It seems likely that there is ongoing work and discussion about aligning or extending Phoenix/OpenInference semantic conventions to better support OpenAI's Responses API, especially as the structure of agent SDK outputs (like `{"type": ..., "text": ..., "annotation": ...}`) can differ from standard chat completions and may be harder to read in Phoenix. There are open feature requests and issues about supporting OpenAI agent response payloads and unifying conventions for different OpenAI APIs, but as of now, there is no explicit confirmation in the data sources that this has been fully implemented or released yet. The current behavior you observe is likely due to the agent SDK wrapping outputs to match the Responses API format, as you described, and this is a known area for improvement in Phoenix and OpenInference instrumentation for better readability and consistency in the UI.
Would you like more detail on the current state of these discussions or on possible workarounds?
Sources:
- <https://github.com/Arize-ai/phoenix/issues/7710|https://github.com/Arize-ai/phoenix/issues/7710>
- <https://github.com/Arize-ai/openinference/blob/main/python/instrumentation/openinference-instrumentation-openai/src/openinference/instrumentation/openai/_response_attributes_extractor.py|https://github.com/Arize-ai/openinference/blob/main/python/instrumentation/openinference-instrumentation-openai/src/openinference/instrumentation/openai/_response_attributes_extractor.py>
