The bridge: a schema becomes C++

[matt-edmondson]

All three sections of Holotype's docs/generated-cpp-target.md now generate from a .schema.json rather than from an AST built by hand. ktsu.Coder already had acceptance tests that built those headers as ASTs; these start one step further back, from the file a person would actually write, and so are what says the schema can describe the thing at all.

Schema.Cpp turns a schema into a ktsu.Coder AST and hands it to CppGenerator, which owns every question about how C++ is spelled. Nothing in this project writes a brace.

Where it lives, and why

It cannot ship inside ktsu.Schema: that library publishes net8.0 and ktsu.Coder does not. SchemaGenerator.Register is how it is found by the language a schema names, and SchemaTool/Program.cs is the worked example of a host doing it — which makes that seam load-bearing rather than hypothetical.

The options record is smaller than expected

Almost everything comes from the schema. A unit is a semantic type, so the generator emits the class. A class is a struct. The standard library spells a string, a sequence, a view and an absent value.

What's left is the short list a program supplies for itself:

Option	Because standard C++ has
Vector2 / Vector3 / Vector4	no fixed-shape numeric vector (std::array makes the arity a type argument, and a schema's vector is one value with named components rather than a short sequence)
ColorRgb / ColorRgba	no colour
Handle	no identifier-with-a-generation
Result	no fallible return
DateTime	an answer that is rarely the one wanted — system_clock::time_point is a duration since an epoch, not a calendar date

A target that has them says how it spells them and which header they come from — a generated file naming holo::Vector3 without including the header that declares it does not compile, and nothing but whoever supplied the name knows which header that is. A target that has not is refused those schema types by name, naming the option to set.

ExistingTypes is the other direction: a semantic type the target already hand-wrote is named rather than generated a second time beside it, which is how the exemplar gets holo::Kilograms instead of a freshly emitted Kilograms.

One deliberate deviation from the document

An enum goes to namespace scope in a header of its own rather than nested in the struct.

The document nests it, on the grounds that two components in one file might both declare a kind — true of the TOML front end, where an enum belonged to the component that declared it. In ktsu.Schema an enum is a top-level element any class may name, so nesting it in the one class that uses it today would move the type the moment a second class used it, breaking every caller that spelled RigidBody::BodyKind because of an unrelated edit. Field offsets are unaffected, so generated_compiles.cpp still holds. The acceptance test asserts this difference rather than absorbing it.

Three defects found by reading the output, not by a test going green

• A file naming a generated type did not include the header that declares it. A component using an enum, a class or a semantic type did not compile.
• A float default was written as the schema holds it. C++ reads 1 as an int and 1.0 as a double, and a braced initialiser refuses either for narrowing — which is the whole reason to brace it. Now 1.0f.
• An enum default was constructed rather than named: BodyKind{ BodyKind::Dynamic } where the document has BodyKind::Dynamic. The schema holds a value's name, so qualifying it is the generator's job.

And two in code I had just written: a file with no namespace configured indexed an empty array, and the banner cited a source file the schema does not remember. Schema now keeps SourceFileName beside SourceDirectory — not serialized, for the same reason the directory is not.

One gap left open, and asserted

SchemaFunction cannot say a call leaves the receiver unchanged, so the document's const on find() is not emitted. The four conventions cover whether an argument is read-only and say nothing about the receiver. AQueryCannotYetSayItLeavesTheReceiverUnchanged asserts the gap so it is visible rather than merely absent, and fails when it is closed. The AST is ready for it (FunctionDeclaration.IsReadOnly emits the trailing const), so closing it is a property on SchemaFunction and one line in the builder.

Testing

46 tests in Schema.Cpp.Test — three acceptance tests against the document's sections, the complete §4 type table a row at a time, every refusal, the naming rules, and the registration seam. 397 in Schema.Test, 201 in SchemaEditor.Test. Whole solution builds with 0 warnings on both net10.0 and net9.0.

🤖 Generated with Claude Code

https://claude.ai/code/session_01AhoPJ5AbxP8QEBNxPQYEPk

---

Generated by Claude Code

[sonarqubecloud]

Quality Gate passed

Issues
0 New issues
0 Accepted issues

Measures
0 Security Hotspots
96.4% Coverage on New Code
0.0% Duplication on New Code

See analysis details on SonarQube Cloud