This repository was archived by the owner on Apr 14, 2026. It is now read-only.
Commit 3a5f803
feat: multi-pass type name resolution (oapi-codegen#2213)
* feat: multi-pass type name resolution
I've been meaning to use this approach for a long time, because
the attempts at avoiding type collisions via structure suffixes or
prefixes work sporadically, at best.
Conflict resolution is fundamentally a global problem, not a local
problem when doing recursive traversal, so this PR splits the code
generation into two parts. First, the OAPI document structure is
traversed, and all the schemas that we generate are gathered up
into a list of candidates, then we do global conflict resolution
across the space of all schemas. This allows us to preserve the
functionality of things affected by schema name - `$ref`, `required`
properties, and so forth.
This fixes issue oapi-codegen#1474 (client response wrapper type colliding with
a component schema of the same name and improves issue oapi-codegen#200 handling
(same name across schemas, parameters, responses, requestBodies,
headers).
The new system is gated behind the existing `resolve-type-name-collisions`
output option. When disabled, behavior is unchanged, oapi-codegen exits
with an error. This flag is default false, so there is no behavior change
to oapi-codegen unless it's specified. All current test files regenerate
without any differences.
Added a comprehensive test which reproduces the scenarios in all the
PR's and Issues below, and adds a few more, to make sure that references
to renamed targets are also correct..
Key changes:
- gather.go: walks entire spec collecting schemas with location metadata
- resolve_names.go: assigns unique names via context suffix, per-schema
disambiguation, and numeric fallback strategies
- Component schemas are privileged and keep bare names on collision
- Client response wrapper types now participate in collision detection
- Removed ComponentType/DefinedComp from Schema struct
- Removed FixDuplicateTypeNames and related functions from utils.go
Obsoletes issues:
- oapi-codegen#1474 Schema name vs client wrapper (CreateChatCompletionResponse)
- oapi-codegen#1713 Schema name vs client wrapper (CreateBlueprintResponse)
- oapi-codegen#1450 Schema name vs client wrapper (DeleteBusinessResponse)
- oapi-codegen#2097 Path response type vs schema definition (Status)
- oapi-codegen#255 Endpoint path vs response type (QueryResponse)
- oapi-codegen#899 Duplicate types from response wrapper vs schema (AccessListResponse)
- oapi-codegen#1357 Schema vs operationId response (ListAssistantsResponse, OpenAI spec)
- oapi-codegen#254 Cross-section: requestBodies vs schemas (Pet)
- oapi-codegen#407 Cross-section: requestBodies vs schemas (myThing)
- oapi-codegen#1881 Cross-section: requestBodies with multiple content types
Obsoletes PRs:
- oapi-codegen#292 Parameter structures params postfix (superseded by context suffix)
- oapi-codegen#1005 Fix generate equals structs (superseded by multi-pass resolution)
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
EOF
)
* Fix linter issues
* fix: apply collision strategies in global phases to prevent oscillation
When multiple content types map to the same short suffix (e.g.,
application/json, application/merge-patch+json, and
application/json-patch+json all mapping to "JSON"), the per-group
strategy cascade caused an infinite oscillation: context suffix
appended "RequestBody", then content type suffix appended "JSON",
then context suffix fired again because the name no longer ended
in "RequestBody", ad infinitum.
Fix by restructuring resolveCollisions to apply each strategy as
a global phase: exhaust one strategy across ALL colliding groups
(re-checking for new collisions after each pass) before advancing
to the next strategy. This ensures numeric fallback is reached
when earlier strategies cannot disambiguate.
Also fix resolvedNameForComponent to accept an optional content
type for exact matching, so each media type variant of a
requestBody or response gets its own resolved name instead of all
variants receiving whichever name the map iterator returns first.
Adds Pattern H test case (TMF622 scenario from PR oapi-codegen#2213): a
component that exists in both schemas and requestBodies where the
requestBody has 3 content types that all collapse to the "JSON"
short name.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* Regenerate files with new code
* test: add Pattern I for inline response with x-go-type $ref properties
Add test case from oapi-codegen-exp#14: an inline response object whose
properties $ref component schemas with x-go-type: string. In the
experimental rewrite (libopenapi), this caused duplicate type declarations
because libopenapi copies extensions from $ref targets. V2 (kin-openapi)
handles this correctly, but the test guards against future regressions.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* fix: resolve type name mismatches for multi-content-type responses
When a component response has multiple JSON content types (e.g.,
application/json, application/json-patch+json, application/merge-patch+json),
three bugs produced uncompilable code:
1. Duplicate inline type declarations: GenerateGoSchema was called with a
path of [operationId, responseName] (or [responseName] in the component
phase), which doesn't include the content type. When multiple content
types share oneOf schemas with inline elements, they produce
identically-named AdditionalTypes. If the schemas differ, the result is
conflicting declarations; if they're the same, duplicate declarations.
Fix: include a mediaTypeToCamelCase(contentType) segment in the schema
path when jsonCount > 1, giving each content type's inline types unique
names. This is guarded by jsonCount > 1 for backward compatibility.
2. Undefined types in unmarshal code: GetResponseTypeDefinitions used
RefPathToGoType to resolve $ref paths like
#/components/responses/200Resource_Patch, but without content type
context. resolvedNameForComponent fell into a non-deterministic prefix
match, returning an arbitrary per-content-type base name that didn't
match any defined type when mediaTypeToCamelCase was appended.
Fix: add resolvedNameForRefPath helper that parses the $ref path and
delegates to resolvedNameForComponent with the specific content type
for exact matching.
3. Mismatched types in response struct fields: same root cause as bug 2 —
the struct field types were derived from the same non-deterministic
resolution path.
Additionally, promote AdditionalTypes from GenerateTypesForResponses and
GenerateTypesForRequestBodies to the top-level type list, matching the
existing pattern in GenerateTypesForSchemas. Without this, inline types
(e.g., oneOf union members, nested objects with additionalProperties)
defined inside response/requestBody schemas were silently dropped from
the output.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* Address greptile's comment
* fix: propagate user name overrides through codegen
The collision resolver had two bugs when resolve-type-name-collisions
was enabled:
1. x-go-name was ignored: generateCandidateName() never consulted the
x-go-name extension, so schemas/responses/requestBodies/parameters/
headers with explicit Go name overrides would lose them during
collision resolution.
2. Client wrapper names bypassed the name normalizer:
generateCandidateName() used UppercaseFirstCharacter(operationID)
instead of SchemaNameToTypeName(operationID), so configured
normalizers (e.g. ToCamelCaseWithInitialisms) were not applied to
client response wrapper type names.
Changes:
- Add GoNameOverride field to GatheredSchema, populated from x-go-name
on the parent container (schema, response, requestBody, parameter,
header) when the component is not a $ref
- Add extractGoNameOverride() helper to read x-go-name from extensions
- Add Pinned field to ResolvedName; pinned names are returned as-is
from generateCandidateName() and skipped by all collision resolution
strategies (context suffix, content type, status code, param index,
numeric fallback)
- Fix client wrapper candidate name to use SchemaNameToTypeName()
- Add unit tests for GoNameOverride population and pinning behaviour
- Add integration test patterns K/L/M (x-go-name on schema, response,
and requestBody) and regenerate
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>1 parent efb2df3 commit 3a5f803
21 files changed
Lines changed: 5664 additions & 586 deletions
File tree
- internal/test
- issues/issue-200
- name_conflict_resolution
- pkg/codegen
| Original file line number | Diff line number | Diff line change | |
|---|---|---|---|
| |||
255 | 255 | | |
256 | 256 | | |
257 | 257 | | |
258 | | - | |
| 258 | + | |
259 | 259 | | |
260 | 260 | | |
261 | 261 | | |
| |||
This file was deleted.
This file was deleted.
This file was deleted.
0 commit comments