feat(api): add OpenAI Responses API compatibility endpoint

docs/EN/source/index.rst

@@ -54,6 +54,7 @@ Documentation List
    Reward Model Deployment <tutorial/reward_model>
    OpenAI api Usage <tutorial/openai>
    Anthropic Messages API <tutorial/anthropic>
+   OpenAI Responses API <tutorial/openai_responses>
    Function Calling <tutorial/function_calling>
    Reasoning Parser <tutorial/reasoning_parser>
    APIServer Parameters <tutorial/api_server_args_zh>

docs/EN/source/tutorial/openai_responses.rst

@@ -0,0 +1,106 @@
+.. _openai_responses_api:
+
+OpenAI Responses API (Experimental)
+===================================
+
+LightLLM can expose a ``/v1/responses`` endpoint that speaks the OpenAI
+Responses API wire protocol. This is useful if you have client code written
+against the newer OpenAI Python/TypeScript SDK (``client.responses.create``),
+Codex CLI, or the OpenAI Agents SDK and want to point it at a locally hosted
+open-source model without rewriting the client.
+
+Enabling
+--------
+
+The ``/v1/responses`` endpoint is always exposed; no extra flag and no extra
+Python dependency are required:
+
+.. code-block:: bash
+
+    python -m lightllm.server.api_server \
+        --model_dir /path/to/model \
+        --port 8088
+
+Using it from the OpenAI SDK
+----------------------------
+
+.. code-block:: python
+
+    from openai import OpenAI
+
+    client = OpenAI(
+        base_url="http://localhost:8088/v1",
+        api_key="dummy",
+    )
+    resp = client.responses.create(
+        model="any-name",  # echoed back; LightLLM serves the loaded model
+        input="hello",
+        max_output_tokens=1024,
+    )
+    print(resp.output_text)
+
+Streaming works the same way the OpenAI SDK expects:
+
+.. code-block:: python
+
+    with client.responses.stream(
+        model="any-name",
+        input="Count from 1 to 5.",
+        max_output_tokens=256,
+    ) as stream:
+        for event in stream:
+            if event.type == "response.output_text.delta":
+                print(event.delta, end="", flush=True)
+
+Function tools round-trip:
+
+.. code-block:: python
+
+    resp = client.responses.create(
+        model="any-name",
+        input="What's the weather in Paris?",
+        tools=[{
+            "type": "function",
+            "name": "get_weather",
+            "description": "Get the weather for a location",
+            "parameters": {
+                "type": "object",
+                "properties": {"location": {"type": "string"}},
+                "required": ["location"],
+            },
+        }],
+    )
+    for item in resp.output:
+        if item.type == "function_call":
+            print(item.name, item.arguments)
+
+Supported features
+------------------
+
+- Text generation (streaming and non-streaming)
+- ``instructions`` as a top-level system prompt
+- Role-based ``input`` items with ``input_text`` / ``output_text`` content parts
+- Function tools (definitions, tool calls, and ``function_call_output``
+  follow-up turns)
+- ``tool_choice`` (``auto`` / ``none`` / ``required`` / ``{type:"function", name:...}``)
+- Vision inputs via ``input_image`` content parts
+- Usage reporting (``input_tokens`` / ``output_tokens`` / ``total_tokens``)
+
+Known limitations
+-----------------
+
+- Built-in tools (``web_search``, ``file_search``, ``code_interpreter``,
+  ``computer_use_preview``) are not supported and are silently dropped from
+  the request.
+- Stateful fields (``store``, ``previous_response_id``) are accepted but
+  ignored; LightLLM is stateless and does not persist responses.
+- Structured outputs via ``text.format.json_schema`` are accepted but the
+  ``text`` field is ignored — use the equivalent Chat Completions
+  ``response_format`` via the ``/v1/chat/completions`` endpoint if you need
+  schema-constrained generation.
+- ``reasoning.summary`` is passed through from the underlying model's
+  ``reasoning_content`` when present, but LightLLM does not synthesise
+  summaries on its own.
+- Model name is accepted but ignored; LightLLM always serves the model
+  loaded via ``--model_dir`` and echoes the requested name back in the
+  response.

lightllm/server/api_http.py

@@ -277,6 +277,17 @@ async def anthropic_messages(raw_request: Request) -> Response:
     return await anthropic_messages_impl(raw_request)
 
 
+@app.post("/v1/responses")
+async def openai_responses(raw_request: Request) -> Response:
+    if get_env_start_args().run_mode in ["prefill", "decode", "nixl_prefill", "nixl_decode"]:
+        return create_error_response(
+            HTTPStatus.EXPECTATION_FAILED, "service in pd mode dont recv reqs from http interface"
+        )
+    from .api_openai_responses import openai_responses_impl
+
+    return await openai_responses_impl(raw_request)
+
+
 @app.get("/v1/models", response_model=ModelListResponse)
 @app.post("/v1/models", response_model=ModelListResponse)
 async def get_models(raw_request: Request):