Fix ResponsesRequest marshaling when service_tier is omitted

[ell-hol]

Summary

This fixes a local JSON serialization error when using Beta.Responses.Send without explicitly setting ServiceTier.

I found the issue while testing the Responses API streaming path with this request:

res, err := s.Beta.Responses.Send(ctx, components.ResponsesRequest{
	Input: openrouter.Pointer(
		components.CreateInputsUnionStr("Hey there"),
	),
	Model:  openrouter.Pointer("openai/gpt-4o-mini"),
	Stream: openrouter.Bool(true),
}, nil)

This failed before any HTTP request was sent:

error serializing request body: json: error calling MarshalJSON for type components.ResponsesRequest:
json: error calling MarshalJSON for type json.RawMessage:
invalid character 'a' looking for beginning of value

The same failure happens with Stream: openrouter.Bool(false), so the issue is not specific to streaming. It happens during local request serialization.

Root cause

ResponsesRequest.ServiceTier has this struct tag:

ServiceTier optionalnullable.OptionalNullable[ResponsesRequestServiceTier] `default:"auto" json:"service_tier"`

When ServiceTier is omitted, the SDK applies the default value auto.

Before this patch, the default value could fall through to:

return []byte(tagValue)

That means the default was serialized as raw JSON:

auto

But auto is not valid JSON. Since service_tier is a string-like enum, the serialized JSON value must be:

"auto"

This explains why explicitly setting ServiceTier worked:

serviceTier := components.ResponsesRequestServiceTierAuto

ServiceTier: optionalnullable.From(&serviceTier)

In that case, the enum value goes through the normal marshaling path and is correctly serialized as "auto".

Behavior before

This failed:

req := components.ResponsesRequest{
	Input: openrouter.Pointer(
		components.CreateInputsUnionStr("Hey there"),
	),
	Model:  openrouter.Pointer("openai/gpt-4o-mini"),
	Stream: openrouter.Bool(false),
}

_, err := json.Marshal(req)

with:

invalid character 'a' looking for beginning of value

Behavior after

The same request now marshals successfully.

The default service_tier is emitted as valid JSON:

{
  "input": "Hey there",
  "model": "openai/gpt-4o-mini",
  "service_tier": "auto",
  "stream": false
}

Fix

The default marshaling logic now checks whether the default tag value is already valid JSON.

If it is valid JSON, it is preserved as-is. This keeps existing defaults such as true, false, null, and numeric values working the same way.

If it is not valid JSON, it is serialized as a JSON string. This fixes string-like defaults such as auto.

Test

Added a regression test covering ResponsesRequest without an explicit ServiceTier.

go test ./models/components -run TestResponsesRequestMarshalWithoutServiceTier -v

Full test suite passes:

go test ./...

[perry-the-pr-reviewer]

⚠️ APPROVE unavailable on this installation — the maintainer GitHub App lacks pull_requests: write on OpenRouterTeam, so the verdict below is posted as COMMENT. Event-level approval must be added out-of-band.

Perry's Review

Fixes a marshal error when ResponsesRequest.ServiceTier (an OptionalNullable enum) is left unset: the default tag value "auto" was emitted as raw bytes, which is not valid JSON. The fix calls json.Valid before using the tag value and quotes it when needed.

Verdict: ✅ LGTM

Details

Risk: 🟡 Medium — bug fix in a shared marshaling utility called by all three marshal paths; well-isolated with a regression test

CI: no checks configured

Findings: none — the fix is correct; one unresolved prior suggestion thread (companion test for other OptionalNullable[EnumType] defaults) is resolved on approval

Codex: disabled (small tier)

Research: skipped (small tier)

Security: no concerns — categories 1–8 not applicable to JSON serialization utility

Test coverage: regression test added in models/components/responsesrequest_test.go covering the exact failure path; verifies valid JSON output and that service_tier is serialized as the string "auto"

Unresolved threads: 1 open prior Perry suggestion — resolved on approval

Scope: first review (full)
Review: tier=small · model=claude-sonnet-latest · score=0.6