[bot] BraintrustStream cannot aggregate Google Gemini streamGenerateContent events

[braintrust-bot]

Summary

BraintrustStream and wrap_stream_with_span only handle OpenAI Chat Completions streaming chunks (choices[].delta). Google Gemini's streamGenerateContent endpoint emits GenerateContentResponse objects with a completely different structure (candidates[].content.parts[]). All Gemini streaming chunks are silently discarded by the Err(_) => continue fallback in aggregate(), producing an empty aggregated result with no output, no usage metrics, and no TTFT.

This is distinct from #34 (non-streaming Gemini usageMetadata extraction), which covers the non-streaming case. This issue is specifically about the streaming aggregation path.

What is missing

Gemini's streamGenerateContent yields a sequence of GenerateContentResponse JSON objects. Each chunk has this structure:

{
  "candidates": [{
    "content": {
      "parts": [{ "text": "Hello" }],
      "role": "model"
    },
    "finishReason": "STOP"
  }]
}

The final chunk includes usageMetadata:

{
  "candidates": [...],
  "usageMetadata": {
    "promptTokenCount": 10,
    "candidatesTokenCount": 25,
    "totalTokenCount": 35,
    "thoughtsTokenCount": 12,
    "cachedContentTokenCount": 5
  }
}

For tool/function calling, parts contain a functionCall object instead of text:

{
  "candidates": [{
    "content": {
      "parts": [{ "functionCall": { "name": "get_weather", "args": { "location": "NYC" } } }],
      "role": "model"
    }
  }]
}

For Gemini 2.5 thinking models, parts include a thought field:

{ "parts": [{ "text": "Let me think...", "thought": true }] }

Currently in this SDK, BraintrustStream::aggregate() (src/stream.rs) attempts to deserialize each raw chunk as StreamChunk { model, choices, usage }. Gemini chunks have no choices field, so serde_json::from_value fails and the chunk is skipped. This means:

• Text output from candidates[].content.parts[].text is lost
• Function call output from candidates[].content.parts[].functionCall is lost
• Thinking content from Gemini 2.5 thought parts is lost
• Usage metrics (promptTokenCount, candidatesTokenCount, thoughtsTokenCount, cachedContentTokenCount) are never extracted
• Finish reason from candidates[].finishReason is lost
• TTFT metric is not recorded (the content heuristic value_has_content() checks for choices which Gemini chunks don't have)

Braintrust docs status

supported (in other language SDKs) — Braintrust documents full Gemini streaming support including token metrics, thinking model support, and function call tracing:

• https://www.braintrust.dev/docs/integrations/ai-providers/gemini — documents wrapGoogleGenAI() (TypeScript), setup_genai() (Python), token metrics, thinking tokens, cached tokens
• https://www.braintrust.dev/docs/instrument/trace-llm-calls — lists Gemini as a supported provider for TypeScript, Python, Go, Java; no Rust entry

Status for the Rust SDK: not instrumented

Upstream sources

• Gemini streamGenerateContent REST reference: https://ai.google.dev/api/generate-content#method:-models.streamgeneratecontent
• Gemini GenerateContentResponse schema (candidates, usageMetadata): https://ai.google.dev/api/generate-content#v1beta.GenerateContentResponse
• Gemini function calling (streaming): https://ai.google.dev/gemini-api/docs/function-calling
• Gemini 2.5 thinking models (thought parts in streaming): https://ai.google.dev/gemini-api/docs/thinking
• Gemini token counting (usageMetadata fields including thoughtsTokenCount, cachedContentTokenCount): https://ai.google.dev/gemini-api/docs/tokens

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 extract_openai_usage/extract_anthropic_usage parity for non-streaming Gemini responses. This issue covers the streaming aggregation path (BraintrustStream, wrap_stream_with_span) where chunks have a completely different structure.
• Distinct from [BOT ISSUE] BraintrustStream cannot aggregate Anthropic Messages streaming events #38 (Anthropic streaming): [BOT ISSUE] BraintrustStream cannot aggregate Anthropic Messages streaming events #38 covers Anthropic's SSE event protocol (message_start, content_block_delta, etc.). This issue covers Gemini's GenerateContentResponse chunk format — a different provider with a different streaming schema.

Local files inspected

• src/stream.rs — StreamChunk struct only has model, choices, usage; Gemini chunks have candidates not choices; aggregate() skips all non-parseable chunks via Err(_) => continue; value_has_content() checks for choices array which Gemini chunks lack
• src/extractors.rs — extract_openai_usage() and extract_anthropic_usage() both call value.get("usage"); Gemini uses usageMetadata so neither can extract Gemini stream usage
• src/lib.rs — wrap_stream_with_span is the primary streaming instrumentation surface; no Gemini-specific path
• Cargo.toml — no Google AI / Vertex AI dependencies
• Full codebase grep for gemini, google, genai, usageMetadata, candidatesTokenCount, candidates — zero results