schema

schema builder — maintainer notes

This document is the long-form companion to the schema builder code.

The source files keep godoc concise; complex invariants, design trade-offs, and known quirks live here.

---

Table of contents

• §build-entry — Build modes and dispatch entry points
• §dispatch-table — buildNamedType's underlying-shape table
• §dissolve-named — when a named type is unwrapped instead of $ref'd
• §special-types — applyStdlibSpecials, applySpecialType, UUID heuristic
• §textmarshal-order — buildFromTextMarshal precedence
• §aliases — TransparentAliases vs RefAliases vs default-expand
• §discovery — Models / ExtraModels / discovered, dedup layers
• §struct — buildFromStruct two-pass shape
• §allof — buildAllOf, buildNamedAllOf, scanEmbeddedFields
• §embedded — embed routing, struct/interface specials asymmetry
• §embed-depth — ambiguous-embed diagnostic mechanism
• §method-mangler — interface-method JSON-name derivation
• §user-overrides — explicit user-driven type/format overrides at decl-site and field-site
• §ref-override — applyToRefField, the allOf-on-$ref shape, refOverrideCollector, applyPattern
• §simple-schema-mode — the SimpleSchema build mode for OAS v2 non-body params and response headers
• §classifier-walkers — per-call-site classifier walkers and findAnnotationArg's single-word filter
• §quirks — known behavioural caveats
  • §quirks-resolved — ✅ fixed in this refactor
  • §quirks-open — 🟡 deferred / 🟦 documented behaviour

---

§build-entry — Build modes and dispatch entry points

Builder.Build has three modes selected by Option:

Option	Sets	Entry point	Output destination
WithDefinitions(map)	s.definitions	buildFromDecl	map[s.Name]
WithType(tpe, tgt)	s.inputType, s.target	buildFromType(inputType, target)	tgt (caller-owned, full Schema)
WithSimpleSchema(tpe, tgt, in)	s.inputType, s.target, s.simpleSchema, s.paramIn	buildFromType(inputType, target) + exit validator	tgt (caller-owned, SimpleSchema-shaped)

WithDefinitions and the WithType/WithSimpleSchema pair are mutually exclusive; Build panics on misuse. WithDefinitions is the spec-orchestrator entry point for top-level type declarations. WithType produces a full OAS v2 Schema for body parameters, response bodies, and any other site that owns a Typable and accepts the full schema vocabulary. WithSimpleSchema produces an OAS v2 SimpleSchema for non-body parameter sites and response headers — see §simple-schema-mode for the contract and the exit-validator's role.

buildFromDecl does three things in order:

1. Consume the decl's doc-comment block (may short-circuit on swagger:ignore).
2. Defer annotateSchema (x-go-name, x-go-package extensions).
3. Intercept stdlib special types (§special-types) before kind-dispatch. Necessary because stdlib decls can reach buildFromDecl via the discovery chain (e.g. type X = time.Time pulls time.Time itself into discovered).
4. Dispatch on s.Decl.ObjType(): *types.Named → buildFromType(ti.Type, …); *types.Alias → buildDeclAlias; otherwise warn-and-skip.

---

§dispatch-table — buildNamedType's underlying-shape table

buildNamedType handles a *types.Named reaching as a field/embed/etc. type (not a top-level decl). The body is a table indexed by Underlying() kind, with each arm following the same three-step shape:

1. shape-local pre-check (e.g. UnsupportedBuiltinType for *types.Basic)
2. classifier walk — short-circuit on a match, may recurse on Underlying
3. FindModel pivot via resolveRefOr / resolveRefOrErr — hit ⇒ makeRef, miss ⇒ shape-specific fallback

The arms:

Underlying	Classifier	On-miss fallback
*types.Struct	classifierNamedStructStrfmt	silent — nil
*types.Interface	—	missingSource error
*types.Basic	classifierNamedBasic (cascade)	SwaggerSchemaForType(name)
*types.Array / *types.Slice	classifierNamedArrayLike(forSlice)	inline element via buildFromType
*types.Map	—	silent — nil

buildNamedArrayLike is the unified Array/Slice helper; the slice arm passes forSlice=true so classifierNamedArrayLike can honour the bsonobjectid slice-only special case.

---

§dissolve-named — when a named type is unwrapped instead of $ref'd

After the prelude but before the Underlying() switch, buildNamedType has a short-circuit:

if s.Decl.Spec.Assign.IsValid() || (titpe.TypeArgs() != nil && titpe.TypeArgs().Len() > 0) {
    return s.buildFromType(titpe.Underlying(), target)
}

Two disjuncts, both meaning "this named type has no source-level TypeSpec of its own to $ref — emit its structural form inline."

Disjunct 1 — s.Decl.Spec.Assign.IsValid()

This is the TransparentAliases plumbing mechanism. When the outer Builder was constructed for an alias-syntax decl (type X = Y), Spec.Assign is the position of the = token (valid). Every named type reached during this Build is then inlined rather than $ref'd — which is what TransparentAliases semantically means.

It also covers the gotypesalias=0 legacy mode (type X = Y surfacing as *types.Named directly), but that's an incidental side-benefit.

Example: under TransparentAliases=true, a body parameter aliased as type AliasBody = Payload causes schema.Builder to dissolve AliasBody → walk Payload's underlying struct → emit {type:object, properties:{…}} inline on the body param. Without the disjunct, buildNamedType would emit {$ref: "#/definitions/Payload"} — wrong for the dissolve intent.

Disjunct 2 — titpe.TypeArgs().Len() > 0

Generic instantiations like GenericSlice[int]. The instantiated *types.Named has no source-level TypeSpec; only the generic declaration (GenericSlice[T any] []T) has one. Unwrapping to Underlying() substitutes type params with concrete types ([]int) so the schema reflects the substituted shape.

Fixture: fixtures/enhancements/generic-instantiation/.

---

§special-types — applyStdlibSpecials, applySpecialType, UUID heuristic

Three layers, all in special_types.go:

• applySpecialType(obj, target, recognizers...) — the engine. Iterates recognizers and applies the first match. Each recognizer is identity-based (exact Pkg.Path + Name match) except recognizeUUID which is fuzzy (case-insensitive name match).

• applyStdlibSpecials(obj, target) — the canonical safe set {recognizeAny, recognizeTime, recognizeError, recognizeRawMessage}. All four are identity-based and cannot misfire on user types, so this helper is called uniformly at every site that handles a *types.TypeName.

• recognizeUUID — opt-in via applySpecialType's variadic. Currently used only by buildFromTextMarshal because the upstream IsTextMarshaler gate guarantees the type renders as text, making the fuzzy name match safe.

The seven call sites of applyStdlibSpecials (buildFromDecl, buildDeclAlias RHS, buildAlias, buildNamedType, buildNamedEmbedded interface arm, processEmbeddedType named arm, buildNamedAllOf struct arm) previously varied their recognizer subsets. Unification preserved goldens — the narrow subsets were historical accumulation, not semantic. Identity checks cannot misfire, so passing the full safe set everywhere is correct by construction.

---

§textmarshal-order — buildFromTextMarshal precedence

The function is entered from buildFromType's shortcut hasNamedCore(tpe) && IsTextMarshaler(tpe). Pipeline:

1. peel pointers (self-recurse)
2. route aliases through buildAlias (honour TransparentAliases / RefAliases)
3. type-assert to *types.Named (fallback: {string, ""})
4. classifier (swagger:strfmt) — explicit user intent wins
5. stdlib trio via applySpecialType(recognizeError, recognizeTime, recognizeRawMessage, recognizeUUID)
6. PkgForType-miss bail (gates only the generic fallback below)
7. generic fallback — {string, ""} + x-go-type: pkg.Name

The user-intent-first rule (step 4 before steps 5–6) is the same shape applied in buildNamedAllOf and elsewhere. The order matters because, e.g., a UUID-named type carrying swagger:strfmt date should emit {string, date} — the classifier wins.

Fixtures: fixtures/enhancements/text-marshal/explicit_override/ demonstrates classifier-beats-heuristic; text-marshal/uuid_wrapping_time/ demonstrates heuristic-still-fires when no override exists.

Why stdlib recognizers run before the PkgForType bail

PkgForType looks tpe up in s.app.AllPackages. Stdlib packages (time, encoding/json) often aren't there — the scanner only registers packages it was asked to scan. Previously the stdlib recognizers ran after the PkgForType bail, so stdlib types reaching here when their package wasn't in AllPackages silently emitted {}. The new order recognizes stdlib types via tio.Pkg() alone (no PkgForType call needed) before the bail.

---

§aliases — TransparentAliases vs RefAliases vs default-expand

Alias handling has two axes: decl shape (what does the alias's own definition look like?) and use-site shape (what does a field / element / body site that references the alias produce?). The same contract governs the schema, parameters and responses builders — both at the decl and at the use site — so the rules described here apply uniformly across all three layers (see parameters/README.md and responses/README.md for the layer-specific dispatches that consume this contract).

Decl shape — buildDeclAlias

buildDeclAlias handles top-level alias declarations. Independent flags with deterministic precedence:

TransparentAliases	RefAliases	Outcome
true	(any)	Dissolve — buildFromType(rhs, target). No LHS definition. Wins outright.
false	true	$ref — makeRef to the RHS's named target.
false	false (default)	Expand — buildFromType(Underlying, target). LHS gets a structural definition.

swagger:model on an alias decl forces decl-level registration unconditionally — even under TransparentAliases, an annotated decl reaches definitions (with a structural shape via the Spec.Assign.IsValid() dissolve-named branch). The trade-off is that under Transparent the annotated decl exists but nothing references it at use sites — the orphan-annotated-decl shape.

The dissolve case propagates to nested named types via the Spec.Assign.IsValid() disjunct in buildNamedType (§dissolve-named).

swagger:strfmt <format> and swagger:type <go-type> on an alias decl are honoured at the decl entry — swagger:strfmt at the top of buildDeclAlias, swagger:type via classifierNamedTypeOverride in buildFromDecl. Both write the canonical shape ({type: string, format: <format>} or the Swagger shape for the named Go type) and short-circuit the underlying-kind dispatch.

Use-site shape — buildAlias (and the parameters / responses analogues)

buildAlias is the use-site handler — called whenever an alias-typed value is encountered as a struct field, slice / array element, allOf member, body parameter or response body. The rule is annotation-gated and identical across the three builders:

Mode	swagger:model on the alias?	Use-site shape
TransparentAliases=true	(irrelevant)	Dissolves to the unaliased target. No definition for the alias. Wins outright.
Default / RefAliases=true	yes	$ref: <AliasName> — the alias surfaces in definitions with shape per mode (decl shape table above).
Default / RefAliases=true	no	Dissolves to the unaliased target via types.Unalias (full chain collapse in one step). The alias produces no definition entry.

The use-site $ref target is decided by annotation alone (with Transparent as the override) — the mode flags only shape the alias decl's own definition. A struct field typed with an annotated alias produces the same $ref: <AliasName> under both Default and Ref; the difference shows up downstream in the alias's own definition (Expand structural under Default vs chain $ref under Ref).

The same applies inside allOf composition: swagger:allOf on an embed governs the composition shape (allOf vs flat inline); swagger:model on an embedded alias governs the identity of the allOf member's $ref target (alias name vs unaliased target). They are orthogonal.

SimpleSchema reach contexts

Non-body parameters and response headers are SimpleSchema targets and cannot carry $ref (OpenAPI 2.0 constraint). At those sites the alias always expands to the unaliased target regardless of annotation. The annotation gate has no effect for SimpleSchema because the question "$ref to what" never arises.

Top-level alias parameters and responses

When swagger:parameters or swagger:response is on an alias declaration, the alias is transparent re: model creation in all three modes — neither the alias nor any chain link of its backing struct surfaces in definitions. The fields of the unaliased target become the operation's parameters or the response's body / headers. This clause has no schema-builder analogue: swagger:parameters / swagger:response declare a parameter set or response, not a model, and never promote their backing chain to the spec's definitions section.

Rhs() vs Underlying()

• tpe.Rhs() — immediate right-hand side of the alias declaration. For type X = Y, Rhs is Y as-is (may itself be a *types.Alias or *types.Named).
• tpe.Underlying() — peels through aliases and named types to reach the structural form (*types.Struct, *types.Basic, …).
• types.Unalias(tpe) — peels through aliases only, leaving Named types intact. Used at use sites for full chain dissolve in one step.

Example: type X = Y; type Y = Z; type Z = int gives X.Rhs() == Y (*types.Alias), X.Underlying() == int (*types.Basic), and types.Unalias(X) == int (*types.Basic) as well.

The branches use them intentionally:

• Dissolve (decl) uses rhs — dissolves one layer at a time; if the RHS is itself a named/aliased type, build from it (may recurse).
• Expand (decl) uses Underlying() — fully expand the structural shape onto the LHS definition.
• Dissolve (use site) uses types.Unalias — collapses the whole alias chain in one step at the field / element / body position.

---

§discovery — Models / ExtraModels / discovered, dedup layers

The scanner exposes two model indexes (in scanner.app):

• Models — decls carrying swagger:model (or response/parameter) annotations. The orchestrator builds these unconditionally.
• ExtraModels — decls discovered transitively that aren't annotated but need a top-level definition. Promoted to Models at consumption (joinExtraModels calls MoveExtraToModel).

The spec orchestrator also maintains an in-flight queue s.discovered populated via Builder.AppendPostDecl:

• makeRef(decl, …) appends decl to Builder.postDecls.
• spec.Builder.buildDiscoveredSchema runs schema Builds and harvests each Builder's PostDeclarations() into s.discovered.
• buildDiscovered then drains s.discovered until empty.

ScanCtx.AddDiscoveredModel(decl) is the explicit hook for the ExtraModels side (replacing the now-deprecated FindModel side effect).

Dedup layers

Three layers prevent the same decl from being built twice:

1. AppendPostDecl — per-Builder dedup keyed by EntityDecl.Ident.
2. AddDiscoveredModel — no-op when the decl is already in Models (avoids Models ↔ ExtraModels bouncing).
3. spec.Builder.buildDiscovered — per-pass dedup keyed by decl.Names() string.

Without (3), the TestCoverage_RefAliasChain shape regression: the same alias-target decl, appended via multiple makeRef call sites, got queued twice in one pass, and the second Build read the half-built schema and appended another set of allOf entries — doubling them.

---

§struct — buildFromStruct two-pass shape

buildFromStruct emits the schema for a named Go struct in two passes over the same *types.Struct. The split is required because allOf composition (driven by swagger:allOf on embedded fields) can change which schema receives the property map.

1. Pass 1 — scanEmbeddedFields. Walks every anonymous field. Each embed is classified:
   • plain embed (no swagger:allOf, embed is not an *types.Alias): its fields are merged into the outer schema directly via buildEmbedded. Returned target == schema.
   • allOf embed (swagger:allOf present, or the embed is an alias): the embedded type becomes an entry in schema.AllOf, and a fresh schema is allocated as the target for the property map. Returned target == fresh schema.
2. Pass 2 — buildStructFields. Iterates non-anonymous exported fields and emits each via applyFieldCarrier.

If hasAllOf is true and the fresh target ended up with properties, the target itself is appended to schema.AllOf so the inline properties live as their own compound member alongside the embedded $refs.

Why target.Typed("object", "") always fires

The line runs unconditionally after target selection. Reason: a struct with zero exported fields and zero embeds still emits as {type: object, properties: {}} — distinguishable from a missing schema. SimpleSchema (forthcoming with M1) introduces the only path where this line would not fire; the current code is Full-Schema only.

User-classifier short-circuit

classifierStructPreBuildType runs at the very top of buildFromStruct. It consumes only swagger:type on the decl's own comment group. On match, the struct walk is skipped entirely — the schema is whatever swagger:type X resolves to via SwaggerSchemaForType. See §user-overrides for the cascade and the unknown-leaf fallthrough handled by the caller (buildFromDecl), not here.

---

§allof — buildAllOf, buildNamedAllOf, scanEmbeddedFields

OAS 2.0 allOf composition surfaces in three places in the schema builder: scanEmbeddedFields (decides which embeds become allOf members), buildAllOf (peels and dispatches an allOf member type), and buildNamedAllOf (resolves a named-type allOf member through the same user-classifier-first precedence the rest of the builder uses).

scanEmbeddedFields — embed classification

Walks *types.Struct's anonymous fields. Three signals decide classification per embed:

• swagger:ignore — embed skipped entirely.
• json:"-" — embed skipped (parity with v1's JSON tag handling).
• swagger:allOf (via fieldDoc.IsAllOfMember) or the embedded type is *types.Alias — embed becomes an allOf member; remaining properties land on a fresh target schema.
• otherwise — plain embed; properties merge into the outer schema.

The swagger:allOf arg, when present, is recorded as x-class: <arg> on the outer schema (fd.AllOfClass). This is the discriminator hint downstream go-swagger consumes.

buildAllOf — three-arm peel

Strips pointers (recurses), routes *types.Named through buildNamedAllOf, routes *types.Alias through buildAlias. Any other input is dropped silently with a logger.UnsupportedTypeKind warning — parity with v1, which had no surface for non-Named / non-Alias allOf members.

buildNamedAllOf — symmetric arm dispatch

Struct and interface underlyings share the same precedence shape:

1. classifier first — classifierAliasTargetStrfmt(ftpe, tgt). On match the named type is emitted as {string, <format>} and the walk terminates. Same shape as §textmarshal-order's step 4.
2. stdlib specials — applyStdlibSpecials(tio, tgt, skipExt). Identity-based, cannot misfire. Catches time.Time, error, json.RawMessage, any/interface{} if any of them reach as an allOf member.
3. model lookup — Ctx.GetModel(pkg, name). Missing decl is a build error (the allOf member must be resolvable).
4. HasModelAnnotation() → makeRef — annotated types become $ref entries. The struct/interface body is built lazily via discovered.
5. inline build — buildFromStruct / buildFromInterface on the underlying. Used when the type is reachable but not swagger:model-annotated.

Both arms route through the same tgt := NewTypable(schema, 0, skipExtensions) target, avoiding the v1 asymmetry where the struct arm used classifierAliasTargetStrfmt (decl-fetched internally) and the interface arm used a comment-group-keyed variant (classifierAliasOwnDocStrfmt, since deleted) that pre-fetched the decl. The unified target lets both arms run classifier-first without doing the model lookup upfront — earlier classification means no orphan ExtraModels side effect for strfmt-tagged interface types.

Error message on missing decl

The missing-decl error is now uniform across arms: "can't find source for named allOf member %s: %w". Previous phrasing differentiated struct vs interface; no test asserts on the text, the change is golden-neutral.

---

§embedded — embed routing, struct/interface specials asymmetry

buildEmbedded is the entry point for a struct's embedded fields that the embed classifier (§allof §scanEmbeddedFields → plain-embed arm) routed for inline merge into the outer schema. It splits three ways: pointers peel (recurse), *types.Named descends into buildNamedEmbedded, *types.Alias goes through buildAlias (so alias-resolution honours TransparentAliases / RefAliases).

buildNamedEmbedded — the two-arm specials asymmetry

The interface arm runs applyStdlibSpecials(o, target, skipExt) before model lookup. The struct arm does not.

The asymmetry is deliberate:

• Interface embed is the common shape that surfaces error promoted into a struct (type Err struct { error }). The stdlib specials catch it and emit {string} with the x-go-type: error hint, matching the field-level recognizer behaviour.
• Struct embed of time.Time is uncommon — the fixture corpus has none. Adding applyStdlibSpecials to the struct arm would change behaviour only for code paths we don't currently exercise, and the change would surface as a golden delta on the day someone adds such a fixture. Until then the conservative choice is to preserve v1 parity: stdlib structs embedded as *types.Struct reach GetModel and build through the normal buildFromStruct path. If time.Time ends up embedded and the package isn't in AllPackages, missingSource fires — same as v1.

AddDiscoveredModel pairing

Both arms call s.Ctx.AddDiscoveredModel(decl) before recursing. Reason: embedded user types appear as their own top-level definition even when not annotated swagger:model. The discovered queue picks them up on the next build pass. Parity with v1.

processEmbeddedType — interface-side allOf composition

Called from buildNamedInterface when walking the embedded elements of an interface's underlying. Three shapes:

• *types.Named — runs applyStdlibSpecials (dummy target swallows the write so the recognizer can short-circuit without contaminating schema), then routes through buildNamedInterface.
• *types.Interface — anonymous embedded interface. Builds into a side schema; only when the result is non-empty (Ref || Properties || AllOf) is it appended to the outer schema.AllOf.
• *types.Alias — same non-empty guard, builds via buildAlias so alias modes are honoured.

The non-empty guard is what makes interface{} and other zero-content interfaces invisible at the allOf seam — they don't contribute an {} entry to the outer schema.

---

§embed-depth — ambiguous-embed diagnostic mechanism

Builder.embedDepth (incremented around buildNamedEmbedded's recursive descents via defer s.enterEmbed()()) tracks the depth of embedded-type recursion in a single buildFromStruct / buildFromInterface pass.

applyFieldCarrier uses it to distinguish:

• embedDepth == 0 — legitimate explicit override (S.Foo redefining the JSON name of an embedded E.Foo). No diagnostic.
• embedDepth > 0 + prior at deeper depth — Go's depth rule (shallower wins) already disambiguates. No diagnostic.
• embedDepth > 0 + prior at same-or-shallower depth + different Go name — peer ambiguity Go itself would refuse to promote. Emits CodeAmbiguousEmbed (SeverityWarning). Last-write-wins behaviour is preserved; only the signal is added.

Fixture: fixtures/enhancements/diagnostics/types.go covers all three cases.

---

§method-mangler — interface-method JSON-name derivation

Interface methods can't carry struct tags. To pick a JSON property name for a method, Builder.methodMangler (a mangling.NameMangler from go-openapi/swag) applies the acronym-aware lower-first transform: CreatedAt → createdAt, ID → id, ExternalID → externalId.

Principled asymmetry vs the struct path

The struct-field path does NOT mangle: resolvers.ParseJSONTag returns the json-tag value when present, otherwise the Go field name verbatim. So CreatedAt string with no tag emits property CreatedAt, while CreatedAt() string on an interface emits property createdAt.

This asymmetry is intentional, not a quirk:

• Struct fields mirror real serialization. encoding/json uses the tag-or-verbatim rule at runtime; codescan's emitted spec is what json.Marshal would actually produce. Adding an auto-mangle on the struct side would silently disagree with the user's running program.
• Interface methods don't have a "natural" serialization. Go's encoding/json can't marshal embedded interface methods at all without a custom MarshalJSON. There is no runtime ground truth to mirror, so codescan invents a sensible default JSON name. The mangler is the documentation convention, not a serialization mirror.

The "one size fits all" mangler on interfaces will not always be what the author wanted — a future global opt-out (skip-jsonify-interfaces or similar) is on the roadmap.

swagger:name X is verbatim

When the author provides swagger:name X on an interface method, X is emitted exactly as written — the mangler is bypassed entirely. The carrier code (fields.go:methodCarrier) checks fd.JSONName first; only when empty does it call s.interfaceJSONName(fld.Name()).

This contract matters for non-camelCase user input — a user who writes swagger:name UserIdentifier wants UserIdentifier, not userIdentifier. The regression-detector for this is fixtures/enhancements/interface-name-verbatim/ + integration/coverage_interface_name_verbatim_test.go: PascalCase, snake_case, SCREAMING_CASE, and hyphenated user inputs all assert on the exact spelling reaching the spec.

Constructor invariant

The mangler is initialised in NewBuilder; &Builder{…} literals that bypass the constructor will nil-panic in interfaceJSONName.

---

§user-overrides — explicit user-driven overrides

Three classifier annotations let the author bend the type-driven default. They live at two scopes — the decl-site (on the type declaration's doc comment) and the field-site (on a struct field or interface method's doc comment) — and the schema builder applies each at a distinct point in the build pipeline.

Where each override is consumed

Annotation	Scope	Consumed by	Applied at
swagger:type X	decl-site	classifierNamedTypeOverride(s.Decl.Comments, ps)	buildFromDecl — before kind-dispatch
swagger:type X	decl-site, reached via field reference	same classifier, walked from the field's *types.Named decl	buildNamedType / buildNamedArrayLike / classifierNamedBasic arms
swagger:type X	field-site	fieldDoc.TypeOverride populated by scanFieldDoc	applyFieldCarrier — after buildFromType
swagger:strfmt X	decl-site	per-arm classifiers (classifierTextMarshal, classifierNamedStructStrfmt, …)	dispatch-table arms
swagger:strfmt X	field-site	fieldDoc.StrfmtName	applyFieldCarrier
json tag ,string	field-site	fieldCarrier.isString	applyFieldCarrier

Ordering inside applyFieldCarrier (last-write-wins)

buildFromType(propType, ps)
    → isString   sets {type: string, format: <kept>}
    → StrfmtName sets {type: string, format: <X>}
    → TypeOverride  resets ps; runs SwaggerSchemaForType(X) or falls back to Underlying()
    → applyBlockToField  consumes everything else (description, validations, extensions)

A field that picks up multiple overrides resolves by the last write. Mixing ,string and swagger:strfmt and swagger:type is misuse in source — the precedence simply prevents accidental contradictions from corrupting the schema mid-build.

Decl-site swagger:type fallthrough — known leaf vs unknown leaf

classifierNamedTypeOverride tries SwaggerSchemaForType(name, tgt):

• known leaf (object, string, integer, boolean, number): classifier returns (handled=true, recurse=false). The override terminates the dispatch — the schema is left typed as the user asked.
• unknown leaf (array, badvalue, …): SwaggerSchemaForType errors, classifier returns (handled=true, recurse=true). The caller falls back to s.Decl.ObjType().Underlying() so item shapes are filled from the Go-level shape. For type X json.RawMessage // swagger:type array, the Underlying is []byte and the result is {type: array, items: {integer, uint8}}.

The same (handled, recurse) contract applies at field-reference sites via classifierNamedArrayLike (the wrapper-decl path that inlines into the field's schema) and at the decl-site via buildFromDecl (the wrapper's own top-level definition).

Two scopes, two effects, independent precedence

The same swagger:type annotation at the decl-site decides what the type's definitions entry emits; at the field-site it decides what one specific field emits, regardless of its Go type's natural shape.

A field referencing a wrapper-decl honours both overrides — the wrapper's own definition reflects its decl-site override, and the referring field reflects its own field-site override (which wins locally by ordering). Both layers compose without mutating each other.

scanFieldDoc and the AnnType filter

scanFieldDoc is the field-level walker that pre-extracts every classifier signal in one pass over ParseBlocks(afld.Doc). Most annotations (ignore, name, strfmt, allOf) go through the lexer's firstIdent arg-classifier, which already produces a single-token arg — no filter needed.

AnnType is the exception: its arg-classifier TrimSpaces the whole rest of the line, so prose noise like swagger:type so the scanner emits … reaches b.AnnotationArg() as a multi-word string. scanFieldDoc filters those out with an inline strings.ContainsAny(name, " \t") check, mirroring the filter inside findAnnotationArg (walker_classifiers.go). The enhancements/named-basic fixture documents the v1 trap this filter protects against.

Recognizer skipExt plumbing

applySpecialType and applyStdlibSpecials take a skipExt bool parameter that gates any vendor-extension writes the recognizers would otherwise emit. Currently only recognizeError writes one (x-go-type: error); the other recognizers (recognizeTime, recognizeAny, recognizeRawMessage, recognizeUUID) are purely type / format mutations and don't consult skipExt. All eight schema-internal call sites pass s.skipExtensions so the recognizer subsystem honours the same SkipExtensions flag as the rest of the builder.

---

§ref-override — field-level overrides on a $ref'd field

applyToRefField handles the case where a struct field's Go type resolves to a named type whose schema lives in definitions (ps.Ref set). Field-level sibling content — description, pattern, enum, example, x-* extensions — cannot ride alongside $ref per JSON-Schema-draft-4: the ref predates and replaces siblings. The correct shape is an allOf compound:

{
  "description": "...",
  "allOf": [
    { "$ref": "#/definitions/Parent" },
    { "...override validations and extensions..." }
  ]
}

Per-keyword landing rules

• required: writes to enclosing.Required (it's a parent-side concern, not a sibling of the $ref).
• Description rides the outer allOf compound when any field-level content is present, including just the description itself. The DescWithRef option (below) covers the only-description edge case.
• Validations (maximum, pattern, enum, …) land on allOf[1] — the override schema arm.
• Vendor extensions (x-* via the extensions: raw block) are lifted onto the outer compound, not nested inside allOf[1]. Reason: x-* siblings of $ref should live at the same level as scanner-derived metadata (x-go-name, x-go-package) for consistency. See refOverrideCollector's flag explanation below.

The DescWithRef toggle and the description-only case

scanner.Options.DescWithRef controls how a description-only override is emitted:

• DescWithRef=true: a $ref'd field whose only field-level decoration is a description produces a single-arm allOf compound. The description rides the outer parent so JSON-Schema consumers see it alongside the $ref.
• DescWithRef=false (the default — matches v1 strict behaviour): the description is dropped and a bare $ref is emitted. Users who want the description preserved via the JSON-Schema-correct compound shape opt in explicitly.

When validations or user-authored extensions are present, the allOf wrap is mandatory regardless of the flag — the override would be lost otherwise.

refOverrideCollector — accumulate-then-decide

The collector accumulates field-level overrides into a scratch schema so applyToRefField can pick the final shape after the Walker has finished firing. Two flags track what was collected:

• collectedValidation — a JSON-Schema validation keyword fired (maximum, pattern, enum, …). When true, the override arm (allOf[1]) is emitted carrying the validation.
• collectedExtension — a vendor extension fired. When true, the collected extensions are lifted onto the outer compound (not the override arm) so x-* keys live alongside the scanner-derived x-go-name / x-go-package.

Splitting the collector out of applyToRefField keeps the per-shape Walker callbacks short and the orchestrator's cognitive complexity in check. The Walker fires; the collector records; the outer function shapes.

applyPattern — best-effort RE2 hint

applyPattern stores a regex pattern unconditionally — JSON Schema regex's grammar is broader than Go's RE2 (lookaheads, named groups, etc.) and a user may rely on a downstream validator that accepts the wider syntax. A best-effort regexp.Compile check runs alongside: if the pattern is invalid against RE2, a SeverityWarning diagnostic surfaces the issue without dropping the value.

The diagnostic rides on CodeInvalidAnnotation rather than a dedicated CodeInvalidPattern. Reason: it's the closest existing class for "the value is grammatically valid but semantically off." A dedicated code can land alongside a broader pattern-hygiene pass when one materialises.

---

§simple-schema-mode — SimpleSchema build mode for non-body params and response headers

OAS v2 distinguishes a full Schema (body parameters, response bodies, top-level definitions) from a SimpleSchema that applies to non-body parameter sites and response headers:

• A SimpleSchema is a parameter with in: {query, path, header, formData} (anything except in: body) or a response header.
• SimpleSchema vocabulary is a restricted subset of Schema — $ref, allOf / oneOf / anyOf / not, properties / additionalProperties, object-level required, discriminator, readOnly, xml, externalDocs are all forbidden.
• The notion disappears in OAS v3.

References: https://swagger.io/specification/v2/#parameter-object and https://swagger.io/specification/v2/#header-object.

The schema builder offers WithSimpleSchema(tpe, tgt, in) as the third Build mode (parallel to WithType / WithDefinitions). The caller signals SimpleSchema-context explicitly via the option; the builder no longer infers it from tgt.In(). See §build-entry for the option table.

Allowed keyword surface

Common between parameters and headers:

Keyword	Notes
type	{string, number, integer, boolean, array}; file for params only
format	same vocabulary as full Schema
items	required when type: array; nested SimpleSchema, recursive
collectionFormat	{csv, ssv, tsv, pipes, multi}, default csv — SimpleSchema-only
default
maximum, exclusiveMaximum, minimum, exclusiveMinimum, multipleOf
maxLength, minLength, pattern
maxItems, minItems, uniqueItems
enum
x-* vendor extensions

Parameter-only extras: allowEmptyValue (only when in ∈ {query, formData} — forbidden on path and header); type: file (only when in: formData).

Response headers exclude file and allowEmptyValue entirely.

"Catch at exit" contract

The schema builder does not pre-filter inputs in SimpleSchema mode. *types.Struct and *types.Interface are allowed to enter the resolution pipeline because they can legitimately resolve to a SimpleSchema-legal primitive:

• time.Time → {string, date-time} (stdlib recognizer)
• TextMarshaler → {string, format} (textmarshal shortcut)
• json.RawMessage → empty {} (any)
• user-driven overrides (swagger:strfmt, swagger:type, the swagger:alias per-type opt-in via §classifier-walkers' classifierNamedBasic) win the cascade as they always do

The exit validator (validateSimpleSchemaOutcome in simpleschema.go) inspects the resolved target after buildFromType returns:

• accept when shape.Type is in {"", string, number, integer, boolean, array}, plus file if in == "formData" (empty means "any" — json.RawMessage ends up here).
• otherwise emit CodeUnsupportedInSimpleSchema (severity SeverityWarning) and reset the target.

The reset wipes the target back to empty {} (no Type, no Format, no Ref) rather than degrading to {type: string} — empty is honest about the failed resolution and avoids silently mistyping a complex shape as a string.

SimpleSchemaProbe interface

The target must expose three methods so the validator can inspect the post-resolution shape and reset it on violation:

type SimpleSchemaProbe interface {
    SimpleSchemaShape() *oaispec.SimpleSchema
    HasRef() bool
    ResetForViolation()
}

Implemented structurally by paramTypable in internal/builders/parameters and (forthcoming with M2) headerTypable in internal/builders/responses. Consumers don't need to import the schema package — the interface is satisfied by method set.

A target that doesn't implement SimpleSchemaProbe is trusted: the validator no-ops. A nil SimpleSchemaShape is also trusted — the caller chose SimpleSchema mode for something that can't surface a violation, by intent.

Knock-on cleanups this contract enables

Once WithSimpleSchema carries the mode flag explicitly, the builder can stop sniffing tgt.In():

• classifierNamedBasic's primitive-inline arm (see §classifier-walkers) now keys on s.simpleSchema instead of the v1-era isAliasParam(tgt) predicate. That closes the in: header omission documented as a resolved quirk below (§quirks-resolved).
• The swagger:alias per-type author override stays orthogonal — it bypasses the model-ref pipeline regardless of mode.

Walker keyword gating under SimpleSchema mode

The single source of truth for the SimpleSchema vocabulary is internal/builders/handlers.IsSimpleSchemaKeyword(name). The package-level simpleSchemaAllowed map enumerates every keyword legal on a non-body parameter, response header, or items chain within either, per OAS v2's Parameter Object and Header Object tables.

schemaBoolHandler consults this predicate when s.simpleSchema == true:

• Full-Schema-only Bool keywords (readOnly, discriminator) trigger a CodeUnsupportedInSimpleSchema SeverityWarning diagnostic and the write is skipped. Even if the path lands on a throwaway scratch schema (the common case under SimpleSchema mode — paramTypable.Schema() returns nil for non-body), the diagnostic still surfaces the misuse to the author.
• required: is accepted by the predicate (it IS SimpleSchema-legal as a parameter-level boolean) but the schema's Bool handler skips it silently under SimpleSchema mode anyway. The parameter-level write — param.Required = val — lives in parameters/walker.go:paramRequiredBool. Headers don't carry required: at all. The schema walker's full-Schema semantics (object-level required-array via enclosing.Required[name]) don't fit the SimpleSchema shape.
• Number / Integer / String / Raw / Extension dispatchers are unchanged: all the keywords they handle are SimpleSchema-legal.

The IsSimpleSchemaKeyword set is locked down by unit tests in the handlers package — any future addition (or removal) of a SimpleSchema keyword has to update the test alongside the map, so the contract can't drift silently.

---

§classifier-walkers — per-call-site classifier walkers and findAnnotationArg's single-word filter

The schema builder dispatches user-classifier annotations (swagger:strfmt, swagger:type, swagger:enum, swagger:default, swagger:allOf, swagger:alias) via per-call- site walker functions in walker_classifiers.go. The shape is deliberate:

• One walker per call site. Each walker is named for the call- site context (e.g. classifierTextMarshal, classifierNamedBasic, classifierNamedArrayLike). Its godoc documents which swagger:<kind> classifier annotations it consumes — explicit per-site contract instead of a single catch-all dispatcher.
• Reads through the ParseBlocks cache. Each walker calls s.ParseBlocks(cg) so a single CommentGroup is parsed exactly once per build regardless of how many walkers inspect it.
• Site-local writes. The walker performs the call-site's writes directly onto the typable target — both lookup and side effect encapsulated.

Correctness first; if a future re-read finds genuine redundancy, the factorisation is mechanical and safer last.

findAnnotationArg and the single-word filter

findAnnotationArg(cg, kind) returns the first positional argument of the first Block of the given annotation kind, filtered to non-empty single-word arguments:

if strings.ContainsAny(arg, " \t") {
    continue
}

The single-word filter only matters for annotation kinds whose lexer arg-classifier doesn't already split on whitespace — namely AnnType (via argTypeRef) and AnnDefaultName (via argDefaultValue). Kinds that go through firstIdent in the lexer (AnnStrfmt, AnnName, AnnAllOf, AnnModel, AnnResponse, AnnEnum) already produce single-token args, but the filter is harmless on those.

The check matches v1's de-facto \S+-anchored capture, which silently rejected prose lines that happened to open with swagger:<kind> followed by a sentence. The fixtures/enhancements/named-basic fixture documents this trap with a swagger:type so the scanner emits ... prose line preceding the real swagger:type string annotation — without the filter, the prose's "so" would pre-empt the real arg "string".

findAnnotationArg reads through the per-Builder ParseBlocks cache, so the lookup is parse-once per CommentGroup and the ParseAll-aware multi-annotation case still surfaces every annotation of interest.

Walker inventory

Walker	Call site	Consumes
classifierTextMarshal	buildFromTextMarshal end-of-pipe	swagger:strfmt
classifierNamedTypeOverride	buildFromType named fallback, buildFromStruct pre-pass	swagger:type
classifierNamedBasic	buildNamedBasic	cascade: swagger:strfmt → swagger:enum → swagger:default → swagger:type → swagger:alias (the alias arm doubles as the SimpleSchema-mode primitive-inline branch — see §simple-schema-mode)
classifierNamedArrayLike	buildNamedArray / buildNamedSlice	swagger:strfmt, swagger:type
classifierAliasTargetStrfmt	buildNamedAllOf (struct + interface arms)	swagger:strfmt
classifierStructPreBuildType	buildFromStruct top	swagger:type
classifierNamedStructStrfmt	buildNamedStruct strfmt-first branch	swagger:strfmt
scanFieldDoc	field-level FieldWalker (applyFieldCarrier)	swagger:ignore, swagger:name, swagger:strfmt, swagger:type (with the same single-word filter), swagger:allOf

---

§quirks — known behavioural caveats

Entries are split into two groups: Resolved in this refactor lists quirks fixed in the current pass with a short note on how, so a future reader can find the change and the rationale. Still open lists quirks the refactor either deliberately did not touch (deferred to v2 / design call required) or documents as intentional behaviour.

---

§quirks-resolved — fixed in this refactor

✅ recognizeRawMessage now emits an empty schema (was {type: object})

json.RawMessage is []byte underneath but JSON-marshals as arbitrary JSON. The recognizer now emits an empty schema ({}), which in Swagger 2.0 / JSON Schema means "any type" — the most faithful representation of RawMessage's contract. Previous behaviour emitted {type: object} as a narrower approximation.

Fix: recognizeRawMessage arm in applySpecialType rewritten to call _ = target.Schema() (the "any" pattern used by recognizeAny) instead of target.Typed("object", ""). Golden delta captured in go123_special_spec.json's Message property.

✅ recognizeError x-go-type extension now honours SkipExtensions

The error arm of applySpecialType writes x-go-type: error in addition to typing the target as {string, ""}. This is for downstream tooling that wants to detect the Go-error origin. The write is now gated by the skipExt argument threaded into applySpecialType / applyStdlibSpecials from s.skipExtensions, so SkipExtensions=true suppresses it like any other vendor extension.

Fix: added skipExt bool parameter to applySpecialType and applyStdlibSpecials; gated the target.AddExtension(...) call in the recognizeError arm. Eight schema-internal call sites updated to pass s.skipExtensions. Golden-neutral (no existing fixture combines the error shape with SkipExtensions=true).

✅ Field-level swagger:type on json.RawMessage fields

A struct field of type json.RawMessage with a field-level swagger:type object / swagger:type array annotation now produces the user-specified shape instead of silently emitting the recognizer default. Pre-fix: scanFieldDoc only consumed ignore / name / strfmt / allOf, so field-level swagger:type was dropped before applyBlockToField ran.

Fix: added TypeOverride to fieldDoc; scanFieldDoc consumes AnnType with a single-word filter (mirroring findAnnotationArg, since AnnType uses TrimSpace and can carry prose on noise lines); applyFieldCarrier applies the override after buildFromType — tries SwaggerSchemaForType(name, …) first, falls back to buildFromType(c.propType.Underlying(), …) on unknown leaves like "array" so item shapes are computed from the Go type. Fixture fixtures/enhancements/raw-message-override/ (case C); golden enhancements_raw_message_override.json.

✅ Wrapper-decl swagger:type honoured at top-level definition

A named wrapper of json.RawMessage (or any other type the recognizer would otherwise short-circuit) decorated with swagger:type on the decl now emits the user-specified shape at its own top-level definition, not just at field reference sites. Pre-fix: only the field-reference path consulted the wrapper's swagger:type; the top-level definition emitted an empty schema because buildFromDecl dispatched on ti.Type (the RHS) and the RHS recognizer fired before any wrapper-side classifier could.

Fix: buildFromDecl now calls classifierNamedTypeOverride on s.Decl.Comments before the kind-dispatch. Known leaves (object, string, …) terminate; unknown leaves (array) fall back to s.Decl.ObjType().Underlying() so items / properties are filled from the Go-level shape. Isolation fixture fixtures/enhancements/wrapper-decl-type-override/ — BareWrapperObject / BareWrapperArray.

✅ in: header parameters now inline named-basic types

Pre-fix, classifierNamedBasic's primitive-inline arm only fired for in: query / in: path / in: formData — header was silently omitted from the isAliasParam predicate (carried over verbatim from v1's parsers.IsAliasParam). So a header parameter typed as a named string type SessionID string resolved through FindModel → makeRef and emitted {$ref: "#/definitions/SessionID"} — invalid under OAS v2 SimpleSchema, which forbids $ref on non-body parameter sites.

Fix: the isAliasParam(tgt) In()-sniffing predicate replaced with the M1 s.simpleSchema flag (set by WithSimpleSchema). The parameter bridge now wires WithSimpleSchema for all four non-body locations uniformly (query / path / header / formData), so the primitive-inline arm covers header automatically. The predicate function deletes (sole consumer).

Two paths through the arm remain orthogonal: the SimpleSchema flag is caller-driven (parameter/header build mode); swagger:alias on the decl is a per-type author override. Either triggers SwaggerSchemaForType(underlying basic name) on the typable.

Isolation fixture fixtures/enhancements/header-named-basic/ and test internal/integration/coverage_header_named_basic_test.go pin the post-fix shape. Responses won't pick up the same fix until M2 wires WithSimpleSchema on header-field builds; the response edges fixture covers a different (strfmt-tagged) shape already.

---

§quirks-open — still open

🟡 Named-strfmt + swagger:model combo (deferred)

When the author combines swagger:strfmt with swagger:model on the same type, the FIELD reference inlines as {string, format} (via the strfmt classifier) but the TOP-LEVEL definition body is still emitted from walking the underlying struct.

Reproduction. Fixture fixtures/enhancements/named-struct-tags-ref/types.go declares PhoneNumber with both swagger:strfmt phone and swagger:model, used by Contact.Phone. The golden enhancements_named_struct_tags-ref.json captures the observable inconsistency:

• Field site: {type: "string", format: "phone"} — strfmt wins.
• Top-level definition: {type: "object", properties: {CountryCode, Number}} — the struct walk wins; the strfmt annotation is ignored at decl time.

The author asked for "named strfmt" (a reusable PhoneNumber definition rendered as a formatted string) but gets an inconsistent pair: the field says string, the definition says object.

Attempted fix and reasons it reverted. The first attempt (referred to as "Option 1") would have:

1. Detected swagger:strfmt on the decl in buildDeclNamed and emitted {string, fmt} instead of walking the struct body.
2. In buildNamedStruct, when the target also has swagger:model, emitted $ref instead of inlining the strfmt.

This was reverted before merge because:

• Pre-existing fixtures in fixtures/goparsing/classification/transitive/mods/aliases.go use the same swagger:strfmt + swagger:model combination on defined-from-time.Time types (e.g. SomeTimeType time.Time). The existing tests (TestAliasedTypes, TestAliasedModels) assert the inline baseline (scantest.AssertProperty(..., "string", ...)) rather than a $ref. Option 1 flips these to $ref, requiring coordinated test updates.
• The decl-level StrfmtName check also over-fires on slice / array / map underlyings: type SomeTimesType []time.Time with swagger:strfmt date-time should emit {array, items: {string, date-time}}, not flatten to {string}. A correct fix would gate the check on struct-underlying first, then symmetrically consider whether buildNamedSlice / buildNamedArray / buildNamedMap should also route through $ref under the swagger:model combination.

The surface area is wider than the Option 1 code change suggested, and the existing test coverage of the combination is entangled with the inconsistency itself.

Why deferred. The combination is niche, the footgun is narrow (you get what you asked for on one side of the indirection, not both), and v2's annotation redesign can reshape the contract without carrying this legacy. A focused decision on "named strfmt" semantics belongs in the v2 design, not a bug-fix pass.

The named-struct-tags-ref fixture and its golden are checked in as a deliberate marker — the golden captures the observable inconsistency (inline field + struct-body definition) so future work on this decision has a failing test to anchor against.

🟦 interface{} literals (documented behaviour)

A bare interface{} field hits the *types.Interface arm of buildFromType (anonymous, not Named). It produces an empty schema. The user-named type X interface{} is a *types.Named with empty Underlying() and emits as $ref to a definition with no properties — JSON-equivalent to "any object" in v1. Behavioural; changing it would break consumers.

🟦 Generic declarations vs instantiations (documented behaviour)

Generic declarations (e.g. GenericSlice[T any] []T) are processed but their schemas are essentially empty — the type parameter T is filtered out by UnsupportedBuiltinType as a *types.TypeParam. Generic instantiations (e.g. a field of type GenericSlice[int]) emit correctly with the substituted underlying via the TypeArgs short-circuit (§dissolve-named). No bug — generic decls without a concrete instantiation simply have no representable schema.

🟡 Cross-package definition-name collisions silently overwrite

buildFromDecl writes the top-level schema as s.definitions[s.Name] = schema, keyed only by the Go identifier (decl.Names()[0]). When two packages in a single scan declare a type with the same identifier — pkg/a.User and pkg/b.User — both map to definitions["User"] and the second build silently overwrites the first. The output spec carries only one User, with no record of the collision and no signal of which package won.

The existing nameByJSON (propOwner) map in field emission is not a defense against this case: it tracks JSON property names within a single struct's field set plus its embeds (for the ambiguous-embed diagnostic), not type-level identifier conflicts across packages.

Target shape

A proper fix needs three pieces:

1. Detection — at write time, recognise the case "definition key already exists with non-empty schema and originates from a different package" (use x-go-package, or stash origin in the Builder).
2. Diagnostic — emit CodeNameConflict (severity SeverityWarning minimum, possibly SeverityError under strict mode) carrying both (pkg, name) pairs.
3. Policy — open design call:
   • a. Rename — prefix loser(s) with a stable short-package (e.g. a_User, b_User). Stable but ugly; needs all $refs to follow the rename — cross-cutting.
   • b. Skip + warn — keep the first writer, drop subsequent ones, emit a warning. Predictable but lossy.
   • c. Fail the build — under strict mode, treat as an error. Forces the author to rename in source. Cleanest semantics, most disruptive.

Why deferred

Each policy choice changes the contract for downstream code generators (go-swagger, oapi-codegen, …) — they have assumptions about definitions keys matching exported Go names. The "rename" path additionally requires every $ref writer in the builders to consult a rename map; the surface is wide.

For multi-package scans where the author controls both packages, the workaround today is to scope scans to one package per spec, or to rename one of the colliding types at the source. A future strict-mode flag (e.g. Options.StrictNameConflicts) could enable option (c) without breaking existing scans.

🟡 Stale x-go-enum-desc after a field-level enum override

When a field uses a type marked swagger:enum TypeName and carries its own enum: … override, v1 mutates the schema in place: it replaces Enum, strips the inherited x-go-enum-desc, and trims the matching description suffix. This is lossy — the per-value docs contributed by TypeName are silently discarded.

Concretely, given (fixture fixtures/enhancements/enum-overrides/, case E):

// swagger:enum PriorityE
type PriorityE string

const (
    PriorityELow  PriorityE = "low"    // low-priority requests
    PriorityEMed  PriorityE = "medium" // medium-priority requests
    PriorityEHigh PriorityE = "high"   // high-priority requests
)

type NotificationE struct {
    // Inline enum provides a narrower set than the const block.
    //
    // enum: urgent, normal
    Priority PriorityE `json:"priority"`
}

v1 emits:

priority:
  type: string
  enum: [urgent, normal]   # the override wins
  description: "Inline enum provides a narrower set than the const block."
  # x-go-enum-desc removed by clearStaleEnumDesc
  # PriorityE's per-value doc lines silently dropped from description

The cleanup runs reactively from schemaValidations.SetEnum (typable.go) via clearStaleEnumDesc (extensions.go). It treats any x-go-enum-desc present at SetEnum time as inherited (and therefore stale once Enum is replaced), deletes it, and trims the matching suffix off Description. The TrimSuffix dance is fragile — it relies on the enum-desc pipeline having appended the doc lines as a literal suffix — but it works under v1's emission discipline.

Target shape (allOf composition)

OpenAPI 2.0 supports allOf for schema composition, so the cleaner model does not have to wait for OAS 3. The replacement shape is:

# PriorityE promoted to a top-level definition:
definitions:
  PriorityE:
    type: string
    enum: [low, medium, high]
    description: |
      low: low-priority requests
      medium: medium-priority requests
      high: high-priority requests
    x-go-enum-desc: |
      low: low-priority requests
      medium: medium-priority requests
      high: high-priority requests

  NotificationE:
    type: object
    properties:
      priority:
        description: "Inline enum provides a narrower set than the const block."
        allOf:
          - $ref: '#/definitions/PriorityE'   # inherited enum + per-value docs
          - enum: [urgent, normal]            # the override

Each branch keeps its own concern:

• the $ref branch carries PriorityE's full schema (values + docs + x-go-enum-desc), untouched and reusable by every field that references PriorityE;
• the inline branch carries the narrowing override only.

No mutation of the inherited schema, no TrimSuffix dance. Validator semantics for enum-narrowing allOf aren't perfectly uniform across tools, but for the documentation / code-gen use cases codescan feeds (go-swagger, oapi-codegen, redoc, …) this composition preserves both layers cleanly.

Prerequisites for the migration (both currently missing)

1. Promote unannotated swagger:enum types to top-level definitions so the $ref branch has a target. Today they exist only as inlined fragments on each referring field.
2. Move override detection from SetEnum (validation hook) to the field-emission path, so the override is composed alongside the inherited schema instead of mutating it after the fact.

Until both land, clearStaleEnumDesc stays in place. The TODO in extensions.go flags it as the replacement target.