dxkit-schema

dxkit-schema

This skill owns the model-schema drift gate: dxkit statically extracts every declared data model (ORM entities, tagged structs, spec schemas) and fails a PR that changes one in a breaking way — a removed field, a changed type, an optional field made required, a removed model. Additive changes surface as warnings or information, never blocks.

It is thin orchestration over the deterministic CLI. It never re-implements extraction or diffing: it runs schema / schema diff / guardrail check, reads their structured output, and supplies judgment + code edits. The determinism stays in the CLI; the agent supplies the reasoning.

Language coverage. Recognition is marker-based per language pack — precision over recall. TypeScript/JavaScript: decorator-marked entity classes (TypeORM/MikroORM @Entity, sequelize-typescript @Table, NestJS-mongoose @Schema, type-graphql @ObjectType/@InputType) plus BaseEntity heritage; field facts from the type annotations. Python: Django models.Model, SQLAlchemy Base/DeclarativeBase, pydantic BaseModel, SQLModel, and @dataclass; Django/SQLAlchemy field constructors supply types and null=/nullable= optionality. Go: a struct tagged json:/gorm:/db:/bson: IS a wire contract; the tag supplies the wire name, omitempty and pointer types read as optional. Java: JPA @Entity/@Table/@Embeddable/@MappedSuperclass classes; @Column(nullable = …) supplies optionality and name = … the wire name — an unannotated column is an honest unknown (JPA defaults to nullable; dxkit never fabricates required). Kotlin: JPA annotations, kotlinx @Serializable data classes (incl. positional @SerialName("wire") names), and Exposed Table objects (fluent column chains — varchar("x", 50).nullable() reads type + optionality); String? nullability is read from the grammar, @Column overrides it. C#: [Table]/[Keyless]/[Owned] entities AND classes referenced by a DbContext's DbSet<T> properties; partial-class declarations assemble into one entity (codegen splits are never drift); string? NRT annotations supply optionality, positional [Column("x")]/[JsonPropertyName("x")] the wire names. Ruby (Rails): db/schema.rb is the model source — one entity per create_table, named by the TABLE (the wire contract), null: false ⇒ required and absence ⇒ nullable (the Rails default); the ActiveRecord class marker is discovery-only while a schema file exists. Rust: #[derive(Serialize/Deserialize)] structs; Option<T> optionality is precise, #[serde(rename = "x")] supplies wire names (rename_all/default container attributes are not read). Any other language — and unmarked DTOs in covered languages — participate via a spec: point .dxkit/policy.json:schema.specs at an OpenAPI (components.schemas) or JSON Schema document and its models are gated identically, no pack extraction needed.

What the gate does not see (tell the user honestly when relevant): unmarked plain interfaces/classes; dynamically-built schemas (disclosed in the inventory's dynamic list, not diffable); constraint-level changes (max length, enum members, validation rules, defaults); a type alias whose underlying type changed (comparison is lexical — resolving aliases would need a type-checker); serialization renames outside Go tags; migration files / raw SQL; whether anything actually consumes a removed field.