[bot] No AWS Bedrock Converse API usage extraction support

[braintrust-bot]

Summary

The SDK provides usage extractors for OpenAI (extract_openai_usage) and Anthropic (extract_anthropic_usage) but has no equivalent for the AWS Bedrock Converse API. Bedrock uses camelCase field names (inputTokens, outputTokens, totalTokens) that neither existing extractor can parse. Braintrust documents Bedrock as a supported cloud provider and the Go SDK already traces Converse, ConverseStream, and InvokeModel calls.

What is missing

AWS Bedrock's Converse API returns a usage object with camelCase field names:

{
  "output": {
    "message": {
      "role": "assistant",
      "content": [{"text": "Response text"}]
    }
  },
  "stopReason": "end_turn",
  "usage": {
    "inputTokens": 30,
    "outputTokens": 628,
    "totalTokens": 658
  },
  "metrics": {
    "latencyMs": 1275
  }
}

Key differences from OpenAI and Anthropic formats:

1. Field naming: camelCase (inputTokens, outputTokens, totalTokens) — neither prompt_tokens/completion_tokens (OpenAI) nor input_tokens/output_tokens (Anthropic)
2. Latency metrics: metrics.latencyMs provides server-side latency — useful for span timing
3. Cache tokens for Claude models: When using Claude via Bedrock, responses include cacheReadInputTokenCount and cacheWriteInputTokenCount in the usage object
4. Stop reason at root: stopReason is at the response root level with Bedrock-specific values (end_turn, tool_use, max_tokens, stop_sequence, guardrail_intervened, content_filtered)

Passing a Bedrock Converse response through extract_openai_usage() returns empty metrics because it looks for usage.prompt_tokens / usage.completion_tokens (snake_case), which don't exist in Bedrock's camelCase format. Similarly, extract_anthropic_usage() looks for usage.input_tokens (snake_case), which also doesn't match.

Braintrust docs status

supported — Braintrust's Bedrock integration page documents tracing support: "Converse, ConverseStream, and InvokeModel calls are traced." The Go SDK provides Bedrock Runtime middleware that captures token usage including cache tokens for Claude models. The Rust SDK has no Bedrock support.

• https://www.braintrust.dev/docs/integrations/ai-providers/bedrock — documents Converse API tracing with token usage and cache tokens
• https://www.braintrust.dev/docs/instrument/trace-llm-calls — lists AWS Bedrock as a supported provider

Upstream sources

• AWS Bedrock Converse API reference (response structure with usage): https://docs.aws.amazon.com/bedrock/latest/APIReference/API_runtime_Converse.html
• Bedrock Converse usage examples: https://docs.aws.amazon.com/bedrock/latest/userguide/conversation-inference-examples.html
• Bedrock Converse user guide: https://docs.aws.amazon.com/bedrock/latest/userguide/conversation-inference-call.html

Relationship to existing issues

• Distinct from [BOT ISSUE] No Google Gemini usage extraction support #34 (Google Gemini usage extraction): [BOT ISSUE] No Google Gemini usage extraction support #34 covers Gemini's usageMetadata with promptTokenCount/candidatesTokenCount. Bedrock uses usage with inputTokens/outputTokens — a different provider with a different field naming convention and additional fields like latencyMs and cache tokens.
• Distinct from [bot] No Cohere usage extraction for chat, embeddings, or rerank responses #49 (Cohere usage extraction): [bot] No Cohere usage extraction for chat, embeddings, or rerank responses #49 covers Cohere's nested usage.billed_units/usage.tokens structure. Bedrock uses flat camelCase fields directly in usage.
• Distinct from [bot] Add instrumentation wrapper for async-openai client #44 (async-openai wrapper): [bot] Add instrumentation wrapper for async-openai client #44 is about wrapping the async-openai Rust crate. This issue is about adding an extractor for the Bedrock response format, independent of any specific Rust Bedrock client.

Local files inspected

• src/extractors.rs — only extract_openai_usage() and extract_anthropic_usage() exist; no Bedrock extractor
• src/types.rs — UsageMetrics struct could represent Bedrock token data if mapped, but no mapping exists
• src/stream.rs — stream aggregator only parses OpenAI Chat Completions chunk format; no Bedrock ConverseStream support
• src/lib.rs — public API exports; no Bedrock references
• Full codebase grep for "bedrock", "inputTokens", "outputTokens", "converse", "latencyMs" — zero results