[Bug] OpenAiChatModel internalCall missing reasoningContent in metadata

[CL-LSFY]

Description

When using the non-streaming API (internalCall), the reasoningContent field is not included in the AssistantMessage metadata, but the streaming API (internalStream) correctly includes it.

Affected Versions

• 1.1.2
• 1.1.4
• 2.0.0-SNAPSHOT (main branch)

Root Cause

In OpenAiChatModel.java, the internalCall method builds metadata without reasoningContent:

// internalCall (non-streaming) - missing reasoningContent
Map<String, Object> metadata = Map.of(
    "id", chatCompletion.id() != null ? chatCompletion.id() : "",
    "role", choice.message().role() != null ? choice.message().role().name() : "",
    "index", choice.index() != null ? choice.index() : 0,
    "finishReason", getFinishReasonJson(choice.finishReason()),
    "refusal", StringUtils.hasText(choice.message().refusal()) ? choice.message().refusal() : "",
    "annotations", choice.message().annotations() != null ? choice.message().annotations() : List.of(Map.of()));

But internalStream (streaming) correctly includes it:

// internalStream (streaming) - has reasoningContent
Map<String, Object> metadata = Map.of(
    "id", id,
    "role", roleMap.getOrDefault(id, ""),
    "index", choice.index() != null ? choice.index() : 0,
    "finishReason", getFinishReasonJson(choice.finishReason()),
    "refusal", StringUtils.hasText(choice.message().refusal()) ? choice.message().refusal() : "",
    "annotations", choice.message().annotations() != null ? choice.message().annotations() : List.of(),
    "reasoningContent", choice.message().reasoningContent() != null ? choice.message().reasoningContent() : "");

Expected Behavior

The internalCall method should also include reasoningContent in the metadata, consistent with the streaming API.

Proposed Fix

Add reasoningContent to the metadata map in internalCall:

Map<String, Object> metadata = Map.of(
    "id", chatCompletion.id() != null ? chatCompletion.id() : "",
    "role", choice.message().role() != null ? choice.message().role().name() : "",
    "index", choice.index() != null ? choice.index() : 0,
    "finishReason", getFinishReasonJson(choice.finishReason()),
    "refusal", StringUtils.hasText(choice.message().refusal()) ? choice.message().refusal() : "",
    "annotations", choice.message().annotations() != null ? choice.message().annotations() : List.of(),
    "reasoningContent", choice.message().reasoningContent() != null ? choice.message().reasoningContent() : "");

Workaround

Currently, users can use the streaming API (stream()) with .last().block() to get the reasoningContent, but this is not ideal for non-streaming use cases.

Related

This affects models like Qwen3.5-plus and DeepSeek-R1 that return reasoning content in API responses.