#3552 bug: `resolve_schema_references` does not recurse into `additionalProperties`, leaving `$ref`s unresolved Issue created by edgarrmondragon Summary

resolve_schema_references

(in

singer_sdk/singerlib/schema.py

) recursively resolves

$ref

values inside

properties

,

patternProperties

,

items

,

anyOf

,

allOf

, and

oneOf

— but not

additionalProperties

. Any

$ref

nested under

additionalProperties

is left unresolved in the returned schema. When

OpenAPISchema.fetch_schema

is used and a component schema contains

additionalProperties: { "$ref": "..." }

, the resulting schema is not self-contained. Downstream validators (e.g.

jsonschema

) then fail with

PointerToNowhere

because the

$ref

points into

#/components/schemas/…

which no longer exists in the extracted standalone schema. Steps to reproduce Given an OpenAPI spec with a component like: "WorkspaceResource": { "type": "object", "properties": { "_links": { "type": "object", "additionalProperties": { "$ref": "#/components/schemas/Link" } } } } Fetching the schema via

OpenAPISchema.fetch_schema("WorkspaceResource")

returns: { "type": "object", "properties": { "_links": { "type": "object", "additionalProperties": { "$ref": "#/components/schemas/Link" # ← still unresolved } } } } Validating a real record against this schema with

jsonschema

then raises:

referencing.exceptions.PointerToNowhere: '/components/schemas/Link' does not exist within {...}

This manifests in practice when running the SDK's own built-in tap tests (

get_tap_test_class

) against a tap that uses

OpenAPISchema

—

test_tap_stream_record_matches_stream_schema

fails for every stream whose resource schema has

additionalProperties: { "$ref": "..." }

. Root cause

_resolve_schema_references

in

singer_sdk/singerlib/schema.py

defines a

_SchemaKey

class and recurses only into the keywords listed there: class _SchemaKey: ref = "$ref" items = "items" properties = "properties" pattern_properties = "patternProperties" any_of = "anyOf" all_of = "allOf" one_of = "oneOf"

additionalProperties

is absent. The JSON Schema specification also allows

$ref

to appear in

not

, `if`/`then`/`else`,

prefixItems

,

contains

,

propertyNames

, and `unevaluatedProperties`/`unevaluatedItems`, which are similarly unhandled — but

additionalProperties

is by far the most common in real-world OpenAPI specs. Expected behaviour

resolve_schema_references

should recurse into

additionalProperties

(and ideally the other missing applicator keywords) so the returned schema is fully self-contained. Proposed fix Add

additionalProperties

to `_resolve_schema_references`: if "additionalProperties" in schema and isinstance(schema["additionalProperties"], dict): schema["additionalProperties"] = _resolve_schema_references( schema["additionalProperties"], resolver, visited_refs, ) A more complete fix would also cover

not

,

if

,

then

,

else

,

prefixItems

,

contains

,

propertyNames

,

unevaluatedProperties

, and

unevaluatedItems

. Environment • Singer SDK version:

0.53.5

• Python version:

3.14.3

• Reproduced with a tap using

OpenAPISchema

+

StreamSchema

against a Spring HATEOAS API whose resource schemas use

additionalProperties: { "$ref": "..." }

for HAL link maps. meltano/sdk