Add support for OpenAI /v1/responses endpoint

[ggsmith842]

Problem:
Currently, Docker Model Runner supports the OpenAI provider endpoints for the Chat Completions API. OpenAI has indicated that users should begin migrating away from Chat Completions to the new Responses API. Adding support for the /v1/responses endpoint to ensure ongoing compatibility and give users access to new OpenAI features.

Proposal:

• Implement support for the OpenAI /v1/responses endpoint in the provider layer.
• Update API routing and model selection logic as needed.
• Maintain backward compatibility with existing /chat/completions and /embeddings behavior.
• Document usage and migration steps.

OpenAI Ref:

• migrate-to-responses

[ggsmith842]

I'm happy to take a stab at this.

[ilopezluna]

Hello @ggsmith842! A few weeks ago we added this endpoint:

model-runner/pkg/responses/handler.go

Line 69 in 9a168e7

func (h *HTTPHandler) handleCreate(w http.ResponseWriter, r *http.Request) {

. Have you had a chance to try it? Did you run into any issues?

We might be missing proper documentation for this endpoint, would you be interested in taking a look? Here are the docs: https://github.com/docker/docs/tree/042c4c4d13b9a7a5160c1cbea83b2a3bc9643f82/content/manuals/ai/model-runner

[ggsmith842]

Thanks for the info @ilopezluna! I should have added what I was running into that led me to open the issue.

I was trying to use the OpenAI Agents SDK with DMR. Everything works as expected using OpenAIChatCompletionsModel but using OpenAIResponsesModel returns an error. I assumed it was due to lack of support for the Responses API.

from openai import AsyncOpenAI
from agents import (
    Agent,
    Runner,
    OpenAIResponsesModel,
)

DOCKER_ENDPOINT_URL = "http://127.0.0.1:12434/engines/v1"
MODEL = "ai/gemma4"

client = AsyncOpenAI(base_url=DOCKER_ENDPOINT_URL, api_key="dmr")
model = OpenAIResponsesModel(model=MODEL, openai_client=client)
agent = Agent(
    name="History Tutor",
    instructions="You answer history questions concisely.",
    model=model,
)

result = await Runner.run(agent, "Hello! Are you working correctly?")
print(result.final_output)

Error Message

Error getting response: not found. (request_id: None)

NotFoundError                             Traceback (most recent call last)
Cell In[9], [line 1](vscode-notebook-cell:?execution_count=9&line=1)
----> [1](vscode-notebook-cell:?execution_count=9&line=1) result = await Runner.run(agent, "Hello! Are you working correctly?")
      2 print(result.final_output)

File ~/Documents/agents/.venv/lib/python3.14/site-packages/agents/run.py:223, in Runner.run(cls, starting_agent, input, context, max_turns, hooks, run_config, error_handlers, previous_response_id, auto_previous_response_id, conversation_id, session)
    173 """
    174 Run a workflow starting at the given agent.
    175 
   (...)    219     type of the output.
    220 """
    222 runner = DEFAULT_AGENT_RUNNER
--> [223](https://<...>/agents/.venv/lib/python3.14/site-packages/agents/run.py:223) return await runner.run(
    224     starting_agent,
    225     input,
    226     context=context,
    227     max_turns=max_turns,
    228     hooks=hooks,
    229     run_config=run_config,
    230     error_handlers=error_handlers,
    231     previous_response_id=previous_response_id,
    232     auto_previous_response_id=auto_previous_response_id,
    233     conversation_id=conversation_id,
    234     session=session,
...
-> [1698](https://<...>/agents/.venv/lib/python3.14/site-packages/openai/_base_client.py:1698)         raise self._make_status_error_from_response(err.response) from None
   1700     break
   1702 assert response is not None, "could not resolve response (should never happen)"

NotFoundError: not found

[ilopezluna]

Thanks for the details @ggsmith842! could you try:
DOCKER_ENDPOINT_URL = "http://127.0.0.1:12434/v1"

Seems like we are not registering the DOCKER_ENDPOINT_URL = "http://127.0.0.1:12434/engines/v1" and that's why you get the not found error

[ggsmith842]

Same error as above using http://127.0.0.1:12434/v1

[ilopezluna]

I noticed that the endpoint seems to be missing when Docker Model Runner is installed via Docker Desktop. Is that the case for you?

docker model version
Client:
 Version:    v1.1.33
 OS/Arch:    darwin/arm64

Server:
 Version:    v1.1.29
 Engine:     Docker Desktop

If that’s the case, as a workaround you could pull the project and run:
MODEL_RUNNER_PORT=8080 make run
This will start DMR listening on port 8080.

Then, try:

curl -s -w "\nHTTP_STATUS: %{http_code}\n" http://localhost:8080/v1/responses \
  -H "Content-Type: application/json" \
  -d '{"model": "smollm2", "input": "Say hello in one word"}'

You should get a response like this (HTTP 200):

{"id":"resp_...","output_text":"Say Hello!", ...}

HTTP_STATUS: 200

[ericcurtin]

Note I think we initially tried to implement /v1/responses in the DMR layer (but I didn't look at the code, so would have to confirm). I think llama.cpp and vLLM have decent /v1/responses support now, so we can just use that.

[ggsmith842]

I noticed that the endpoint seems to be missing when Docker Model Runner is installed via Docker Desktop. Is that the case for you?

docker model version
Client:
 Version:    v1.1.33
 OS/Arch:    darwin/arm64

Server:
 Version:    v1.1.29
 Engine:     Docker Desktop

If that’s the case, as a workaround you could pull the project and run: MODEL_RUNNER_PORT=8080 make run This will start DMR listening on port 8080.

Then, try:

curl -s -w "\nHTTP_STATUS: %{http_code}\n" http://localhost:8080/v1/responses \
  -H "Content-Type: application/json" \
  -d '{"model": "smollm2", "input": "Say hello in one word"}'

You should get a response like this (HTTP 200):

{"id":"resp_...","output_text":"Say Hello!", ...}

HTTP_STATUS: 200

Yea using Docker Desktop. Not a big deal what I'm doing but felt like something that would need to be implemented in case support for ChatCompletions goes away.