[bot] ADK integration sets OTel semconv attributes instead of `braintrust.metrics`/`braintrust.metadata`

[braintrust-bot]

Summary

The ADK (trace/contrib/adk) integration records token usage as raw OpenTelemetry semantic-convention attributes (gen_ai.usage.prompt_tokens, gen_ai.usage.completion_tokens, gen_ai.usage.total_tokens) instead of the braintrust.metrics JSON blob that the Braintrust platform expects. It also never sets braintrust.metadata (model name, temperature, etc.) or time_to_first_token. This means ADK model-call spans are invisible to Braintrust's token-usage dashboards, cost tracking, and metadata filters.

Current behaviour

In trace/contrib/adk/traceadk.go lines 379-388, the AfterModel callback sets token counts as individual OTel attributes:

if resp.UsageMetadata != nil {
    if resp.UsageMetadata.PromptTokenCount > 0 {
        span.SetAttributes(attribute.Int64("gen_ai.usage.prompt_tokens", int64(resp.UsageMetadata.PromptTokenCount)))
    }
    if resp.UsageMetadata.CandidatesTokenCount > 0 {
        span.SetAttributes(attribute.Int64("gen_ai.usage.completion_tokens", int64(resp.UsageMetadata.CandidatesTokenCount)))
    }
    if resp.UsageMetadata.TotalTokenCount > 0 {
        span.SetAttributes(attribute.Int64("gen_ai.usage.total_tokens", int64(resp.UsageMetadata.TotalTokenCount)))
    }
}

No braintrust.metadata attribute is ever set.

Expected behaviour

Token usage should be serialised into a braintrust.metrics JSON attribute (matching the pattern used by every other integration — OpenAI, Anthropic, GenAI, Bedrock, Genkit, Eino, LangChainGo):

{
  "prompt_tokens": 100,
  "completion_tokens": 42,
  "tokens": 142,
  "time_to_first_token": 0.35
}

Model metadata should be recorded in braintrust.metadata:

{
  "model": "gemini-2.0-flash"
}

References

• Every other HTTP-middleware integration uses braintrust.metrics — e.g. trace/contrib/openai/chatcompletions.go, trace/contrib/anthropic/messages.go, trace/contrib/genai/generatecontent.go.
• The Genkit integration (trace/contrib/genkit/tracegenkit.go) is the closest analogue (typed-SDK callback, Google models) and correctly sets both attributes.