Summary

The OpenAI Responses API streaming tracer only processes response.completed events. The two other terminal event types — response.failed and response.incomplete — are silently ignored. When a streaming response ends with failure or truncation, the span captures no output, no error status, and no usage metrics (only time_to_first_token).

What is missing

The OpenAI Responses API defines three terminal streaming events (docs):

In trace/contrib/openai/responses.go lines 132–147, only response.completed is matched:

if msgType == "response.completed" {
    // ...
    completedMsg = msg
}

When the stream ends with response.failed:

• The response object in the event contains an error field with code and message
• The status field is "failed"
• None of this is captured — the span ends with no output, no error status, and no error details

When the stream ends with response.incomplete:

• The response object contains incomplete_details explaining why (e.g. max_output_tokens)
• The output array may contain partial results
• The usage object is present with final token counts
• None of this is captured — partial output and usage metrics are lost

By contrast, the non-streaming path (parseResponse) handles the full response body regardless of status, so it correctly captures failed/incomplete responses.

Braintrust docs status

not_found — The Braintrust OpenAI integration docs do not mention handling of response.failed or response.incomplete streaming events specifically.

Upstream sources

• OpenAI Responses API streaming events: platform.openai.com/docs/api-reference/responses/streaming-events
• OpenAI Responses API create: platform.openai.com/docs/api-reference/responses/create — documents status field values (completed, failed, incomplete)
• OpenAI streaming guide: platform.openai.com/docs/guides/streaming-responses

Local repo files inspected

• trace/contrib/openai/responses.go — parseStreamingResponse() at lines 109–162: only response.completed event type is handled
• trace/contrib/openai/responses.go — handleResponseCompletedMessage() at lines 177–229: shared handler that would work for failed/incomplete if called
• trace/contrib/openai/traceopenai.go — router and shared utilities
• trace/internal/middleware.go — generic middleware handles HTTP-level errors but not stream-level API failures (streaming responses return HTTP 200)
• trace/contrib/openai/chatcompletions.go — Chat Completions streaming for comparison (aggregates deltas, so partial output is always captured)