…odegen#2335)

Two bugs were interacting to silently drop user intent:

1. In GenerateGoSchema, the `allOf` branch early-returned before the
   `x-go-type` check, so an outer `allOf` schema's own `x-go-type` was
   ignored. The documented contract ("completely override the definition
   of this schema") was violated whenever the schema used `allOf`. Fix:
   move the `x-go-type` check above the `allOf` branch.

2. In mergeOpenapiSchemas, extensions from all allOf members were
   blindly unioned into the merged schema's extensions. When a member's
   extensions included an identity-bound directive like `x-go-type`,
   that directive would leak up and become the merged schema's alias
   target -- e.g. `Client` with `x-go-type: OverlayClient` composed with
   `{properties:{id}}` via `allOf` would emit `type ClientWithId =
   OverlayClient`, silently dropping the Id field.

   The naive fix of dropping all extensions breaks the common
   "$ref + sibling extension" idiom (see issue oapi-codegen#1957), where allOf is
   used not for composition but as a workaround for OpenAPI 3.0's
   $ref-sibling restriction to attach extensions to a ref.

   Fix: in mergeSchemas, detect which of the two uses applies by
   checking whether any allOf member is extension-only. If so (decorator
   idiom), extensions flow through as before. If not (real composition),
   drop extensions from the merged result -- each source schema's
   extensions belong to that schema, not to the composed type.

Regression test lives in internal/test/extensions/allof-merge-extensions/.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

…nicely (#2087)

* demonstration of overlays and allOf not playing together nicely

* Reorganize bug test case

Relocate the overlay+allOf regression reproduction from
examples/anyof-allof-oneof/ to internal/test/extensions/allof-merge-extensions/.

The examples/ directory is for happy-path usage, not regression artifacts.
The submitter's overlay.yaml, overlays.go, and overlays_test.go are moved
via git rename (preserving per-line blame); package renamed to match new
directory. examples/anyof-allof-oneof/ reverts to a clean anyOf/allOf/oneOf
usage example.

The bug is tracked at #2335
and is unrelated to overlays — it's in the allOf+x-go-type interaction in
pkg/codegen/schema.go and pkg/codegen/merge_schemas.go.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

* fix: honor x-go-type on allOf schemas, stop extension leakage (#2335)

Two bugs were interacting to silently drop user intent:

1. In GenerateGoSchema, the `allOf` branch early-returned before the
   `x-go-type` check, so an outer `allOf` schema's own `x-go-type` was
   ignored. The documented contract ("completely override the definition
   of this schema") was violated whenever the schema used `allOf`. Fix:
   move the `x-go-type` check above the `allOf` branch.

2. In mergeOpenapiSchemas, extensions from all allOf members were
   blindly unioned into the merged schema's extensions. When a member's
   extensions included an identity-bound directive like `x-go-type`,
   that directive would leak up and become the merged schema's alias
   target -- e.g. `Client` with `x-go-type: OverlayClient` composed with
   `{properties:{id}}` via `allOf` would emit `type ClientWithId =
   OverlayClient`, silently dropping the Id field.

   The naive fix of dropping all extensions breaks the common
   "$ref + sibling extension" idiom (see issue #1957), where allOf is
   used not for composition but as a workaround for OpenAPI 3.0's
   $ref-sibling restriction to attach extensions to a ref.

   Fix: in mergeSchemas, detect which of the two uses applies by
   checking whether any allOf member is extension-only. If so (decorator
   idiom), extensions flow through as before. If not (real composition),
   drop extensions from the merged result -- each source schema's
   extensions belong to that schema, not to the composed type.

Regression test lives in internal/test/extensions/allof-merge-extensions/.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

* Implement Greptile suggestions

Three review findings addressed on top of the previous fix commit:

1. Narrow the non-decorator extension clearing from "wipe everything"
   to targeted delete() calls for x-go-type, x-go-type-name, and
   x-go-type-import. User-defined schema-level extensions and any other
   non-identity directives now survive allOf composition -- we only
   have concrete evidence of incorrect aliasing for the identity-bound
   three.

2. Tighten isExtensionOnlySchema via a zero-value approach. Zero out
   Extensions plus the documentary/metadata fields that don't affect
   the generated Go type (Title, Description, Default, Example,
   ExternalDocs, Deprecated, ReadOnly, WriteOnly, AllowEmptyValue,
   XML) and the kin-openapi Origin tracker, then reflect.DeepEqual
   against the zero Schema. Shorter, and default-safe against future
   kin-openapi field additions: any new structural field would be
   non-zero and correctly disqualify decorator classification.

3. Extend the regression test to cover the harder Fix 2 path. The
   original scenario had x-go-type on both Client and ClientWithId,
   which Fix 1 alone makes pass. New schemas BaseOnly/DerivedNoOverride
   exercise the case where only the base schema has an x-go-type
   overlay -- the derived schema must still be emitted as its own
   struct (with both Name and Extra fields) rather than aliased to
   OverlayBaseOnly. The struct literal in TestBaseOnlyOverlay won't
   compile under the buggy codegen because OverlayBaseOnly has no
   Extra field, giving the test a compile-time check in addition to
   the runtime assertion.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

* Defensively clone extensions before deleting identity keys

The delete() calls in mergeSchemas operated on schema.Extensions
directly. In the current code this is safe: valueWithPropagatedRef
returns a shallow struct copy whose Extensions map is shared with
the caller, but mergeOpenapiSchemas always reallocates Extensions
before merging, and the n>=2 guard ensures that reallocation has
happened by the time we reach the delete. The delete thus operates
on a fresh map, not the caller's shared one.

That safety is an invariant held by two separate pieces of code
(the n>=1 short-circuit, and mergeOpenapiSchemas' make()), neither
of which is obvious from reading the delete. Cloning the map before
mutating makes the non-aliasing explicit locally and keeps the code
correct even if either piece changes (e.g. an allocation-skipping
optimization in mergeOpenapiSchemas).

maps.Clone(nil) returns nil and delete(nil, ...) is a no-op, so no
nil-guard is needed.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

---------

Co-authored-by: Marcin Romaszewicz <marcinr@gmail.com>
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Address greptile review feedback and adapt to upstream main:

- pkg/codegen/schema.go: simplify the allOf branch. Upstream main now
  catches x-go-type at the schema level via the combined extensions
  early return, so the PR's allOf-specific x-go-type override block is
  unreachable; remove it. Keep x-go-type-skip-optional-pointer
  propagation as a conditional override (only when the parent sets it
  explicitly), which preserves the decorator-idiom behavior MergeSchemas
  relies on for issue oapi-codegen#1957. Read from combined extensions for parity
  with the rest of GenerateGoSchema.

- pkg/codegen/test_specs/x-go-type-pet-allof.yaml: add trailing newline.

…1610)

* fix: allow x-go-type and x-go-type-skip-optional-pointer for allOf

This patch ensures that if a schema is defined using allOf the main
schema entry's x-go-type and x-go-type-skip-optional-pointer values
are considered instead of a random allOf member's.

* Add unit tests

* update PR for upstream changes

Address greptile review feedback and adapt to upstream main:

- pkg/codegen/schema.go: simplify the allOf branch. Upstream main now
  catches x-go-type at the schema level via the combined extensions
  early return, so the PR's allOf-specific x-go-type override block is
  unreachable; remove it. Keep x-go-type-skip-optional-pointer
  propagation as a conditional override (only when the parent sets it
  explicitly), which preserves the decorator-idiom behavior MergeSchemas
  relies on for issue #1957. Read from combined extensions for parity
  with the rest of GenerateGoSchema.

- pkg/codegen/test_specs/x-go-type-pet-allof.yaml: add trailing newline.

---------

Co-authored-by: Jamie Tanna <jamie.tanna@elastic.co>
Co-authored-by: Marcin Romaszewicz <mromaszewicz@users.noreply.github.com>
Co-authored-by: Marcin Romaszewicz <marcinr@gmail.com>