[BOT ISSUE] BraintrustStream cannot aggregate Anthropic Messages streaming events

[braintrust-bot]

Summary

The BraintrustStream streaming aggregator only handles OpenAI Chat Completions chunk format (choices[].delta). Anthropic's Messages streaming API uses a structurally different SSE event protocol (message_start, content_block_delta, message_delta, etc.) that the aggregator cannot parse. All Anthropic streaming chunks are silently discarded, producing an empty aggregated result.

What is missing

Anthropic's Messages streaming API emits typed SSE events:

• message_start — contains the initial Message object with usage.input_tokens
• content_block_start — opens a content block (text, tool_use, or thinking)
• content_block_delta — incremental content: text_delta.text, input_json_delta.partial_json, or thinking_delta.thinking
• content_block_stop — closes a content block
• message_delta — final stop_reason and cumulative usage.output_tokens
• message_stop — stream termination

These events have no choices array and no delta sub-field inside choices. The current aggregation path in BraintrustStream::aggregate() (src/stream.rs:727-807) parses each chunk as a StreamChunk expecting choices: Vec<StreamChoice>. Anthropic events fail deserialization and hit the Err(_) => continue fallback at line 740-741, silently dropping all data.

This means:

• Text content from content_block_delta events is lost
• Tool use arguments streamed via input_json_delta are lost
• Extended thinking content from thinking_delta events is lost
• Usage metrics from message_start and message_delta are lost
• Stop reason from message_delta is lost

Braintrust's other SDKs handle Anthropic streaming via dedicated wrappers:

• TypeScript: wrapAnthropic
• Python: wrap_anthropic
• Go: tracing middleware
• Ruby: Braintrust::Trace::Anthropic.wrap
• Java/C#: interceptors/extensions

All of these capture streamed Anthropic responses including content, tool use, and usage metrics.

This is distinct from existing issues:

• [BOT ISSUE] No support for OpenAI Responses API format in streaming or usage extraction #31 covers the OpenAI Responses API streaming format (different provider, different format)
• [BOT ISSUE] Streaming aggregator drops tool calls from Chat Completions deltas #30 covers tool call aggregation within the supported OpenAI Chat Completions format
• [BOT ISSUE] Streaming usage metrics drop reasoning tokens and token detail breakdowns #32 covers usage detail loss within the supported streaming format

Braintrust docs status

Braintrust documents Anthropic streaming support across all non-Rust SDKs: "Braintrust handles streaming, metric collection (including cached tokens), and other details." Status: supported in Braintrust platform via other SDK wrappers, not instrumented in this Rust SDK.

• https://www.braintrust.dev/docs/integrations/ai-providers/anthropic

Upstream sources

• Anthropic Messages streaming docs: https://docs.anthropic.com/en/api/messages-streaming — full event type specification
• Anthropic streaming event types: message_start, content_block_start, content_block_delta, content_block_stop, message_delta, message_stop, ping, error
• Anthropic Python SDK streaming: https://github.com/anthropics/anthropic-sdk-python — reference implementation of event aggregation

Local files inspected

• src/stream.rs — StreamChunk struct (line 639-647) only has model, choices, usage fields; StreamChoice/StreamDelta (lines 651-664) expect OpenAI format; aggregate() (lines 727-807) parses only choices[].delta; error fallback at line 740-741 silently skips unparseable chunks
• src/extractors.rs — extract_anthropic_usage() handles the non-streaming Anthropic response format correctly, but is not invoked during streaming aggregation
• src/lib.rs — wrap_stream_with_span is exported as the primary streaming instrumentation surface