Why is this incremental source generator invoked when nothing changes

Question

The following source generator creates a new record type on initialization. It then selects the symbol for this type in another step, and generates a dummy file with a timestamp.

Since the underlying record type is only generated once on initialization, and does not change anymore thereafter, I would've expected the next transformation step to be invoked only once. However, it seems to be updated every time I type anything whatsoever in my IDE (Rider 2022.3.1), ie test2.g.cs is re-generated each time with an updated timestamp.

Why does that happen - and how can I prevent these unnecessary updates?

public class TestGenerator : IIncrementalGenerator
{
    public void Initialize(IncrementalGeneratorInitializationContext context)
    {
        // Create test1.g.cs on initialization
        context.RegisterPostInitializationOutput(context => context.AddSource("test1.g.cs", "public record TestRecord();"));

        // Find the type symbol for the previously generated record, and generate test2.g.cs based on this
        context.RegisterSourceOutput(
            context.SyntaxProvider.CreateSyntaxProvider(
                static (context, _) => context is RecordDeclarationSyntax s && s.Identifier.Text.Equals("TestRecord"),
                static (context, _) => context.SemanticModel.GetDeclaredSymbol(context.Node) as INamedTypeSymbol
                ).Where(x => x is not null),
            (context, symbol) => context.AddSource("test2.g.gs", $"// Found: {symbol is not null} at {DateTime.UtcNow}")
            );
    }
}

Answer 1

Turns out this was caused by referencing the analyzer project directly.

When packaging it into a NuGet package and referencing that, caching works as expected.

Answer 2

Don't include symbols in the pipeline. They root compilations in memory, and won't compare equal across different compilations.

Instead, you should be creating a data model (with proper equality semantics - preferably with records). The data model should only include what you really need for generation, not including symbols, syntax nodes, compilation.