[bot] Bedrock Converse `reasoningContent` blocks not normalized — extended thinking traces break UI schema

[braintrust-bot]

<!-- provider-gap-audit: bedrock-reasoningcontent-not-normalized -->

Summary

InstrumentationSemConv.normalizeBedrockMessage() — the function that normalizes Bedrock Converse content blocks so they carry an explicit "type" field the Braintrust UI schema requires — recognizes text, toolUse, toolResult, and image blocks, but does not recognize reasoningContent, the official Bedrock Converse block type for extended thinking / Chain-of-Thought output (Claude 3.7+ and Amazon Nova reasoning models via additionalModelRequestFields.reasoning_config).

This is the same code path issue #85 describes as working correctly for tool use ("Non-streaming Converse calls handle tool use correctly — normalizeBedrockMessage() ... already recognizes toolUse, toolResult, and image"), but that issue is scoped to the streaming (ConverseStream) tool-use path and does not mention reasoningContent at all — this is a distinct gap in the shared, non-streaming-safe normalization helper used by both request and response tagging.

What is missing

In braintrust-sdk/src/main/java/dev/braintrust/instrumentation/InstrumentationSemConv.java, normalizeBedrockMessage() (lines 467–501):

if (block.has("text")) {
    normalized.put("type", "text");
    changed = true;
} else if (block.has("toolUse")) {
    normalized.put("type", "tool_use");
    changed = true;
} else if (block.has("toolResult")) {
    normalized.put("type", "tool_result");
    changed = true;
} else if (block.has("image")) {
    normalized.put("type", "image");
    changed = true;
}

There is no else if (block.has("reasoningContent")) branch. When a Bedrock Converse (or ConverseStream, once #85 is fixed) response contains a reasoningContent block ({"reasoningContent": {"reasoningText": {"text": "...", "signature": "..."}}}), it is left without a type field — unlike every sibling content-block type in the same array — so it doesn't match the schema the UI validates against for thread rendering (per this same method's own doc comment: "both the OpenAI and Anthropic schemas the UI validates against require an explicit type field").

This function is shared by both:

• tagBedrockRequest() (line 389: inputArray.add(normalizeBedrockMessage(msg))) — affects multi-turn conversations that echo back a prior assistant reasoningContent block (required by some models to preserve the reasoning trace across turns)
• tagBedrockResponse() (line 408: outputArray.add(normalizeBedrockMessage(message))) — affects the model's own reasoning output on the current turn

Braintrust docs status

unclear — the Braintrust docs list Bedrock as a supported provider (https://www.braintrust.dev/docs/integrations/ai-providers) but do not specifically document extended-thinking/reasoning content handling for Bedrock Converse.

Upstream sources

• AWS Bedrock Converse API — reasoningContent block, part of the ContentBlock union: https://docs.aws.amazon.com/bedrock/latest/APIReference/API_runtime_ContentBlock.html
• Amazon Nova extended reasoning guide (documents reasoningContent/reasoningText response shape): https://docs.aws.amazon.com/nova/latest/userguide/extended-thinking.html
• Using Claude 3.7 Sonnet extended thinking via the Bedrock Converse API (additionalModelRequestFields.reasoning_config, response reasoningContent): https://seobs.medium.com/using-sonnet-3-7-reasoning-with-the-converse-api-on-aws-bedrock-6c838c685ad9

Local files inspected

• braintrust-sdk/src/main/java/dev/braintrust/instrumentation/InstrumentationSemConv.java — lines 339–429 (tagBedrockRequest/tagBedrockResponse), lines 460–501 (normalizeBedrockMessage: only text/toolUse/toolResult/image handled)
• braintrust-sdk/instrumentation/aws_bedrock_2_30_0/src/test/java/dev/braintrust/instrumentation/awsbedrock/v2_30_0/BraintrustAWSBedrockTest.java — no test exercises a reasoning/extended-thinking model response
• Confirmed via issue [bot] Bedrock ConverseStream drops tool use and non-text content blocks in streaming spans #85 that this shared helper is treated as the "correct" reference implementation for non-streaming content-block normalization, but it does not cover reasoningContent