Hosted NuGet remove / rollback / takeover restore packages.lock.json to the nuget.org catalog packageHash, which isn't NuGet's contentHash for signed packages, so every later dotnet restore fails NU1403

[mikolalysenko]

[agent] Found by the scheduled NuGet / dotnet bug-hunt routine (ledger #320).

Summary

The v5.0 hosted unwind (remove, rollback, and the hosted → vendored takeover that vendor --revert later lands on) restores every packages.lock.json entry of the patched id to "nuget.org's contentHash". It reads that value from the nuget.org catalog packageHash (UpstreamClient::nuget_content_hash, crates/socket-patch-core/src/patch/redirect/upstream/client.rs:498). The two hashes are different things:

• The catalog packageHash is the SHA-512 of the whole .nupkg file as served, including .signature.p7s.
• NuGet's lock contentHash (and .nupkg.metadata) is the package's content hash. For a signed package, NuGet computes it with the signature excluded (PackageArchiveReader.GetContentHash → SignedPackageArchiveUtility).

nuget.org has repository-signed every package since 2018, so for practically every real package the restored lock pins a hash that no download will ever match. Every later dotnet restore, locked or not, fails NU1403: Package content hash validation failed … The package is different than the last restore.

Live values for Newtonsoft.Json 13.0.3:

value	hash
lock contentHash written by dotnet restore (correct, original)	HrC5BXdl00IP9zeV+0Z848QWPAoCr9P3bDEZguI+gkLcBKAOxix/tLEAAHC+UvDNPv4a2d18lOReHMOagPa+zQ==
catalog packageHash (catalog0/data/2023.03.08.07.46.17/newtonsoft.json.13.0.3.json)	mbJSvHfRxfX3tR/U6n1WU+mWHXswYc+SB/hkOpx8yZZe68hNZGfymJu0cjsaJEkVzCMqePiU6LdIyogqfIn7kg==
sha512(base64) of the flat-container .nupkg file	mbJSvH…kg== (equals the catalog value)
lock after socket-patch remove / rollback	mbJSvH…kg== ❌

The unit tests in upstream/nuget.rs mock the catalog with the same UPSTREAM hash the lock fixture uses, so they can't catch this mismatch.

Impact

Undoing a hosted NuGet patch breaks the project's restore: the command reports success / hosted_reverted, nuget.config comes back byte-exact, and then dotnet restore fails NU1403 on every machine until someone deletes or regenerates the lock by hand. The same bad hash becomes the vendor ledger's recorded original on a hosted → vendored takeover, so vendor --revert later writes it back too.

Repro (Linux, .NET SDK 8.0.131, main 045d7ec, real api.nuget.org)

This uses the Backend stand-in and fixture from crates/socket-patch-cli/tests/e2e_nuget_dotnet_build.rs: app.csproj with RestorePackagesWithLockFile, a nuget.org-only nuget.config, and a real dotnet restore of Newtonsoft.Json 13.0.3.

cp packages.lock.json lock.orig
socket-patch scan --mode hosted --json --yes --api-url $URI --org test-org \
  --api-token fake-token --patch-server-url $URI           # redirected: 1
# lockfile discovery only recognises production feed URLs, so point the
# stand-in's source at patch.socket.dev (what a real hosted run writes):
sed -i "s#$URI#https://patch.socket.dev#" nuget.config

socket-patch remove pkg:nuget/newtonsoft.json@13.0.3 --json
# status success, events: [{"action":"removed","errorCode":"hosted_reverted",
#   "reason":"hosted lockfile pin restored to the upstream registry on remove"}]
cmp nuget.config <original config>     # byte-exact
diff lock.orig packages.lock.json
# <  "contentHash": "HrC5BXdl…a+zQ=="
# >  "contentHash": "mbJSvHfR…n7kg=="
# (and a trailing newline is added)

# fresh checkout, cold global packages folder:
dotnet restore                 # rc 1: error NU1403 … Newtonsoft.Json.13.0.3
dotnet restore --locked-mode   # rc 1: error NU1403

The same result (lock = mbJSv…, NU1403) also comes from:

• rollback pkg:nuget/newtonsoft.json@13.0.3: hosted.reverted: [purl], editedFiles: 2, status: success.
• hosted project → vendor --vendor-source service … (takeover: lock re-pinned to the vendored nupkg) → vendor --revert: status: success, nuget.config unwired, lock = mbJSv….

Reproduced 5 times (remove with an LF, a CRLF and a BOM nuget.config; rollback; takeover → vendor --revert).

Expected vs actual

• Expected: CLI_CONTRACT.md says remove "a hosted-only match restores its upstream registry entry", and "Takeover reconciliation" says "vendor --revert lands back on upstream registry state". SOCKET_NUGET_URL documents that the restored contentHash "comes from" the catalog packageHash, but the restored entry has to be one NuGet accepts: the original contentHash, which dotnet restore would write.
• Actual: a hash NuGet never produces for a signed package, so restore fails NU1403.

Possible fixes: compute the content hash the way NuGet does (download the nupkg, drop .signature.p7s per SignedPackageArchiveUtility, then hash). Or, when the catalog says the package is signed, refuse with redirect_revert_failed and point to git checkout -- packages.lock.json instead of writing a wrong pin.

Matrix

OS	SDK	path	result
Linux	8.0.131	hosted remove (LF / CRLF / BOM config)	fail NU1403
Linux	8.0.131	hosted rollback <purl>	fail NU1403 (locked and plain restore)
Linux	8.0.131	hosted → vendor takeover → vendor --revert	fail (lock = catalog hash)
macOS / Windows, SDK 6/9/10	—	—	not run. The hash comes from nuget.org and NuGet's content-hash algorithm is the same on every OS and SDK, so the mismatch is platform-independent.

First bad version

Release v4.0.0 isn't affected: there, remove / rollback on a manifest-less hosted project exits manifest_not_found, so there's no hosted unwind. The NuGet unwind arrived with de316b4 (#358, 2026-10-01), which added crates/socket-patch-core/src/patch/redirect/upstream/nuget.rs.

Suspect code

• crates/socket-patch-core/src/patch/redirect/upstream/client.rs:498-552: nuget_content_hash returns the catalog packageHash.
• crates/socket-patch-core/src/patch/redirect/upstream/nuget.rs:301: writes that value into every lock entry of the id.
• crates/socket-patch-core/src/patch/redirect/upstream/nuget.rs:13-14 (module doc), and SOCKET_NUGET_URL in CLI_CONTRACT.md:1040, which describe the same assumption.

[mikolalysenko]

[agent] Triaged as priority:p3 (NuGet). Not a duplicate: #593 is about rewriting entries at other versions, and this is the hash source (catalog packageHash vs the signature-excluded contentHash) used by the hosted unwind. No open PR covers it yet.

---

Generated by Claude Code