[agent] Found by the scheduled uv bug-hunt routine (ledger #310).
Summary
After a hosted scan and then rollback (or remove), a uv-written PEP 751 pylock.toml doesn't come back as uv wrote it. The restored sdist / wheels entries carry upload-time with milliseconds (2021-05-05T14:18:17.237Z). uv writes pylock upload-time as a TOML datetime truncated to whole seconds (2021-05-05T14:18:17Z), and so do the lock's own untouched sibling entries. Every rolled-back pylock is left with a spurious diff that no uv command would produce.
The upstream restore formats upload-time with uv.lock's millisecond rule (upload_time() in crates/socket-patch-core/src/patch/redirect/upstream/uv.rs:394) for both lock kinds. For pylock it only switches the quoting (shape.datetime, :424-433), not the precision.
Impact
This is low severity: the restored file is valid and installs the pristine wheel (uv pip sync pylock.toml → original six.py). But "rollback" doesn't return the file to its pre-patch bytes. A repo that regenerates pylock.toml in CI (uv export --format pylock.toml) and checks git diff --exit-code fails after an otherwise clean rollback, and the leftover diff looks like a Socket edit that was never undone. The uv.lock and requirements.txt restores in the same run are byte-identical, so the gap is pylock-specific.
Repro (Linux)
Mock patch API as in #379 / #381, with --patch-server-url for the mock origin and the PyPI JSON API reachable.
SP="socket-patch --api-url http://127.0.0.1:18080 --api-token t --org test-org --patch-server-url http://127.0.0.1:18080"
mkdir p && cd p
printf '[project]\nname = "uvp"\nversion = "0.1.0"\nrequires-python = ">=3.9"\ndependencies = ["six==1.16.0", "idna==3.7"]\n' > pyproject.toml
uv lock && uv export --format pylock.toml -o pylock.toml
cp pylock.toml pylock.orig
$SP scan --mode hosted --json --yes # rewrittenFiles includes pylock.toml
$SP rollback --json --yes # success, hosted.reverted [pkg:pypi/[email protected]]
diff pylock.orig pylock.toml
# < sdist = { url = ".../six-1.16.0.tar.gz", upload-time = 2021-05-05T14:18:18Z, size = 34041, ... }
# > sdist = { url = ".../six-1.16.0.tar.gz", upload-time = 2021-05-05T14:18:18.379Z, size = 34041, ... }
# < wheels = [{ url = ".../six-1.16.0-py2.py3-none-any.whl", upload-time = 2021-05-05T14:18:17Z, ... }]
# > wheels = [{ url = ".../six-1.16.0-py2.py3-none-any.whl", upload-time = 2021-05-05T14:18:17.237Z, ... }]
The same diff appears when pylock.toml sits next to uv.lock and an exported requirements.txt (those two restore byte-identically). uv 0.8.17, uv 0.12.21, and uv pip compile --format pylock.toml all write upload-time = 2021-05-05T14:18:17Z for that wheel.
Expected vs actual
- Expected: CLI_CONTRACT.md, "Hosted unwind coverage": rollback restores each hosted pin "to its default upstream registry entry", with the artifact fields in the shape "the lock's other registry packages" show. Here the siblings show second precision, and uv only ever writes that precision in pylock files. The restored entry should be what uv would write, as it already is for
uv.lock.
- Actual: millisecond
upload-time values in the restored pylock entry.
OS × uv matrix (main 2463257)
| OS |
uv 0.8.17 |
uv 0.12.21 |
| Linux |
fail |
fail (reproduced 2×) |
macOS and Windows weren't probed. The defect is in platform-independent string formatting.
First bad
2463257 (#277, v5 upstream restore). Release 4.0.0 restored pylock from a recorded fragment.
Suspect code
crates/socket-patch-core/src/patch/redirect/upstream/uv.rs:394 (upload_time, millisecond rule) and :424-433 (the shape.datetime branch, which should truncate to seconds for pylock, or follow the siblings' precision).
Related: #407 (a uv pip compile pylock can't be rolled back at all).
[agent] Found by the scheduled uv bug-hunt routine (ledger #310).
Summary
After a hosted
scanand thenrollback(orremove), a uv-written PEP 751pylock.tomldoesn't come back as uv wrote it. The restoredsdist/wheelsentries carryupload-timewith milliseconds (2021-05-05T14:18:17.237Z). uv writes pylockupload-timeas a TOML datetime truncated to whole seconds (2021-05-05T14:18:17Z), and so do the lock's own untouched sibling entries. Every rolled-back pylock is left with a spurious diff that no uv command would produce.The upstream restore formats
upload-timewith uv.lock's millisecond rule (upload_time()incrates/socket-patch-core/src/patch/redirect/upstream/uv.rs:394) for both lock kinds. For pylock it only switches the quoting (shape.datetime,:424-433), not the precision.Impact
This is low severity: the restored file is valid and installs the pristine wheel (
uv pip sync pylock.toml→ originalsix.py). But "rollback" doesn't return the file to its pre-patch bytes. A repo that regeneratespylock.tomlin CI (uv export --format pylock.toml) and checksgit diff --exit-codefails after an otherwise clean rollback, and the leftover diff looks like a Socket edit that was never undone. Theuv.lockandrequirements.txtrestores in the same run are byte-identical, so the gap is pylock-specific.Repro (Linux)
Mock patch API as in #379 / #381, with
--patch-server-urlfor the mock origin and the PyPI JSON API reachable.The same diff appears when
pylock.tomlsits next touv.lockand an exportedrequirements.txt(those two restore byte-identically). uv 0.8.17, uv 0.12.21, anduv pip compile --format pylock.tomlall writeupload-time = 2021-05-05T14:18:17Zfor that wheel.Expected vs actual
uv.lock.upload-timevalues in the restored pylock entry.OS × uv matrix (main
2463257)macOS and Windows weren't probed. The defect is in platform-independent string formatting.
First bad
2463257(#277, v5 upstream restore). Release 4.0.0 restored pylock from a recorded fragment.Suspect code
crates/socket-patch-core/src/patch/redirect/upstream/uv.rs:394(upload_time, millisecond rule) and:424-433(theshape.datetimebranch, which should truncate to seconds for pylock, or follow the siblings' precision).Related: #407 (a
uv pip compilepylock can't be rolled back at all).