fix(responses): fix regression in support for mcp tool require_approval argument (#3731)

# What does this PR do?

It prevents a tool call message being added to the chat completions
message without a corresponding tool call result, which is needed in the
case that an approval is required first or if the approval request is
denied. In both these cases the tool call messages is popped of the next
turn messages.

Closes #3728

## Test Plan
Ran the integration tests
Manual check of both approval and denial against gpt-4o

Signed-off-by: Gordon Sim <[email protected]>

Author: grs

llama_stack/providers/inline/agents/meta_reference/responses/openai_responses.py

@@ -269,7 +269,7 @@ async def _create_streaming_response(
             response_tools=tools,
             temperature=temperature,
             response_format=response_format,
-            inputs=input,
+            inputs=all_input,
         )
 
         # Create orchestrator and delegate streaming logic

llama_stack/providers/inline/agents/meta_reference/responses/streaming.py

@@ -175,6 +175,8 @@ async def create_response(self) -> AsyncIterator[OpenAIResponseObjectStream]:
             ):
                 yield stream_event
 
+            messages = next_turn_messages
+
             if not function_tool_calls and not non_function_tool_calls:
                 break
 
@@ -187,9 +189,7 @@ async def create_response(self) -> AsyncIterator[OpenAIResponseObjectStream]:
                 logger.info(f"Exiting inference loop since iteration count({n_iter}) exceeds {self.max_infer_iters=}")
                 break
 
-            messages = next_turn_messages
-
-        self.final_messages = messages.copy() + [current_response.choices[0].message]
+        self.final_messages = messages.copy()
 
         # Create final response
         final_response = OpenAIResponseObject(
@@ -232,9 +232,11 @@ def _separate_tool_calls(self, current_response, messages) -> tuple[list, list,
                                     non_function_tool_calls.append(tool_call)
                                 else:
                                     logger.info(f"Approval denied for {tool_call.id} on {tool_call.function.name}")
+                                    next_turn_messages.pop()
                             else:
                                 logger.info(f"Requesting approval for {tool_call.id} on {tool_call.function.name}")
                                 approvals.append(tool_call)
+                                next_turn_messages.pop()
                         else:
                             non_function_tool_calls.append(tool_call)