refactor

# What does this PR do?


## Test Plan

llama_stack/core/server/server.py

@@ -184,7 +184,17 @@ async def lifespan(app: StackApp):
 
 def is_streaming_request(func_name: str, request: Request, **kwargs):
     # TODO: pass the api method and punt it to the Protocol definition directly
-    return kwargs.get("stream", False)
+    # Check for stream parameter at top level (old API style)
+    if "stream" in kwargs:
+        return kwargs["stream"]
+
+    # Check for stream parameter inside Pydantic request params (new API style)
+    if "params" in kwargs:
+        params = kwargs["params"]
+        if hasattr(params, "stream"):
+            return params.stream
+
+    return False
 
 
 async def maybe_await(value):
@@ -289,21 +299,34 @@ def create_dynamic_typed_route(func: Any, method: str, route: str) -> Callable:
     if method == "post":
         # Annotate parameters that are in the path with Path(...) and others with Body(...),
         # but preserve existing File() and Form() annotations for multipart form data
-        new_params = (
-            [new_params[0]]
-            + [
-                (
+        def should_embed(param: inspect.Parameter) -> bool:
+            """Determine if Body should use embed=True or embed=False.
+
+            For OpenAI-compatible endpoints (param name is 'params'), use embed=False
+            so the request body is parsed directly as the model (not nested).
+            This allows OpenAI clients to send standard OpenAI format.
+            For other endpoints, use embed=True for SDK compatibility.
+            """
+            param_type = param.annotation
+            # HACK
+            # Check if it's a Pydantic BaseModel and param name is 'params' (OpenAI-compatible)
+            if inspect.isclass(param_type) and issubclass(param_type, BaseModel) and param.name == "params":
+                return False
+            return True
+
+        original_params = new_params[1:]  # Skip request parameter
+        new_params = [new_params[0]]  # Keep request parameter
+
+        for param in original_params:
+            if param.name in path_params:
+                new_params.append(
                     param.replace(annotation=Annotated[param.annotation, FastapiPath(..., title=param.name)])
-                    if param.name in path_params
-                    else (
-                        param  # Keep original annotation if it's already an Annotated type
-                        if get_origin(param.annotation) is Annotated
-                        else param.replace(annotation=Annotated[param.annotation, Body(..., embed=True)])
-                    )
                 )
-                for param in new_params[1:]
-            ]
-        )
+            elif get_origin(param.annotation) is Annotated:
+                new_params.append(param)  # Keep existing annotation
+            else:
+                embed = should_embed(param)
+                new_params.append(param.replace(annotation=Annotated[param.annotation, Body(..., embed=embed)]))
 
     route_handler.__signature__ = sig.replace(parameters=new_params)