[agent] Found by the scheduled Bundler (RubyGems) bug-hunt routine (ledger #316).
Summary
Bundler marks a dependency declared inside a source … do block with ! in the lock's DEPENDENCIES (colorize (= 0.8.1)!), even when that block's source is rubygems.org. The hosted scan of such a gem works: it nests the patch-registry block inside the user's block, and the fresh frozen install gets the patched bytes. The unwind (rollback, or remove <purl>) then restores the Gemfile byte for byte, so the declaration is back inside the user's source "https://rubygems.org" do block. But the unwind always drops the ! from the lock's DEPENDENCIES line. The restored pair no longer matches what Bundler writes, so BUNDLE_FROZEN=true bundle install fails with exit 16 ("Your lockfile needs to be updated, but it can't be because frozen mode is set"). rollback still reports success with hosted.reverted: [pkg:gem/colorize@0.8.1].
Impact
After undoing a patch, every CI or deployment (frozen) install of the project breaks until someone runs an unfrozen bundle install and commits the lock. The unwind is supposed to give back the original, installable pair.
Repro (Linux, Ruby 3.3.6, Bundler 4.0.22 or 2.6.9 with bundle lock --add-checksums)
Patch API and patch registry mocked on loopback (the run-13 mock from ledger #316); rubygems.org is the real upstream.
mkdir app && cd app
printf 'source "https://rubygems.org"\n\ngem "rake"\nsource "https://rubygems.org" do\n gem "colorize", "0.8.1"\nend\n' > Gemfile
bundle lock && cp Gemfile.lock /tmp/orig.lock # DEPENDENCIES: colorize (= 0.8.1)!
socket-patch scan --mode hosted --yes --api-url $MOCK --org org --api-token fake
BUNDLE_FROZEN=true BUNDLE_PATH=vb bundle install # exit 0, patched bytes (OK)
socket-patch rollback --api-url $MOCK --org org --api-token fake --patch-server-url $MOCK
diff /tmp/orig.lock Gemfile.lock
# < colorize (= 0.8.1)!
# ---
# > colorize (= 0.8.1)
git diff Gemfile # empty: the Gemfile came back exactly
BUNDLE_FROZEN=true BUNDLE_PATH=vb2 bundle install # exit 16
socket-patch remove pkg:gem/colorize@0.8.1 leaves the same diff, and a frozen install then fails with exit 16 too.
Expected vs actual
- Expected: CLI_CONTRACT.md's "Hosted unwind coverage" gem row says the unwind undoes the
source "<patch registry>" do … end block and "the pair comes back". The ! should only be dropped when the restored declaration no longer sits in a user source block (the case where hosted mode added it). Here the declaration's own block still pins the source, so the original ! must stay, and the lock should come back byte-identical to the pre-scan one.
- Actual: the
! is always removed (the row says "the DEPENDENCIES pin loses its !" without exception), which leaves a pair that Bundler's frozen mode rejects.
Matrix
| OS |
Ruby |
Bundler |
Shape |
Result |
| Linux |
3.3.6 |
4.0.22 |
top-level source + source "https://rubygems.org" do gem … end |
reproduces (×2) |
| Linux |
3.3.6 |
4.0.22 |
every gem inside one source "https://rubygems.org" do block |
reproduces |
| Linux |
3.3.6 |
4.0.22 |
source … do + nested group :default do |
reproduces |
| Linux |
3.3.6 |
2.6.9 (CHECKSUMS added) |
every gem inside one source block |
reproduces |
| Linux |
3.3.6 |
4.0.22 |
remove <purl> instead of rollback |
reproduces |
| Linux |
3.3.6 |
4.0.22 |
plain top-level declaration (control) |
pass (lock byte-restored) |
macOS and Windows weren't probed. The lock surgery is OS-independent.
First bad version
Not bisected. This is current main d47eab3; the v5 upstream restore introduced the hosted unwind.
Suspect code
crates/socket-patch-core/src/patch/redirect/upstream/gem.rs:354: for a non-transitive gem, entry.strip_suffix('!') unconditionally rewrites the DEPENDENCIES line to the unpinned form. It doesn't check whether the restored Gemfile still declares the gem inside a source … do block (or with a source: / git: / path: option), and in that case Bundler keeps the !.
Backlog review — 2026-10-08
Priority: P1 → P2. A rubygems source-block dependency loses its marker and frozen installation fails after rollback; visible conditional breakage.
[agent] Found by the scheduled Bundler (RubyGems) bug-hunt routine (ledger #316).
Summary
Bundler marks a dependency declared inside a
source … doblock with!in the lock'sDEPENDENCIES(colorize (= 0.8.1)!), even when that block's source is rubygems.org. The hosted scan of such a gem works: it nests the patch-registry block inside the user's block, and the fresh frozen install gets the patched bytes. The unwind (rollback, orremove <purl>) then restores the Gemfile byte for byte, so the declaration is back inside the user'ssource "https://rubygems.org" doblock. But the unwind always drops the!from the lock'sDEPENDENCIESline. The restored pair no longer matches what Bundler writes, soBUNDLE_FROZEN=true bundle installfails with exit 16 ("Your lockfile needs to be updated, but it can't be because frozen mode is set").rollbackstill reportssuccesswithhosted.reverted: [pkg:gem/colorize@0.8.1].Impact
After undoing a patch, every CI or deployment (frozen) install of the project breaks until someone runs an unfrozen
bundle installand commits the lock. The unwind is supposed to give back the original, installable pair.Repro (Linux, Ruby 3.3.6, Bundler 4.0.22 or 2.6.9 with
bundle lock --add-checksums)Patch API and patch registry mocked on loopback (the run-13 mock from ledger #316); rubygems.org is the real upstream.
socket-patch remove pkg:gem/colorize@0.8.1leaves the same diff, and a frozen install then fails with exit 16 too.Expected vs actual
source "<patch registry>" do … endblock and "the pair comes back". The!should only be dropped when the restored declaration no longer sits in a usersourceblock (the case where hosted mode added it). Here the declaration's own block still pins the source, so the original!must stay, and the lock should come back byte-identical to the pre-scan one.!is always removed (the row says "theDEPENDENCIESpin loses its!" without exception), which leaves a pair that Bundler's frozen mode rejects.Matrix
source "https://rubygems.org" do gem … endsource "https://rubygems.org" doblocksource … do+ nestedgroup :default dosourceblockremove <purl>instead ofrollbackmacOS and Windows weren't probed. The lock surgery is OS-independent.
First bad version
Not bisected. This is current main
d47eab3; the v5 upstream restore introduced the hosted unwind.Suspect code
crates/socket-patch-core/src/patch/redirect/upstream/gem.rs:354: for a non-transitive gem,entry.strip_suffix('!')unconditionally rewrites theDEPENDENCIESline to the unpinned form. It doesn't check whether the restored Gemfile still declares the gem inside asource … doblock (or with asource:/git:/path:option), and in that case Bundler keeps the!.Backlog review — 2026-10-08
Priority: P1 → P2. A rubygems source-block dependency loses its marker and frozen installation fails after rollback; visible conditional breakage.