Sub-languages

The annotation body grammar is not a single language — it’s a top-level keyword grammar that embeds several smaller languages inside specific body keywords. Each embedded language has its own shape rules.

This document catalogs the embedded languages and how they fit together. For the per-keyword surface, see keywords.md; for the formal grammar that hosts them, see grammar.md.

Prose classification

Comment lines that don’t match any keyword head OR YAML fence OR annotation marker are classified as prose — free-form text. The lexer splits prose into two token kinds:

TITLE — the first paragraph of prose, expected to fit on a short summary line.
DESC — every prose paragraph after the title (or following a blank line within the first paragraph).

Three heuristics decide the title-vs-desc boundary, evaluated in order. The first to fire wins:

Blank-line split. Any blank line inside the prose run ends the title paragraph and starts the description.
Closing punctuation. If the first prose line ends with Unicode punctuation (., ?, !, …, :, …), the title is just that one line; everything after becomes description.
Markdown ATX heading. If the first prose line matches markdown’s # Heading shape, the # markers are stripped and the remaining text becomes the title.

When no heuristic fires, the entire prose run is title (the schema builder later collapses to a description-only schema when appropriate).

`Package <name>` prefix strip

The swagger:meta annotation’s title comes from the package doc comment, which by Go convention starts with Package <name>. The spec builder strips that prefix before publishing:

// Package petstore Petstore API.
//
// Description of the petstore service.
//
// swagger:meta
package petstore

Produces info.title = "Petstore API." (the Package petstore prefix stripped) and info.description = "Description of the petstore service."

Only the capital-P Package form is recognised — author prose like “package this carefully” is not chopped.

Comment-marker noise stripping

Block-comment routes (/* swagger:route … */) typically carry indented continuation lines:

/* swagger:route POST /pets pets createPet

	Create a pet based on the parameters.

	Consumes:
		- application/json
*/
func CreatePet() {}

The lexer strips the leading whitespace (\t, *, /, |) per line via trimContentPrefix before classification.

Markdown semantics that survive

Dash lists in descriptions are preserved verbatim. A line starting with - foo lands in the description as "- foo" (not "foo").
--- lines open a YAML fence — see YAML extensions below.

Flex-list

Body keywords that publish a flat list of tokens (schemes:, consumes:, produces:) accept multiple surface forms uniformly. The unified reader is Property.AsList().

Accepted forms

# Inline, comma-separated
Schemes: http, https

# Multi-line, indented bare lines
Schemes:
  http
  https

# Multi-line, YAML-style dash markers
Schemes:
  - http
  - https

# Inline value plus indented continuation
Schemes: http
  - https

# All combinations of the above
Consumes: application/json, application/xml
  - application/protobuf

All five forms produce the same ["http", "https"] (or ["application/json", "application/xml", "application/protobuf"]) output.

Algorithm

For each input line — Property.Value first (if non-empty), then each line of Property.Body:

Trim surrounding whitespace.
Drop a leading - YAML marker if present.
Re-trim whitespace.
Comma-split.
Trim each token; drop empties.

Aggregate into a single slice in source order.

What flex-list does NOT touch

Enum values (enum: ...) — their elements may themselves be complex (JSON arrays, quoted strings with commas). enum: keeps its raw-value path; the value coercion layer handles array / comma-list / multi-line shapes per the schema type.
Parameters chunks — the + name: chunk grammar is not a simple token list; see §parameters.
YAML structural bodies — securityDefinitions:, extensions:, infoExtensions: parse the body as YAML directly; their structure isn’t a flat list. See §yaml-extensions.

Parameters

The Parameters: body in swagger:route and swagger:operation carries a sequence of parameter declarations separated by + name: chunks (the + is the chunk-start sigil; - is accepted as an alias for forward compatibility with proper YAML).

Chunk shape

Parameters:
  + name: id
    in: path
    type: integer
    description: the item identifier
    required: true
  + name: limit
    in: query
    type: integer
    minimum: 1
    maximum: 100
    default: 20
  + name: body
    in: body
    type: User
    required: true

Per-chunk fields

The fields are classified into head fields (consumed by the orchestrator to populate the *spec.Parameter shell) and validation fields (lowered to grammar properties and dispatched through the standard validation pipeline).

Head fields:

Field	Lands on	Notes
`name:`	`parameter.name`	Required. Identifies the parameter.
`in:`	`parameter.in`	One of `path` / `query` / `header` / `body` / `formData`. `form` accepted as an alias for `formData`.
`type:`	`parameter.type` (for SimpleSchema) or determines the body $ref	For non-body: one of `string` / `integer` / `number` / `boolean` / `array`. For body: a Go ident referring to a `swagger:model`-declared type, optionally with `[]` array prefixes (`[][]Pet`). `bool` accepted as an alias for `boolean`.
`format:`	`parameter.format` or `parameter.schema.format`	Free-form string. Applied after validation dispatch so it doesn’t interfere with default/example coercion.
`description:`	`parameter.description`	Free-form prose.
`required:`	`parameter.required`	Boolean.
`allowempty:` / `allowemptyvalue:`	`parameter.allowEmptyValue`	Boolean.

Validation fields: any other recognised keyword — min, max, minLength, maxLength, minItems, maxItems, pattern, unique, collectionFormat, default, example, enum. These are looked up via grammar.Lookup (which accepts canonical names + aliases) and dispatched through the standard handlers seam.

Empty chunks and unknown keys

A bare + (or -) sigil with no follow-up content emits a CodeInvalidAnnotation diagnostic and is dropped. The legacy parser silently emitted an empty Parameter{} object — current behaviour rejects it.
Unknown keys (typos like defualt:) emit CodeInvalidAnnotation and drop. The legacy parser silently discarded them.

Body parameters

When in: body, the orchestrator looks up type: as either:

A primitive (string, integer, number, boolean, array, object) — emits a typed schema with the primitive on parameter.schema.type.
A Go ident — emits a $ref to #/definitions/<Ident>. With [] prefixes, wraps the ref in nested array schemas.

Validation properties on a body chunk apply to the schema, gated by the schema’s resolved type via checkShape. A min: 0 on a body chunk with type: Pet (object) emits CodeShapeMismatch and drops; a min: 0 with type: integer lands on the schema’s minimum.

Validation on SimpleSchema (non-body) parameters

For in: other than body, validation properties apply directly to the parameter (not to a sub-schema). Type-gating still applies: minLength on type: integer emits a diagnostic and drops.

Responses

The Responses: body in swagger:route carries one response declaration per line. Each line has the shape:

<code>: <token>*

where <code> is default (case-insensitive) or a decimal HTTP status code, and <token> is either a tag:value form or an untagged token.

Recognised tags

Tag	Value shape	Lands on
`body:`	Go ident with optional `[]` prefixes (`body:[]Pet`)	A `$ref` to `#/definitions/<name>`, array-wrapped per `[]` count
`response:`	Go ident referring to a `swagger:response`-declared type	A `$ref` to `#/responses/<name>`
`description:`	Free-form prose (rest of line)	`response.description`

Untagged token rules

The first untagged token defaults to a response ref. The orchestrator resolves it against the operation’s responses map first, then falls back to definitions — if found in definitions (not responses), it’s silently promoted to a body ref.
Subsequent untagged tokens accumulate into the description.

Examples

Responses:
  200: User the user as returned                  # untagged → response="User", desc="the user as returned"
  200: body:User the user                         # body ref + description
  200: response:userResponse the user             # named response ref
  201: body:Pet the created pet
  404: description: not found
  default: response:genericError
  default: body:[]ErrorList the error list        # array-wrapped body ref

Diagnostics

Unknown tag (200: weird:value) — emits CodeInvalidAnnotation and drops the line.
Duplicate body/response tags on one line (200: body:Pet response:errors) — emits CodeInvalidAnnotation; the line drops.
Space-separated body Foo (instead of body:Foo) — detected as a likely typo and dropped with diagnostic. The legacy parser silently treated it as response="body" (a dangling ref to a non-existent response).
Unresolvable response ref — when a response name appears in neither responses nor definitions, the line drops with diagnostic. The legacy parser emitted a dangling $ref.

Empty value lines

A line like 204: with nothing after the colon produces a Response with the code and an empty description. This is intentional — some authors want a 204 No Content with no body and no description.

YAML extensions

Several body keywords parse their body as YAML directly:

extensions: and infoExtensions: — a YAML map of x-* entries.
securityDefinitions: — a YAML map matching OAS v2’s securityDefinitions shape.
externalDocs: — a YAML map with description and url keys.

Extension typing

Extension values are NOT coerced to strings — they preserve their YAML-typed form: bool, float64, string, []any, or map[string]any for nested structures.

Extensions:
  x-feature-flags:
    - alpha
    - beta
  x-rate-limit:
    requests: 100
    window: 60
  x-internal: true
  x-version: 0.5

Produces (extract):

"x-feature-flags": ["alpha", "beta"],
"x-rate-limit": {"requests": 100, "window": 60},
"x-internal": true,
"x-version": 0.5

`x-*` name gating

Keys that don’t start with x- or X- emit a CodeInvalidAnnotation diagnostic and drop. The build still succeeds. Authors who relied on the legacy “hard error on non-x-*” behaviour see a diagnostic + a clean spec missing the typo’d key.

Extensions:
  x-good: 1
  not-good: 2   # → diagnostic, dropped

YAML body delimitation

The YAML extension bodies use indentation to delimit. A line that returns to the indentation level of the keyword head — or introduces a sibling keyword — terminates the body. The grammar also recognises --- fence pairs around the body (matching the swagger:operation YAML shape) and absorbs them silently.

Security requirements

The security: body (in swagger:meta, swagger:route, and swagger:operation) carries OAuth-style security requirements where each line is one requirement.

Shape

Each line: schemeName: scope1, scope2, …

schemeName matches a scheme declared in securityDefinitions.
Scope list is comma-separated; trimmed; empties dropped.
An empty scope list (schemeName:) means “this scheme is required, no scopes.” Common for apiKey and basic.

Example

Security:
  api_key:
  oauth2: read, write
  oauth2: admin

Produces:

"security": [
  {"api_key": []},
  {"oauth2": ["read", "write"]},
  {"oauth2": ["admin"]}
]

Each requirement is a single-key map; the array is an OR relationship (the request satisfies security if it matches ANY entry).

Contact / License

Inline single-line meta keywords with structured value parsing.

Contact

The contact: value carries up to three components: name, email, URL. Recognised forms:

Contact: Name <email@example.com> https://example.com
Contact: Name <email@example.com>
Contact: https://example.com
Contact: <email@example.com>

The grammar splits the value on the first URL prefix it finds (https://, http://, ftps://, ftp://, wss://, ws://), then parses the prefix portion as Name <email> via Go’s net/mail.ParseAddress.

A malformed Name <email> head (e.g., unbalanced angle brackets) surfaces as an error from Block.Contact(); the meta builder propagates it as a build failure.
An empty contact line produces an empty Contact value (no error, no diagnostic — equivalent to omitting the keyword).

Aliases: contact info, contact-info.

License

The license: value is split similarly:

License: Apache 2.0 https://www.apache.org/licenses/LICENSE-2.0
License: MIT
License: https://opensource.org/licenses/Custom

Same URL-prefix detection. Everything before the URL is the license name; the URL (when present) is the license URL. Either part may be empty.

License does NOT use mail.ParseAddress — the name is taken as raw text up to the URL boundary.

Sub-language interactions

Two interaction points worth flagging:

Block-comment continuation lines and the parameters/responses sub-languages. A /* swagger:route … */ block with Parameters: inside requires the chunk-start sigils (+ / - ) to be at the start of the trimmed line. Block-comment continuation noise (\t, *) is stripped first; if your editor inserts a * continuation marker, the lexer handles it transparently.
Flex-list and description: on a parameter chunk. description: is a head field, not a list — it does NOT comma-split. Authors who write description: foo, bar get a single description "foo, bar", not two descriptions. (This was a real ambiguity in older versions of go-swagger; the current grammar resolves it cleanly.)