A model generator turns each component schema (refName) into one\ndefinition in one file: entry → projection → schema-type router → one\nsnippet class per schema type. That structure is invariant across\ntarget libraries — only naming policy and per-type syntax vary. So do\nnot write the structure: copy it.
The skeleton/ directory next to this file is a complete, compiling,\nengine-tested model generator that renders a placeholder syntax\n(m.object({...})). Author by transplant, not from scratch:
skeleton/ to your package location; run\ndeno test --allow-env --allow-sys --allow-read — 6 green tests\nprove the machinery before you touch anything.name in deno.json; MyLib → YourLib in class\nnames and filenames; myLibEntry → yourLibEntry; then\nsrc/lib.ts — LIB_MODULE (emitted module specifier) and LIB\n(imported symbol).mod.test.ts to\nyour target syntax. The structural assertions (files exist, shared\nrefs dedup to ONE definition, import headers stitched, recursion\nannotated) must pass UNCHANGED — if one breaks, you broke machinery,\nnot syntax.Every slot is a // SLOT(name): comment. Everything outside a SLOT is\nengine machinery — modifying it is almost always a mistake.
| Slot | \nFile | \nDecision | \n
|---|---|---|
library | \nsrc/lib.ts | \nemitted module + symbol, single point | \n
naming, identifier-kind, export-path | \nsrc/base.ts | \nidentity policy (from refName ONLY) | \n
string, string-constraints | \nsrc/MyLibString.ts | \nstring / enum / literal syntax; formats, min/maxLength | \n
number, integer, boolean, unknown, void | \nsrc/MyLibScalars.ts | \nscalar syntax; numeric constraints | \n
array | \nsrc/MyLibArray.ts | \nlist syntax | \n
object-properties, object-intersection, object-empty, visibility | \nsrc/MyLibObject.ts | \nobject syntax; properties+record composition; readOnly/writeOnly policy | \n
record | \nsrc/MyLibObject.ts | \nadditionalProperties map syntax | \n
union | \nsrc/MyLibUnion.ts | \noneOf/anyOf; discriminated form | \n
lazy | \nsrc/MyLibRef.ts | \ndeferred-reference form for cycles | \n
recursion-annotation | \nsrc/MyLibProjection.ts | \ntype annotation breaking circular inference | \n
modifiers | \nsrc/modifiers.ts | \noptional/nullable syntax and wrap order | \n
enrichments | \nsrc/enrichments.ts | \nconfig seam (default: opt-out) | \n
MyLibRef puts only the peer's\nNAME in the value tree; the ModelDriver resolves the definition\n(cache hit → reuse, miss → construct) and stitches the cross-file\nimport. Inline-expanding a ref, or hand-writing its import, is how\nshared models duplicate.context.modelDepth > 0)\nrenders via SLOT(lazy) and bumps the depth; the projection then sees\n> 1 and sets settings.identifier.typeName (SLOT\nrecursion-annotation) so the emitted export const doesn't die of\ncircular inference (TS7022/7024). Self-recursion only — mutual\nrecursion is not detected.required comes from\nthe PARENT object's required list and flows into each property\nleaf's modifiers; nullable sits on the node itself. Both render\nexactly once, in applyModifiers, at the leaf — no other owner, and\nnever while building stored fields.true/empty schema →\nthe unknown fallback; properties + additionalProperties together →\nSLOT(object-intersection).z.object({...}), .and(z.record(...)) for both-forms), so the\nobject SLOTs compose freely. In a head+value language (Kotlin) the\ntwo positions DIVERGE, and a position-blind toString() cannot serve\nboth (compiler-verified 2026-08-04, kotlin-debug rig): properties-only\ndeclares as a data class parameter list, and in type position must\nrender a NAME — synthesize the named sibling declaration and\nreference it (name derived from the schema's own stackTrail, no\nnaming param threaded through the router; collisions policed by a\ndocument-wide claim registry that throws per-item, since the name\nshares a PACKAGE with every component class — gen-kotlin-jackson\ntoSynthesizedName.ts + synthesizedNames.ts; a parameter list in\ntype position parses as a function type and fails, and widening to\nMap<String, Any?> discards the type — capitulation, not a\nsolution); record-only and empty\nmust not take a data-class head at all (data class X() is illegal —\ntheir declaration kind is typealias); both-forms has a declaration\nform (data class plus a @field:JsonAnySetter @get:JsonAnyGetter\ncatch-all map property) but no anonymous type form. Decide the\nidentifier KIND and the value together from the same schema guards\n(gen-kotlin-jackson shape.ts is the worked example) — never from\nthe name alone, and never by making one toString() answer both\npositions.z.discriminatedUnion(\"type\", [...])). In a language without union\ntypes (Kotlin) a qualifying discriminated union becomes a named\nsealed declaration, and the member models must declare the\nsupertype — a member may be BUILT before its union is ever seen, so\nmembership comes from a document-wide scan (parent → member\ninversion, WeakMap-memoized) consulted at member construction, never\nfrom the union's own walk. Non-qualifying unions render the honest\nwire type (JsonNode), not Any. Full pattern: the Kotlin lang\nskill §8c.handleKey — 'first-name' renders\nquoted; never assume keys are identifiers.readOnly/writeOnly are captured per property in\nMyLibObjectProperties.visibility. Default policy ignores them; if\nthe target needs them, annotate the value (e.g. .readonly()) or\nemit request/response variants via variant threading — decide at\nSLOT(visibility). Caller options are the third route: declare the\nsecond type parameter of toTsModelProjectionBase, let the calling\ngenerator pass { options } on the insert, and fold them into the name.custom values\npass through untouched.TypeSystemString needs format + enums; objects expose\nobjectProperties/recordProperties). Add fields freely; remove\nnone — removal breaks insertNormalizedModel consumers and fails the\nSchemaToValueFn check.The shipped mod.test.ts runs the REAL pipeline (toArtifacts) over a\nfixture with an enum, an array-of-ref, a shared ref (×2 → one\ndefinition), optional + nullable, a record, and a self-recursive model.\nIt is your regression gate: green before you start, green after every\nslot. Read failures in this order: import header first (a missing\nimport means a string swallowed a snippet), then the body, then\ndeno lint (the skmtc/* rules are wired in deno.json).
| Symptom | \nFix | \n
|---|---|
| Shared model duplicated per consumer | \nA ref was rendered/expanded instead of flowing through MyLibRef | \n
| Stack overflow on recursive schema | \nThe modelDepth branch in MyLibRef was removed or bypassed | \n
| Emitted file dies of TS7022/7024 | \nSLOT(recursion-annotation) not set for the target | \n
.optional() doubled or missing | \nModifiers applied outside applyModifiers, or a second owner added | \n
Enum with null member renders 'null' | \nKeep the literal() null-guard from MyLibString | \n
| Peer generator can't consume yours | \nschemaToValueFn/createIdentifier statics or TypeSystem contract fields removed | \n
Lint fires no-template-imports/no-adhoc-tostring | \nTarget syntax leaked outside a toString() body — move it into the SLOT | \n
data class NameMap<String, Any?> (head glued to a type) in output | \nDeclaration kind and value were decided separately — see the four-forms bullet in §3; kind+value must come from the same schema guards | \n
Engine semantics (the one law, memoization, enrichments, variants,\nnaming rules) live in skmtc-generator — read it first. TS-layer\nspecifics (register shapes, identifier kinds, import machinery,\nList/FunctionParameter) live in skmtc-lang-typescript. This\nskill owns only the model SHAPE. The skeleton is TypeScript-emitting;\nfor a Kotlin model generator, keep this skill's shape and edge-case\nrules but take call shapes from the Kotlin lang skill (no Kotlin\nskeleton yet). Operation generators are a different shape — load\nskmtc-operation; accumulators are covered by neither (clone\ngen-msw/gen-express per skmtc-generator §2).