[BOT ISSUE] Spring AI OpenAI SSE reassembly produces malformed tool_calls — deltas appended instead of merged by index

[braintrust-bot]

Summary

The Spring AI OpenAI SSE reassembly in BraintrustSpringAI.reassembleOpenAISSE() incorrectly handles tool_calls deltas during streaming. Each streaming chunk's delta.tool_calls array is appended as-is to the result array instead of being merged by the index field. This produces a malformed tool_calls array with N partial objects instead of one properly assembled tool_call per invocation.

This is distinct from #78 (which covers Anthropic SSE non-text content blocks). This issue affects the OpenAI SSE path specifically and involves a different bug: incorrect delta-merge semantics for tool_calls rather than missing content block types.

For comparison, the direct OpenAI SDK instrumentation in this repo (TracingHttpClient in openai_2_8_0) uses the official ChatCompletionAccumulator which correctly merges tool_call deltas by index.

What is missing

In BraintrustSpringAI.reassembleOpenAISSE() (lines 594–605):

if (delta.has("tool_calls")) {
    if (!choice.get("message").has("tool_calls")) {
        ((ObjectNode) choice.get("message"))
                .set("tool_calls", mapper.createArrayNode());
    }
    for (var tc : delta.get("tool_calls")) {
        ((ArrayNode) choice.get("message").get("tool_calls"))
                .add(tc);
    }
}

How OpenAI streaming sends tool_calls

OpenAI sends tool_calls incrementally across multiple chunks using an index field to identify which tool_call is being updated:

1. First chunk: {"tool_calls": [{"index": 0, "id": "call_abc", "type": "function", "function": {"name": "get_weather", "arguments": ""}}]}
2. Subsequent chunks: {"tool_calls": [{"index": 0, "function": {"arguments": "{\"lo"}}]}
3. More chunks: {"tool_calls": [{"index": 0, "function": {"arguments": "cation\": \"NYC\"}"}}]}

The correct reassembly should find the existing tool_call with the matching index and concatenate function.arguments. Instead, the current code appends each chunk as a new element, producing:

"tool_calls": [
  {"index": 0, "id": "call_abc", "type": "function", "function": {"name": "get_weather", "arguments": ""}},
  {"index": 0, "function": {"arguments": "{\"lo"}},
  {"index": 0, "function": {"arguments": "cation\": \"NYC\"}"}}
]

instead of the expected:

"tool_calls": [
  {"index": 0, "id": "call_abc", "type": "function", "function": {"name": "get_weather", "arguments": "{\"location\": \"NYC\"}"}}
]

Additional missing delta fields

The method (lines 587–606) only processes delta.content and delta.tool_calls. The following delta fields are also dropped:

• refusal — safety refusal text, sent incrementally like content when the model declines a request
• reasoning_content — reasoning text for o-series models in streaming

Braintrust docs status

• Braintrust docs at https://www.braintrust.dev/docs/integrations/ai-providers/openai document "Function calling and tools" as a supported feature: supported
• Braintrust docs at https://www.braintrust.dev/docs/instrument/trace-llm-calls state "Streaming responses are fully supported — Braintrust automatically collects streamed chunks and logs the complete response as a single span": supported
• Spring AI integration is not specifically documented on Braintrust docs: not_found

Upstream sources

• OpenAI streaming format: https://platform.openai.com/docs/api-reference/chat/create — streaming tool_calls use an index field for delta-merge identification; function.arguments is sent incrementally and must be concatenated
• OpenAI function calling: https://platform.openai.com/docs/guides/function-calling — tool_calls are a core feature of the Chat Completions API
• Spring AI OpenAI integration: https://docs.spring.io/spring-ai/reference/api/chat/openai-chat.html — supports streaming and function calling

Local files inspected

• braintrust-sdk/instrumentation/springai_1_0_0/src/main/java/dev/braintrust/instrumentation/springai/v1_0_0/BraintrustSpringAI.java — lines 554–632 (reassembleOpenAISSE: tool_calls appended at lines 600–603 instead of merged by index; no handling for refusal or reasoning_content deltas)
• braintrust-sdk/instrumentation/openai_2_8_0/src/main/java/dev/braintrust/instrumentation/openai/v2_8_0/TracingHttpClient.java — direct SDK path uses ChatCompletionAccumulator which handles tool_calls merging correctly
• braintrust-sdk/instrumentation/springai_1_0_0/src/test/java/dev/braintrust/instrumentation/springai/v1_0_0/BraintrustSpringAITest.java — no streaming tests exercise tool_calls responses