[bot] OpenAI audio transcription and speech spans missing braintrust.output_json #131

Gap description

TracingHttpClient correctly creates a span for every OpenAI HTTP call, including calls to the Audio API (POST /v1/audio/transcriptions, POST /v1/audio/speech, POST /v1/audio/translations). However, InstrumentationSemConv.tagOpenAIResponse() only sets braintrust.output_json when the response body contains a choices array (Chat Completions) or an output array (Responses API). Audio endpoint responses use neither field:

Because neither shape is recognized, braintrust.output_json is never populated for these calls. The span is created and braintrust.input_json is set, but the output is silently dropped.

Relevant code

InstrumentationSemConv.java — tagOpenAIResponse() (checks for choices / output only):

// Only choices-based and output-based shapes are handled today.
// Audio responses ({"text":"..."}) fall through without setting output_json.

TracingHttpClient.java — tagSpanFromBuffer() calls InstrumentationSemConv.tagLLMSpanResponse() with the raw response bytes, but the downstream parser discards unrecognized shapes.

Suggested fix

In InstrumentationSemConv.tagOpenAIResponse(), add a branch for the text-keyed transcription/translation shape:

} else if (root.has("text") && !root.has("choices") && !root.has("output")) {
    // Audio transcription / translation response
    span.setAttribute("braintrust.output_json", responseJson);
}

For binary speech responses the content-type will be an audio MIME type (audio/mpeg, audio/opus, etc.); those should be skipped or represented with a placeholder rather than passing raw bytes as a string attribute.

Impact

Users calling client.audio().transcriptions().create(...) or .translations().create(...) today see an LLM span with no output, making it impossible to review the transcribed text in the Braintrust UI.