feat: add OpenAI Responses API model implementation #975

notgitika · 2025-10-03T17:27:50Z

Description

This PR implements OpenAI Responses API as a separate model provider supporting streaming, structured output and tool calling.

Note: this commit does not include the extra capabilities that the Responses API supports such as the built-in tools and the stateful conversation runs.

Related Issues

#253

Documentation PR

TODO; coming next: will be adding a section for the Responses API within the existing openai model page

Type of Change

New feature

Testing

Added a unit test file similar to the existing test_openai.py model provider. Reusing the integ tests in the same file test_model_openai.py using pytest.parameterize with OpenAIResponses model so that I could test that the funcitonality is the same between the two models.

I ran everything in the CONTRIBUTING.md file.

hatch run integ-test
================== 81 passed, 68 skipped, 49 warnings in 106.56s (0:01:46) ==========

hatch run test-integ tests_integ/models/test_model_openai.py -v

============= 18 passed, 2 skipped in 13.44s =============

pre-commit run --all-files

I ran hatch run prepare --> yes, all tests pass

Checklist

I have read the CONTRIBUTING document
I have added any necessary tests that prove my fix is effective or my feature works
I have updated the documentation accordingly
I have added an appropriate example to the documentation to outline the feature, or no new docs are needed
My changes generate no new warnings
Any dependent changes have been merged and published

By submitting this pull request, I confirm that you can use, modify, copy, and redistribute this contribution, under the terms of your choice.

src/strands/models/openai_responses.py

dbschmigelski · 2025-10-09T17:23:33Z

src/strands/models/openai_responses.py

+        """
+        match event["chunk_type"]:
+            case "message_start":
+                return {"messageStart": {"role": "assistant"}}


reach out to @zastrowm , but I believe we can make this more readable by now returning the typed StreamEvents rather than the dictionaries

I was looking into this briefly but did not implement this. I will talk to him about this and maybe we can implement this in the next iteration.

src/strands/models/openai_responses.py

fsajjad · 2025-10-24T16:33:59Z

Hey Team, this is one of the feature my customer request and concerned as blocker from adopting Strands Agent. They mentioned Strands currently only support deprecated chatcompletion API and not responses API. When can we expect this to be merged? and will this also enable streaming structured output?

dbschmigelski

A major version bump is proposed in #1370.

Adding a comment here that we would need to consider when the user has v1 installed compared to v2.

For example

import pydantic
from packaging import version

# Detect the version once at the module level
PYDANTIC_V2 = version.parse(pydantic.VERSION) >= version.parse("2.0.0")

if PYDANTIC_V2:
    from pydantic import ConfigDict
    def get_model_fields(model):
        return model.model_fields
else:
    def get_model_fields(model):
        return model.__fields__

codecov · 2026-01-21T06:19:37Z

Codecov Report

❌ Patch coverage is 91.41631% with 20 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
src/strands/models/openai_responses.py	91.41%	4 Missing and 16 partials ⚠️

📢 Thoughts on this report? Let us know!

dbschmigelski · 2026-01-21T14:53:19Z

src/strands/models/openai_responses.py

+        # Validate OpenAI SDK version - Responses API requires v2.0.0+
+        openai_version = Version(get_package_version("openai"))
+        if openai_version < _MIN_OPENAI_VERSION:
+            raise ImportError(


We should be able to do this on the import rather than in the constructor, no?

dbschmigelski · 2026-01-21T14:56:25Z

src/strands/models/openai_responses.py

+
+            # Yield tool calls if any
+            for call_info in tool_calls.values():
+                mock_tool_call = type(


mock_tool_call?

dbschmigelski · 2026-01-21T14:56:36Z

src/strands/models/openai_responses.py

+_MIN_OPENAI_VERSION = Version("2.0.0")
+
+# Maximum file size for media content in tool results (20MB)
+MAX_MEDIA_SIZE_BYTES = 20 * 1024 * 1024


why public?

dbschmigelski · 2026-01-21T15:00:10Z

src/strands/models/openai_responses.py

+        async with openai.AsyncOpenAI(**self.client_args) as client:
+            try:
+                response = await client.responses.create(**request)
+            except openai.BadRequestError as e:


it looks like we can just move to one big try catch to avoid duplicating these lines

except openai.BadRequestError as e: if hasattr(e, "code") and e.code == "context_length_exceeded": logger.warning("OpenAI Responses API threw context window overflow error") raise ContextWindowOverflowException(str(e)) from e raise except openai.RateLimitError as e: logger.warning("OpenAI Responses API threw rate limit error") raise ModelThrottledException(str(e)) from e

dbschmigelski · 2026-01-21T15:04:58Z

src/strands/models/openai_responses.py

+                        if event.type == "response.output_text.delta":
+                            # Text content streaming
+                            if not has_text_content:
+                                yield self._format_chunk({"chunk_type": "content_start", "data_type": "text"})


Does the reasoning API support reasoning content? Which would mean we'd need something like _stream_switch_content in

sdk-python/src/strands/models/openai.py

Line 612 in 7604e98

chunks, data_type = self._stream_switch_content("reasoning_content", data_type)

dbschmigelski · 2026-01-21T15:09:55Z

src/strands/models/openai_responses.py

+                )()
+
+                yield self._format_chunk({"chunk_type": "content_start", "data_type": "tool", "data": mock_tool_call})
+                yield self._format_chunk({"chunk_type": "content_delta", "data_type": "tool", "data": mock_tool_call})


Here we are only ever emitting one delta per tool? Is this correct? Wouldn't we expect multiple?

dbschmigelski · 2026-01-21T15:14:31Z

src/strands/models/openai_responses.py

+                yield self._format_chunk({"chunk_type": "content_delta", "data_type": "tool", "data": mock_tool_call})
+                yield self._format_chunk({"chunk_type": "content_stop", "data_type": "tool"})
+
+            finish_reason = "tool_calls" if tool_calls else "stop"


Based on your logic above finish_reason = "tool_calls" if tool_calls else "stop" we are never going t hit the length case in

case "message_stop": match event["data"]: case "tool_calls": return {"messageStop": {"stopReason": "tool_use"}} case "length": return {"messageStop": {"stopReason": "max_tokens"}} case _: return {"messageStop": {"stopReason": "end_turn"}}

dbschmigelski · 2026-01-21T15:17:11Z

src/strands/models/openai_responses.py

+                    (),
+                    {
+                        "prompt_tokens": getattr(final_usage, "input_tokens", 0),
+                        "completion_tokens": getattr(final_usage, "output_tokens", 0),


What is the reason behind manipulating these names input_tokens->prompt_tokens

Do we need to add something like what we do in openai v1 https://github.com/strands-agents/sdk-python/blob/main/src/strands/models/openai.py#L430C33-L430C46

dbschmigelski · 2026-01-21T15:18:30Z

src/strands/models/openai_responses.py

+        ...
+
+
+class OpenAIResponsesModel(Model):


Is it possible to use system tools with the current implementation?

dbschmigelski · 2026-01-21T15:22:05Z

src/strands/models/openai_responses.py

+
+        for message in messages:
+            role = message["role"]
+            if role == "system":


Why do we need this? this is format request and Messages should never be able to contain a Role of type system?

notgitika had a problem deploying to manual-approval October 3, 2025 17:28 — with GitHub Actions Failure

JackYPCOnline reviewed Oct 3, 2025

View reviewed changes

src/strands/models/openai_responses.py Show resolved Hide resolved

dbschmigelski requested changes Oct 9, 2025

View reviewed changes

pgrayy mentioned this pull request Jan 8, 2026

[FEATURE] Missing support for native web_search_preview for OpenAI #663

Open

dbschmigelski reviewed Jan 9, 2026

View reviewed changes

dbschmigelski mentioned this pull request Jan 9, 2026

ci: update openai requirement from <1.110.0,>=1.68.0 to >=1.68.0,<2.15.0 #1370

Closed

notgitika added 2 commits January 21, 2026 06:13

feat: add OpenAI Responses API model implementation

2b53376

fix: address comments and refactor

8eea5d9

notgitika force-pushed the gitikavj/add-openai-responses-model branch from 7c05e84 to 8eea5d9 Compare January 21, 2026 06:16

github-actions bot added the size/xl label Jan 21, 2026

notgitika requested a deployment to manual-approval January 21, 2026 06:17 — with GitHub Actions Waiting

feat: add conditional to check v1 vs v2 ; add tests to increase coverage

e09f874

github-actions bot added size/xl and removed size/xl labels Jan 21, 2026

notgitika requested a deployment to manual-approval January 21, 2026 06:41 — with GitHub Actions Waiting

notgitika requested review from JackYPCOnline and dbschmigelski January 21, 2026 14:43

dbschmigelski reviewed Jan 21, 2026

View reviewed changes

feat: add OpenAI Responses API model implementation #975

Are you sure you want to change the base?

feat: add OpenAI Responses API model implementation #975

Conversation

notgitika commented Oct 3, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Description

Related Issues

Documentation PR

Type of Change

Testing

Checklist

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

fsajjad commented Oct 24, 2025

Uh oh!

dbschmigelski left a comment

Choose a reason for hiding this comment

Uh oh!

codecov bot commented Jan 21, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Codecov Report

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

notgitika commented Oct 3, 2025 •

edited

Loading

codecov bot commented Jan 21, 2026 •

edited

Loading