add OpenAI responses support

[M-Hietala]

Summary

This PR adds instrumentation support for the OpenAI Responses API (structured outputs) to the opentelemetry-instrumentation-openai-v2 library, following the same monkeypatching pattern used for chat completions.

Background

The OpenAI SDK introduced the Responses API (client.responses.create) for structured outputs in version 1.66.0. This API was not previously instrumented, meaning calls to it would not generate telemetry data (spans, logs, or metrics).

Changes

This PR instruments both synchronous and asynchronous versions of the Responses API:

from openai import OpenAI
from opentelemetry.instrumentation.openai_v2 import OpenAIInstrumentor

OpenAIInstrumentor().instrument()

client = OpenAI()

# Now automatically instrumented!
response = client.responses.create(
    model="gpt-4o-mini",
    input="Write a short poem on open telemetry.",
)

client.conversations.create()

items = client.conversations.items.list(conversation_id=conversation.id)
# Print all the items
for item in items:
    display_conversation_item(item)

Implementation Details

Version Checking:

• Added _is_responses_api_supported() function to detect if OpenAI SDK >= 1.66.0
• Instrumentation only wraps responses API when supported version is detected
• Chat completions instrumentation is always enabled (no version requirement)
• Uses packaging.version for reliable version comparison

New wrapper functions in patch.py:

• responses_create() - Wraps synchronous Responses.create method
• async_responses_create() - Wraps asynchronous AsyncResponses.create method
• _set_responses_attributes() - Sets span attributes for responses
• _record_responses_metrics() - Records metrics for responses API calls

Instrumentation hooks in __init__.py:

• Added conditional wrap_function_wrapper calls for openai.resources.responses.responses.Responses.create
• Added conditional wrap_function_wrapper calls for openai.resources.responses.responses.AsyncResponses.create
• Added corresponding conditional unwrap calls in _uninstrument() method

Telemetry Captured

The instrumentation captures (when responses API is available):

• Spans with attributes including operation name, model, response ID, service tier, and token usage
• Span events for input/output messages (when OTEL_INSTRUMENTATION_GENAI_CAPTURE_MESSAGE_CONTENT=true)
• Metrics for operation duration and token usage (input/output tokens)

Tests

Added comprehensive test coverage with version-aware skipping:

• test_responses.py - Tests for synchronous responses API with/without content capture (skipped if OpenAI < 1.66.0)
• test_async_responses.py - Tests for asynchronous responses API with/without content capture (skipped if OpenAI < 1.66.0)
• 'test_conversations.py' - Tests for synchronous conversations API with/without content capture (skipped if OpenAI < 1.101.0)
• 'test_async_conversations.py' - Tests for asynchronous conversations API with/without content capture (skipped if OpenAI < 1.101.0)

Documentation

Updated documentation to include responses API examples:

• README.rst - Added usage example showing both chat completions and responses API
• Module docstring in __init__.py - Added responses API example

Bug Fixes

• Fixed ChatCompletion imports to use openai.types.chat instead of openai.resources.chat.completions

Testing

Verified that:

• All methods are correctly wrapped after instrumentation
• All methods are correctly unwrapped after uninstrumentation
• Spans capture correct attributes (model, tokens, service tier)
• Events capture input/output based on content capture setting
• Metrics are recorded for duration and token usage
• Implementation follows existing code patterns and style
• Version checking correctly detects supported/unsupported OpenAI versions (1.66.0 threshold)
• Tests are automatically skipped when OpenAI version doesn't support responses API
• ChatCompletion imports are correct and use the proper type location

Compatibility

• OpenAI SDK: >= 1.26.0 (minimum version), >= 1.66.0 (for responses API support)
• Python: >= 3.9
• OpenTelemetry API: ~= 1.37

Backward Compatibility

This implementation maintains full backward compatibility. Users with OpenAI SDK versions < 1.66.0 will continue to have chat completions instrumented while responses API instrumentation is gracefully skipped.

Original prompt

Add support to the openai v2 instrumentation library for the openai responses API. Use the same pattern (monkeypatching) as is done for the chat completions.

---

💡 You can make Copilot smarter by setting up custom instructions, customizing its development environment and configuring Model Context Protocol (MCP) servers. Learn more Copilot coding agent tips in the docs.

[linux-foundation-easycla]

• ✅ login: Copilot / name: copilot-swe-agent[bot] (28e348e, 34d589b, 3653a91, 75f1c8d, 80d670e, 91a2bde, df8e7f9)
• ❌ - login: @johanste / name: Johan Stenberg (MSFT) . The commit (339c4b3) is not authorized under a signed CLA. Please click here to be authorized. For further assistance with EasyCLA, please submit a support request ticket.
• ❌ - login: @M-Hietala / name: M-Hietala . The commit (3f5b74c, 6ac62a2, 8268099, eb3e8f7) is not authorized under a signed CLA. Please click here to be authorized. For further assistance with EasyCLA, please submit a support request ticket.

[shuwpan]

Just a thought, the patch.py is getting bigger and bigger...anyone think we should split it up ? I assume we will keep adding things in the future. so...just wondering...

[JWinermaSplunk]

Agree with the previous comment, responses itself could probably be a separate file. Also didn't look too deeply into it, but I think there is a bit of code here that can be moved into helper functions (to prevent duplicate code, decrease clutter, etc).

[iamemilio]

Just a thought, the patch.py is getting bigger and bigger...anyone think we should split it up ? I assume we will keep adding things in the future. so...just wondering...

There is a style check in one of the test suites that enforces a 1000 line limit for any file otherwise a set of changes will fail to score a 10/10 and the test will get marked fail. I believe a refactor will be necessary to pass CI for these changes.

[iamemilio]

Suggested change

	response = await async_openai_client.responses.create(
	input=input_value, model=llm_model_value
	response = await async_openai_client.responses.create(
	input=input_value,
	model=llm_model_value,
	stream_options={"include_usage": True}

This will make sure that token usage gets returned, so that we can test that the instrument captures it. That said, I do see usage in your cassettes, so this is more of a nit.

[JWinermaSplunk]

Hi, is there any progress on this?

[iamemilio]

I no longer have the capacity to work on this, I'm sorry. Can we mark this is stale so someone else can pick it up? It has been stale for a very long time.

[JWinermaSplunk]

Hi @iamemilio,

I don't mind picking it up, I was just wondering if there has been any communication from the original author, @M-Hietala?

[iamemilio]

None that I am aware of, no :(

[xrmx]

Closing in favor of #4166