Merge branch 'fix_duplicate' into 'mgx_ops'

Fix duplicate

See merge request pub/MetaGPT!259

metagpt/roles/di/data_analyst.py

@@ -31,7 +31,7 @@ class DataAnalyst(RoleZero):
     task_type_desc: str = TASK_TYPE_DESC
 
     tools: list[str] = ["Plan", "DataAnalyst", "RoleZero", "Browser"]
-    custom_tools: list[str] = ["machine learning", "web scraping", "Terminal"]
+    custom_tools: list[str] = ["web scraping", "Terminal"]
     custom_tool_recommender: ToolRecommender = None
     experience_retriever: ExpRetriever = KeywordExpRetriever()
 
@@ -72,6 +72,9 @@ class DataAnalyst(RoleZero):
         Args:
             instruction: The specific task description for which the code needs to be written.
         """
+        if self.planner.plan:
+            logger.info(f"Current task {self.planner.plan.current_task}")
+
         counter = 0
         success = False
         await self.execute_code.init_code()

metagpt/roles/di/role_zero.py

@@ -18,7 +18,9 @@ from metagpt.prompts.di.role_zero import (
     CMD_PROMPT,
     JSON_REPAIR_PROMPT,
     QUICK_THINK_PROMPT,
+    REGENERATE_PROMPT,
     ROLE_INSTRUCTION,
+    THOUGHT_GUIDANCE,
 )
 from metagpt.roles import Role
 from metagpt.schema import AIMessage, Message, UserMessage
@@ -155,6 +157,7 @@ class RoleZero(Role):
             plan_status=plan_status,
             current_task=current_task,
             instruction=instruction,
+            thought_guidance=THOUGHT_GUIDANCE,
             latest_observation=memory[-1].content,
         )
         memory = await self.parse_browser_actions(memory)
@@ -168,6 +171,8 @@ class RoleZero(Role):
             )
             self.command_rsp = await self.llm_cached_aask(req=req, system_msgs=self.system_msg, state_data=state_data)
 
+        self.command_rsp = await self._check_duplicates(req, self.command_rsp)
+
         self.rc.memory.add(AIMessage(content=self.command_rsp))
         return True
 
@@ -260,6 +265,19 @@ class RoleZero(Role):
 
         return rsp_msg
 
+    async def _check_duplicates(self, req: list[dict], command_rsp: str):
+        past_rsp = [mem.content for mem in self.rc.memory.get(self.memory_k)]
+        if command_rsp in past_rsp:
+            # Normal response with thought contents are highly unlikely to reproduce
+            # If an identical response is detected, it is a bad response, mostly due to LLM repeating generated content
+            # In this case, ask human for help and regenerate
+            # TODO: switch to llm_cached_aask
+            logger.warning(f"Duplicate response detected: {command_rsp}")
+            regenerate_req = req + [UserMessage(content=REGENERATE_PROMPT)]
+            regenerate_req = self.llm.format_msg(regenerate_req)
+            command_rsp = await self.llm.aask(regenerate_req)
+        return command_rsp
+
     async def _parse_commands(self) -> Tuple[List[Dict], bool]:
         """Retrieves commands from the Large Language Model (LLM).
 

metagpt/roles/di/team_leader.py

@@ -72,6 +72,9 @@ class TeamLeader(RoleZero):
         Publish a message to a team member, use member name to fill send_to args. You may copy the full original content or add additional information from upstream. This will make team members start their work.
         DONT omit any necessary info such as path, link, environment, programming language, framework, requirement, constraint from original content to team members because you are their sole info source.
         """
+        self._set_state(-1)  # each time publishing a message, pause to wait for the response
+        if send_to == self.name:
+            return  # Avoid sending message to self
         # Specify the outer send_to to overwrite the default "no one" value. Use UserMessage because message from self is like a user request for others.
         self.publish_message(
             UserMessage(content=content, sent_from=self.name, send_to=send_to, cause_by=RunCommand), send_to=send_to