[bot] Streaming aggregator drops `service_tier` and `system_fingerprint` from OpenAI Chat Completions chunks

[braintrust-bot]

Summary

The BraintrustStream streaming aggregator only extracts model from OpenAI Chat Completions streaming chunks, silently dropping service_tier and system_fingerprint. These are documented fields on every ChatCompletionChunk object. The StreamMetadata type already has an extra: HashMap<String, Value> that could hold these values, but the aggregation logic always passes an empty map.

What is missing

OpenAI Chat Completions streaming chunks include these fields alongside model:

{
  "id": "chatcmpl-abc123",
  "object": "chat.completion.chunk",
  "created": 1694268190,
  "model": "gpt-4o",
  "service_tier": "scale",
  "system_fingerprint": "fp_44709d6fcb",
  "choices": [...]
}

Currently in the SDK:

1. StreamChunk (src/stream.rs:687-694) only deserializes three fields — model, choices, and usage. The service_tier, system_fingerprint, id, and created fields are silently discarded during serde deserialization.

2. aggregate() (src/stream.rs:983) creates metadata with an always-empty extra map:

   let metadata = StreamMetadata::new(model, HashMap::new());

   The StreamMetadata type already supports extra fields via its extra: HashMap<String, Value>, but the aggregation logic never populates it.

Why these fields matter

• service_tier ("default" | "scale" | "flex") — determines which pricing tier processed the request. Scale tier is ~2x the cost of default; flex tier is ~50% cheaper. Without this field, cost attribution in Braintrust spans may use incorrect pricing assumptions. This is especially relevant as OpenAI's newer flex tier becomes more widely used.
• system_fingerprint (string) — identifies the backend configuration the model ran with. Used alongside the seed parameter to detect when backend changes affect determinism. Important for reproducibility analysis and debugging behavior regressions.

Braintrust docs status

unclear — Braintrust's OpenAI integration page documents automatic token metrics capture but does not explicitly mention service_tier or system_fingerprint capture from streaming responses. The tracing guide states that "Model parameters (model name, temperature, etc.)" and "Token usage and costs" are captured, but does not enumerate response-level metadata fields.

• https://www.braintrust.dev/docs/integrations/ai-providers/openai — documents wrapOpenAI / wrap_openai automatic tracing
• https://www.braintrust.dev/docs/guides/tracing — mentions model parameters and token usage capture

Upstream sources

• OpenAI ChatCompletionChunk type in Python SDK defines service_tier and system_fingerprint: https://github.com/openai/openai-python/blob/main/src/openai/types/chat/chat_completion_chunk.py
• OpenAI Chat Completions streaming events reference: https://developers.openai.com/api/reference/resources/chat/subresources/completions/streaming-events
• OpenAI Chat Completions create reference (documents service_tier parameter and response field): https://developers.openai.com/api/reference/resources/chat/subresources/completions/methods/create

Relationship to existing issues

• Distinct from [bot] Anthropic usage extractor drops service_tier and inference_geo execution metadata #45 (Anthropic service_tier and inference_geo): [bot] Anthropic usage extractor drops service_tier and inference_geo execution metadata #45 covers the Anthropic non-streaming usage extractor dropping string metadata from the usage object. This issue is about the OpenAI streaming aggregator dropping chunk-level response metadata — a different provider, different API surface (streaming aggregator vs. non-streaming extractor), and different field locations (response root vs. usage sub-object).
• Distinct from [BOT ISSUE] No support for OpenAI Responses API format in streaming or usage extraction #31 (OpenAI Responses API format): [BOT ISSUE] No support for OpenAI Responses API format in streaming or usage extraction #31 covers the entirely different Responses API streaming event format. This issue is about metadata loss within the supported Chat Completions streaming format.
• Distinct from [bot] Streaming aggregator collapses multiple choices (n>1) into a single output #48 (multi-choice collapse): [bot] Streaming aggregator collapses multiple choices (n>1) into a single output #48 is about choice content aggregation logic. This is about chunk-level metadata fields outside choices.

Local files inspected

• src/stream.rs:687-694 — StreamChunk struct only has model, choices, usage fields
• src/stream.rs:983 — aggregate() always creates StreamMetadata::new(model, HashMap::new())
• src/stream.rs:488-535 — StreamMetadata struct has extra: HashMap<String, Value> that could hold these values
• src/stream.rs:840-1009 — full aggregate() function; no code path reads service_tier, system_fingerprint, id, or created from chunks