NuGet: references to assemblies in runtimes folder not added

This is what I've finally figured out/guessed (because as best I can tell there's no official documentation for some of this)

• Files added to the /runtimes folder aren't automatically added as references to the target project.
• The /ref and /runtime folder should be used in conjunction with each other and only for the .NET Core target. As best I can .NET Framework targets apparently don't support these folders.
• The /ref folder is for compile time references and anything added here will be added as a reference to the target project.
• Assemblies in the /ref folder don't need to have an implementation - every public API could just throw a not implemented exception. In practice however you typically just take a copy of one of the implementation assemblies and declare it as the compile time API.
• I've read (but haven't tested myself) that assemblies in the /ref folder must be "Any CPU" builds. You can use CorFlags utility to patch an implementation assembly for this if necessary.
• The /runtimes folder is used to provide an implementation assemblies for any references included in the /ref folder. These assemblies are used at runtime and during deployment.
• The /runtimes folder can include additional assemblies that are only required at runtime and don't need to be seen by the client project. These additional assemblies won't be included as references in the target project but will be available for run/deployment.
• As mentioned by others, the files in the /runtimes folder aren't copied to the output folder of the build. Instead config files are placed there that tell the runtime how to locate the /runtimes files from the NuGet cache.
• For .NET Framework targets (ie: net461) just use the /lib folder as there's no other runtimes for .NET aside from Windows anyway.

Putting this all together, my original example, should have looked like this:

/lib/net461/myassembly.dll                       (net461/Windows Compile and Runtime)
/runtimes/osx/lib/netcoreapp2.0/myassembly.dll   (netcore/OSX Runtime)
/runtimes/win/lib/netcoreapp2.0/myassembly.dll   (netcore/Win Runtime)
/ref/netcoreapp2.0/myassembly.dll                (netcore/* Compile Time)

I spent a fair amount of time trying your project on OSX in both Visual Studio for Mac and VS Code. I'll try to stick with factual observations without getting into "why don't you do X instead".

• The runtimes/{rid}/lib/{tfm}/*.dll paths look ok
• target="lib/{tfm}/..." assemblies are automatically referenced, runtimes/... are not
• Using target framework of netstandard seems like it would make your package work in both netcoreapp and netstandard projects (e.g. use target="lib/netstandard1.6/..."). Compare with this
• runtimes/ seems to be intended for platform-dependent assemblies you'll load at runtime. For example, 32/64-bit native assemblies in runtimes/win-x64/native/ and runtimes/win-x86/native/) loaded with AssemblyLoadContext (another post by McMaster)
• Using separate slns for Windows and OSX, or separate platform-specific projects that reference platform-agnostic projects (like Xamarin) would obviate some of the configuration wrangling
• I found no documentation on target="ref/...", but you can add Explicit Assembly <references> (inside the nuspec <metadata> block)
• Packaged assemblies won't appear in the output directory, but when prepared for distribution with dotnet publish they'll be included:
  

.NET Core and .NETSTANDARD don't copy dependencies to output directory, they are mapped using deps.json which points to relative paths from local NuGet cache.

This has been a very useful thread to get more information and hints on how to create a NuGet package that references native DLLs, and is consumed in both .NET Framework as well as .NET Core / modern .NET libraries / applications.

My experience so far has been that if this library (let's call it library A) only targets .NET Standard, consuming this library in a .NET Core / 5.0 or 6.0 application does lead to the native assemblies being pulled in correctly from the runtimes folder. In a .NET Framework 4.7 application however, this does not appear to be the case. Unless the runtime is explicitly specified when compiling, e.g.:

    dotnet build ... --runtime win-x86

When using library A in a .NET Core or .NET 5.0/6.0 application however, this runtime identifier is not required - all runtimes are made available and the right one is selected at runtime.

If you want library A to be consumed in applications that also target .NET Framework, and you don't want the user to have to specify the runtime explicitly, then it seems to be necessary to:

• Target both .NET Standard and .NET Framework

• Ensure that the native assemblies end up in the following folder structure in the NuGet package:

    "lib/{tfm}/..."
  

While the .NET documentation referenced by tm1 earlier here talks about how to get this to work using nuspec files, it is less clear how to do so in the SDK .csproj format. I managed to do this in the NLoptNet project, see the relevant .csproj file here. Final relevant point (in addition to the two bullets above):

• Use "<None Include" rather than "<None Update" to add the native assemblies

So far - this works, but there is one quirk - as you can see here it generates warning MSB3246 when consuming library A in a .NET Framework application. See also this Stack Overflow post. This leads me to believe that maybe the above is not the right approach, and therefore to some questions:

• Is this the intended way to consume library A in .NET Core, modern .NET and .NET Framework applications?
• Should one always specify the runtime identifier when using dotnet build / dotnet test?

Can you try to target .NET Standard 2.0 instead of net461 and netcoreapp2.0? Libraries built against netstandard2.0 should work with .NET Core 2.0 and .NET Framework 4.6.1: https://docs.microsoft.com/en-us/dotnet/standard/net-standard

Are you using the new csproj format? If so it has built in support for multiple target frameworks.

For example running dotnet pack against a .csproj file with this content:

<Project Sdk="Microsoft.NET.Sdk">    
  <PropertyGroup>
    <TargetFrameworks>net461;netcoreapp2.1;netstandard2.0</TargetFrameworks>
  </PropertyGroup>    
</Project>

will result in a .nupkg that works for .NET Framework 4.6.1, .NET Core 2.1 and .NET Standard 2.0.

Various trick can then be used to include specific parts for each framework depending on what's available.

I'm trying to solve the same problem. The solution proposed by you works fine, but there is one question ... The case of Win and net46 is clear. And now I need to add a reference to the assembly in the project for a netcoreapp for the Win and for Linux. The problem is that this is a DIFFERENT assembly with the SAME name. Those my package looks like this:

/lib/net461/myassembly1.dll                       (net461/Windows Compile and Runtime)
/runtimes/ubuntu/lib/netcoreapp2.0/myassembly2.dll   (netcore/Ubuntu Runtime)
/runtimes/win/lib/netcoreapp2.0/myassembly1.dll   (netcore/Win Runtime)
/ref/netcoreapp2.0/???

Update: Actually, the myassembly1.dll and myassembly2.dll are both called myassembly.dll. But to show that one is assembled for Windows, and the second one for Linux, I will leave here such a name.

The most interesting thing is that I tried to put any assembly in the folder ref, and it works on both Windows and Linux. This version works on both systems

/lib/net461/myassembly1.dll   
/runtimes/ubuntu/lib/netcoreapp2.0/myassembly2.dll 
/runtimes/win/lib/netcoreapp2.0/myassembly1.dll  
/ref/netcoreapp2.0/myassembly1.dll

And this too

/lib/net461/myassembly1.dll   
/runtimes/ubuntu/lib/netcoreapp2.0/myassembly2.dll 
/runtimes/win/lib/netcoreapp2.0/myassembly1.dll  
/ref/netcoreapp2.0/myassembly2.dll

But I think this is not right and I was wrong somewhere.