[OpenAI] Responses API support for Python

[codefromthecrypt]

Is your feature request related to a problem? Please describe.

The OpenAI platform just released an alternative to the completions endpoint called responses

https://platform.openai.com/docs/api-reference/responses

They are now using this by default in the playground (which generates content) as well as in their libraries like the OpenAI Agents SDK.

Soon, users will complain as they've been pushed into an endpoint that we don't have trace support for.

Describe the solution you'd like

Logs, metrics and tracing for responses like we do for completions.

TODO feature checkboxes

Describe alternatives you've considered

Tell people to not use responses

Additional context

announcement: https://community.openai.com/t/introducing-the-responses-api/1140929

local inference support of core features

• litellm https://github.com/BerriAI/litellm/releases/tag/v1.67.0-stable
• llama-stack https://github.com/meta-llama/llama-stack/releases/tag/v0.2.4
• masaic-ai https://github.com/masaic-ai-platform/open-responses/releases/tag/v0.1.3-M2
• ollama Support for OpenAI Responses API (for Codex CLI compatibility) ollama/ollama#10309
• vllm [Feature]: Support openai responses API interface vllm-project/vllm#14721

other python SDKs

• langtrace https://github.com/Scale3-Labs/langtrace-python-sdk/releases/tag/3.8.13
• logfire https://github.com/pydantic/logfire/releases/tag/v3.8.0
• openinference https://github.com/Arize-ai/openinference/releases/tag/python-openinference-instrumentation-openai-v0.1.26
• openlit https://github.com/openlit/openlit/releases/tag/py-1.33.16
• openllmetry Feature: Instrument OpenAI's new Responses API #2782 traceloop/openllmetry#2787

[codefromthecrypt]

FYI I updated this as now there are two released python instrumentation libraries that claim to support this (langtrace and logfire), as well litellm claims to support the server-side.

Note: A project owner will need to further edit this issue description to break down what support means. For example, if it means the same features as needed by openai codex, or something else. Also, if there are inaccuracies or anything that becomes out of date in the issue description, a lead likely want to edit in corrections.

[codefromthecrypt]

note there are now 4 independent python instrumentation that support responses and 3 different local inference platforms (though not ollama, yet)

[codefromthecrypt]

note: I've verified that at least at a basic level, llama-stack and open-responses emulate the responses API when proxying to ollama.

There may be other instrumentation and platform updates besides this, but I wanted to give this one as I can personally verify they work.