The source generator model

The source generator model

A source generator is a Roslyn extension that runs DURING compilation: it reads the code being compiled and adds new syntax trees to the same compilation. The generated code compiles alongside yours — one assembly, no runtime reflection, no code after build.

[Generator]
public sealed class HelloGen : IIncrementalGenerator
{
    public void Initialize(IncrementalGeneratorInitializationContext ctx)
    {
        ctx.RegisterPostInitializationOutput(static init =>
            init.AddSource("Marker.g.cs",
                "public sealed class HelloAttribute : System.Attribute { }"));

        var targets = ctx.SyntaxProvider.ForAttributeWithMetadataName(
            "HelloAttribute",
            static (node, _) => node is Microsoft.CodeAnalysis.CSharp.Syntax.ClassDeclarationSyntax,
            static (ctx2, _) => ctx2.TargetSymbol.Name);

        ctx.RegisterSourceOutput(targets, static (spc, name) =>
            spc.AddSource($"{name}.g.cs",
                $"public static class Hello_{name} {{ public const string Who = \"{name}\"; }}"));
    }
}

What this changes architecturally: anything built with runtime reflection (DI registries, serializers, mappers, enum-to-string tables) can instead be computed once at compile time. That removes startup cost, becomes trim/AOT-safe (Module 19), and surfaces errors at build time.

What generators CANNOT do: modify existing user code (they only ADD sources), run arbitrary per-user logic from configuration files at dev time, or replace the build. They see the same compilation the compiler sees — nothing more.