[bot] Bedrock ConverseStream drops tool use and non-text content blocks in streaming spans

Bắt đầu với TeeingSubscriber.onNext, parseDeltaText và buildConverseJson trong braintrust-sdk/instrumentation/aws_bedrock_2_30_0/src/main/java/dev/braintrust/instrumentation/awsbedrock/v2_30_0/BraintrustBedrockInterceptor.java. So sánh cách xử lý sự kiện streaming với normalizeBedrockMessage() trong braintrust-sdk/src/main/java/dev/braintrust/instrumentation/InstrumentationSemConv.java và xem lại các bài kiểm thử Bedrock hiện có. Hoàn thành khi nội dung tool-use trong streaming, các mã định danh, tên và dữ liệu đầu vào được biểu diễn trong đầu ra của span, với coverage trong BraintrustAWSBedrockTest.java.

Summary

The Bedrock ConverseStream instrumentation in TeeingSubscriber only accumulates delta.text from contentBlockDelta events. When the model returns tool use content blocks during streaming (a documented and supported Bedrock feature), the tool call data is silently dropped from the span output.

Non-streaming Converse calls handle tool use correctly — normalizeBedrockMessage() in InstrumentationSemConv already recognizes toolUse, toolResult, and image content block types and adds appropriate type annotations. The gap is exclusively in the streaming path.

What is missing

1. TeeingSubscriber only parses text deltas (lines 273–289)

case "contentBlockDelta" -> {
    String t = parseDeltaText(payload);   // only looks for delta.text
    if (t != null) {
        text.append(t);
        // ...
    }
}

parseDeltaText() (lines 322–339) searches exclusively for delta.text. Bedrock ConverseStream sends delta.toolUse for tool use blocks (containing incremental input JSON), which is silently ignored.

2. contentBlockStart events are not handled

The TeeingSubscriber switch (lines 273–289) has no case for contentBlockStart. This event carries toolUse.toolUseId and toolUse.name — metadata needed to reconstruct the tool call. Without it, even if tool use deltas were captured, the tool name and ID would be lost.

3. buildConverseJson always produces a single text block (lines 385–410)

gen.writeArrayFieldStart("content");
gen.writeStartObject();
gen.writeStringField("text", text);   // always one text block
gen.writeEndObject();
gen.writeEndArray();

The synthetic response JSON always contains exactly one {"text": "..."} content block, regardless of what the model actually returned. When the model requests a tool call, the span output will show empty text instead of the tool invocation.

Impact

When a streaming Bedrock call triggers tool use (common in agentic workflows):

• The span's output_json contains [{"role":"assistant","content":[{"text":"","type":"text"}]}] — as if the model returned nothing
• The stopReason is captured as "tool_use", contradicting the empty output
• Tool call names, IDs, and argument JSON are permanently lost from the trace
• Non-streaming Converse calls with tool use are correctly captured

Braintrust docs status

• Bedrock is listed as a supported cloud provider at https://www.braintrust.dev/docs/integrations/ai-providers: supported
• Tool calling in Bedrock streaming is not specifically documented on Braintrust docs: not_found

Upstream sources

• Bedrock ConverseStream API reference: https://docs.aws.amazon.com/bedrock/latest/APIReference/API_runtime_ConverseStream.html — documents toolConfig parameter and streaming tool use events (contentBlockStart with toolUse, contentBlockDelta with toolUse input)
• Bedrock tool use examples: https://docs.aws.amazon.com/bedrock/latest/userguide/tool-use-examples.html — shows ConverseStream with tool use
• Bedrock streaming event sequence: Events include contentBlockStart (carries start.toolUse with toolUseId and name), contentBlockDelta (carries delta.toolUse with input JSON fragments), and contentBlockStop

Local files inspected

• braintrust-sdk/instrumentation/aws_bedrock_2_30_0/src/main/java/dev/braintrust/instrumentation/awsbedrock/v2_30_0/BraintrustBedrockInterceptor.java — lines 273–289 (TeeingSubscriber.onNext: only contentBlockDelta→text, messageStop, metadata handled; no contentBlockStart), lines 322–339 (parseDeltaText: only delta.text), lines 385–410 (buildConverseJson: hardcoded single text block)
• braintrust-sdk/src/main/java/dev/braintrust/instrumentation/InstrumentationSemConv.java — lines 400–434 (normalizeBedrockMessage: non-streaming path correctly handles toolUse, toolResult, image content block types)
• braintrust-sdk/instrumentation/aws_bedrock_2_30_0/src/test/java/dev/braintrust/instrumentation/awsbedrock/v2_30_0/BraintrustAWSBedrockTest.java — no streaming test exercises tool use responses