[BOT ISSUE] Spring AI Anthropic SSE reassembly drops tool_use and thinking content blocks in streaming

[braintrust-bot]

Summary

The Spring AI Anthropic SSE reassembly in BraintrustSpringAI.reassembleAnthropicSSE() only handles text delta events. When streaming Anthropic responses include tool use (input_json_delta) or extended thinking (thinking_delta) content blocks, these are silently dropped from the reconstructed span output. All content blocks are hardcoded as {"type": "text", "text": "..."}, losing the actual block type and non-text content.

This is distinct from #71 (which covers missing provider/embedding coverage). The Anthropic backend IS supported for Spring AI, but its streaming output reconstruction is lossy for non-text content blocks.

For comparison, the direct Anthropic SDK instrumentation in this repo (TracingHttpClient in anthropic_2_2_0) uses the official MessageAccumulator which correctly handles all content block types.

What is missing

In BraintrustSpringAI.reassembleAnthropicSSE() (lines 646–734):

1. content_block_delta only handles text (line 696)

case "content_block_delta" -> {
    int index = event.has("index") ? event.get("index").asInt() : 0;
    textByIndex.putIfAbsent(index, new StringBuilder());
    if (event.has("delta") && event.get("delta").has("text")) {
        textByIndex.get(index).append(event.get("delta").get("text").asText());
    }
}

Missing delta types:

• thinking_delta — has delta.thinking (not delta.text). Extended thinking content is dropped.
• input_json_delta — has delta.partial_json (not delta.text). Tool use input JSON is dropped.
• citations_delta — has delta.citation. Citation data is dropped.

2. Content block reconstruction hardcodes type: "text" (lines 717–725)

textByIndex.entrySet().stream()
        .sorted(Map.Entry.comparingByKey())
        .forEach(entry -> {
            var block = mapper.createObjectNode();
            block.put("type", "text");          // always "text"
            block.put("text", entry.getValue().toString());
            contentBlocks.add(block);
        });

The original content block type from content_block_start (e.g., tool_use, thinking) is not preserved. Tool use blocks should include id, name, and input; thinking blocks should use type: "thinking".

Impact

• Tool use in streaming: When the model returns a tool_use content block, the content_block_start event (containing id, name) is ignored beyond creating an empty buffer. The input_json_delta events are dropped. The final output has an empty {"type": "text", "text": ""} where a tool call should be.
• Extended thinking in streaming: Thinking content (thinking_delta events) is dropped. The span output loses all chain-of-thought reasoning.
• Non-streaming calls are unaffected — the full response JSON is captured correctly.

Braintrust docs status

• Braintrust docs at https://www.braintrust.dev/docs/integrations/ai-providers/anthropic document tool use and extended thinking as supported features: supported
• Spring AI integration is not specifically documented on Braintrust docs: not_found

Upstream sources

• Anthropic streaming events: https://docs.anthropic.com/en/api/messages-streaming — documents thinking_delta, input_json_delta, citations_delta event types alongside text_delta
• Anthropic extended thinking: https://docs.anthropic.com/en/docs/build-with-claude/extended-thinking — extended thinking uses thinking content blocks in streaming
• Anthropic tool use streaming: https://docs.anthropic.com/en/docs/build-with-claude/tool-use/overview — tool use returns tool_use content blocks with input_json_delta in streaming
• Spring AI Anthropic integration: https://docs.spring.io/spring-ai/reference/api/chat/anthropic-chat.html — supports tool calling and streaming

Local files inspected

• braintrust-sdk/instrumentation/springai_1_0_0/src/main/java/dev/braintrust/instrumentation/springai/v1_0_0/BraintrustSpringAI.java — lines 646–734 (reassembleAnthropicSSE: only delta.text handled at line 696; content blocks hardcoded as type: "text" at line 722)
• braintrust-sdk/instrumentation/anthropic_2_2_0/src/main/java/dev/braintrust/instrumentation/anthropic/v2_2_0/TracingHttpClient.java — lines 340–368 (direct SDK path uses MessageAccumulator which handles all event types 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 use or extended thinking responses