[agent] Found by the scheduled Composer bug-hunt routine (ledger #321).
Summary
When a vendored Composer unit's uuid dir (.socket/vendor/composer/<uuid>/) is missing, for example because a checkout lost it or someone deleted it, socket-patch repair and a re-run of vendor download the patched copy again. Neither writes back the unit's socket-patch.vendor.json marker. The rebuild then reports success, but git status shows the marker as deleted. Committing that "repair" leaves the vendored unit without the marker that CLI_CONTRACT.md says every unit carries. Later vendor and repair runs (already_vendored, success []) never put it back.
The Cargo backend handles the same rebuild path correctly: it restores the marker when it's missing (crates/socket-patch-core/src/vendor/cargo.rs:928-940, pinned by test_wired_deleted_uuid_dir_rebuild_restores_marker). The Composer artifact-only rebuild branch doesn't.
Impact
Low. The install itself is fine: composer install from the repaired tree installs the patched bytes, and vex never reads the marker. But the repair leaves the committed vendored unit incomplete. External tools that rely on the marker ("belt-and-braces for tools that have the tree but not the lockfile") lose it, and the repair diff a user is asked to commit contains a surprise deletion.
Repro (Linux, Composer 1.10.28 and 2.10.3, PHP 8.3, main 61cfb9b)
The patch service is a local mock: POST /v0/orgs/acme/patches/package returns a granted result with an sha512 SRI, and GET serves a single-top-dir dist zip of the patched package. A manifest plus blob is staged under .socket/ for pkg:composer/acme/tool@1.0.0, from an inline package repository with a file:// zip dist.
composer update -n # installs acme/tool 1.0.0
printf '/vendor/\n' > .gitignore
git init -q && git add -A && git commit -qm init
socket-patch vendor --json --cwd . --api-url http://127.0.0.1:8765 --api-token fake --org acme
git add -A && git commit -qm vendored
git ls-files .socket/vendor | grep vendor.json
# .socket/vendor/composer/4d5e6f7a-…/socket-patch.vendor.json <- committed
rm -rf .socket/vendor/composer/4d5e6f7a-8b9c-4a1b-8c2d-0123456789ab
socket-patch repair --cwd . --api-url http://127.0.0.1:8765 --api-token fake --org acme
# Redownloaded pkg:composer/acme/tool@1.0.0 (.socket/vendor/composer/4d5e6f7a-…/acme/tool@1.0.0)
# Repair complete.
git status --short
# D .socket/vendor/composer/4d5e6f7a-8b9c-4a1b-8c2d-0123456789ab/socket-patch.vendor.json
socket-patch vendor --json --cwd . … # -> already_vendored
socket-patch repair --json --cwd . … # -> success, events: []
git status --short # marker still deleted
A vendor re-run instead of repair takes the same path. The JSON shows rebuilt with reason: vendor_artifact_missing, redownloaded: true, and the marker is still missing afterwards.
Expected vs actual
- Expected: CLI_CONTRACT.md ("Each vendored unit also carries an informational
socket-patch.vendor.json marker … next to the artifact"). repair "redownload[s] missing/corrupt vendored artifacts". The doc comment on write_marker_or_warn (crates/socket-patch-core/src/vendor/state.rs:808-812) says "Every backend's fresh and rebuild paths report it through here". After the rebuild the unit should be complete again (copy + marker), as it is for Cargo.
- Actual: only the copy under
<uuid>/acme/tool@1.0.0/ is recreated. The marker stays deleted for good.
Matrix
| OS |
Composer |
Reproduces |
| Linux |
1.10.28 (PHP 8.3) |
yes |
| Linux |
2.10.3 (PHP 8.3) |
yes (×2) |
| macOS / Windows |
n/a |
untested. The code path is pure Rust filesystem logic, independent of OS and Composer version. |
First bad version: not a regression. v4.0.0 behaves the same way (rebuilt, marker left deleted).
Suspect code
crates/socket-patch-core/src/vendor/composer_lock.rs:311-345: the "Wired but the committed copy is missing/stale: rebuild the ARTIFACT only" branch calls composer_service_copy and mirror_filters::heal_or_warn, then returns. It never calls write_marker_or_warn, which only runs on the fresh path (composer_lock.rs:437-439). Porting Cargo's "restore the marker only when it is missing" block (cargo.rs:928-940) into this branch should fix it. test_wired_missing_copy_rebuilds_artifact_only (composer_lock.rs:1643) could assert the marker as well.
No probe runs: this is OS-independent.
[agent] Found by the scheduled Composer bug-hunt routine (ledger #321).
Summary
When a vendored Composer unit's uuid dir (
.socket/vendor/composer/<uuid>/) is missing, for example because a checkout lost it or someone deleted it,socket-patch repairand a re-run ofvendordownload the patched copy again. Neither writes back the unit'ssocket-patch.vendor.jsonmarker. The rebuild then reports success, butgit statusshows the marker as deleted. Committing that "repair" leaves the vendored unit without the marker that CLI_CONTRACT.md says every unit carries. Latervendorandrepairruns (already_vendored,success []) never put it back.The Cargo backend handles the same rebuild path correctly: it restores the marker when it's missing (
crates/socket-patch-core/src/vendor/cargo.rs:928-940, pinned bytest_wired_deleted_uuid_dir_rebuild_restores_marker). The Composer artifact-only rebuild branch doesn't.Impact
Low. The install itself is fine:
composer installfrom the repaired tree installs the patched bytes, andvexnever reads the marker. But the repair leaves the committed vendored unit incomplete. External tools that rely on the marker ("belt-and-braces for tools that have the tree but not the lockfile") lose it, and therepairdiff a user is asked to commit contains a surprise deletion.Repro (Linux, Composer 1.10.28 and 2.10.3, PHP 8.3, main
61cfb9b)The patch service is a local mock:
POST /v0/orgs/acme/patches/packagereturns agrantedresult with an sha512 SRI, andGETserves a single-top-dir dist zip of the patched package. A manifest plus blob is staged under.socket/forpkg:composer/acme/tool@1.0.0, from an inlinepackagerepository with afile://zip dist.A
vendorre-run instead ofrepairtakes the same path. The JSON showsrebuiltwithreason: vendor_artifact_missing, redownloaded: true, and the marker is still missing afterwards.Expected vs actual
socket-patch.vendor.jsonmarker … next to the artifact").repair"redownload[s] missing/corrupt vendored artifacts". The doc comment onwrite_marker_or_warn(crates/socket-patch-core/src/vendor/state.rs:808-812) says "Every backend's fresh and rebuild paths report it through here". After the rebuild the unit should be complete again (copy + marker), as it is for Cargo.<uuid>/acme/tool@1.0.0/is recreated. The marker stays deleted for good.Matrix
First bad version: not a regression. v4.0.0 behaves the same way (
rebuilt, marker left deleted).Suspect code
crates/socket-patch-core/src/vendor/composer_lock.rs:311-345: the "Wired but the committed copy is missing/stale: rebuild the ARTIFACT only" branch callscomposer_service_copyandmirror_filters::heal_or_warn, then returns. It never callswrite_marker_or_warn, which only runs on the fresh path (composer_lock.rs:437-439). Porting Cargo's "restore the marker only when it is missing" block (cargo.rs:928-940) into this branch should fix it.test_wired_missing_copy_rebuilds_artifact_only(composer_lock.rs:1643) could assert the marker as well.No probe runs: this is OS-independent.