[cyphercodes]

References to other Issues or PRs

Fixes #6938

Have you read the Contributing Guidelines?

Yes.

Brief description of what is fixed or changed

AssignableExprPrep now initializes singular intermediate message fields before descending into nested body/path field expressions. This prevents generated gateway handlers from dereferencing a nil parent message when a body selector traverses through a message field to a oneof field.

A regression test covers the input.input_value shape from the issue and verifies the parent message initialization is emitted before the oneof wrapper initialization.

Other comments

Tests run:

• go test ./internal/descriptor ./protoc-gen-grpc-gateway/...

[johanbrandhorst]

Thanks for this PR! Could you make a change to one of our example protobuf files that would reproduce this bug, so we can protect against it breaking again in the future? See examples/internal/proto.

[cyphercodes]

Updated to address the review request by adding a protobuf fixture that reproduces the nested body path through an intermediate message to a oneof field. I also excluded internal/descriptor/testdata from the default generation template so the fixture can live in-tree without being emitted by the OpenAPI generator, which currently requires top-level body fields.

Verification:

• buf lint internal/descriptor/testdata/v1/nested_body_oneof.proto
• make proto
• go test ./internal/descriptor
• buf generate --path internal/descriptor/testdata/v1/nested_body_oneof.proto --template /tmp/grpc_gateway_only.buf.gen.yaml plus assertion that generated gateway code contains protoReq.Input == nil, protoReq.Input = &ExecutionInput{}, and protoReq.Input.Input == nil
• git diff --check

[johanbrandhorst]

This... is not what I asked for? I asked for you to ideally update one of the existing protobuf example files, but you've added a new protobuf file as testdata, and then promptly excluded it from the generation path? Please remove the changes to buf.gen.yaml and the new protobuf file, it would be much better if you just edited a_bit_of_everthing.proto or one of the others.

[dantedoyl]

@cyphercodes Hi! Any updates?

[cyphercodes]

Thanks for the ping. I checked the existing-example route requested in the last review: adding a nested body path such as body: "single_nested.nested_value" in examples/internal/proto/examplepb/a_bit_of_everything.proto makes make proto fail earlier in protoc-gen-openapiv2 with:

panic: body of request "ABitOfEverythingService" is not a top level field: 'single_nested.nested_value'

The current Bazel failure is the expected generated BUILD.bazel diff for the descriptor testdata fixture that was rejected, not a separate source failure.

I don't want to broaden this PR into OpenAPI nested-body support without maintainer direction. Which direction would you prefer?

1. allow a descriptor/gateway-only fixture outside default OpenAPI generation,
2. point me at an existing proto path that avoids OpenAPI generation, or
3. expand this PR to include the OpenAPI generator support too.

[johanbrandhorst]

I'm confused by your answer. This fixes an issue in the gateway generator, right? If so, we should definitely just go ahead and fix that same issue in the openapiv2 generator and include an example in a_bit_of_everything.proto.

[dantedoyl]

@cyphercodes Hi! Any updates?

[cyphercodes]

Updated this to match the requested existing-example direction.

Changes:

• removed the buf.gen.yaml exclusion and deleted the descriptor-only testdata proto
• added a nested body/oneof case to examples/internal/proto/examplepb/a_bit_of_everything.proto
• taught protoc-gen-openapiv2 to render nested body field paths instead of rejecting them as non-top-level
• regenerated the gateway/openapi artifacts and added the example server stub

Verification:

• go test ./internal/descriptor ./protoc-gen-openapiv2/internal/genopenapi
• go install ./protoc-gen-grpc-gateway ./protoc-gen-openapiv2 ./protoc-gen-openapiv3
• buf generate --path examples/internal/proto/examplepb/a_bit_of_everything.proto
• buf generate --template ./protoc-gen-openapiv3/buf.gen.yaml --path ./examples/internal/proto/examplepb/a_bit_of_everything.proto
• go test ./examples/internal/proto/examplepb ./examples/internal/server ./protoc-gen-grpc-gateway/internal/gengateway ./protoc-gen-openapiv2/internal/genopenapi
• go test ./...
• git diff --check

[dantedoyl]

@johanbrandhorst Hi! Could you review the changes?

[dantedoyl]

@johanbrandhorst ping)

[cyphercodes]

Updated to address the latest review and generated-artifact failure. I restored the explanatory HttpRule.body comment with an added note for nested gateway body paths, then refreshed the generated OpenAPI/clients artifacts from make generate.

Verification:

• make generate
• go test ./protoc-gen-openapiv2/internal/genopenapi ./examples/internal/proto/examplepb ./examples/internal/server ./protoc-gen-grpc-gateway/internal/gengateway
• git diff --check --cached

[johanbrandhorst]

Looks like there's still a bazel error

[cyphercodes]

Updated the Bazel BUILD files that bazel run //:gazelle reported as stale for the generated nested-body oneof client/model files.

Verification:

• go test ./protoc-gen-openapiv2/internal/genopenapi ./examples/internal/proto/examplepb ./examples/internal/server ./protoc-gen-grpc-gateway/internal/gengateway
• /root/go/bin/bazelisk run //:gazelle && git add -N . && git diff --exit-code -- . ':(exclude)bazel-grpc-gateway-6948'
• git diff --check HEAD~1..HEAD

[johanbrandhorst]

Thanks!