Update v1/responses to be more OpenAI-compatible. (#9624)

python/sglang/srt/entrypoints/http_server.py

@@ -299,7 +299,23 @@ app.add_middleware(
 
 @app.exception_handler(HTTPException)
 async def validation_exception_handler(request: Request, exc: HTTPException):
-    """Enrich HTTP exception with status code and other details"""
+    """Enrich HTTP exception with status code and other details.
+
+    For /v1/responses, emit OpenAI-style nested error envelope:
+    {"error": {"message": "...", "type": "...", "param": null, "code": <status>}}
+    """
+    # adjust fmt for responses api
+    if request.url.path.startswith("/v1/responses"):
+        nested_error = {
+            "message": exc.detail,
+            "type": HTTPStatus(exc.status_code).phrase,
+            "param": None,
+            "code": exc.status_code,
+        }
+        return ORJSONResponse(
+            content={"error": nested_error}, status_code=exc.status_code
+        )
+
     error = ErrorResponse(
         object="error",
         message=exc.detail,
@@ -312,7 +328,10 @@ async def validation_exception_handler(request: Request, exc: HTTPException):
 # Custom exception handlers to change validation error status codes
 @app.exception_handler(RequestValidationError)
 async def validation_exception_handler(request: Request, exc: RequestValidationError):
-    """Override FastAPI's default 422 validation error with 400"""
+    """Override FastAPI's default 422 validation error with 400.
+
+    For /v1/responses, emit OpenAI-style nested error envelope; for other endpoints keep legacy format.
+    """
     exc_str = str(exc)
     errors_str = str(exc.errors())
 
@@ -321,6 +340,16 @@ async def validation_exception_handler(request: Request, exc: RequestValidationE
     else:
         message = exc_str
 
+    if request.url.path.startswith("/v1/responses"):
+        # adapt specially, for v1/responses API only (notice the error key is different)
+        nested_error = {
+            "message": message,
+            "type": HTTPStatus.BAD_REQUEST.phrase,
+            "param": None,
+            "code": HTTPStatus.BAD_REQUEST.value,
+        }
+        return ORJSONResponse(status_code=400, content={"error": nested_error})
+
     err = ErrorResponse(
         message=message,
         type=HTTPStatus.BAD_REQUEST.phrase,

python/sglang/srt/entrypoints/openai/protocol.py

@@ -22,6 +22,8 @@ from openai.types.responses import (
     ResponseFunctionToolCall,
     ResponseInputItemParam,
     ResponseOutputItem,
+    ResponseOutputMessage,
+    ResponseOutputText,
     ResponseReasoningItem,
 )
 from openai.types.responses.response import ToolChoice
@@ -881,6 +883,26 @@ class ResponsesResponse(BaseModel):
     tool_choice: str = "auto"
     tools: List[ResponseTool] = Field(default_factory=list)
 
+    # OpenAI compatibility fields. not all are used at the moment.
+    # Recommend checking https://platform.openai.com/docs/api-reference/responses
+    error: Optional[dict] = None
+    incomplete_details: Optional[dict] = None  # TODO(v) support this input
+    instructions: Optional[str] = None
+    max_output_tokens: Optional[int] = None
+    previous_response_id: Optional[str] = None
+    reasoning: Optional[dict] = (
+        # Unused. No model supports this. For GPT-oss, system prompt sets
+        # the field, not server args.
+        None  # {"effort": Optional[str], "summary": Optional[str]}
+    )
+    store: Optional[bool] = None
+    temperature: Optional[float] = None
+    text: Optional[dict] = None  # e.g. {"format": {"type": "text"}}
+    top_p: Optional[float] = None
+    truncation: Optional[str] = None
+    user: Optional[str] = None
+    metadata: Optional[Dict[str, Any]] = None
+
     @classmethod
     def from_request(
         cls,
@@ -895,6 +917,41 @@ class ResponsesResponse(BaseModel):
         usage: Optional[UsageInfo],
     ) -> "ResponsesResponse":
         """Create a response from a request."""
+
+        # Determine if the output is plain text only to set text.format
+        def _is_text_only(
+            items: List[
+                Union[
+                    ResponseOutputItem, ResponseReasoningItem, ResponseFunctionToolCall
+                ]
+            ]
+        ) -> bool:
+            if not items:
+                return False
+            for it in items:
+                # tool call -> not pure text.
+                if isinstance(it, ResponseReasoningItem) or isinstance(
+                    it, ResponseFunctionToolCall
+                ):
+                    return False
+                try:
+                    if isinstance(it, ResponseOutputText):
+                        continue
+                    elif isinstance(it, ResponseOutputMessage):
+                        if not it.content:
+                            continue
+                        for c in it.content:
+                            if not isinstance(c, ResponseOutputText):
+                                return False
+                    else:
+                        # Unknown type, not considered text-only
+                        return False
+                except AttributeError:
+                    return False
+            return True
+
+        text_format = {"format": {"type": "text"}} if _is_text_only(output) else None
+
         return cls(
             id=request.request_id,
             created_at=created_time,
@@ -905,6 +962,23 @@ class ResponsesResponse(BaseModel):
             parallel_tool_calls=request.parallel_tool_calls or True,
             tool_choice=request.tool_choice,
             tools=request.tools,
+            # fields for parity with v1/responses
+            error=None,
+            incomplete_details=None,
+            instructions=request.instructions,
+            max_output_tokens=request.max_output_tokens,
+            previous_response_id=request.previous_response_id,  # TODO(v): ensure this is propagated if retrieved from store
+            reasoning={
+                "effort": request.reasoning.effort if request.reasoning else None,
+                "summary": None,  # unused
+            },
+            store=request.store,
+            temperature=request.temperature,
+            text=text_format,  # TODO(v): Expand coverage per https://platform.openai.com/docs/api-reference/responses/list
+            top_p=request.top_p,
+            truncation=request.truncation,
+            user=request.user,
+            metadata=request.metadata or {},
         )
 
 

python/sglang/srt/entrypoints/openai/serving_responses.py

@@ -123,6 +123,39 @@ class OpenAIServingResponses(OpenAIServingChat):
 
         self.background_tasks: dict[str, asyncio.Task] = {}
 
+    # error helpers dedicated for v1/responses
+    def create_error_response(
+        self,
+        message: str,
+        err_type: str = "invalid_request_error",
+        status_code: int = 400,
+        param: Optional[str] = None,
+    ) -> ORJSONResponse:
+        nested_error = {
+            "message": message,
+            "type": err_type,
+            "param": param,
+            "code": status_code,
+        }
+        return ORJSONResponse(content={"error": nested_error}, status_code=status_code)
+
+    def create_streaming_error_response(
+        self,
+        message: str,
+        err_type: str = "BadRequestError",
+        status_code: int = 400,
+    ) -> str:
+        return json.dumps(
+            {
+                "error": {
+                    "message": message,
+                    "type": err_type,
+                    "param": None,
+                    "code": status_code,
+                }
+            }
+        )
+
     def _request_id_prefix(self) -> str:
         return "resp_"
 
@@ -834,6 +867,13 @@ class OpenAIServingResponses(OpenAIServingChat):
 
         async for ctx in result_generator:
 
+            # Only process context objects that implement the `is_expecting_start()` method,
+            # which indicates they support per-turn streaming (e.g., StreamingHarmonyContext).
+            # Contexts without this method are skipped, as they do not represent a new turn
+            # or are not compatible with per-turn handling in the /v1/responses endpoint.
+            if not hasattr(ctx, "is_expecting_start"):
+                continue
+
             if ctx.is_expecting_start():
                 current_output_index += 1
                 sent_output_item_added = False