feat: skip empty fields in mcp tool call + improve error handling and response

application/llm/google_ai.py

@@ -151,15 +151,8 @@ class GoogleLLM(BaseLLM):
 
             if role == "assistant":
                 role = "model"
-            elif role == "system":
-                continue
             elif role == "tool":
-                continue
-            elif role not in ["user", "model"]:
-                logging.warning(
-                    f"GoogleLLM: Converting unsupported role '{role}' to 'user'"
-                )
-                role = "user"
+                role = "model"
 
             parts = []
             if role and content is not None:

application/llm/handlers/base.py

@@ -205,7 +205,6 @@ class LLMHandler(ABC):
                     except StopIteration as e:
                         tool_response, call_id = e.value
                         break
-
                 updated_messages.append(
                     {
                         "role": "assistant",
@@ -222,17 +221,36 @@ class LLMHandler(ABC):
                 )
 
                 updated_messages.append(self.create_tool_message(call, tool_response))
-
             except Exception as e:
                 logger.error(f"Error executing tool: {str(e)}", exc_info=True)
-                updated_messages.append(
-                    {
-                        "role": "tool",
-                        "content": f"Error executing tool: {str(e)}",
-                        "tool_call_id": call.id,
-                    }
+                error_call = ToolCall(
+                    id=call.id, name=call.name, arguments=call.arguments
                 )
+                error_response = f"Error executing tool: {str(e)}"
+                error_message = self.create_tool_message(error_call, error_response)
+                updated_messages.append(error_message)
 
+                call_parts = call.name.split("_")
+                if len(call_parts) >= 2:
+                    tool_id = call_parts[-1]  # Last part is tool ID (e.g., "1")
+                    action_name = "_".join(call_parts[:-1])
+                    tool_name = tools_dict.get(tool_id, {}).get("name", "unknown_tool")
+                    full_action_name = f"{action_name}_{tool_id}"
+                else:
+                    tool_name = "unknown_tool"
+                    action_name = call.name
+                    full_action_name = call.name
+                yield {
+                    "type": "tool_call",
+                    "data": {
+                        "tool_name": tool_name,
+                        "call_id": call.id,
+                        "action_name": full_action_name,
+                        "arguments": call.arguments,
+                        "error": error_response,
+                        "status": "error",
+                    },
+                }
         return updated_messages
 
     def handle_non_streaming(
@@ -263,13 +281,11 @@ class LLMHandler(ABC):
                 except StopIteration as e:
                     messages = e.value
                     break
-
             response = agent.llm.gen(
                 model=agent.gpt_model, messages=messages, tools=agent.tools
             )
             parsed = self.parse_response(response)
             self.llm_calls.append(build_stack_data(agent.llm))
-
         return parsed.content
 
     def handle_streaming(

application/llm/handlers/google.py

@@ -17,7 +17,6 @@ class GoogleLLMHandler(LLMHandler):
                 finish_reason="stop",
                 raw_response=response,
             )
-
         if hasattr(response, "candidates"):
             parts = response.candidates[0].content.parts if response.candidates else []
             tool_calls = [
@@ -41,7 +40,6 @@ class GoogleLLMHandler(LLMHandler):
                 finish_reason="tool_calls" if tool_calls else "stop",
                 raw_response=response,
             )
-
         else:
             tool_calls = []
             if hasattr(response, "function_call"):
@@ -61,14 +59,16 @@ class GoogleLLMHandler(LLMHandler):
 
     def create_tool_message(self, tool_call: ToolCall, result: Any) -> Dict:
         """Create Google-style tool message."""
-        from google.genai import types
 
         return {
-            "role": "tool",
+            "role": "model",
             "content": [
-                types.Part.from_function_response(
-                    name=tool_call.name, response={"result": result}
-                ).to_json_dict()
+                {
+                    "function_response": {
+                        "name": tool_call.name,
+                        "response": {"result": result},
+                    }
+                }
             ],
         }