xgae 0.1.19__tar.gz → 0.2.0__tar.gz

{xgae-0.1.19 → xgae-0.2.0}/CHANGELOG.md

@@ -1,7 +1,21 @@
 # Release Changelog
+## [0.2.0] - 2025-9-10
+### Added
+- Agent Engine release 0.2
+- Example: Langgraph ReactAgent release 0.2
+### Fixed
+- Agent Engine: call mcp tool fail, call 'ask' tool again and again
+- Example Langgraph ReactAgent: retry on 'ask', user_input is ask answer
+
+
+## [0.1.20] - 2025-9-9
+### Added
+- Example: Langgraph ReactAgent add final_result_agent 
+
+
 ## [0.1.19] - 2025-9-8
 ### Added
-- Example: Langgraph react agent release V1, full logic but no final result agent and tool select agent
+- Example: Langgraph ReactAgent release V1, full logic but no final result agent and tool select agent
 
 
 # Release Changelog

{xgae-0.1.19 → xgae-0.2.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: xgae
-Version: 0.1.19
+Version: 0.2.0
 Summary: Extreme General Agent Engine
 Requires-Python: >=3.13
 Requires-Dist: colorlog==6.9.0

{xgae-0.1.19 → xgae-0.2.0}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "xgae"
-version = "0.1.19"
+version = "0.2.0"
 description = "Extreme General Agent Engine"
 readme = "README.md"
 requires-python = ">=3.13"

xgae-0.2.0/src/examples/agent/langgraph/react/agent_base.py

@@ -0,0 +1,32 @@
+from typing import Any, Dict, List, TypedDict, Optional
+from xgae.engine.engine_base import XGATaskResult
+
+class EvaluateResult(TypedDict, total=False):
+    task_result: Dict[str, Any]
+    task_process: Dict[str, Any]
+    function_call: Dict[str, Any]
+
+class AgentContext(TypedDict, total=False):
+    task_id: str
+    session_id: str
+    user_id: str
+    agent_id: str
+    thread_id: str
+
+
+class TaskState(TypedDict, total=False):
+    """State definition for the agent orchestration graph"""
+    llm_messages: List[Dict[str, Any]]
+    user_inputs: List[str]
+    next_node: str
+    system_prompt: str
+    custom_tools: List[str]
+    general_tools: List[str]
+    task_result: XGATaskResult
+    final_result: XGATaskResult
+    eval_result: EvaluateResult
+    retry_count: int
+    task_no: int
+    agent_context: AgentContext
+
+

{xgae-0.1.19 → xgae-0.2.0}/src/examples/agent/langgraph/react/react_agent.py

@@ -1,67 +1,50 @@
 import logging
 import os
 
-from typing import Any, Dict, List, Annotated, Sequence, TypedDict, Optional, AsyncGenerator
+from typing import Any, Dict, List, Optional, AsyncGenerator
 from uuid import uuid4
 
 from langfuse.callback import CallbackHandler
 from langfuse import Langfuse
 
-from langchain_core.messages import BaseMessage, HumanMessage, AIMessage
 
 from langgraph.graph import END, START, StateGraph
 from langgraph.types import interrupt, Command
 from langgraph.checkpoint.memory import MemorySaver
-from langgraph.graph.message import add_messages
 from langgraph.config import get_stream_writer
 
 from xgae.utils.misc import read_file
 from xgae.utils import log_trace
 
-from xgae.engine.engine_base import XGATaskResult, XGAResponseMessage
+from xgae.engine.engine_base import XGATaskResult
 from xgae.engine.mcp_tool_box import XGAMcpToolBox
 from xgae.engine.task_engine import XGATaskEngine
 
-
-class AgentContext(TypedDict, total=False):
-    task_id: str
-    session_id: str
-    user_id: str
-    agent_id: str
-    thread_id: str
-
-
-class TaskState(TypedDict, total=False):
-    """State definition for the agent orchestration graph"""
-    messages: Annotated[Sequence[BaseMessage], add_messages]  # for message persistent
-    user_input: str
-    next_node: str
-    system_prompt: str
-    custom_tools: List[str]
-    general_tools: List[str]
-    task_result: XGATaskResult
-    final_result: XGATaskResult
-    iteration_count: int
-    agent_context: AgentContext
-
+from examples.agent.langgraph.react.agent_base import AgentContext, TaskState, EvaluateResult
+from examples.agent.langgraph.react.result_eval_agent import TaskResultEvalAgent
 
 class XGAReactAgent:
     MAX_TASK_RETRY = 2
+    QUALIFIED_RESULT_SCORE = 0.7
 
     def __init__(self):
-        self.tool_box = XGAMcpToolBox(custom_mcp_server_file="mcpservers/custom_servers.json")
         self.graph = None
-        self.graph_config = None
-        self.graph_langfuse = Langfuse(enabled=False)
 
+        self.graph_config = None
+        self.graph_langfuse = None
         self.task_engine: XGATaskEngine = None
 
+        self.tool_box = XGAMcpToolBox(custom_mcp_server_file="mcpservers/custom_servers.json")
+        self.result_eval_agent = TaskResultEvalAgent()
+
+
     async def _create_graph(self) -> StateGraph:
         try:
             graph_builder = StateGraph(TaskState)
 
             # Add nodes
             graph_builder.add_node('supervisor', self._supervisor_node)
+            graph_builder.add_node('prompt_optimize', self._prompt_optimize_node)
             graph_builder.add_node('select_tool', self._select_tool_node)
             graph_builder.add_node('exec_task', self._exec_task_node)
             graph_builder.add_node('final_result', self._final_result_node)
@@ -72,12 +55,14 @@ class XGAReactAgent:
                 'supervisor',
                 self._next_condition,
                 {
-                    'select_tool': 'select_tool',
-                    'exec_task': 'exec_task',
-                    'end': END
+                    'select_tool'       : 'select_tool',
+                    'exec_task'         : 'exec_task',
+                    'prompt_optimize'   : 'prompt_optimize',
+                    'end'               : END
                 }
             )
 
+            graph_builder.add_edge('prompt_optimize', 'select_tool')
             graph_builder.add_edge('select_tool', 'exec_task')
             graph_builder.add_edge('exec_task', 'final_result')
 
@@ -86,8 +71,8 @@ class XGAReactAgent:
                 self._next_condition,
                 {
                     'supervisor': 'supervisor',
-                    'exec_task': 'exec_task',
-                    'end': END
+                    'exec_task' : 'exec_task',
+                    'end'       : END
                 }
             )
             
@@ -99,55 +84,105 @@ class XGAReactAgent:
             logging.error("Failed to create XGARectAgent Graph: %s", str(e))
             raise
 
+
     def _search_system_prompt(self, user_input: str) -> str:
         # You should search RAG use user_input, fetch COT or Prompt for your business
         system_prompt = None if "fault" not in user_input else read_file("templates/example/fault_user_prompt.txt")
         return system_prompt
 
+
     async def _supervisor_node(self, state: TaskState) -> Dict[str, Any]:
-        user_input = state.get("user_input", "")
+        user_input = state['user_inputs'][0]
+        eval_result = state.get('eval_result', None)
+
         system_prompt = self._search_system_prompt(user_input)
+        is_system_prompt = True if system_prompt is not None else False
 
         general_tools = [] if system_prompt else ["*"]
         custom_tools = ["*"] if system_prompt  else []
 
-        next_node = "select_tool" if system_prompt else "exec_task"
+        task_plan_score = None
+        if eval_result and 'task_plan' in eval_result and 'score' in eval_result['task_plan']:
+            task_plan_score = eval_result['task_plan'].get('score', 1.0)
+
+        function_call_score = None
+        if eval_result and 'function_call' in eval_result and 'score' in eval_result['function_call']:
+            function_call_score = eval_result['function_call'].get('score', 1.0)
+
+        super_state = {}
+        if task_plan_score and task_plan_score < self.QUALIFIED_RESULT_SCORE:
+            next_node = "prompt_optimize"
+            super_state = self._prepare_task_retry(state)
+            logging.warning(f"****** ReactAgent TASK_RETRY: task_plan_score={task_plan_score} < {self.QUALIFIED_RESULT_SCORE} , "
+                            f"Start Optimize Prompt ...")
+        elif function_call_score and function_call_score < self.QUALIFIED_RESULT_SCORE:
+            next_node = "select_tool"
+            super_state = self._prepare_task_retry(state)
+            logging.warning(f"****** ReactAgent TASK_RETRY: function_call_score={function_call_score} < {self.QUALIFIED_RESULT_SCORE} , "
+                            f"Select Tool Again ...")
+        elif eval_result is not None:  # retry condition is not satisfied, end task
+            next_node = "end"
+        else:
+            next_node = "select_tool" if is_system_prompt else "exec_task"
+
+        logging.info(f"ReactAgent supervisor_node: is_system_prompt={is_system_prompt}, next_node={next_node}")
+
+        super_state['next_node']        = next_node
+        super_state['system_prompt']    = system_prompt
+        super_state['custom_tools']     = custom_tools
+        super_state['general_tools']    = general_tools
+
+        return super_state
 
+
+    async def _prompt_optimize_node(self, state: TaskState) -> Dict[str, Any]:
+        system_prompt = state['system_prompt']
+        logging.info("ReactAgent prompt_optimize_node: optimize system prompt")
+        # @todo optimize system prompt in future
         return {
             'system_prompt' : system_prompt,
-            'next_node'     : next_node,
-            'general_tools' : general_tools,
-            'custom_tools'  : custom_tools,
         }
 
+
     def _select_custom_tools(self, system_prompt: str) -> list[str]:
+        # @todo select mcp tool based on system prompt in future
         custom_tools = ["*"] if system_prompt  else []
         return custom_tools
 
+
     async def _select_tool_node(self, state: TaskState) -> Dict[str, Any]:
         system_prompt = state.get('system_prompt',None)
         general_tools = []
+
+        logging.info("ReactAgent select_tool_node: select tool based on system_prompt")
         custom_tools = self._select_custom_tools(system_prompt)
         return {
             'general_tools' : general_tools,
             'custom_tools'  : custom_tools,
         }
 
+
     async def _exec_task_node(self, state: TaskState) -> Dict[str, Any]:
-        user_input = state['user_input']
+        user_input = state['user_inputs'][0]
         system_prompt = state.get('system_prompt',None)
         general_tools = state.get('general_tools',[])
         custom_tools = state.get('custom_tools',[])
+        retry_count = state.get('retry_count', 0)
+        task_no = state.get('task_no', 0)
         is_system_prompt = True if system_prompt is not None else False
 
         trace_id = self.graph_langfuse.get_trace_id()
+        llm_messages = []
         try:
-            logging.info(f"🔥 XGATaskEngine run_task: user_input={user_input}, general_tools={general_tools}, "
+            logging.info(f"🔥 ReactAgent exec_task_node: user_input={user_input}, general_tools={general_tools}, "
                          f"custom_tools={custom_tools}, is_system_prompt={is_system_prompt}")
+
+            # if langgraph resume , must use same task engine
             if self.task_engine is None:
                 self.task_engine = XGATaskEngine(
                     task_id         = state['agent_context']['task_id'],
-                    session_id=state['agent_context'].get('session_id', None),
+                    task_no         = task_no,
+                    session_id      = state['agent_context'].get('session_id', None),
                     user_id         = state['agent_context'].get('user_id', None),
                     agent_id        = state['agent_context'].get('agent_id', None),
                     tool_box        = self.tool_box,
@@ -155,6 +190,7 @@ class XGAReactAgent:
                     custom_tools    = custom_tools,
                     system_prompt   = system_prompt
                 )
+                retry_count += 1
 
             chunks = []
             stream_writer = get_stream_writer()
@@ -164,40 +200,60 @@ class XGAReactAgent:
                 stream_writer({"engine_message": chunk})
 
             task_result = self.task_engine.parse_final_result(chunks)
+            llm_messages = self.task_engine.get_history_llm_messages()
+            task_no += 1  # a task use unique task_no, no matter retry n times
         except Exception as e:
             logging.error(f"XReactAgent exec_task_node: Failed to execute task: {e}")
             task_result = XGATaskResult(type="error", content="Failed to execute task")
 
-        iteration_count = state.get('iteration_count', 0) + 1
         return {
-            'task_result' : task_result,
-            'iteration_count': iteration_count,
+            'task_result'   : task_result,
+            'retry_count'   : retry_count,
+            'llm_messages'  : llm_messages.copy(),
+            'task_no'       : task_no,
         }
 
+
     async def _final_result_node(self, state: TaskState) -> Dict[str, Any]:
-        user_input = state['user_input']
-        iteration_count = state['iteration_count']
+        user_inputs = state['user_inputs']
         task_result = state['task_result']
+        llm_messages = state['llm_messages']
+        agent_context = state['agent_context']
+        system_prompt = state.get('system_prompt', None)
+        retry_count = state['retry_count']
+
+        is_system_prompt = True if system_prompt is not None else False
 
         next_node = "end"
-        if task_result['type'] == "error" and iteration_count < self.MAX_TASK_RETRY:
-            next_node = "supervisor"
-            final_result = None
-        elif task_result['type'] == "ask":
-            final_result = task_result
+        final_result = task_result
+        eval_result = None
+        if task_result['type'] == "ask":
             logging.info(f"XReactAgent final_result_node: ASK_USER_QUESTION: {task_result['content']}")
-            user_input = interrupt({
-                'final_result' : final_result
+            ask_input = interrupt({
+                'final_result' : task_result
             })
-            logging.info(f"XReactAgent final_result_node: ASK_USER_ANSWER: {user_input}")
+            logging.info(f"XReactAgent final_result_node: ASK_USER_ANSWER: {ask_input}")
             next_node = "exec_task"
-        else:
-            final_result = task_result
-
+            user_inputs.insert(0, ask_input)
+            final_result = None
+        elif is_system_prompt and retry_count < self.MAX_TASK_RETRY:
+            trace_id = self.graph_langfuse.get_trace_id()
+            session_id = agent_context.get('session_id', None)
+            task_input = ", ".join(reversed(user_inputs))
+            eval_result = await self.result_eval_agent.eval_result(task_input, system_prompt, task_result,
+                                                                   llm_messages, trace_id, session_id)
+            if 'task_result' in eval_result and 'score' in eval_result['task_result']:
+                score =  eval_result['task_result'].get('score', 1.0)
+                if score < self.QUALIFIED_RESULT_SCORE:
+                    next_node = "supervisor"
+        
+        logging.info(f"ReactAgent final_result_node: next_node={next_node}")
+        
         return {
-            'user_input' : user_input,
-            'next_node' : next_node,
-            'final_result' : final_result,
+            'user_inputs'    : user_inputs,
+            'next_node'     : next_node,
+            'final_result'  : final_result,
+            'eval_result'   : eval_result
         }
 
 
@@ -266,13 +322,15 @@ class XGAReactAgent:
 
         except Exception as e:
             log_trace(e, f"XReactAgent generate: user_input={user_input}")
-            yield XGATaskResult(type="error", content=f"React Agent error: {e}")
+            yield {'type': "error", 'content': f"React Agent generate error: {e}"}
 
 
-    async def _prepare_graph_start(self, user_input, agent_context: AgentContext):
+    async def _prepare_graph_start(self, user_input, agent_context: AgentContext)->TaskState:
         if self.graph is None:
             self.graph = await self._create_graph()
 
+        self._clear_graph()
+
         agent_context = agent_context or {}
         task_id = agent_context.get("task_id", f"xga_task_{uuid4()}")
         agent_context["task_id"] = task_id
@@ -281,20 +339,14 @@ class XGAReactAgent:
         session_id = agent_context.get('session_id', task_id)
         agent_context['session_id'] = session_id
 
-        graph_input = {
-            'messages'          : [HumanMessage(content=f"information for: {user_input}")],
-            'user_input'        : user_input,
-            'next_node'         : None,
-            'agent_context'     : agent_context,
-            'iteration_count'   : 0
-        }
 
         langfuse_handler = self._get_langfuse_handler(agent_context)
         callbacks = None
         if langfuse_handler:
             callbacks = [langfuse_handler]
             self.graph_langfuse = langfuse_handler.langfuse
-
+        else:
+            self.graph_langfuse = Langfuse(enabled=False)
 
         self.graph_config = {
             'recursion_limit': 100,
@@ -304,6 +356,14 @@ class XGAReactAgent:
             'callbacks': callbacks
         }
 
+        graph_input = {
+            'user_inputs'       : [user_input],
+            'next_node'         : None,
+            'agent_context'     : agent_context,
+            'retry_count'       : 0,
+            'task_no'           : 0
+        }
+
         return graph_input
 
 
@@ -325,3 +385,21 @@ class XGAReactAgent:
         return langfuse_handler
 
 
+    def _clear_graph(self):
+        self.graph_config = None
+        self.graph_langfuse = None
+        self.task_engine: XGATaskEngine = None
+
+
+    def _prepare_task_retry(self, state: TaskState)-> Dict[str, Any]:
+        self.task_engine = None
+        user_inputs = state['user_inputs']
+        task_input = ", ".join(reversed(user_inputs))
+
+        return {
+            'user_inputs'   : [task_input],
+            'llm_messages'  : [],
+            'task_result'   : None,
+            'final_result'  : None,
+            'eval_result'   : None,
+        }

xgae-0.2.0/src/examples/agent/langgraph/react/result_eval_agent.py

@@ -0,0 +1,125 @@
+import json
+import logging
+import re
+
+from typing import Any, Dict, List, Optional
+
+from xgae.engine.engine_base import XGATaskResult
+from xgae.utils.misc import read_file
+from xgae.utils.llm_client import LLMClient, LangfuseMetadata
+
+
+class TaskResultEvalAgent:
+    def __init__(self):
+        self.model_client = LLMClient()
+        self.prompt_template: str = read_file("templates/example/result_eval_template.txt")
+
+
+    async def eval_result(self,
+                          task_input: str,
+                          task_plan: str,
+                          task_result: XGATaskResult,
+                          llm_messages: List[Dict[str, Any]],
+                          trace_id: Optional[str] = None,
+                          session_id: Optional[str] = None)-> Dict[str, Any]:
+        prompt = self._build_prompt(task_input, task_plan, task_result, llm_messages)
+        messages = [{"role": "user", "content": prompt}]
+
+        langfuse_metadata = self._create_llm_langfuse_meta(trace_id, session_id)
+
+        response = await self.model_client.acompletion(messages, langfuse_metadata)
+        response_text = await self.model_client.get_response_result(response)
+
+        cleaned_text = re.sub(r'^\s*```json|```\s*$', '', response_text, flags=re.MULTILINE).strip()
+        eval_result = json.loads(cleaned_text)
+
+        result_score = eval_result.get('task_result', {}).get('score', -1)
+        plan_score = eval_result.get('task_plan', {}).get('score', -1)
+        function_score = eval_result.get('function_call', {}).get('score', -1)
+
+        logging.info(f"FINAL_RESULT_SCORE: task_result_score={result_score}, "
+                     f"task_plan_score={plan_score}, function_call_score={function_score}")
+        return eval_result
+
+
+    def _build_prompt(self, task_input: str, task_plan: str, task_result: XGATaskResult, llm_messages: List[Dict[str, Any]])-> str:
+        prompt = self.prompt_template.replace("{task_input}", task_input)
+        prompt = prompt.replace("{task_result}", str(task_result))
+        llm_process = ""
+        function_process = ""
+        llm_step = 1
+        function_step = 1
+        for llm_message in llm_messages:
+            content = llm_message.get('content', '')
+            if "tool_execution" in content:
+                function_process += f"{function_step}. \n"
+                tool_exec = json.loads(content)
+                func_call = tool_exec['tool_execution']
+                func_call.pop('xml_tag_name')
+                clear_content = json.dumps(func_call, indent=2)
+                function_process += clear_content
+                function_process += "\n"
+                function_step += 1
+            else:
+                llm_process += f"{llm_step}. \n"
+                llm_process += content
+                llm_process += "\n"
+                llm_step += 1
+
+        prompt = prompt.replace("{task_plan}", task_plan)
+        prompt = prompt.replace("{llm_process}", llm_process)
+        prompt = prompt.replace("{function_process}", function_process)
+
+        return prompt
+
+
+    def _create_llm_langfuse_meta(self, trace_id:str, session_id: str)-> LangfuseMetadata:
+        generation_name = "xga_agent_final_result_completion"
+
+        return LangfuseMetadata(
+            generation_name     = generation_name,
+            existing_trace_id   = trace_id,
+            session_id          = session_id
+        )
+
+
+
+if __name__ == "__main__":
+    import asyncio
+    from xgae.utils.setup_env import setup_logging
+    setup_logging()
+
+    async def main():
+        final_result_agent = TaskResultEvalAgent()
+
+        task_plan = read_file("templates/example/fault_user_prompt.txt")
+        user_input = "locate 10.2.3.4 fault and solution"
+
+        answer = ("Task Summary: The fault for IP 10.2.3.4 was identified as a Business Recharge Fault (Code: F01), "
+                  "caused by a Phone Recharge Application Crash. The solution applied was to restart the application. "
+                  "Key Deliverables: Fault diagnosis and resolution steps. Impact Achieved: Service restored.")
+        task_result:XGATaskResult = {'type': "answer",  'content': answer}
+
+        llm_messages: List[Dict[str, Any]] = [{
+            'content':
+                    """<function_calls> 
+                        <invoke name="get_alarm_type">
+                                       <parameter name="alarm_id">alm0123</parameter>
+                         </invoke>
+                    </function_calls>'""",
+             'role': "assistant"
+             },{
+            'content': """{"tool_execution": {
+                                "function_name": "get_alarm_type", 
+                                "xml_tag_name": "get-alarm-type", 
+                                "arguments": {"alarm_id": "alm0123"}, 
+                                "result": {"success": true, "output": "1", "error": null}}}""",
+            'role': 'assistant'
+            }]
+
+        return await final_result_agent.eval_result(user_input, task_plan, task_result, llm_messages)
+
+
+    final_result = asyncio.run(main())
+    final_result_json = json.dumps(final_result, ensure_ascii=False, indent=2)
+    print(f"FINAL_RESULT：   {final_result_json} ")

{xgae-0.1.19 → xgae-0.2.0}/src/examples/agent/langgraph/react/run_react_agent.py

@@ -1,5 +1,7 @@
 import asyncio
 
+from uuid import uuid4
+
 from xgae.utils.setup_env import setup_logging
 
 from examples.agent.langgraph.react.react_agent import XGAReactAgent, AgentContext
@@ -9,18 +11,18 @@ async def main():
     is_stream = True  # two mode agent experience
     task_no = 0
     user_inputs = [
-        # "5+5", # For no tool call
-        # "locate 10.2.3.4 fault and solution",  # For custom tool
-        "locate fault and solution",   # For human append input
+        #"5+5", # For no tool call
+        "locate 10.2.3.4 fault and solution",  # For custom tool
+        #"locate fault and solution",   # For human append input
     ]
 
     for user_input in user_inputs:
         agent = XGAReactAgent()
         task_no += 1
         context: AgentContext = {
-            'task_id': f"agent_task_{task_no}",
-            'user_id': f"agent_user_{task_no}",
-            'agent_id': f"agent_{task_no}",
+            'task_id': f"agent_task_{uuid4()}",   # can be set with request_id, must be unique
+            'user_id': "agent_user_1",
+            'agent_id': "agent_1",
         }
 
         is_resume = False

{xgae-0.1.19 → xgae-0.2.0}/src/xgae/engine/task_engine.py

@@ -18,6 +18,7 @@ from xgae.engine.responser.responser_base import TaskResponserContext, TaskRespo
 class XGATaskEngine:
     def __init__(self,
                  task_id: Optional[str] = None,
+                 task_no: Optional[int] = None,
                  session_id: Optional[str] = None,
                  user_id: Optional[str] = None,
                  agent_id: Optional[str] = None,
@@ -28,7 +29,8 @@ class XGATaskEngine:
                  tool_exec_parallel: Optional[bool] = None,
                  llm_config: Optional[LLMConfig] = None,
                  prompt_builder: Optional[XGAPromptBuilder] = None,
-                 tool_box: Optional[XGAToolBox] = None):
+                 tool_box: Optional[XGAToolBox] = None
+                 ):
         self.task_id    = task_id if task_id else f"xga_task_{uuid4()}"
         self.session_id = session_id
         self.user_id    = user_id
@@ -50,7 +52,7 @@ class XGATaskEngine:
         self.use_assistant_chunk_msg = to_bool(os.getenv('USE_ASSISTANT_CHUNK_MSG', False))
         self.tool_exec_parallel = True if tool_exec_parallel is None else tool_exec_parallel
 
-        self.task_no = -1
+        self.task_no = (task_no - 1) if task_no  else -1
         self.task_run_id :str = None
         self.task_prompt :str = None
         self.task_langfuse: XGATaskLangFuse = None
@@ -213,7 +215,7 @@ class XGATaskEngine:
         auto_count = continuous_state.get("auto_continue_count")
         langfuse_metadata = self.task_langfuse.create_llm_langfuse_meta(auto_count)
 
-        llm_response = await self.llm_client.create_completion(llm_messages, langfuse_metadata)
+        llm_response = await self.llm_client.acompletion(llm_messages, langfuse_metadata)
         response_processor = self._create_response_processer()
 
         async for chunk in response_processor.process_response(llm_response, llm_messages, continuous_state):
@@ -315,7 +317,7 @@ class XGATaskEngine:
     def get_history_llm_messages (self) -> List[Dict[str, Any]]:
         llm_messages = []
         for message in self.task_response_msgs:
-            if message['is_llm_message']:
+            if message['is_llm_message'] and message['type'] != "assistant_chunk":
                 llm_messages.append(message)
 
         response_llm_contents = []