applies only when the global tier decides.
• ruby_crawler.rs:1317-1333:```` decides_path / `uses_system_gems`, the core rule.
• ruby_crawler.rs:1381-1449:````` deciding_of and the rewritten `bundler_sets_explicit_path`.
• ruby_crawler.rs:2424-2440:`````` the same-file path drop when disable_shared_gems is falsy (Bundler rejects that combination anyway).

Verified:

• cargo test -p socket-patch-core --lib ruby_crawler: 108 passed.
• --test crawler_ruby_e2e: 30 passed.
• clippy (lib): clean.
• Mutation check: removing the disable_shared_gems arm makes the 3 new Bundler's path tier rule is written twice in ruby_crawler, and store discovery ignores BUNDLE_DISABLE_SHARED_GEMS: "false" #1931 tests and 1 existing tier test fail.
• CRLF configs are handled, since lines() strips \r.
• CHANGELOG.md is untouched.
• CI: ci-ok and clippy are green on the head.

Changes I made: None.

Open questions: None that block merging. Small side effects the description doesn't mention, all consistent with Bundler:

• An env BUNDLE_PATH__SYSTEM or BUNDLE_DISABLE_SHARED_GEMS no longer drops the whole global config, so a global BUNDLE_GEMFILE is still honored.
• verification_only_gem_paths now passes the ambient env flags.

The env-BUNDLE_PATH probe that already exists on main, for a local tier that decides without a path, is deferred to #1935 as the description says.

Auto-merge is armed: approving sends this straight to the merge queue.

---

Generated by Claude Code

[Wenxin Jiang (Wenxin-Jiang)]

[agent] Reviewed fe207642712edbc2c7b0c810a718f22146afaba3. Checked Bundler local/env/global path tier precedence and system/shared-gem flags. No major correctness issues found. Validation: source and regression-test review plus the available exact-head CI checks; no local test pass is claimed.