Repository navigation
Description
[agent] Found by the scheduled Bundler (RubyGems) bug-hunt routine (ledger #316).
Summary
When vendored mode rewires a gem declaration in a CRLF Gemfile, the new gem "x", "<ver>", path: "…" line is written with a bare \n. Every other line keeps \r\n. So after socket-patch vendor the Gemfile has mixed line endings:
source "https://rubygems.org"\r\n
\r\n
gem "rack", "3.2.7", path: ".socket/vendor/gem/<uuid>/rack-3.2.7"\n <- was "...\"~> 3.1\"\r\n"
vendor --revert, rollback and remove restore the recorded original line (with its \r), so the round trip is byte-exact. The vendored state in between is the problem. Bundler parses the file fine, so installs and bundle exec aren't affected.
This is separate from #1808, which is about the CRLF Gemfile.lock (the lock surgery). Its fix PR #1811 only touches the lock path and doesn't change the Gemfile line.
Impact
- A project that commits its
Gemfileas CRLF (.gitattributes-textoreol=crlfwith a non-normalizing checkout, or Windows editors withautocrlf=false) commits a mixed-endingGemfileafter vendoring. The diff shows an EOL change on top of the real edit, and RuboCopLayout/EndOfLine(crlf/nativeon Windows) flags the line. - This is inconsistent with hosted mode, which keeps a CRLF
GemfileCRLF (gemfile_rerun_on_crlf_checkout_never_nests), and with the CRLF-preservation fixes the vendored backends got for other manifests (Vendored pnpm rewrites a CRLF package.json as LF, and vendor --revert does not restore it #662 pnpmpackage.json, Vendored requirements.txt rewrite re-joins every line of a mixed CRLF/LF file with one terminator, so untouched lines change andvendor --revert/removecan't restore the file #1567 requirements.txt).
Severity is low: the file still works. It's a correctness and hygiene defect in a user-owned manifest.
Repro
This is a scratch addition to gem_vendor_fresh_checkout_bundle_install_and_revert in crates/socket-patch-cli/tests/e2e_vendor_gem_build.rs, with a real bundle install from rubygems.org and the patch staged offline with stage_patch_with_vuln:
Gemfile (CRLF): source "https://rubygems.org"\r\n\r\ngem "rack", "~> 3.1"\r\n
bundle config set --local path vendor/bundle && bundle install # Gemfile.lock stays LF
# stage a lib/rack.rb patch for pkg:gem/rack@3.2.7
socket-patch vendor --json --offline # exit 0, applied 1
ruby -e 'p File.binread("Gemfile")'
# => "source \"https://rubygems.org\"\r\n\r\ngem \"rack\", \"3.2.7\", path: \".socket/vendor/gem/<uuid>/rack-3.2.7\"\n"
socket-patch vendor --revert --json # (or rollback / remove <purl>) -> Gemfile byte-exact CRLF again
Expected vs actual
- Expected: a rewritten line keeps the line ending of the line it replaces (or
utils::line_endings::terminatorof the file), as hosted mode does for the sameGemfile. docs/ecosystems.md lists the vendoredGemfile+Gemfile.lockpair edit as supported, with no line-ending caveat, and Pick the line terminator for spliced lines through one line_endings::terminator #815 / Pick inserted line terminators through line_endings::terminator (#815) #1108 set one terminator rule for lines a writer inserts. - Actual: the rewired line ends in
\ninside a CRLF file.
Matrix (Linux, Ruby 3.3.6; the edit is pure text handling, so it doesn't depend on the OS)
| Case | Bundler 2.6.9 | Bundler 4.0.23 |
|---|---|---|
vendor on CRLF Gemfile + LF lock: vendored Gemfile has mixed endings |
fail (×3) | fail (×3) |
… then vendor --revert / rollback / remove: Gemfile and lock byte-exact |
pass | pass |
LF Gemfile (control) |
pass | pass |
macOS / Windows probes are unavailable (push-triggered probe runs don't start, see #316). Tested on main ecd4924 (v5.0.1).
Suspect code
crates/socket-patch-core/src/vendor/gem.rs:1346-1360(plan_gemfile_edit):new_lineis built from scratch (format!("gem {q}{name}{q}, …")) whileoriginal_lineislines[idx]fromtext.split('\n'), so it still carries the\r. The new line drops it unless a trailing comment happens to carry it.apply_gemfile_plan(gem.rs:1610-1633) then joins with\n.- The
RewireOursarm keeps the\r(replacenon the original), and theAppendarm (transitive gem,GemfilePlan::Append { block }) appends an LF-only block. I didn't run theAppendarm here. - A related shape I saw (not filed separately): after a vendored LF pair is re-checked-out as CRLF,
restore_gemfile_line(gem.rs:2403) writes back the recorded LF original into the CRLF file. With Fix vendored gems on CRLF Gemfile.lock (#1808) #1811 applied,vendor --reverton such a checkout then leaves theGemfilemixed as well (…\r\n\r\ngem "rack", "~> 3.1"\n).
Activity
mikolalysenko commentedon Oct 29, 2026
[agent] Triage:
priority:p1(Bundler). Not a duplicate of #1808: that one is the Gemfile.lock surgery (PR #1811), this is the Gemfile line writer. Confirmed on mainecd4924:plan_gemfile_edit(crates/socket-patch-core/src/vendor/gem.rs) builds theRewritenew_linefrom scratch whileoriginal_linekeeps its\r, and theAppendmanaged block is LF-only, so both arms write LF lines into a CRLF Gemfile.Generated by Claude Code
mikolalysenko commentedon Oct 29, 2026
[agent] Claiming this issue (shared root cause: vendored Gemfile writer builds new/appended lines with LF regardless of the file's line endings). Branch: agent/fix-vendored-gemfile-crlf-line. Claim-ID: 2026-10-29T09:20:55Z-e4d7ea
Generated by Claude Code
mikolalysenko commentedon Oct 29, 2026
[agent] Draft fix PR: #1822
Generated by Claude Code
Keep a CRLF Gemfile CRLF when vendoring gems
Test vendoring a CRLF Gemfile end to end
Keep old vendored Gemfile lines CRLF on revert