Publish with `--no-build --no-restore` fails when using AOT in .NET 10

[hojmark]

Description

Separate restore/build and publish steps fails when using AOT in .NET 10. This worked in .NET 9 and is thus a regression.

I have prepared a minimal example at https://github.com/hojmark/dotnet-publish-aot-regression with two branches:

• main (using net10.0 TFM)
• dotnet9 (using net9.0 TFM)

The TFM in the project file is the only difference between the two branches.

Furthermore there's a GitHub workflow that illustrates the issue. The same workflow is successful when run on the dotnet9 branch while it fails on the main branch.

Successful run (.NET 9)

Run dotnet build
  Determining projects to restore...
  Restored /home/runner/work/dotnet-publish-aot-regression/dotnet-publish-aot-regression/dotnet-publish-aot-regression.csproj (in 1.83 sec).
  dotnet-publish-aot-regression -> /home/runner/work/dotnet-publish-aot-regression/dotnet-publish-aot-regression/bin/Debug/net9.0/dotnet-publish-aot-regression.dll

Build succeeded.
    0 Warning(s)
    0 Error(s)

Time Elapsed 00:00:08.48

Run dotnet publish --runtime linux-x64 --self-contained --no-restore --no-build
  Generating native code
  dotnet-publish-aot-regression -> /home/runner/work/dotnet-publish-aot-regression/dotnet-publish-aot-regression/bin/Release/net9.0/linux-x64/publish/

Failed run (.NET 10)

Run dotnet build
  Determining projects to restore...
  Restored /home/runner/work/dotnet-publish-aot-regression/dotnet-publish-aot-regression/dotnet-publish-aot-regression.csproj (in 1.74 sec).
  dotnet-publish-aot-regression -> /home/runner/work/dotnet-publish-aot-regression/dotnet-publish-aot-regression/bin/Debug/net10.0/dotnet-publish-aot-regression.dll

Build succeeded.
    0 Warning(s)
    0 Error(s)

Time Elapsed 00:00:09.01

Run dotnet publish --runtime linux-x64 --self-contained --no-restore --no-build
/home/runner/.nuget/packages/microsoft.dotnet.ilcompiler/10.0.0/build/Microsoft.NETCore.Native.Publish.targets(70,5): error : The PrivateSdkAssemblies ItemGroup is required for _ComputeAssembliesToCompileToNative [/home/runner/work/dotnet-publish-aot-regression/dotnet-publish-aot-regression/dotnet-publish-aot-regression.csproj]

Reproduction Steps

Create dotnet-publish-aot-regression.csproj

<Project Sdk="Microsoft.NET.Sdk">

  <PropertyGroup>
    <OutputType>Exe</OutputType>
    <TargetFramework>net10.0</TargetFramework>
    <PublishAot>true</PublishAot>
    <RuntimeIdentifiers>linux-x64</RuntimeIdentifiers>
  </PropertyGroup>

</Project>

Create Program.cs

System.Console.WriteLine("Hello, World!");

Build (and implicitly restore)

dotnet build

Publish

dotnet publish --runtime linux-x64 --self-contained --no-restore --no-build

Expected behavior

The project is published.

Actual behavior

The project fails to publish with the following error:

/home/runner/.nuget/packages/microsoft.dotnet.ilcompiler/10.0.0/build/Microsoft.NETCore.Native.Publish.targets(70,5): error : The PrivateSdkAssemblies ItemGroup is required for _ComputeAssembliesToCompileToNative [/home/runner/work/dotnet-publish-aot-regression/dotnet-publish-aot-regression/dotnet-publish-aot-regression.csproj]

Regression?

Worked in .NET 9.

Known Workarounds

Skip the --no-restore --no-build flags on the publish command. This avoids the failure, but it also makes separating the build steps (and the flags themselves) pointless.

Configuration

.NET 10.0.100
Ubuntu 24.04.3 x64 (GitHub runner)
Fedora 43 x64 (local workstation)

Other information

Peculiarity
On GitHub’s Ubuntu runners, the publish step consistently fails when using the --no-restore --no-build flags. On my local workstation (Fedora), if I perform a single successful publish without these flags, any subsequent publishes with the flags succeed. This remains true even after deleting the entire project directory and re-cloning the repository. I have verified this on two separate Fedora workstations.

My suspicion is that Ubuntu would behave similarly if state were preserved, but GitHub runners of course provide a clean environment each time.

This suggests that something outside the project directory is being modified after the first successful publish, though I have absolutely no idea what that might be?

Or maybe I’ve just made two faulty observations. :)

[MihuBot]

I'm a bot. Here is a possible related and/or duplicate issue (I may be wrong):

• Instructions for using locally built NativeAOT builds are broken runtime#119166

[hojmark]

This smells related: #46611

I should probably have created this issue in the sdk repo. Maybe someone can move it.

[hojmark]

Just saw this few hour old PR which seems very related: #51765

Inspired by the description I tried to add <SelfContained>true</SelfContained> to my project, and publish with --no-restore --no-build started working!

Perhaps @baronfel can confirm whether his PR is the fix for this issue?

[baronfel]

@hojmark hi! I do strongly suspect that my PR and the issue it links to are the same thing you're reporting here (as well as many other odd corner cases related to separate restore and our more advanced publish methods). In the meantime I'd suggest you use PublishSelfContained instead of SelfContained, though - doing selfcontained on normal builds has negative impacts on dev inner loop experiences (longer builds for simple repeated compilations, etc). Save the self contained stuff for publish time is my advice :)

[hojmark]

Thanks for the info. Unfortunately, using PublishSelfContained results in the same error.

[baronfel]

Good call-out - I'll make sure that PR has an end-to-end test to verify that PublishSelfContained comes online too.